text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
abstract: 'Due to the foreground extinction of the Milky Way, galaxies become increasingly faint as they approach the Galactic Equator creating a “zone of avoidance” (ZOA) in the distribution of optically visible galaxies of about 25%. A “whole-sky” map of galaxies is essential, however, for understanding the dynamics in our local Universe, in particular the peculiar velocity of the Local Group with respect to the Cosmic Microwave Background and velocity flow fields such as in the Great Attractor (GA) region. The current status of deep optical galaxy searches behind the Milky Way and their completeness as a function of foreground extinction will be reviewed. It has been shown that these surveys – which in the mean time cover the whole ZOA (Fig. \[cor\]) – result in a considerable reduction of the ZOA from extinction levels of A$_{\rm B} = 1\fm0$ (Fig. \[ait\]) to A$_{\rm B} = 3\fm0$ (Fig. \[aitc\]). In the remaining, optically opaque ZOA, systematic [H]{} surveys are powerful in uncovering galaxies, as is demonstrated for the GA region with data from the full sensitivity Parkes Multibeam [H]{} survey ($300\deg \le \ell \le 332\deg$, $|b| \le 5\fdg5$, Fig. \[MBGA\]).'
author:
- 'Renée C. Kraan-Korteweg $^1$'
- 'Sebastian Juraszek $^{2,3}$'
title: |
Mapping the Hidden Universe:\
The Galaxy Distribution in the Zone of Avoidance
---
=2em
=17.5cm =24.6 cm =-2.5cm =-1.0cm =-1.0cm 0.15cm
$^1$ Departamento de Astronom[í]{}a, Universidad de Guanajuato, Apartado Postal 144, Guanajuato GTO 36000, Mexico\
[email protected]\
$^2$ School of Physics, University of Sydney, NSW 2006 Australia\
$^3$ ATNF, CSIRO, PO Box 76, Epping, NSW 2121, Australia\
[email protected]\
[**Keywords:**]{} zone of avoidance — surveys — ISM: dust, extinction — large-scale structure of the universe
Introduction
============
In 1988, Lynden-Bell & Lahav for the first time prepared a “whole-sky” distribution of the extragalactic light to map the density enhancements in the local Universe, to compare them to cosmic flow fields and to determine the gravity field on the Local Group. Assuming that light traces mass, this could be realized through a diameter-coded distribution of galaxies larger than ${\rm D} {\,\raisebox{-0.4ex}{$\stackrel{>}{\scriptstyle\sim}$}\,}1\farcm0$ taken from the following galaxy catalogs: the Uppsala General Catalog UGC (Nilson 1973) for the north ($\delta \ge
-2\fdg5$), the ESO Uppsala Catalog (Lauberts 1982) for the south ($\delta \le -17\fdg5$), while the missing strip ($-17\fdg5 < \delta <
-2\fdg5$) was filled with data from the Morphological Catalog of Galaxies MCG (Vorontsov-Velyaminov & Archipova 1963-74).
Because these optical galaxy catalogs are limited to the largest galaxies they become increasingly incomplete close to the Galactic equator where the dust thickens, reducing the apparent diameters of the galaxies. Added to this are the growing number of foreground stars which fully or partially block the view of galaxy images. This is clearly demonstrated in Fig. \[ait\], where such a light distribution is presented in an Aitoff equal-area projection centered on the Galactic plane. The same corrections as advocated in Lahav (1987) have been applied to homogenize the data of the three different galaxy catalogs, [[i.e.]{}]{} ${\rm D_{25} = 1.15 \cdot D_{UGC},
D_{25} = 0.96 \cdot D_{ESO}}$ and ${\rm D_{25} = 1.29 \cdot D_{MCG}}$. A cut-off at ${\rm D} =1\farcm3$ was imposed – the completeness limit of the respective catalogs according to Hudson & Lynden-Bell (1991).
Fig. \[ait\] clearly displays the irregularity in the distribution of galaxies in the nearby Universe with its dynamically important density enhancements such as the Local Supercluster visible as a circle (the Supergalactic Plane) centered on the Virgo cluster at $\ell=284\deg, b=74\deg$, the Perseus-Pisces chain bending into the ZOA at $\ell=95\deg$ and $\ell=165\deg$, the general overdensity in the Cosmic Microwave Background dipole direction ($\ell=280\deg,b=27\deg$, Kogut [[etal.]{}]{} 1993) and the Great Attractor region centered on $\ell=320\deg, b=0\deg$ (Kolatt, Dekel & Lahav 1995) with the Hydra ($270\deg,27\deg$), Antlia ($273\deg,19\deg$), Centaurus ($302\deg,22\deg$) and Pavo ($332\deg,-24\deg$) clusters.
Most conspicuous in this distribution is, however, the very broad, nearly empty band of about 20$\deg$: the Zone of Avoidance. Comparing this band with the 100 dust extinction maps of the DIRBE experiment (Schlegel, Finkbeiner & Davis 1998) we found that the ZOA – the area where our galaxy counts are severely incomplete – is described almost perfectly by the extinction contour ${\rm A_B} = 1\fm0$ (where ${\rm
A_B} = 4.14 \cdot {\rm E(B-V)}$, Cardelli, Clayton & Mathis 1989).
Deep Optical Galaxy Searches
============================
Systematic deeper searches for “partially obscured” galaxies – down to fainter magnitudes and smaller dimensions (${\rm D} {\,\raisebox{-0.4ex}{$\stackrel{>}{\scriptstyle\sim}$}\,}0\farcm1$) than existing catalogs – have been performed with the aim to reduce this ZOA. Using existing sky surveys (POSS I and POSS II in the north and the ESO/SERC surveys in the south), the whole ZOA has in the mean time been visually surveyed for galaxies. Here, examination by eye is still the best technique. A separation of galaxy and star images can as yet not be done by automated measuring machines ([[e.g.]{}]{} COSMOS or APM) on a viable basis below $|b| {\,\raisebox{-0.4ex}{$\stackrel{<}{\scriptstyle\sim}$}\,}10\deg-15\deg$ though surveys by eye are clearly both very trying and time consuming, and maybe not as objective.
The various surveyed regions are displayed in Fig. \[cor\]. Details and results on the uncovered galaxy distributions for the various regions are discussed in:
A: the Perseus-Pisces Supercluster by Pantoja 1997;\
B$_{1-3}$: the northern Milky Way (B$_1$ by Seeberger [[etal.]{}]{} 1994, Seeberger, Saurer & Weinberger 1996, Lercher, Kerber & Weinberger 1996, Saurer, Seeberger & Weinberger 1997, and Seeberger & Saurer 1998 from POSS I; B$_2$ by Marchiotti, Wildauer & Weinberger 1999 from POSS II; B$_3$ by Weinberger, Gajdosik & Zanin 1999 from POSS II);\
C$_{1-3}$: the Puppis region by Saito [[etal.]{}]{} 1990, 1991 \[C$_1$\], the Sagittarius/Galactic region by Roman, Nakanishi & Saito 1998 \[C$_2$\], and the Aquila and Sagittarius region by Roman [[etal.]{}]{} 1996 \[C$_{3}$\];\
D$_{1-5}$: the southern Milky Way (the Hydra to Puppis region \[D$_1$\] by Salem & Kraan-Korteweg in prep., the Hydra/Antlia Supercluster region \[D$_2$\] by Kraan-Korteweg 1999, the Crux region \[D$_3$\] by Woudt 1998, Woudt & Kraan-Korteweg in prep., the GA region \[D$_4$\] by Woudt 1998, Woudt & Kraan-Korteweg in prep, and the Scorpius region \[D$_5$\] by Fairall & Kraan-Korteweg in prep.);\
E: the Ophiuchus Supercluster by Wakamatsu [[etal.]{}]{} 1994, Hasegawa [[etal.]{}]{} 1999;\
F: the northern GP/SGP crossing by Hau [[etal.]{}]{} 1996.
These surveys have uncovered over 50000 previously unknown galaxies and are not biased with respect to any particular morphological type. Although the various optical surveys are based on different plate material and were performed by different groups, the search techniques overall are similar. All ZOA regions have been searched to diameter limits of at least ${\rm D} {\,\raisebox{-0.4ex}{$\stackrel{>}{\scriptstyle\sim}$}\,}0\farcm2$. This is considerably deeper than the previously regarded “whole-sky” catalogs with their completeness limits of ${\rm D} {\,\raisebox{-0.4ex}{$\stackrel{>}{\scriptstyle\sim}$}\,}1\farcm3$. How can these catalogs be merged to arrive at a well-defined whole-sky galaxy distribution with a reduced ZOA?
Completeness of Optical Galaxy Searches
=======================================
In order to merge the various deep optical ZOA surveys with existing galaxy catalogs, Kraan-Korteweg (1999) and Woudt (1998) have analyzed the completeness of their ZOA galaxy catalogs – the Hy/Ant \[D$_2$\], Crux \[D$_3$\] and GA \[D$_4$\] region – as a function of the foreground extinction.
By studying the apparent diameter distribution as a function of the extinction E(B-V) (Schlegel [[etal.]{}]{} 1998) as well as the location of the flattening in the slope of the cumulative diameter curves $\log {\rm D}$-$\log{\rm N}$ for various extinction intervals ([[cf.]{}]{} Fig. 5 and 6 in Kraan-Korteweg 1999), we conclude that our optical ZOA surveys are complete to an apparent diameter of ${\rm D} = 14\arcsec$ – where the diameters correspond to an isophote of 24.5 mag/arcsec${^2}$ (Kraan-Korteweg 1999) – for extinction levels less than ${\rm A_B} = 3\fm0$.
How about the intrinsic diameters, [[i.e.]{}]{} the diameters galaxies would have if they were unobscured? A spiral galaxy seen through an extinction of ${\rm A_B} = 1\fm0$ will, for example, be reduced to $\sim 80\%$ of its unobscured size. Only $\sim 22\%$ of a (spiral) galaxy’s original dimension is seen when it is observed through ${\rm
A_B} = 3\fm0$. In 1990, Cameron derived analytical descriptions to correct for the obscuration effects by artificially absorbing the intensity profiles of unobscured galaxies. These corrections depend quite strongly on morphological type due to the difference in surface brightness profiles and mean surface brightness between early-type galaxies and spiral galaxies. Applying these corrections, we find that at ${\rm A_B} = 3\fm0$, an obscured spiral or elliptical galaxy at our [*apparent*]{} completeness limit of ${\rm D} =
14\arcsec$ would have an intrinsic diameter of ${\rm D^o} \approx
60\arcsec$ or ${\rm D^o} \approx 50\arcsec$, respectively. At extinction levels higher than ${\rm A_B} = 3\fm0$, an elliptical galaxy with ${\rm D^o} = 60\arcsec$ would appear smaller than the completeness limit ${\rm D} = 14\arcsec$ and might have gone unnoticed. The here discussed optical galaxy catalog should therefore be complete to ${\rm
D^o} \ge 60\arcsec$ for galaxies of all morphological types down to extinction levels of ${\rm A_B} \le 3\fm0$ with the possible exception of extremely low-surface brightness galaxies. Only intrinsically very large and bright galaxies – particularly galaxies with high surface brightness – will be recovered in deeper extinction layers. This completeness limit could be confirmed by independently analyzing the diameter vs. extinction and the cumulative diameter diagrams for extinction-corrected diameters.
We can thus supplement the ESO, UGC and MCG catalogs – which are complete to ${\rm D} = 1\farcm3$ – with galaxies from optical ZOA galaxy searches that have ${\rm D^o} \ge 1\farcm3$ and ${\rm A_B}
\le 3\fm0$. As our completeness limit lies well above the ESO, UGC and MCG catalogs, we can assume that the other similarly performed optical galaxy searches in the ZOA should also be complete to ${\rm D^o} =
1\farcm3$ for extinction levels of ${\rm A_B} \le 3\fm0$.
In Fig. \[aitc\] we have then taken the first step in arriving at an improved whole-sky galaxy distribution with a reduced ZOA. In this Aitoff projection we have plotted all the UGC, ESO, MCG galaxies that have [*extinction-corrected*]{} diameters ${\rm D^o} \ge 1\farcm3$ (remember that galaxies adjacent to the optical galaxy search regions are also affected by absorption though to a lesser extent: ${\rm
A_B} \le 1\fm0$), and added the galaxies from the various optical surveys with ${\rm D^o} = 1\farcm3$ and ${\rm A_B} \le 3\fm0$ for which positions and diameters were available. The regions for which these data are not yet available are marked in Fig. \[aitc\]. As some searches were performed on older generation POSS I plates, which are less deep compared to the second generation POSS II and ESO/SERC plates, an additional correction was applied to those diameters, [[i.e.]{}]{} the same correction as for the UGC galaxies which also are based on POSS I survey material (${\rm D_{25} = 1.15 \cdot
D_{POSS I}}$).
A comparison of Fig. \[ait\] with Fig. \[aitc\] demonstrates convincingly how the deep optical galaxy searches realize a considerable reduction of the ZOA; we can now trace the large-scale structures in the nearby Universe to extinction levels of ${\rm A_B} = 3\fm0$. Inspection of Fig. \[aitc\] reveals that the galaxy density enhancement in the GA region is even more pronounced and a connection of the Perseus-Pisces chain across the Milky Way at $\ell=165\deg$ more likely. Hence, these supplemented whole-sky maps certainly should improve our understanding of the velocity flow fields and the total gravitational attraction on the Local Group.
Redshift follow-ups of well-defined samples of ZOA galaxies will be important in analyzing the large-scale structures in redshift-space. Systematic redshift surveys have been performed for various ZOA regions and revealed a number of dynamically important structures such as
– the rich, massive ($\sim 2-5 \cdot 10^{15}{\mbox{${\rm M}_\odot$}}$) cluster A3627 at ($\ell,b,v)\sim(325\deg,-7\deg,4882~$) which seems to constitute the previously unrecognized but predicted density peak at the bottom of the potential well of the Great Attractor (Kraan-Korteweg [[etal.]{}]{} 1996)\
– the 3C129 cluster at ($\ell,b,v)\sim(160\deg,0\deg,5500$ ) connecting Perseus-Pisces and A569 across the Galactic Plane (Chamaraux [[etal.]{}]{} 1990, Pantoja [[etal.]{}]{} 1997)\
– and the Ophiuchus supercluster at ($\ell,b,v)\sim(0\deg,8\deg,8500$ ) behind the Galactic Center (Wakamatsu [[etal.]{}]{} 1994, Hasegawa [[etal.]{}]{} 1999).
Optical galaxy searches, however, fail in the most opaque part of the Milky Way, the region encompassed by the ${\rm A_B} = 3\fm0$ contour in Fig. \[aitc\] – a sufficiently large region to hide further dynamically important galaxy densities. Here, systematic [HI]{} surveys have proven very powerful as the Galaxy is transparent to the 21-cm line radiation of neutral hydrogen and [HI]{}-rich galaxies can readily be found through detection of their redshifted 21-cm emission.
H Galaxy Searches in the ZOA
============================
In March 1997, a systematic blind [HI]{} survey began in the most opaque part of the southern Milky Way ($213\deg \le \ell \le 33\deg$; $|b|
\le 5\fdg5$) with the Multibeam (MB) receiver (13 beams in the focal plane array) at the 64m Parkes telescope. The ZOA is being surveyed along constant Galactic latitudes in 23 contiguous fields of length $\Delta\ell=8\deg$. The ultimate goal is 25 scans per field where adjacent strips will be offset in latitude by $\Delta b = 1\farcm5$ for homogeneous sampling. With an effective integration time of 25 min/beam we obtain a 3$\sigma$ detection limit of 25mJy. The survey covers a velocity range of $-1200 {\,\raisebox{-0.4ex}{$\stackrel{<}{\scriptstyle\sim}$}\,}v {\,\raisebox{-0.4ex}{$\stackrel{<}{\scriptstyle\sim}$}\,}12700$ with a channel spacing of 13.2 per channel, and will be sensitive to normal spiral galaxies well beyond the Great Attractor region.
So far, a shallow survey based on 2 out of the foreseen 25 driftscan passages has been analyzed ([[cf.]{}]{} Kraan-Korteweg, Koribalski & Juraszek 1998, Henning [[etal.]{}]{} 1999). 107 galaxies were catalogued with peak [HI]{}-flux densities of ${\,\raisebox{-0.4ex}{$\stackrel{>}{\scriptstyle\sim}$}\,}$80 mJy (${\rm rms} = 15$ mJy after Hanning smoothing) and their detection show no dependence on Galactic latitude, nor the amount of foreground obscuration through which they have been detected.
Four cubes centered on the Great Attractor region ($300\deg \ge \ell
\ge 332\deg$, $|b| \le 5\fdg5$) of the full-sensitivity survey have been analyzed (Juraszek [[etal.]{}]{} 1999) and uncovered 236 galaxies above the $3\sigma$ detection level of 25 mJy. 70% of the detections had no previous identification.
In the left panel of Fig. \[MBGA\], a sky distribution centered on the GA region displays all galaxies with redshifts ${\rm v} \le
10000$ . Next to redshifts from the literature (circles; LEDA), redshifts from the follow-up observations of Kraan-Korteweg and collaborators in the Hy/Ant-Crux-GA ZOA surveys (dashed area) are plotted. They clearly reveal the prominence of the cluster A3627 at $(\ell,b,v) = (325\deg,-7\deg,4882$ , Kraan-Korteweg [[etal.]{}]{} 1996) close to the core of the GA region at $(\ell,b,v) =
(320\deg,0\deg,4500$ ) as predicted by Kolatt [[etal.]{}]{} (1995). Adding now the new detections from the systematic blind [HI]{} MB-ZOA survey (box), we for the first time can trace structures all the way across the Milky Way. The new picture seems to suggest that the GA overdensity is a “great-wall” like structure starting at the Pavo cluster, having its core at the A3627 cluster and then bending over towards shorter longitudes across the ZOA.
This becomes even clearer in the right panel of Fig. \[MBGA\] where the galaxies are displayed in a redshift cone out to ${\rm v} \le
10000$ for the longitude range $300\deg \le \ell \le 332\deg$ analyzed so far of the MB full-sensitivity data. The A3627 cluster is clearly the most massive galaxy cluster uncovered by the various surveys in the GA region and therefore the most likely candidate for the previously unidentified but predicted density-peak at the bottom of the potential well of the GA overdensity.
Finding a hitherto uncharted further cluster of galaxies at the heart of the GA would have serious implications for our current understanding of this massive overdensity in the local Universe. Various indications suggest that PKS1343-601, the second brightest extragalactic radio source in the southern sky ($f_{20cm} = 79$ Jy, McAdam 1991), might form the center of yet another highly obscured rich cluster, particularly as it also shows significant X-ray emission ([[cf.]{}]{}Kraan-Korteweg & Woudt 1999 for further details). At ($\ell, b) \sim
(310\deg,2\deg)$, this radio galaxy lies behind an obscuration layer of about 12 magnitudes of extinction in the B-band, hence optical surveys are ineffective. Still, West & Tarenghi (1989) observed this source and identified it – with an extinction-corrected diameter of ${\rm D^o} \approx 4\arcmin$ and a recession velocity of v = 3872 – as a giant elliptical galaxy. Giant ellipticals generally reside at the cores of clusters.
Interestingly enough, the [HI]{} MB survey does uncover a significant excess of galaxies at this position in velocity space ([[cf.]{}]{} Fig. \[MBGA\]). However, we do not see a “finger of God”, the characteristic signature of a cluster in redshift space. Could it be that too many central cluster galaxies are missed by the [HI]{} observations because spiral galaxies generally avoid the cores of clusters? The existence of this possible cluster still remains a mystery. Meanwhile, this prospective cluster has been imaged in the near infrared (Woudt [[etal.]{}]{} in progress), where extinction effects are less severe compared to the optical, and which should uncover early-type cluster members if they are there. The forthcoming results should then unambiguously settle the question whether another cluster forms part of the GA supercluster.
Acknowledgements {#acknowledgements .unnumbered}
================
The collaborations with our colleagues in the optical surveys and redshift follow-up observations, C. Balkowski, V. Cayatte, A.P. Fairall, C. Salem, P.A. Woudt, and the HIPASS ZOA team members R.D. Ekers, A.J. Green, R.F. Haynes, P.A. Henning, M. J. Kesteven, B. Koribalski, R.M. Price, E. Sadler and L. Staveley-Smith are greatly appreciated.
References {#references .unnumbered}
==========
Cameron L.M. 1990 A&A 233, 16
Chamaraux P., Cayatte V., Balkowski C., Fontanelli P. 1990 A&A 229, 340
Cardelli J.A., Clayton G.C., Mathis J.S. 1989 ApJ 345, 245
Fairall A.P., Kraan-Korteweg R.C., in prep. \[D$_5$\] Hasegawa T., Wakamatsu K., Malkan M. [[etal.]{}]{} 1999 MNRAS, in press \[E\] Hau G.K.T., Ferguson H.C., Lahav O. [[etal.]{}]{} 1996 MNRAS 277, 125 \[F\] Henning P.A., Staveley-Smith L., Kraan-Korteweg R.C., Sadler E.M. 1999 PASA 16, 35
Hudson M., Lynden-Bell D. 1991 MNRAS 252, 219
Juraszek S., Staveley-Smith L., Kraan-Korteweg R.C., Green A.J., Ekers R.D., Henning P.A., Koribalski B.S., Sadler E.M., Schröder A.C., in prep.
Kogut A., Lineweaver C., Smoot G.F [[etal.]{}]{} 1993 ApJ 419, 1
Kolatt T., Dekel A., Lahav O. 1995 MNRAS 275, 797 Kraan-Korteweg R.C. 1999 A&ASS, submitted
Kraan-Korteweg R.C., Woudt P.A. 1999 PASA 16, 53
Kraan-Korteweg R.C., Woudt P.A., Cayatte V. [[etal.]{}]{} 1996 Nature 379, 519 Kraan-Korteweg R.C., Koribalski B., Juraszek S. 1998 in Looking Deep in the Southern Sky, eds. R. Morganti, W. Couch, Springer, 23
Lahav O. 1987 MNRAS 225, 213
Lauberts, A. 1982 The ESO/Uppsala Survey of the ESO (B) Atlas, ESO, Garching
Lercher G., Kerber F., Weinberger R. 1996 A&ASS 117, 369 \[B$_1$\] Lynden-Bell D., Lahav, O. 1988 in Large-Scale Motions in the Universe, eds. V.C. Rubin and G.V. Coyne, Princeton: Princeton University, 199
Marchiotti W., Wildauer H., Weinberger R. 1999 in progress \[B$_2$\] McAdam W.B. 1991 PASA 9, 255
Nilson P. 1973 Uppsala General Catalog of Galaxies, Uppsala, University of Uppsala
Pantoja C.A., Altschuler D.R., Giovanardi C., Giovanelli R. 1997 AJ 113, 905 \[A\] Roman A.T., Nakanishi K., Tomita A., Saito M. 1996 PASJ 48, 679 \[C$_3$\] Roman A.T., Nakanishi K., Saito M. 1998 PASJ 50, 37 \[C$_2$\] Saito M., Ohtani A., Asomuna A. [[etal.]{}]{} 1990 PASJ 42, 603 \[C$_1$\] Saito M., Ohtani A., Baba A. [[etal.]{}]{} 1991 PASJ 43, 449 \[C$_1$\] Salem C., Kraan-Korteweg R.C., in prep. \[D$_1$\] Saurer W., Seeberger R., Weinberger R. 1997 A&ASS 126, 247 \[B$_1$\] Schlegel D.J., Finkbeiner D.P., Davis M. 1998 ApJ 500, 525
Seeberger R., Saurer W., Weinberger R., Lercher G. 1994 in Unveiling Large-Scale Structures Behind the Milky Way, eds. C. Balkowski, R.C. Kraan-Korteweg, ASP Conf. Ser. 67, 81 \[B$_1$\] Seeberger R., Saurer W., Weinberger R. 1996 A&ASS 117, 1 \[B$_1$\]\] Seeberger R., Saurer W. 1998 A&ASS 127, 101 \[B$_1$\] Vorontsov-Velyaminov B., Archipova V. 1963-74 Morphological Catalog of Galaxies, Parts 2 to 5, Moscow, Moscow University
Wakamatsu K., Hasegawa T., Karoji H. [[etal.]{}]{} 1994 in Unveiling Large-Scale Structures Behind the Milky Way, eds. C. Balkowski, R.C. Kraan-Korteweg, ASP Conf. Ser. 67, 131 \[E\] Weinberger R., Gajdosik M., Zanin C. 1999 A&ASS 137, 293 \[B$_5$\] West, R.M., Tarenghi M. 1989 A&A 223, 61
Woudt P.A. 1998 Ph.D. thesis, Univ. of Cape Town. \[D$_3$, D$_4$\] Woudt P.A., Kraan-Korteweg R.C., A&ASS, in prep. \[D$_3$\] Woudt P.A., Kraan-Korteweg R.C., A&ASS, in prep. \[D$_4$\]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Active regions are open wounds in the Sun’s surface. Seismic oscillations from the interior pass through them into the atmosphere, changing their nature in the process to fast and slow magneto-acoustic waves. The fast waves then partially reflect and partially mode convert to upgoing and downgoing Alfvén waves. The reflected fast and downgoing Alfvén waves then re-enter the interior through the active regions that spawned them, infecting the surface seismology with signatures of the atmosphere. Using numerical simulations of waves in uniform magnetic fields, we calculate the upward acoustic and Alfvénic losses in the atmosphere as functions of field inclination and wave orientation as well as the Time-Distance ‘travel time’ perturbations, and show that they are related. Travel time perturbations relative to quiet Sun can exceed 40 seconds in 1 kG magnetic field. It is concluded that active region seismology is indeed significantly infected by waves leaving and re-entering the interior through magnetic wounds, with differing travel times depending on the orientation of the wave vector relative to the magnetic field. [This presages a new directional-time-distance seismology.]{}'
author:
- |
Paul S. Cally[^1] and Hamed Moradi[^2]\
Monash Centre for Astrophysics and School of Mathematical Sciences, Monash University, Victoria, Australia 3800
bibliography:
- 'fred.bib'
title: Seismology of the Wounded Sun
---
Sun: helioseismology – Sun: oscillations – Sun: magnetic fields
Introduction
============
Classical local helioseismology [@GizBir05aa] uses Doppler velocity or intensity data from near the solar surface to probe the Sun’s interior: thermal, magnetic, and flow properties may all be inferred with varying degrees of reliability. The overlying atmosphere, the chromosphere and corona, are generally assumed to not be important in this process, though a separate ‘coronal seismology’ [@NakVer05aa; @De-05aa; @SteZaiNak12aa] using observations from coronal heights and aimed at determining the physical characteristics of coronal structures has developed in recent years. It is well-appreciated that the Sun’s internal waves manifest in the atmosphere as well, particularly in active regions [@BogJud06aa; @JefMcIArm06aa; @KhoCal13aa], but any back-reaction on the interior seismology has heretofore been largely ignored [though see @KhoCol09aa for an explanation of acoustic halos that relies on reflected fast waves]. We argue that this neglect can lead to serious errors in interpretation of seismic data.
![Schematic diagram ($x$-$z$ plane) depicting fast/slow conversion/transmission at the equipartition level $a=c$, and fast-to-Alfvén conversion in a nebulous region near the fast wave reflection height. The magnetic field lines are shown as background. Conversion to upward or downward propagating Alfvén waves depends in the relative orientation of the magnetic field and the wave propagation direction. Both panels should be regarded as projections, with the fast wave raypath having a component in the $y$-direction; otherwise, there is no fast-Alfvén coupling. (Adapted from @KhoCal12aa) \[fig:Alfschem\]](\figdir/Alfschem){width="\hsize"}
The most widely used local helioseismic technique is Time-Distance Helioseismology [TD; @DuvJefHar93aa; @GizBir05aa]. By correlating observations of perturbations at different times and positions on the solar surface, a causative link is inferred and a travel time between pairs of points determined. Comparing these travel times with those calculated for a standard model, one infers the presence of wave-speed anomalies beneath the surface that may be due to such features as temperature variations or flows. Utilizing the full gamut of wave paths sampled by one’s data set, inversions may be performed for subsurface structure. However, ‘travel time’ is not necessarily as straightforward a concept as we might hope. Proper group travel time is difficult to measure with any precision, and in practice phase variations are used as a travel time proxy.
This approach seems well-founded in quiet Sun, where magnetic effects may be ignored, but is subject to several uncertainties and complications in and around active regions dominated by magnetism in the surface layers and above. To set the scene, the quiet Sun’s internal $p$-modes ($p$ for pressure) are essentially acoustic waves travelling in an unmagnetised stratified medium, with dispersion relation $\omega^2-\omega_c^2-c^2k_h^2=c^2k_z^2$, where $\omega$ is the frequency, $c$ is the sound speed (which increases with depth), $k_h$ and $k_z$ are the horizontal and vertical components of the wave vector, and $\omega_c$ is the acoustic cutoff frequency [commonly defined by $\omega_c^2=(c^2/4H^2)(1-2\, dH/dz)$ where $H$ is the density scale height, but see @SchFle98aa for a critique of such formulæ]. Vertical propagation is therefore limited to the cavity where the left hand side is positive. At depth, where $c^2k_h^2\gg\omega_c^2$ typically, the lower turning point is around the Lamb depth $\omega^2=c^2 k_h^2$, whilst at the upper turning point $\omega^2=\omega_c^2+c^2 k_h^2\geq\omega_c^2$. The introduction of strong magnetic field in surface layers has several effects to alter this picture:
1. Magnetic field alters the density and thermal structure of the plasma by supplying additional magnetic forces, requiring the plasma pressure to adjust accordingly (the ‘indirect’ effect). Changes in the sound speed naturally affect wave travel times.
2. Magnetohydrodynamic (MHD) wave modes become available: fast, slow, and Alfvén. At depth, where the Alfvén speed $a$ is much less than the sound speed $c$, the $p$-modes are effectively fast waves. But on passing through the equipartition layer $a=c$ they partially transmit as acoustic waves (now slow) and partially convert to magnetically dominated fast waves [@Cal06aa; @SchCal06aa; @Cal07aa], depending on the attack angle between the wave vector and the magnetic field. In sunspot umbrae, $a=c$ is typically situated several hundred kilometres below the surface, whilst in penumbrae it is around the surface [@MatSolLag04aa]. In the regions surrounding sunspots, $a=c$ may crudely be equated with the level of the magnetic canopy in the low atmosphere.
3. The transmitted slow waves above $a=c$ are essentially field-guided acoustic waves. Unlike their non-magnetic cousins though, they may still propagate vertically at frequencies below the acoustic cutoff (around 5 mHz), provided $\omega>\omega_c\cos\theta$, where $\theta$ is the field inclination from the vertical [@BelLer77aa]; the so-called ‘magneto-acoustic portals’ [@JefMcIArm06aa]. This ramp effect has recently been shown computationally to be far more important than radiation in modifying the acoustic cutoff [@HegHanDe-11aa].
4. Whether or not the slow wave escapes into the solar atmosphere in strong field regions, the fast wave certainly does. It is immune to the acoustic cutoff effect, but nevertheless has its own upper turning point substantially higher in the atmosphere. For simplicity, assume that this is high enough that $a\gg c$ and therefore ignore sound speed altogether. The dispersion relation is then $\omega^2=a^2(k_h^2+k_z^2)$, indicating reflection where $\omega^2=a^2k_h^2$. This is an important point: the fast wave reflects where its horizontal phase speed $\omega/k_h$ coincides with the local Alfvén speed. After it reflects, the fast wave re-enters the solar interior wave field, and therefore is part of the seismology of the Sun. Its journey through the atmosphere must therefore have some effect on wave timings.
However, that is not the full story. The fast wave itself may partially mode-convert to upward or downward propagating Alfvén waves, depending on magnetic field orientation relative to the wave vector [@CalGoo08aa; @CalHan11aa; @HanCal12aa; @KhoCal11aa; @KhoCal12aa; @Fel12aa], thereby removing energy from the seismic field and potentially altering its phase (see the schematic diagram Figure \[fig:Alfschem\]). This conversion typically occurs in a broad region near the upper turning point of the fast wave, but only if the wave is propagating across the vertical plane containing the magnetic field lines; we let $\phi$ denote the angle between the vertical magnetic flux and wave propagation planes. Such a phase change could be wrongly interpreted in TD as a travel time shift [@Cal09ab; @Cal09aa].
So, there are many processes going on here: fast/slow conversion/transmission of upward travelling waves at $a=c$; slow wave reflection or transmission, depending on the acoustic cutoff and the magnetic field inclination $\theta$; fast-to-Alfvén conversion near the fast wave turning point; partial reflection of the fast wave to re-enter the seismic field; further fast/slow conversion/transmission as the fast wave passes downward through $a=c$. In totality, this is too complicated to model analytically. However, knowing the physics of the constituent parts, a full numerical simulation can provide valuable insight. The novel aspect explored here is to simultaneously calculate the TD travel time shifts $\delta\tau$ and upward acoustic and Alfvénic losses in a simulation as functions of $\theta$ and $\phi$ (for selected frequencies and horizontal wavenumbers).
The considerations set out above indicate that active regions are truly wounds in the surface of the Sun. They allow waves and wave energy to escape from the interior cavity wherein they are normally trapped, but also allow their depleted, phase-shifted, and mode-converted remnants to re-enter and infect the local seismology that we rely on for information about the sub-surface. The purpose of this paper is to look for correspondences between acoustic and magnetic wave losses on the one hand, and ‘travel time’ shifts on the other, and thereby to verify that the seismology of the wounded Sun depends on propagation, transmission, and conversion in overlying active region atmospheres. [We introduce a directionally filtered time-distance approach[^3] to simulation data that is directly extensible to real helioseismic data, and thereby envisage a Directional-Time-Distance (DTD) seismology sensitive to magnetic field orientation. In Section \[sec:BVP\], the DTD results are verified using the Boundary Value approach of @Cal09ab.]{}
[ The purpose of this paper is to demonstrate that the above mode conversion scenarios actually operate in the solar context; to survey how they depend on field inclination $\theta$ and wave orientation $\phi$; and to assess the utility of DTD in discerning directional magnetic effects. To this end, we explore only vertical atmospheric stratification, characterized by density scale heights of order 0.15 Mm, and ignore magnetic field inhomogeneity, which in sunspots has scale length of order many Mm. A subsequent paper will address wave simulations in realistic sunspot models using DTD. This is not to say though that flux tubes are not important to wave propagation, though probably they are more relevant to the corona where the far greater gravitational scale height is no longer the primary feature, leaving complex loop structures to dominate [@VanBraVer08aa]. Chromospheric sunspot fields, dominated be a single monopolar spot, may be expected to be smoother. ]{}
Computational Details
=====================
We employ the Seismic Propagation through Active Regions and Convection (SPARC) $6^{\rm th}$ order linear MHD code of @Han07aa, with uniform $B_0=500$ G or 1 kG magnetic fields of inclination $\theta$ from the vertical threading the Convectively Stabilized solar Model (CSM\_B) atmosphere of @SchCamGiz11aa. The ‘observational height’ for vertical velocities used in seismic calculations is 0.3 Mm, above the $a=c$ equipartition layer for both 500 G and 1 kG (0.24 Mm and 0.08 Mm respectively). A stochastic driving plane is placed at depth $z=-0.156$ Mm. The random sources produce a solar wave power spectrum with maximum power in the range 2 – 5.5 mHz, peaking at about 3.2 mHz. Being a linear code, the overall velocity normalization is arbitrary. The standard ${\leavevmode\kern.1em\raise.5ex\hbox{\the\scriptfont0 2}\kern-.1em/\kern-.15em\lower.25ex\hbox{\the\scriptfont0 3}}$ dealiasing rule is applied to avoid spectral blocking, with $|k_x|$ and $|k_y|$ limited to about 1.9 $\rm Mm^{-1}$.
A $140\,{\rm Mm}\times140\,{\rm Mm}\times26.53\,{\rm Mm}$ box is spanned by a $128\times128\times265$ grid of spacings $\Delta x=\Delta y=1.09$ Mm covering heights $-25\,{\rm Mm} \le z \le 1.53\,{\rm Mm}$ with nonuniform $\Delta z$. Periodic lateral boundaries and absorbing PML layers at top (starting at $z_t=1.26$ Mm) and bottom complete the definition of the computational region.
Without loss of generality, the field lines are assumed to lie parallel to the $x$-$z$ plane. As with most such codes, SPARC normally employs an Alfvén speed ‘limiter’ or ‘controller’ (or similar) in the atmosphere to avoid infeasibly small time steps required to satisfy the CFL numerical stability condition [*e.g.*. @Han08aa; @RemSchKno09aa; @CamGizSch11aa]. These commonly cap the Alfvén speed at 20–60 $\rm km\,s^{-1}$. We do not impose such a limiter though, as it corrupts the very processes of fast wave reflection and conversion that we seek to explore [@MorCal13aa]. This is why we adopt lower box heights than usual, so as to keep the peak Alfvén speed manageable. Numerical experiment has allowed us to select box heights sufficient to encompass fast-to-Alfvén conversion without mandating impossibly small time steps. For the 500 G case we use $\Delta t=0.1$ s, and half that for 1 kG.
Data cubes are produced from the simulations consisting of vertical velocity values at each grid point at observation height $z_{\rm obs}=0.3$ Mm over typically $8{\leavevmode\kern.1em\raise.5ex\hbox{\the\scriptfont0 1}\kern-.1em/\kern-.15em\lower.25ex\hbox{\the\scriptfont0 2}}$ hours.
Filtering is applied in Fourier space, both frequency and wavevector. This is more efficient than repeating the simulation for a large number of monochromatic drivers. Wavevector filtering uses a circular Gaussian ball filter with standard variation $\sigma=0.1$ $\rm Mm^{-1}$ centred at a particular horizontal wavevector ${{\bmath{k}}}={{\bmath{k}}}_h$ oriented angle $\phi$ from the $x$-direction. Horizontal wavenumber $k_h=|{{\bmath{k}}}_h|=0.5$, 0.75, or 1.0 $\rm Mm^{-1}$ determines the skip distance and phase speed, and $\phi$ represents the wave propagation direction. Frequency filtering is centred on 3 mHz and 5 mHz with a 0.5 mHz standard deviation.
$k_h$ ($\rm Mm^{-1}$) $v_{ph}$ ($\rm km\,s^{-1}$) Skip Dist (Mm)
----------------------- ----------------------------- ----------------
1.0 21.8 14.7
0.75 29.3 19.5
0.5 43.8 38.8
: Approximate phase speeds and skip distances associated with each of the three selected horizontal wavenumbers $k_h$ in the quiet solar model. \[tab:skip\]
The resultant filtered data cubes are analysed in two ways. First, the acoustic (slow) and magnetic (Alfvén) wave energy fluxes are calculated at $z_f=1.2$ Mm, just below the PML layer, and plotted as contoured functions of field inclination $\theta=0^\circ$, $10^\circ$, $20^\circ$, …, $90^\circ$ and wave orientation $\phi=0^\circ$, $5^\circ$, $10^\circ$, …, $180^\circ$ for each frequency (3 mHz and 5 mHz). The fast wave is evanescent, and so contributes no flux. Similarly, Time-Distance travel time perturbations relative to quiet Sun (same atmospheric model but no magnetic field) are also plotted against $\theta$ and $\phi$.[^4] These are displayed and analysed in Section \[results\].
Results
=======
In this section, wave energy fluxes at the top of the computational box will be compared with ‘travel time’ perturbations relative to the quiet Sun for a range of magnetic field inclinations $\theta$ and orientations $\phi$, field strengths (500 G and 1 kG), horizontal wavenumbers ($k_h=1$, 0.75, and 0.5 $\rm Mm^{-1}$), and wave frequencies (3 mHz and 5 mHz).
Wave Energy Fluxes {#fluxes}
------------------
{width=".497\hsize"} {width=".497\hsize"}\
{width=".497\hsize"} {width=".497\hsize"}
With magnetic field strengths of 500 G and 1 kG, Figure \[fig:Fz\] depicts the top vertical acoustic and magnetic wave energy fluxes at 3 mHz and 5 mHz as functions of field inclination from the vertical $\theta$ and wave orientation as selected by the wavevector space ball filter $\phi$. The (vector) fluxes are calculated according to $${{{\bmath{F}}}}_{\rm ac}=\langle p\,{{\bmath{v}}}\rangle\,, \quad {{{\bmath{F}}}}_{\rm mag}=\langle {{{\bmath{b}}}}{{\bmath{\times}}}({{\bmath{v}}}{{\bmath{\times}}}{{\bmath{B}}}_0)\rangle/\mu_0\,, \label{flux}$$ where ${{\bmath{v}}}$, $p$, and ${{{\bmath{b}}}}$ are the Eulerian velocity, pressure, and magnetic field perturbations, $\mu_0$ is the magnetic permeability, and the angled brackets indicate averaging over time, or over $x$ and $y$. For the most part, only vertical components of these fluxes will be plotted. Three different horizontal wavenumbers are selected: $k_h=1.0$, 0.75, and 0.5 $\rm Mm^{-1}$, corresponding to phase speeds and skip distances set out in Table \[tab:skip\]. The overall flux normalization is arbitrary, but is consistent with the solar-like spectrum of the stochastic driving plane across the frequencies, *i.e.*, having maximum power in the range 2 – 5.5 mHz with peak at about 3.2 mHz.
The features in Figure \[fig:Fz\] are very much in accord with the simplified analyses of @CalGoo08aa and @KhoCal11aa. Specifically, there is negligible acoustic power at low field inclination because of the acoustic cutoff of a little over 5 mHz in the atmosphere. However, once $\omega > \omega_c\cos\theta$, acoustic waves may propagate upward (the ramp effect), and substantial flux is recorded at the top of the box. At 3 mHz, acoustic power peaks at $\theta=50^\circ$– $60^\circ$, whilst the peak is at around $30^\circ$ at 5 mHz. These peak powers may be understood in terms of the ‘attack angle’ $\alpha$ between the 3D wavevector and the magnetic field at the equipartition level [@Cal06aa; @SchCal06aa; @HanCal09aa], with most efficient acoustic transmission occurring at small $\alpha$ in the generalized ray approximation. This also explains the drop-off of acoustic flux with increasing $\phi$. The slight shift in the maximum of acoustic flux away from $\phi=0$ is consistent with the 3 mHz case of @KhoCal11aa.
The magnetic flux also accords with previous (more idealized) modelling [@CalGoo08aa], peaking at higher $\theta$ than the acoustic flux (especially at 5 mHz), and at wave orientations of $60^\circ$ or more. The bias toward $\phi<90^\circ$ is in accord with the schematic scenario of Figure \[fig:Alfschem\] and the detailed cold plasma survey of @CalHan11aa. The corresponding downgoing Alfvén waves at $\phi>90^\circ$ are not registered by these figures. Strictly, there should be no Alfvénic flux at $\phi=0^\circ$ and $180^\circ$ because in those 2D cases the fast and Alfvén waves decouple. However, the finite width of the ball filter in wavevector space, the incomplete separation of the fast and Alfvén waves at $z_f=1.2$ Mm, and other numerical inexactitudes result in weak magnetic fluxes at those orientations. Practical time step constraints preclude us from using a taller box and higher $z_f$, where fast-to-Alfvén conversion is more complete, or applying magnetic fields greater than 1 kG, without invoking some form of Alfvén limiter.
Another feature evident in Figure \[fig:Fz\] is increasing acoustic and magnetic flux with increasing $k_h$ in each case. This is due to the velocity power distribution in the stochastic driver, which is uniform in ${{\bmath{k}}}_h$. That leads to RMS power in $p$ and ${{{\bmath{b}}}}$ that increases linearly with wavenumber, and hence to a similar trend in flux.
![Representative examples of vertical wave energy fluxes (full curve: acoustic; dashed curve: magnetic; dotted curve: total) as a function of height. The top frame pertains to $\theta=60^\circ$, $\phi=80^\circ$, $B_0=1$ kG, and $k_h=1$ $\rm Mm^{-1}$ at 5 mHz, for which significant fast-to-Alfvén conversion is expected (see Figure \[fig:Fz\]). The bottom frame is for the corresponding vertical field case $\theta=0^\circ$ (for which $\phi$ is irrelevant). The full vertical lines indicates the location of the $a=c$ equipartition height, and the vertical dashed line shows the position of the stochastic driving plane.[]{data-label="fig:FzVz"}](figures/fz_th60ph80_5mHz_1kG "fig:"){width=".75\hsize"}\
![Representative examples of vertical wave energy fluxes (full curve: acoustic; dashed curve: magnetic; dotted curve: total) as a function of height. The top frame pertains to $\theta=60^\circ$, $\phi=80^\circ$, $B_0=1$ kG, and $k_h=1$ $\rm Mm^{-1}$ at 5 mHz, for which significant fast-to-Alfvén conversion is expected (see Figure \[fig:Fz\]). The bottom frame is for the corresponding vertical field case $\theta=0^\circ$ (for which $\phi$ is irrelevant). The full vertical lines indicates the location of the $a=c$ equipartition height, and the vertical dashed line shows the position of the stochastic driving plane.[]{data-label="fig:FzVz"}](figures/fz_th0ph80_5mHz_1kG "fig:"){width=".75\hsize"}
Figure \[fig:FzVz\] shows how the acoustic, magnetic, and total fluxes vary with height in two representative cases, the first with highly inclined magnetic field and the second with vertical field. Most fluxes have settled quite well to their asymptotic states by about $z_f=1.2$ Mm, indicating that our computational box is (for the most part) tall enough. There is also clearly a transfer of energy flux from acoustic to magnetic in the inclined field case over several hundred kilometres above $a=c$, as expected [@CalHan11aa]. This is absent in the vertical field case, as there the fast and Alfvén waves are decoupled.
Another reason for the (moderate) differences between the flux maps presented here and those of @CalGoo08aa is the different way that waves are injected from below. In @CalGoo08aa, care is taken to inject only a pure fast wave at $z=-4$ Mm, where it is overwhelmingly acoustic. Any magnetic wave at the top of the computational box could therefore have only originated from mode transmission/conversion. The 2.5D simulations of @KhoCal11aa [@KhoCal12aa] similarly impose a deep fast wave driver (at $-5$ Mm). In the current SPARC simulations though, a shallow more realistic solar driver is placed at $z=-0.156$ Mm, and this excites both acoustic and magnetic oscillations directly (see Figure \[fig:FzVz\]). Consequently, some portion of the magnetic wave energy fluxes at the top may have travelled directly along the fast and Alfvén dispersion relation loci rather than tunnelling from the fast branch. Nevertheless, the correspondence between the flux maps presented here and in @CalGoo08aa is striking.
![Pure field-aligned acoustic and Alfvén fluxes $F_{\parallel\rm ac}$ (left column) and $F_{\rm Alf}$ (right column) for the case $B_0=1$ kG, $\nu=5$ mHz, $k_h=1$ 0.75, and 0.5 $\rm Mm^{-1}$ (top to bottom).[]{data-label="fig:Fpar"}](\figdir/Fparalf_5mHz_1kG){width="\hsize"}
{width=".4\hsize"} {width=".4\hsize"}\
{width=".4\hsize"} {width=".4\hsize"}
As the magnetic field inclination $\theta$ increases toward $90^\circ$, the vertical acoustic and magnetic fluxes obviously diminish since asymptotically, as $z_f\to\infty$, both are identically field-aligned. They therefore give a biased view of energy loss rates; even for purely horizontal magnetic field, the parallel acoustic and Alfvén waves take energy away (horizontally) from the sites of mode conversion. Using vector identities, it is easily shown that ${{{\bmath{F}}}}_{\rm mag} = p_{\rm mag}{{\bmath{v}}}+{{{\bmath{b}}}}{{\bmath{\cdot}}}{{\bmath{v}}}\,{{\bmath{B}}}_0/\mu_0$, where $p_{\rm mag}={{\bmath{B}}}_0{{\bmath{\cdot}}}{{{\bmath{b}}}}/\mu_0$. This comprises the magnetic flux contributions to both the magnetoacoustic and Alfvén waves. To select the ‘pure’ Alfvén contribution, we project both ${{{\bmath{b}}}}$ and ${{\bmath{v}}}$ in the asymptotic Alfvén wave polarization direction $
\hat{{{\bmath{E}}}}_{\rm Alf}=(\cos^2\theta\sin\phi,\cos\phi,\sin\theta\cos\theta\sin\phi)/\sqrt{1-\sin^2\theta\sin^2\phi}
$ (this is the ${{\hat{\bmath{e}}}}_{\rm perp}$ of @KhoCal11aa rotated into our present frame and normalized) to obtain the components $b_{\rm Alf}$ and $v_{\rm Alf}$, and identify $F_{\rm Alf}=-\langle b_{\rm Alf} v_{\rm Alf} B_0\rangle/\mu_0$. Similarly, $F_{\parallel\rm ac}=\langle p\,{{\bmath{v}}}{{\bmath{\cdot}}}\hat{{\bmath{B}}}_0\rangle$ is the pure field-aligned acoustic flux. Figure \[fig:Fpar\] plots these pure parallel acoustic and Alfvén fluxes against $\phi$ and $\theta$.
For horizontal field $\theta=90^\circ$ there is skew-symmetry in Alfvén flux about $\phi=90^\circ$, as there must be, with the Alfvén flux propagating in the positive direction for $\phi<90^\circ$ and the negative direction for $\phi>90^\circ$. As $\theta$ decreases though, the positive flux on $\phi<90^\circ$ is favoured since it follows the field lines sloping upward from the interaction region. On $\phi>90^\circ$ the census height $z_f$ is above the bulk of this interaction region, and so $F_{\rm Alf}$ does not yet exhibit substantial negative values – it will do so as we move lower in the atmosphere. Indeed, if we could reliably measure $F_{\rm Alf}$ on $\phi>90^\circ$ at lower altitudes, these figures would display greater (anti-) symmetry [ (see Section \[sec:BVP\])]{}. The parallel acoustic flux is very similar to the vertical fluxes displayed in the top right panel of Figure \[fig:Fz\], with an additional $\cos\theta$ factor.
Travel Time Perturbations {#TD}
-------------------------
‘Travel time’ perturbations $\delta\tau$ (relative to quiet Sun) are displayed in Figure \[fig:tau\], on a $10^\circ\times10^\circ$ resolution grid. They show clear manifestations of the acoustic cutoff at $\theta=30^\circ$– $40^\circ$ for 5 mHz, and $\theta=50^\circ$– $60^\circ$ at 3 mH where greater inclination is needed to overcome it. Below these inclinations, $\delta\tau$ is small, and typically negative, suggesting a weak speed-up due to the magnetic field. At higher inclinations, substantial travel time perturbations are seen at all orientations $\phi$, most cleanly at 5 mHz.
With only about $3{\leavevmode\kern.1em\raise.5ex\hbox{\the\scriptfont0 1}\kern-.1em/\kern-.15em\lower.25ex\hbox{\the\scriptfont0 2}}$ skips fitting into the computational box for $k_h=0.5$ $\rm Mm^{-1}$ (see Table \[tab:skip\]), the corresponding $\delta\tau$ measurements become very noisy at 3 mHz (not shown). Even the $k_h=0.75$ results appear unreliable at this lower frequency. The large positive $\delta\tau$ at around $\theta=50^\circ$ in the $B_0=500$ G, $\nu=3$ mHz, $k_h=0.75$ $\rm Mm^{-1}$ case should therefore be regarded with some suspicion; it requires confirmation in wider boxes.
There are also clear variations in $\delta\tau$ with $\phi$ at inclinations $\theta$ sufficient for the ramp effect to take hold. In each case, negative travel time perturbations are substantial at small $\sin\phi$ but much reduced at ‘intermediate’ $\phi$, typically around $90^\circ$. Comparison with the flux figures suggests a strong link between both acoustic and magnetic wave energy losses and travel time lags.
For magnetic field inclinations $\theta$ sufficient that $\omega>\omega_c\cos\theta$, the atmosphere is opened up to wave penetration and therefore to both types of mode conversion. With $\phi=0^\circ$, the fast magnetically-dominated waves emerging from $a=c$ return to the surface after reflection near $\omega/k_h=a$ with a different (advanced) phase compared to that of the simply reflecting acoustic waves in quiet Sun. However, if $\phi\ne0^\circ$ (or $180^\circ$), the fast wave loses more energy near its apex to the Alfvén wave, and seeming suffers a further phase retardation that we might interpret as partially cancelling the underlying negative travel time perturbation.
Using acoustic holography to probe sunspot penumbral oscillations, @SchBraCal05aa also detected directionally dependent phase perturbations, finding that phase shift varies with the angle between the line-of-sight and the magnetic field, with equivalent phase travel time variations with viewing angle of order 30 s. These results seem similar in type and magnitude to those found here, though it is difficult to compare precisely since line-of-sight selects fast and slow waves differently, depending on their velocity polarizations with respect to magnetic field. Their result that phase change is minimal along the line of sight is consistent with our understanding that the atmospheric slow (acoustic) wave that will be preferentially selected by that viewing angle has the same phase as the fast wave incident from below on $a=c$ [there is no phase jump in the transmitted wave; see @TraKauBri03aa Eqn. (22)].
Comparison with BVP Results {#sec:BVP}
---------------------------
{width=".8\hsize"}
For the case of a uniform magnetic field and a horizontally invariant atmosphere, the wave propagation equations may be reduced to a $4^\mathrm{th}$ (2D) or $6^\mathrm{th}$ (3D) order system of ordinary differential equations, assuming an $\exp[{\rmn{i}}(k_x x+k_y y-\omega t)]$ dependence on horizontal coordinates and time. Boundary conditions may be applied so as to allow the ingress of only a pure fast wave at the bottom of the computational region (a few Mm below the solar surface), with outgoing radiation or evanescent (as appropriate) conditions on all other waves at top and bottom. This defines an ordinary differential Boundary Value Problem (BVP). Details are set out in @CalGoo08aa, @Cal09ab, and @NewCal10aa.
Figure \[fig:BVP\] displays the results of such a calculation for one of the cases considered above: $B_0=1$ kG, $\nu=5$ mHz, $k_h=1$ $\rm Mm^{-1}$. Travel time perturbations (top left) are calculated according to the prescription of @Cal09ab. These are computed based on the phase of the reflected fast wave when it exits the box at the bottom, compared to the same for quiet Sun. No full seismic skip is constructed, but this assessment of the phase change through the surface layers amounts to the phase change per skip. It is therefore directly comparable to the time-distance results calculated above based on standard TD methodology, and indeed, the correspondence between the results of the two approaches is striking. Experiment with higher magnetic field strength (not shown) indicates that peak travel time perturbations scale roughly linearly with $B_0$, as expected from our Fig. \[fig:tau\] and from Fig. 7 of @Cal09ab. This is sufficient by $B_0=2$ kG for the phase difference $\delta\varphi$ to fold over the $360^\circ$ ambiguity, so that one would not be sure whether it were advanced or retarded. Since $\delta\tau=-\omega^{-1}\delta\varphi$, travel time perturbations may therefore artificially appear positive.
Figure \[fig:BVP\] also displays the acoustic and magnetic fluxes (bottom row) calculated at the top boundary ($z=2$ Mm), though now they are normalized by the injected fast wave flux at the bottom. These are very similar to those obtained above from SPARC simulation, though the peaks of power are shifted a little in $\theta$ and $\phi$, probably because the injected flux in the simulations is not purely fast wave, since the stochastic driver is both shallow and indiscriminate in the type of waves it produces. The greater computational box height for the BVP may also contribute to the difference. It is seen that $F_{\rm ac}$ reaches over 55% and $F_{\rm mag}$ more than 40% in this case, showing that indeed, losses to the upper atmosphere can be substantial.
Total fractional loss $\mathcal{L}$ is also shown (top right). This includes both upgoing and downgoing losses from the fast wave, a quantity difficult to estimate in the SPARC simulations. In other words, the remnant reflected fast wave that ultimately rejoins the interior seismic wavefield has been reduced in power by fraction $\mathcal{L}$, which can be as high as 80%. It is notable that, as expected, $\mathcal{L}$ is more nearly symmetric about $\phi=90^\circ$ than are the upward fluxes alone. This accords with the near-symmetry of $\delta\tau$. The smaller total losses around $\phi=90^\circ$ for fixed $\theta\ga30^\circ$ are clearly reflected in $\delta\tau$. On the other hand, the small losses for $\phi\approx0^\circ$ at high $\theta$ run counter to this nexus, presumably because acoustic and magnetic losses affect phase in different ways.
In passing, we mention that the BVP travel time perturbation map for the case $B_0=0.5$ kG, $\nu=3$ mHz, $k_h=0.75$ $\rm Mm^{-1}$ (not shown) does not display the strong positive feature seen in the bottom left panel of Figure \[fig:tau\] at mid-values of $\theta$ and $\phi$ (there is a positive peak, but it is less than 2 s). This indicates that we must be careful with very noisy TD phase fits. Longer time series or broader wave vector filters may be required.
Discussion and Conclusions
==========================
‘Travel time’ perturbations, as measured by the standard techniques of Time-Distance helioseismology, actually determine phase shifts rather than true (group) travel times. Shifts in phase trivially come about through changes in path and in propagation speed. If Fermat’s principle is valid, then travel time is stationary with respect to variations of path, and so wave speed will be the main culprit. This could consist of perturbations in sound speed (temperature), magnetic field (fast wave speed), and flow (doppler shift). Flow is isolated by contrasting travel times between two points in opposite directions. There is no background flow in our simulations.
However, phase shifts are also produced by wave reflection and mode conversion [@Cal09aa; @TraKauBri03aa]. It is a difficult (and inexact) task to theoretically determine shifts due to a single mode conversion process, but our modelled system consists of several occurring sequentially. It is further complicated by the shallow wave generation layer that excites both acoustic and magnetic waves directly, thereby differing from the model of @CalGoo08aa where care was taken to inject only acoustic waves from the bottom of the computational region [ (see Section \[sec:BVP\])]{}. The stochastic driving layer is however more solar-relevant, and so is of more practical relevance.
To enumerate, the mode transmissions/conversions consist of: fast-to-slow (acoustic) transmission at $a=c$; the resulting slow wave may then either escape upward or be reflected by the acoustic cutoff, depending on frequency and magnetic field inclination; the converted fast wave continues upward to be reflected near where its horizontal phase speed matches the Alfvén speed; conversion to upgoing and downgoing Alfvén waves occurs near this reflection point, depending on field inclination (with or against the wave direction); the downgoing fast wave again passes through $a=c$ and is split into fast and slow; the downgoing Alfvén wave impacts the surface. Each of these in turn leaves a signature on the phase. Altogether, this is far too complex to treat analytically. We are therefore left to analyse simulations by comparing wave fluxes with ‘travel time’ perturbations. Unfortunately, the plotted fluxes do not reveal the full story. Downgoing acoustic and magnetic losses are not easily determined, and in any case are complicated by the excitation layer. Furthermore, phase shifts are not simple functions of conversion or transmission coefficients. Consequently, we should not expect a simple relationship between wave fluxes at the top of the computational domain and travel time discrepancies calculated near the photosphere. Nevertheless, there is clearly a profound link. Especially at 5 mHz, substantial $\delta\tau$ is associated with large fluxes, at field inclinations $\theta$ sufficient to allow acoustic propagation.
The link is perhaps clearer in Figure \[fig:compare\], in which the $\theta=30^\circ$ and $\theta=60^\circ$ fluxes and travel time perturbations with $B_0=1$ kG, $\nu=5$ mHz, $k_h=1$ $\rm Mm^{-1}$ are compared. For $\theta=30^\circ$, acoustic losses dominate, and produce travel time shifts of similar structure in $\phi$. Similarly, at $\theta=60^\circ$, magnetic losses dominate,[^5] and now their structure is mirrored in the $\delta\tau$ plot. Similar behaviour is seen at 3 mHz in Figure \[fig:compare3mHz\].
[ Of course, the results of the Boundary Value Problem (BVP) calculations of Section \[sec:BVP\] are ‘cleaner’ than those from TD, because they are for a monochromatic wave rather than a stochastically excited spectrum of oscillations, and so do not require filtering, nor do they need to be averaged over a finite time series. The TD approach though does presage extension to more realistic atmospheres and field geometries, and importantly is directly applicable to real solar data and not just simulations. The close correspondence between the results of the two methods argues strongly for the viability of directional TD probing of real solar magnetic regions. ]{}
[ Modelling of helioseismic waves is usually carried out in the linear regime, since internal oscillation velocities are invariably highly subsonic. However, our extension of the domain of helioseismology into the chromosphere poses the question of whether linearity is still a valid assumption for the extended waves addressed here. For the most part, the velocities of photospheric oscillations associated with individual p-modes are at most only a few tens of $\rm cm\,s^{-1}$ and do not of themselves grow large enough in the upper chromosphere to be nonlinear. However, it is undoubtedly the case that, regarded as a spectrum of modes, atmospheric acoustic (slow) waves driven by the p-modes in sunspots do steepen and shock with height [@BogCarHan03aa], though the relationship between the photospheric and chromospheric power spectra can be understood broadly in terms of linear theory [@BogJud06aa]. However, this is irrelevant to our considerations. Below the (ramp adjusted) acoustic cutoff frequency, slow waves reflect at too low an altitude for this to occur, and at higher frequencies the slow wave is lost to the seismic wavefield whether it propagates forever upward or is thermalized due to nonlinearities; in either case, the change in phase of the fast wave has already occurred at lower altitudes. By mid-chromospheric heights the plasma $\beta$ is already low and the fast wave is essentially the ‘compressional’ Alfvén wave, travelling at (near) the Alfvén speed, which is now very large. We therefore do not expect significant nonlinearities in the fast waves. The 2D simulations of @BogCarHan03aa identify nonlinear steepening in fast waves only in the neighbourhood of a much more powerful driving piston than is relevant to our p-mode scenario.]{}
Overall, how are our results to be interpreted? At small magnetic field inclination, insufficient to provoke the ramp effect, both fluxes and travel time perturbations relative to quiet Sun are small, suggesting that seismic waves are largely reflected before reaching heights at which they would become involved in mode conversion. In this respect, they do not behave very differently to the quiet Sun case. However, once the ramp effect kicks in at larger $\theta$, wave paths extend into the atmosphere and both ‘travel times’ (wave phases) and energy losses become substantial. This is so even at $\phi=0^\circ$ and $180^\circ$, where Alfveńic losses (essentially) vanish. Further large variations in $\delta\tau$, now positive, are then correlated with the Alfvénic losses as $\sin\phi$ increases. In summary, the current simulations suggest that fast-to-slow conversion at $a=c$ yields large negative travel time shifts, and that subsequent fast-to-Alfvén conversions produce positive shifts superimposed on and therefore partially cancelling the negative shifts. Importantly, $\delta\tau$ displays a very clear directional dependence.
A complication of the current model is that mode conversion is happening at both ends of the skip path, because the atmosphere and magnetic field is horizontally invariant. For sunspot seismology though, normally one end is in the spot and the other in a quiet Sun annular pupil where there is no conversion. This scenario will be explored in a subsequent work. For the moment though, we must realize that the travel time measurements are affected by conversions at both ends of the path, whereas the flux calculations sample only one end. This probably explains why the $\delta\tau$ contour graphs are more symmetric about $90^\circ$ in $\phi$ than are the flux plots, a feature we would not expect to persist in the sunspot/pupil model. Nevertheless, with that caveat in mind, it appears that very significant ‘travel time’ discrepancies of several tens of seconds (depending on field strength, frequency, and wavenumber) are related to phase changes resulting from mode conversion and not true travel time changes (which is not to say that actual travel times have not changed as well due to the routing of fast and slow waves through the atmosphere). This should give pause to helioseismologists attempting to invert TD sunspot data for subsurface structure.
![Comparison of the upward acoustic and magnetic fluxes (left panels) with the travel time perturbations (right panels) against $\phi$ at $\theta=30^\circ$ (top row) and $\theta=60^\circ$ (bottom row), for the case $B_0=1$ kG, $\nu=5$ mHz, $k_h=1$ $\rm Mm^{-1}$, corresponding to the top left panels in each of Figures \[fig:Fz\] and \[fig:tau\]. Acoustic flux: full curve; magnetic flux: dashed curve.[]{data-label="fig:compare"}](figures/compare){width="\hsize"}
![Comparison of the upward acoustic and magnetic fluxes (left panel) with the travel time perturbations (right panel) against $\phi$ at $\theta=70^\circ$, for the case $B_0=1$ kG, $\nu=3$ mHz, $k_h=1$ $\rm Mm^{-1}$. Acoustic flux: full curve; magnetic flux: dashed curve.[]{data-label="fig:compare3mHz"}](figures/compare3mHz){width="\hsize"}
This work was supported by an award under the Merit Allocation Scheme on the NCI National Facility at the ANU, as well as by the Multi-modal Australian ScienceS Imaging and Visualisation Environment (MASSIVE). A portion of the computations was also performed on the gSTAR national facility at Swinburne University of Technology. gSTAR is funded by Swinburne and the Australian Government’s Education Investment Fund.
[^1]: E-mail: [email protected]
[^2]: E-mail: [email protected]
[^3]: A wedge filter was suggested for use in TD by @Gil00aa to measure either rotation or meridional circulation; see his fig. 4.4. @ChoLiaYan09aa employed a similar idea in the construction of an acoustic power map.
[^4]: Travel times are calculated in the standard TD manner [@CouBirKos06aa], after transforming the ball-filtered data back to physical space.
[^5]: The small negative values of $F_{\rm mag}$ near $\phi=0^\circ$ and $180^\circ$ are an indication that the top of the computational box is not quite high enough for the fluxes to have attained their asymptotic values.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In an experiment of oscillatory media, domains and walls are formed under the parametric resonance with a frequency double the natural one. In this bi-stable system, nonequilibrium transition from Ising wall to Bloch wall consistent with prediction is confirmed experimentally. The Bloch wall moves in the direction determined by its chirality with a constant speed. As a new type of moving structure in two-dimension, a traveling loop consisting of two walls and Neel points is observed.'
address: |
Research Institute of Electrical Communication\
Tohoku University, Sendai 980, Japan
author:
- 'Tetsuya Kawagishi, Tsuyosi Mizuguchi and Masaki Sano'
title: |
Points, Walls and Loops\
in Resonant Oscillatory Media
---
Oscillatory media spontaneously form in a wide variety of systems as they are driven away from equilibrium. Examples are seen in many different fields[@physics; @chemistry; @biology]. In nonlinear oscillators having a small number of degrees of freedom, one of the central problem is to clarify synchronization and resonance. However, inspite of the development of theory and experiment, relatively little is known about the response to the external periodic forcing for spatially coupled oscillators. The simplest but unresolved case is parametric resonance of the system where nontrivial domain walls are predicted to exist[@coullet]. In the parametric resonance, when the external forcing has the frequency double the natural frequency, two locked states are possible. The domain wall appears between two different locked states. In this system, sustained motions of walls are predicted including translational motion of wall, and oscillating walls due to the non-variational effect[@coullet; @Mizuguchi] even if two domains are symmetric. This is the crucial difference from domain walls in equilibrium systems whose dynamics is simply a relaxational process governed by a free energy. We present an experimental investigation of parametrically forced oscillatory media. A structural transition of domain walls associated with chirality breaking is elucidated. Furthermore, traveling loops consisting of two types Bloch walls and Neel points are observed.
Oscillatory media are realized by a convective cellular structure called Oscillating Grid Pattern(OGP)[@sano] in liquid crystal convection[@deGennes]. We use a nematic liquid crystal, 4- methoxybezyliden-4’-butylaniline (MBBA) doped with 0.01wt% of ion impurity, tetra-n-butylammonium bromide, to control the electrical conductivity. This liquid crystal is filled in a cell ($ 2cm \times 2cm
\times 50 \mu m $) sandwiched between transparent electrodes. The temperature of the cell is controlled to 25$\pm$0.01$^\circ$C. Applying an alternative (AC) voltage of $V \sim$60\[V\] with frequency $\omega/2\pi \sim$900\[Hz\] , we obtain stationary Grid Pattern(GP). It gives rise a two-dimensional lattice of about 400$\times$400 rectangle convective cells. Nematic director is stationally since the relaxation time is much longer (about 0.2sec) than the period of external AC field.[@deGennes] It is visualized by a shadowgraph under the microscope with poralized light. The shadowgraphic image intensity is directly related to the orientation of molecules of nematics. In this pattern sinks and sources form the centered rectangle net as shown in Fig. \[GP\](a). Slight increase of voltage causes the oscillatory instability of the Grid Pattern with natural frequency $\omega_0$; $\omega_0 / 2 \pi \sim $1.3\[Hz\] which is independent of the external frequency $\omega$. We call this pattern OGP, from the fact that the positions of sinks and sources move oscillatingly. This is a good candidate of two-dimensionally distributed oscillatory media.
To observe behavior of a parametric resonance of the oscillator lattice, we modulate the AC voltage with nearly double the natural frequency, $V_m = 2\sqrt{2}V
(1+r \cos{\omega_e t})cos{\omega t}$, where $r$ is the modulation ratio and $\omega_e = 2 \omega_0 + \Delta \omega$ and the detuning $\Delta \omega$ is small. Typical patterns observed in a phase locked state under parametric resonance are shown in Figs.\[GP\](b) and (c). The dark lines are interfaces between two phase locked states. Across the dark line the phase of the oscillation jumps by $ \pi $ as will be shown later. The interfaces are walls in dynamical systems. Unlike in equilibrium systems, these walls can exhibit transition from stationary to propagating ones by varying the control parameters; $r$ and $\omega_e$. We show a phase diagram in Fig.\[phase-dia\](a) with varying them. Here the AC voltage for convection was fixed at $\omega/ 2 \pi =$928\[Hz\], $V =$66.3\[V\], which corresponds to a little above the onset of the oscillatory instability of GP ($\mu \equiv (V-V_c)/V_c=0.009 \pm 0.001 $ where $V_c$ is the critical voltage for Hopf bifurcation). In the right half part of phase locked region, the wall exhibits stationary spatial periodic patterns (stripe) as shown in Fig.\[GP\](c). Recently this periodic pattern of walls (stripe) are studied theoretically[@Riecke; @Walgaraef; @coullet93]. As we decrease the modulation frequency, the wavelength of the pattern increases, and finally results in isolated walls in the left half part of the phase locked region as shown in Fig.\[GP\](b). Here we focus on this region in which stationary or propagating isolated walls are observed. To elucidate the transition from stationary to propagating walls, we measure the velocity of moving wall as a function of modulation frequency $\Delta \omega$ (Fig.\[phase-dia\](b)) with fixing $r=0.13$ (Fig.\[phase-dia\](b)). It indicates that the transition is a second order consistent with prediction[@coullet].
It was predicted that the spontaneous breaking of chirality is responsible for the transition from stationary to moving walls in nonequilibrium systems[@coullet]. The stationary one is called Ising walls and the moving one is Bloch walls, relying on the analogy to an anisotropic X-Y model [@Magnet]. If the observed stationary interfaces are Ising walls, the amplitude of oscillation must vanish where the phase jumps. Furthermore if moving interfaces are Bloch walls, the amplitude does not vanish at the core, but two domains are connected by rotating the vector of the complex order parameter of the oscillation. Now chirality of the wall is defined by the direction of this rotation, [*i.e.*]{}, right-handed or left-handed[@coullet]. The moving direction of the wall is determined by the chirality. Hence, one can judge wall types by its amplitude at the core and chirality.
In order to clarify the structural transition of walls we perform analysis as followings. The image intensity ${\cal G}(x,y,t)$ was digitized with a resolution of 640$ \times $480 pixels and 256 grey scale levels at frequency of 15Hz. Here we choose $x$-axis to be parallel to the direction of the alignment of nematics.
Under the parametric resonance, OGP has temporal frequency $\omega_0$ (1.3Hz), and lattice wavenumber $ ( k_x , k_y)
= (2\pi/63,2\pi/105)(\mu m^{-1}) $. Thus at the lowest order one can expand ${\cal G} ( x,y,t )$ as follows, $$\begin{aligned}
{\cal G} ( x,y,t ) = [ 1 + a\exp i ( \omega_0 t + \psi ) ]
\exp( ik_x x+ ik_y y ) + c.c. + h.o.t., \label{oscig} \end{aligned}$$ where $ a $ represents the amplitude and $ \psi $ the phase of oscillation mode, and [*h.o.t.*]{} denotes higher order terms. These amplitude and phase are slowly varying functions in space and time, compared with lattice wave number and natural frequency. In order to obtain only the slow variations in the Eq.(1), we use a broad band pass filter centered at $\omega_0$ in frequency and $(k_x,k_y)$ in wavenumber. Thus we obtain filtered signal; $a \exp(i\psi)$, which is the complex order parameter of our system.
Although the oscillation and interfaces are two-dimensional phenomena, to show temporal evolutions of interfaces evidently at first we take a one-dimensional section of oscillators intersecting a wall perpendicularly. Figure \[Ising\] shows the result about a stationary wall. In Figure \[Ising\](a) the solid line represents the spatial variation of the phase, and there exists a jump about $ \pi $ in its center. At the same point the amplitude (dotted line) falls almost to zero. A spatio-temporal plot of amplitude profile in Figure \[Ising\](b) shows evidently that the wall is stationary. Therefore we conclude that stationary walls are Ising walls. On the other hand, in Fig. \[Bloch\] we show the result about the moving wall. The amplitude at the walls are relatively small but does not vanish (dotted line in Fig. \[Bloch\](a)). Phase gradually changes by $\pi$ at the center of wall (solid line in Fig.\[Bloch\](a)) which slope is less steeper than the case of Ising wall. Notice that the wall is moving leftward in spatio-temporal plot of phase profile (Fig. \[Bloch\](b)). Consequently it is concluded that moving walls are Bloch wall. As in an easy-axis ferromagnets, there are two kinds of Bloch walls in this system. The phase decreases from left domain to right one in Fig.\[Bloch\]. The other type of Bloch walls exhibiting rightward motion are also observed, which connects the same domains with increasing the phase. Hence, the motion of Bloch wall is determined by its chirality.
Let us discuss two-dimensional structures and behavior. In two dimension, Bloch walls can form closed loops. The simplest loop is the one consisting of only a single type of Bloch wall which expands invading outer region or shrinks invading inner region depending on its chirality. (see Fig. \[loop\](a),(b)) However interesting motion is expected if the loop consists of two different types of Bloch walls jointed at two points. (Fig. \[loop\](c)) This point is called Neel point or Connecting Point(CP)[@Mizuguchi; @Frish]. Some interesting phenomena are expected because of the interplay between two types of Bloch walls jointed by CP. Among them we report a new type of moving structure in two-dimensional space; a translational motion of a loop. Figure\[CP\](a) shows a snapshot of a moving loop exhibiting translational motion to the right. The loop travels persistently to one direction until it collides with other walls or loops. We analyzed the two- dimensionally distributed oscillators near the upper end of the loop. Figure \[CP\](b) shows the spatial phase variation. It is seen that two different Bloch walls are connected at a kind of branch cut of a $2\pi$ jump. In Fig. \[CP\], the left domain and the right domain are the same since they are connected out side the loop. The branch cut is resulted from this fact. Notice that the center domain is connected with right one by increasing the phase and with left one by decreasing. These increasing and decreasing correspond to left- and right-handed rotation of the vector of complex order parameter. Their directions of motions are opposite to each other, [*i.e.*]{} one invades outward of a loop and the other does inward. Hence, again the motion of Bloch wall is determined uniquely by its chirality. Considering the topology of the loop, two CPs must exist at the upper and lower ends of the loop which correspond to the singular points of the branch cut. The amplitude of Bloch wall does not vanish, nevertheless at only CP it must vanish due to topological constraint. We plot the cross section of the amplitude profile by crossing interfaces twice at a Bloch wall and CP in Fig.\[CP\](c). Evidently Fig. \[CP\](c) shows that the amplitude is relatively small at Bloch wall, and zero at CP. This confirms that the traveling loop consists of two different types of Bloch walls and two CPs as schematically illustrated in Fig.\[loop\](c).
In the traveling loops, CPs move with the Bloch walls. This is in contrast with usual spiral patterns in which the core (CP) remains unmoved.[@Meyer; @Frish; @Nasuno94] A possible reason is following. In the present experiment, the system is fully non- variational because the system consisting of limit cycle oscillators. In fact, the traveling loop is observed numerically in CGL equation with parametric forcing term by appropriate choice of the linear dispersion coefficient and the detuning parameter[@Mizuguchi94]. Recently observed traveling spots in a model of reaction diffusion equation [@Krischer] seems very similar to the present traveling loops. Therefore we believe that the phenomena are generic in (resonant) oscillatory media. The mechanism of the motion of CP may be related to core meandering in spiral patterns. It is an open problem in resonant system. The motion of the loops and connecting point will open an interesting question about dynamics of interfaces; walls, lines and points, appearing in higher dimensional space for nonequilibrium systems[@Kleman].
The authors wish to thank Y.Sawada, Y.Kuramoto, S.Sasa, and H.Sakaguchi for valuable discussions, and to H.Kokubo for helpful suggestion on experiment.
[99]{}
Chain oscillator appears in oscillatory instability of Rayleigh-Benard convection; J. M. Flesselles, V. Croquette and S. Jucquois, Phys. Rev. Lett. [**72**]{}, 2871(1994). Two dimensional oscillating lattice appears in convection of nematic liquid crystal; see e.g. M. Sano, K. Sato and S. Nasuno, Phys. Rev. E (1991).
Oscillator media is well established in un-stirred flow reactor of oscillating chemical reaction; G. S. Skinner and H. L. Swinney, Physica D [**48**]{}, 1 (1991).
Assembly of heart muscule cells or cardiac tissue can be seen as oscillator media; M. Courtemanche, L. Glass and J. P. Keener, Phys. Rev. Lett. [**70**]{}, 2182 (1993); A. T. Winfree, [*When Time Breaks Down*]{}, (Princeton Univ. Press, Princeton, 1987).
P. Coullet, J. Lega, B. Houchmanzadeh and J. Lajzerowicz, Phys. Rev. Lett., [**65**]{}, 1352(1990); P. Coullet, J. Lega and Y. Pomeau, Europhys. Lett. [**15**]{}, 221(1991).
T. Mizuguchi and S. Sasa, Prog. Theor. Phys. (1993).
M. Sano, H. Kokubo, B. Janiaud and K. Sato, Prog. Theor. Phys., [**90**]{}, 1(1993)
P. G. de Gennes,[*The Physics of Liquid Crystals*]{}, (Claredon Press, Oxford, 1974).
J. Lajzerowicz and J. J. Niez, J. Phys. (Paris) Lett. [**40**]{}, L-165 (1979).
H. Riecke, J. Crawford and E. Knobloch, Phys. Rev. Lett., [**61**]{}, 1942 (1988);H. Riecke, M. Silber and L. Krammer, preprint (1993).
D. Walgaraef, Europhys. Lett. [**7**]{}, 495 (1988).
P. Coullet and K. Emilsson, Physica D, [**61**]{}, 119(1992).
K. B. Migler and R. B. Meyer, Phys. Rev. Lett. [**66**]{}, 1485(1991); Phys. Rev. E [**48**]{}, 1218 (1993).
T. Frish, S. Rica, P. Coullet, and J. M. Gilli, Phys. Rev. Lett. [**72**]{}, 1471(1994).
S. Nasuno, N. Yoshimo and S. Kai, preprint (1994).
S. Sasa, T. Mizuguchi and M. Sano, [*Spatio-temporal Patterns in Nonequilibrium Complex Systems*]{} ed. by P.E.Cladis and P.Palfy-Muboray, (Addison-Wesley,1995) pp.331-342.
K. Krischer and A. Mikhailov, Phys. Rev. Lett. 73 (1994) 3165.
M. Kleman, [*Points, Lines and Walls*]{}, (John Wiley Sons, 1983).
[ \[GP\]]{}
[\[phase-dia\]]{}
[\[Ising\]]{}
[\[Bloch\]]{}
[ \[loop\]]{}
[ \[CP\]]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Earth’s tectonic processes regulate the formation of continental crust, control its unique deep water and carbon cycles, and are vital to its surface habitability. A major driver of steady-state plate tectonics on Earth is the sinking of the cold subducting plate into the underlying mantle. This sinking is the result of the combined effects of the thermal contraction of the lithosphere and of metamorphic transitions within the basaltic oceanic crust and lithospheric mantle. The latter of these effects is dependent on the bulk composition of the planet, e.g., the major, terrestrial planet-building elements Mg, Si, Fe, Ca, Al, and Na, which vary in abundance across the Galaxy. We present thermodynamic phase-equilibria calculations of planetary differentiation to calculate both melt composition and mantle mineralogy, and show that a planet’s refractory and moderately-volatile elemental abundances control a terrestrial planet’s likelihood to produce mantle-derived, melt-extracted crusts that sink. Those planets forming with a higher concentration of Si and Na abundances are less likely to undergo sustained tectonics compared to the Earth. We find only 1/3 of the range of stellar compositions observed in the Galaxy is likely to host planets able to sustain density-driven tectonics compared to the Sun/Earth. Systems outside of this compositional range are less likely to produce planets able to tectonically regulate their climate and may be inhospitable to life as we know it.'
author:
- 'Cayman T. Unterborn'
- 'Scott D. Hull'
- Lars Stixrude
- 'Johanna K. Teske'
- 'Jennifer A. Johnson'
- 'Wendy R. Panero'
bibliography:
- 'main.bib'
date: June 2017
title: Stellar Chemical Clues as to The Rarity of Exoplanetary Tectonics
---
Introduction
============
The Earth is unique in our Solar System. It is the only planet with plate tectonics and liquid water on the surface. It is not known, however, the extent to which the Earth is unique among all terrestrial planets beyond our Solar System. The *Kepler* mission’s discoveries establish that Earth-sized planets are common in the Galaxy, with as many as 11% of Sun-like stars hosting planets 1-2 times the radius of the Earth and receiving comparable solar flux [@Marc14; @Fult17]. Together with other discovery campaigns, we know now of many exoplanets with masses and radii consistent with being terrestrial, rock/metal-dominated planets, rather than gas-dominated. The degree to which these planets can maintain surface oceans, plate tectonics or even be considered “Earth-like” is not known and is a complex function of the planet’s composition, formation, and dynamical state [e.g. @Fole15; @Fole16]. We assert that for a planet to be “Earth-like” and habitable, it must be habitable in the same manner as the Earth. At a minimum this means the planet must sustain surface liquid water for millions to billions of years. . Because stable liquid water exists in a relatively narrow range in temperatures and pressures, the planet must have a process to regulate atmospheric temperatures. On the Earth, moderate temperatures are maintained by the incoming solar radiation combined with moderate greenhouse warming from CO$_2$, H$_2$O, and CH$_4$. Therefore, supply and regulation of these gases is key to Earth’s, and thus “Earth-like” planet’s climate. This definition is in contrast to the typical one for “Earth-like," in which a planet is defined simply as one with a bulk density characteristic of being roughly that of a mixture of metal $\pm$ rock.
The Earth’s atmospheric regulatory processes and aqueous chemistry arise from tectonics: the recycling of material between a planet’s surface and mantle. For Earth, the recycling process manifests as special case of tectonics, plate tectonics, in which oceanic crust continuously subducts into the interior and convective upwellings returning some material to the surface. This Earth-scale transport of material produces the buoyant continental crust and releases and sequesters CO$_2$ through weathering of silicate rocks and arc volcanism at subduction zones [@Velb93; @Brad91; @Slee01; @Fole15; @Fole16]. In contrast a terrestrial planet without tectonic processes, such as one with a rigid lid like Mars, or undergoing episodic overturn like Venus, does not have a steady state cycling of material between the surface and interior. Even if Venus had a lower surface temperature, it is unable to regulate atmospheric CO$_{2}$, which may be rapidly released in pulses during overturn or a consequence of its solidification from a magma ocean [@Hama13]. The lack of standing continents and the formation of surface carbonic acid on Venus does not allow CO$_{2}$ to be efficiently buried, and it instead accumulates in the atmosphere, creating a runaway greenhouse.
The planetary controls that lead to mobile-lid regimes, including plate tectonics, and static-lid regimes are a matter of debate even for the well-determined compositional, thermal, and structural parameters found in the Solar System [e.g. @More98]. Models of large, terrestrial planets (so-called Super-Earths) have concluded that plate tectonics could be inevitable due to interior-to-surface heat transfer, surface gravity, and fault strength [@Vale07a; @vanH11], while other models, focused on the fault strength integral to subduction initiation, have come to the opposite conclusion that plate tectonics are unlikely [e.g. @ONei07]. More general models find the tectonic state of a super-Earth is a function of planet size, incident solar radiation, and atmospheric composition [e.g. @Fole12]. Each of these models, however, simply scale the Earth in composition and structure and sought only to understand how changes in the physical parameters of a planet affected tectonics, rather than address the much more complicated (and data-lacking) question of how planet chemical diversity affects tectonics.
An alternate approach is to assess *probabilistically*, rather than *definitively*, the relative likelihood of tectonics on exoplanets in individual systems by examining the effect of planetary chemistry on plate tectonics [@Unte14; @Unte15; @Stam16]. In this study, we quantify the effect of planet composition on a vital aspect of sustaining plate tectonics on terrestrial planets over billion year timescales: the sinking forces of the exocrust into a planet’s mantle. We therefore address a minimum criterion for plate tectonics: whether or not the surface crusts sink into the mantle due to buoyancy forces arising from thermal and chemical differences between the surface crust and the interior mantle. In the most general sense, tectonic processes of both the Earth and Venus are driven by buoyancy forces. The magnitude of this force is proportional to the integrated density difference between the sinking surface material and the surrounding mantle. These density contrasts are due to the composition and thermal state of the surface material compared to the mantle. Those planetary compositions in which crustal material is buoyant even when forced downward through crustal thickening or contraction have no mechanism for cycling crustal material into the interior via subduction. For those planets that have a buoyancy force less than the Earth, though, there is a lower probability of plate tectonics compared to Earth due to the reduced negative buoyancy (that is downward force) available to drive plate tectonics. Those systems are *less likely* to produce planets with lithospheric material of a composition able to sink via buoyancy forces to $\sim$100 km and thereby produce arc volcanism, will not be able to maintain long-lived temperate, atmospheres, nor will they exhibit top-down, buoyancy driven, steady-state crustal recycling as part of tectonic processes.
The Earth’s subducting lithosphere is composed of two parts: a 5-10 km thick basaltic layer lying on top of a 50-80 km thick, cold, and rigid layer of lithospheric mantle [e.g. @Fisc10], which formed as a result of the cooling of the surface of Earth followed by induration from a magma ocean for 50-200 million years. The basaltic layer is formed through eruption at mid-ocean ridges as the result of decompression melting of the passive upwelling of the convecting mantle. The composition of oceanic lithosphere is therefore controlled by the temperature and composition of the mantle.
Once the subducting basaltic crust reaches a depth of 35-50 km below the surface, pressure-induced mineral metamorphism transforms the basaltic rock into rock denser than the surrounding mantle ($\Delta\rho\sim+100$ kg m$^{-3}$) through the formation of garnet as it replaces orthopyroxene with minor spinel [@Fros08 Supplemental Figure \[fig:phases\]]. This basalt-eclogite transition marks the first of two major metamorphic processes in the basaltic layer responsible for the negative buoyancy that causes plates to sink and continue the cycling from surface to mantle and back. The second phase transition occurs at 300 km below the surface, with excess silica (SiO$_2$) undergoing the coesite-stishovite transition, providing an additional density increase, further promoting sinking. The subducting plate is also thermally contracted relative to the surrounding mantle. The resulting density contrast produces an additional downward buoyancy force, controlled by the temperature difference between subducting plate and surrounding mantle.
Pressure- and temperature-induced phase transitions also occur within the mantle surrounding the subducting plate. However, because of the different temperatures of the mantle relative to the plate, these transitions happen at different depths for each. At a depth near 410 km within the Earth, a phase transition occurs in mantle olivine \[(Mg$^{2+}$,Fe$^{2+}$)SiO$_4$\], transforming it to wadsleyite. This phase transition is what delineates Earth’s upper mantle from the transition zone (Supplementary Figure \[fig:phases\]). The depth of transition from olivine to wadsleyite is shallower at lower temperatures such that olivine in the relatively colder subducting plate will transform into the denser wadsleyite at depths shallower than the transition zone. This introduces a wedge of more dense material in the sinking plate above the 410 km transition, further promoting the sinking of subducting plates.
These phase changes, and the relative depths of their occurrence, result in a chemical buoyancy force, $F_{c}$, on the plate. The integrated density contrast between the plate and mantle due to thermal differences, including depth to pressure-induced transitions, results in a thermal buoyancy force, $F_{t}$. High-temperature atmospheres limit the cooling of the plate on the surface [@Fole16], which will reduce the magnitude of $F_{t}$ and lower the likelihood of plate tectonics on these planets. To date, there are no definitive observations of terrestrial exoplanet atmospheres and thus inference of mobile versus stagnant-lid states through atmospheric observation is currently beyond the reach of exoplanetary science.
Together, $F_{c}+F_{t}$ for Earth’s subducting plates contribute a net buoyancy of $\sim-(2-3)* 10^{13}$ N per meter length of subducted plate downward [recalculated from @Kird14]. While the processes responsible for the initiation of subduction are not well-understood on Earth, plates having sufficient negative buoyancy at the depth of the basalt-eclogite transition is a necessary precondition for incipient subduction and prevention of a stagnant lid [@vanH02]. These upper mantle forces, together with the shear traction and induced flow of a slab inciting downward mantle motion, constitute one of two major driving forces of plate tectonics [@Conr02].
An exoplanet’s composition and mineralogy are not directly observable and are instead inferred from mass-radius relations [e.g. @Seag07; @Weis14; @Wolf16; @Unte16]. Given only a planet’s mass and radius, there is considerable degeneracy in planet composition and structure. As noted by @Dorn15, this degeneracy is reduced when the stellar composition of terrestrial planet-building elements (Mg, Si, Fe) is adopted as a proxy for planetary composition. This is well grounded for the Sun-Earth system due to the refractory nature of these elements [@Unte16, Supplementary Table \[tab:comparison\]] and has yet to be tested for Venus as there are no definitive measurements of its interior composition or structure. Relative to solar [@Aspl05], the abundances of the major terrestrial planet-building elements Mg, Si, and Fe vary between 10 and 400% of solar in large stellar surveys [e.g. @Adib12; @Bens14; @Hink14; @Brew16]. Variations in stellar (Mg+2Si)/Fe affect an exoplanet’s core mass fraction and the ratio Mg/Si affects the dominant minerals of the silicate mantle. Decreasing Mg/Si will lead to a shift in the dominance of the upper-mantle mineral Mg$_2$SiO$_4$ (olivine), to MgSiO$_3$ (pyroxene), to SiO$_2$ [quartz; @Unte16; @Unte17]. While oxidation state of a planet will affect each of these simple assumptions somewhat, the extent of these oxidation-reduction reactions is a complex function of disk chemistry and the planet’s initial thermal profile and subsequent evolution, each of which are poorly understood areas in exoplanetary science. To first order, though, the assumption that stellar refractory composition roughly mirrors planetary refractory composition provides testable predictions of planetary mass for those planets orbiting stars of non-Solar composition [@Unte17].
We therefore adopt the stellar compositional diversity observed in the Galaxy as a proxy for the compositional diversity of terrestrial planets. From this starting point, we calculate the composition, mineralogy of the melt-extracted crust and residual mantle as relative contrast in buoyancy between each. We thus quantify how the composition of a rocky planet affects the likelihood of sustaining plate tectonics over billion-year timescales. The melt-extracted crust is created when solid-state convection develops in the bulk silicate planet (BSP) after the differentiation of a core and the adiabatic rise of rock towards the surface leads to partial melting of the mantle, producing a melt-extracted crust of different composition from the initial bulk silicate composition. A key feature of continual plate tectonics over geologic timescales is this melt-extracted crust subducting along with underlying lithospheric mantle into the interior. The relative magnitude of buoyancy contrast between the melt-extracted crust and residual mantle, therefore, will lead to a greater or lesser *likelihood* of plate tectonics in comparison to the Earth.
Results
=======
{width="\linewidth"}
We calculate the melt composition and stable mineralogy of the bulk silicate planet and exocrust, benchmarking our model to the Earth as derived from solar (See Supplementary Materials; Supplementary Figure \[fig:phases\]) for hypothetical planetary compositions derived from two samples of stellar compositions. The first includes 1063 FGK stars from @Adib12, which represent thin and thick-disk stars with a metallicity range from -0.8 $<$ \[Fe/H\] $<$ 0.6, over 100 of which are known to host (mostly gas-dominated) planets. Of this first sample, we thermodynamically model the composition and buoyancy of a subset of 609 stellar compositions that are within the internally consistent MELTS database [@Ghio02], 57 of which are known to host planets. We find a chemical buoyancy ($F_c$) of these systems ranges from $-7.5*10^{12}$ N m$^{-1}$ (sinks) to $4.0*10^{12}$ N m$^{-1}$ (floats) while our model Earth produces buoyancy force of $\sim-2.0*10^{12}$ N m$^{-1}$ (Figure \[fig:histogram\]).
The second sample of stellar compositions comes from the Apache Point Observatory Galactic Evolution Experiment [APOGEE, @Wils10; @Maje15], which is part of the Sloan Digital Sky Survey IV [SDSS-IV; @Blan17]. From the public APOGEE Data Release 13 [DR13, @Alba16], we selected the stars known to host small planets (R$_{\rm {p}}$ $\leq$1.6 R$_{\oplus}$) from the *Kepler* transiting planet survey, totaling 123 stars with a metallicity range from -0.54 $<$ \[Fe/H\] $<$ 0.36. Of these, we thermodynamically model the composition and buoyancy of a subset of 89 stellar compositions that are within the internally consistent MELTS database [@Ghio02]. We find a chemical buoyancy of these systems ranges from $-10*10^{12}$ N m$^{-1}$ (sinks) to $3.6*10^{12}$ N m$^{-1}$ (floats).
Of terrestrial planets with compositions represented by the variation in both stellar datasets, 19% would produce exocrusts entirely buoyant than their BSP throughout the entire upper mantle ($F_{c} > 0$), and therefore unable to subduct (Figure \[fig:histogram\]). Furthermore, 61% of the Adibekyan sample and 41% of the *Kepler* sample produces plates more chemically buoyant than that of our model Earth ($F_{c} > F_{c-Earth}$; Figure \[fig:histogram\]) and therefore *less likely* to host plate tectonics than the Earth.
![Histograms of buoyancy forces calculated using our model for our sample of 609 @Adib12 stars (gray) and 89 **Kepler** planet hosts (purple) with compositions inside the MELTS database [@Ghio02]. Those modeled from stellar composition are shown in red. The buoyancy force and bulk composition of model Earth is shown as a light gray dashed line.[]{data-label="fig:histogram"}](Histogram.pdf){width=".6\linewidth"}
This variation in buoyancy forces is mainly caused by the enrichment of Si and alkali elements (Na, K) within the melt-extracted crust. Partial melting of typical mantle rocks leads to enrichment of silica and alkalis in the melt compared to the parent rock. We find those planetary compositions with lower alkali and silica abundance, such as those crustal materials that form on Earth with basaltic composition, are most likely to sink, while those with greater alkaline and silicic compositions are less likely (Figure \[fig:rocktype\]; Supplementary Figure \[fig:nonEarth\]). This is consistent with Earth’s tectonics, in which low-density and alkali-rich, andesitic continental crust does not subduct [@Cloo93].
![Alkali abundance as a function of silica of the exocrust compositions for our sample of *Kepler* planet-host (diamonds) and @Adib12 stars (circles). Color is a function of the chemical buoyancy as calculated from a thermally equilibrated, 5% melt layer. Points outlined are those with known planets (exoplanets.org). The green box represents the Earth Range of average Earth basaltic compositions of @Gale13. The composition of Mercury’s Northern Volcanic Plains and Intercrater Plains and Heavily Cratered Terrains [diamonds; @Namu16] and Mars’ type 1 and type 2 crusts [squares, including K$_2$O; @McSw03] are included. Venus’ alkali composition has yet to be determined [@Trei13]. Each crustal composition of Mercury and Mars fall within the positively buoyant andesitic field, despite 3 of the 4 measurements containing Na$_2$O abundances within the Earth-like range.[]{data-label="fig:rocktype"}](RockType.pdf){width=".6\linewidth"}
The magnitude of chemical buoyancy is, therefore, closely tied to the abundance of sodium and potassium in the BSP due to their incompatibility in melting processes (Figure \[fig:histogram\]). Because Na is roughly 15 times more abundant than K for the Sun [@Lodd03], we focus on Na as a controlling alkali in the buoyancy calculations. Because all alkalis are moderately volatile in the planetary condensation process [@Lodd03], their abundances are not likely to be mirrored in any orbiting terrestrial planets due to their fractionation relative to the refractory elements (e.g. Mg, Si, Fe) in the planetary formation process. For this model, we assume an incomplete accretion of Na due to volatile loss during planet formation in proportions similar to Sun-Earth fractionation for each star in our sample [$\delta_{ \rm{Na-Sun/Earth}} = \left (\rm{Na/Si}\right)_{\rm{Earth}}/\left(\rm{Na/Si}\right)_{Sun}$ = 0.26 by mole; @McD03; @Aspl05]. Chemical buoyancy is then a function of the total Na in the star as well as the degree of Na volitalization during accretion, $\delta_{\rm {Na}}$: $$\label{eq:fit}
F_{c} \approx -1.15+16.6\left(\rm{[Na/Si]}+log10\left(\delta_{\rm{Na}}/\delta_{\rm{Na-Sun/Earth}}\right)\right) * 10^{12} \rm{N m^{-1}}$$
![Chemical buoyancy force ($F_{c}$) as a function of bulk planetary Mg and Si normalized to Fe (by mole) for values for three values of $\delta_{\rm{Na}}$. Filled circles are those calculations presented here in Figures \[fig:histogram\] and \[fig:rocktype\] for the @Adib12 dataset (black) and *Kepler* hosts (red). Open symbols are those buoyancy forces calculated from the relationship between stellar \[Na/Si\] and $F_{c}$ (Equation \[eq:fit\], Supplementary Figure \[fig:fit\]) for the @Adib12 stars (squares) and *Kepler* hosts (circles). The 1063 stars in the stellar survey of @Adib12 show $\sim$2.5 orders of magnitude difference in total Na, halving or doubling $\delta _{\rm {Na}}$ from the Solar value represents the entire sample either producing none and all plates being negatively chemically buoyant. The index (Mg+2Si)/Fe is both a primary control on the compositional extent of the thermodynamic database [@Ghio02] as well as defining the approximate proportions of core and mantle of an associated terrestrial planet [@Unte16]. []{data-label="fig:Sodium"}](Sodium.pdf){width="\linewidth"}
when \[Na/Si\] is scaled from @Aspl05 (Supplementary Figure \[fig:fit\]), 78% of the variation in $F_c$ can be explained by variation in \[Na/Si\]. For a given stellar \[Na/Si\] value, 95% of $F_c$ values fall within $\pm10^{12}$ N m$^{-1}$, which is a more complex function of the refractory element composition of the planet and must be derived through the methods outlined here. This correlation between $F_c$ and Na abundance allows us to estimate the chemical buoyancy of the 454 stars in the @Adib12 dataset whose mantle compositions were not located within the MELTS database and were generally those stars with bulk Fe/Si $<$ 0.95 [Bulk Earth Fe/Si = 1; @McD03]. For those stars in the Adibekyan dataset, we find 28% are of compositions likely to produce melt-extracted crust that remains buoyant relative to their mantles ($F_c > 0$), with 69% less chemically buoyant than the Earth (Figure \[fig:histogram\]). The Kepler dataset shows a similar trend, with 22% of stars likely to produce crusts with positive chemical buoyancy and 60% with crustal buoyancy forces greater than the Earth. Each of these systems are therefore *less likely* than the Earth or *completely unlikely* to maintain long-term, steady-state tectonics.
When $\delta_{\rm{Na}}$ is twice $\delta_{\rm{Na-Sun/Earth}}$ ($\delta_{\rm{Na}}$ = 0.52), only 1% of stars within the Adibekyan sample and 4% of the Kepler sample produce any plates with a negative chemical buoyancy force ($F_c < 0$), whereas all but fourteen stellar compositions in our sample of 1186 stars produce negatively buoyant plates when $\delta_{\rm{Na}}$ is half of $\delta_{\rm{Na-Sun/Earth}}$ ($\delta_{\rm{Na}}$ = 0.13; Figure \[fig:Sodium\]). The likelihood of producing negative chemical buoyancy in melt-extracted surface plates is therefore very sensitive to $\delta _{\rm {Na}}$ and weakly dependent upon bulk mantle composition. Stars with greater Na abundance than Solar must retain less Na when forming terrestrial planets to create negatively buoyant plates at depth. The same is true for those stars of Solar Na composition with greater $\delta _{\rm {Na}}$ compared to the Earth and Sun ($\delta _{\rm {Na}} >$ 0.26). These high Na planetary systems may initiate subduction, but due to their higher positive chemical buoyancy force, they are unlikely to continue to do so [@Cloo93]. As mantle temperatures increase in this stagnant-lid planet, any volatiles contained within the partially subducted plate will degass into the atmosphere, thus entirely negating any long-term, continuous tectonic regulation of climate.
Our compositionally focused approach is a tool to examine the influence of a host star’s refractory and moderately volatile composition on a key aspect of planetary dynamics: top-down, density-driven plate tectonics. This model attempts to capture only the most general details of terrestrial planetary formation, differentiation, and evolutionary process with the sinking of a surface plate through a planet’s mantle being a consequence of the difference in density between the mantle and crust. This model treat the complex processes of accretion, formation of overlying continents, and prolonged geochemical processing as discrete processes at a fixed pressure, temperature, and oxygen fugacity. While not meant to be strictly prescriptive, these results demonstrate first-order trends and controlling factors in the geochemical consequences of variable compositions of terrestrial planets.
This approach predicts a lack of plate tectonics on Venus and the potential for (transient) plate tectonics on Mars. For planets with significant thermally insulating atmospheres, more heat is retained at the surface, limiting the cooling of warming the plate [@Fole15; @Fole16]. A surface temperature 450 K greater than Earth such as Venus will significantly reduce $F_{t}$, while $F_{c-Venus} \sim F_{c-Earth}$, such that the sum predicts Venus is less likely than Earth to undergo tectonics without having to consider the strength or thickness of surface rocks [@Fole12; @Berc14]. In the case of Mars, the lack of water and rapid heat loss due to its small size limits the likelihood for long-term plate tectonics, although it may have experienced subduction in the past [@Slee94].
One of the most significant neglected variables is the abundance of water and carbon in the planet’s interior, which are sensitive to nebula composition, formation processes and location in the nebula, and subsequent evolution. The volatile abundance in terrestrial planets is not well constrained in planetary formation models, both for our Solar System and exoplanetary systems. While planetary H$_2$O and CO$_2$ abundances may inform the scale of exoplanetary oceans and climate, the planet’s relative abundance of moderately volatile elements such as Na is a first-order control on the likelihood of regulating these important aspects of exoplanet habitability via long-term plate tectonics. This approach, therefore, points to the importance of addressing the thermal and dynamic controls for both the highly and moderately volatile abundances of planets in formation models.
Of the small exoplanets discovered to date (R $\leq$ 1.6 Earth radii), few have both size and mass measurements, from which mean density may be calculated. Mean density is a non-unique function of the relative proportions of core, mantle, and gaseous envelope. Of those mass/radius measurements sufficiently dense to suggest the planets are terrestrial, planetary compositions and structures are indistinguishable from one another due to large observational uncertainties. We offer a complementary approach for those planets inferred to be terrestrial by providing a framework for quantifying an exoplanet’s potential geochemical dynamics. Indeed, these results represent the first observational metric with which to gauge the probability of an exoplanetary system to maintain steady state surface-to-interior geochemical cycling and tectonic behavior with the potential to regulate atmospheric temperatures over long timescales. Furthermore, the connection between stellar Na abundance, degree of volatilization, and surface plate chemical buoyancy point to the importance of adopting a comparative tool to ascertain the likelihood of exoplanetary plate tectonics compared to the Earth. The exoplanetary field needs to refine models of not only the volatile condensation and accretion in planetary systems, but track the fractionation and dynamics of the refractory and moderately volatile elements within the disk as well. Furthermore, geochemical and geophysical experiments and models must be expanded to those compositions relevant to planets of non-Earth/Solar composition. The upcoming James Webb Space Telescope will provide opportunities for time-intensive observational studies to follow up terrestrial exoplanet atmospheric conditions, and our framework provides a strategy to identify those planetary systems most worthy of follow up observations. Bulk planet density is only one factor in determining whether a planet is “Earth-like.” It is only through this holistic approach to characterizing exoplanets, combining data and models from across scientific fields, that we will truly calibrate the likelihood of an extrasolar planet to be behaviorally Earth-like and habitable to life as we know it.
We thank the Cooperative Institute for Dynamic Earth Research (CIDER) for providing the space and intellectual environment for fostering this work as well as funds to support the BurnMan code. This project was supported by NSF-CAREER (EAR-09-55647) to Panero and the Shell Undergraduate Research Experience, OSU Undergraduate Research Scholarship, and Friends of Orton Hall grant to Hull. Funding for the Sloan Digital Sky Survey IV has been provided by the Alfred P. Sloan Foundation, the U.S. Department of Energy Office of Science, and the Participating Institutions. SDSS acknowledges support and resources from the Center for High-Performance Computing at the University of Utah. The SDSS web site is www.sdss.org. SDSS is managed by the Astrophysical Research Consortium for the Participating Institutions of the SDSS Collaboration including the Brazilian Participation Group, the Carnegie Institution for Science, Carnegie Mellon University, the Chilean Participation Group, the French Participation Group, Harvard-Smithsonian Center for Astrophysics, Instituto de Astrofísica de Canarias, The Johns Hopkins University, Kavli Institute for the Physics and Mathematics of the Universe (IPMU) / University of Tokyo, Lawrence Berkeley National Laboratory, Leibniz Institut für Astrophysik Potsdam (AIP), Max-Planck-Institut für Astronomie (MPIA Heidelberg), Max-Planck-Institut für Astrophysik (MPA Garching), Max-Planck-Institut für Extraterrestrische Physik (MPE), National Astronomical Observatories of China, New Mexico State University, New York University, University of Notre Dame, Observatório Nacional / MCTI, The Ohio State University, Pennsylvania State University, Shanghai Astronomical Observatory, United Kingdom Participation Group, Universidad Nacional Autónoma de México, University of Arizona, University of Colorado Boulder, University of Oxford, University of Portsmouth, University of Utah, University of Virginia, University of Washington, University of Wisconsin, Vanderbilt University, and Yale University.
Supplementary Material
======================
Stellar Composition Data
------------------------
We apply this Earth-benchmarked model to two samples of stars with abundances available for 9 elements (Mg, Fe, Si, Ca, Al, Ni, Ti, Cr and Na). First is the survey of F, G and K stars in the Galaxy of @Adib12 and the other is a sample of 123 known planet-hosting stars whose abundances were available within the APOGEE Data Release 13 [DR13 @Alba16]. Where available, only those stars with C/O $<$ 0.8 were chosen. We use 1186 total stars in our model. These stars represent both thin and thick-disk stars with a metallicity range from -0.8 to 0.6 \[Fe/H\], of which 235 have detected planets. Of these stars, 698 have compositions that fall within the thermodynamic databases used in this study (Figure \[fig:histogram\]). For those stars outside of the database, we estimate the crustal Na content and chemical buoyancy force through a linear fit of these values for only those 609 stars that were within the MELTS database and in the @Adib12 dataset (Supplemental Figure \[fig:fit\]).
We note that a separate publication is in progress (Teske et al. in prep) that will discuss the reliability of FGK dwarf star abundances produced by the APOGEE automated analysis pipeline [ASPCAP; @Garc16; Holtzmann et al. in prep, Jonsson et al, in prep]. Initial results (also discussed in Wilson et al. in prep) suggest that, for solar-type stars, the Fe, Mg, and Si abundances produced by ASPCAP are reliable at the $\sim$0.15 dex level. The ASPCAP Na abundances appear to be less reliable, upon comparison with a small sample of values derived from optical data. For consistency, we did not modify the DR13 APOGEE Na abundances to bring them into better agreement with the optical data. We note that Figure 1 shows that the @Adib12 and the APOGEE samples have similar ranges in Na abundances.
Methods Summary
---------------
### Planet Composition Model
There is extensive debate as to the exact nature of the Earth’s composition, from one closer to a volatilized carbonaceous chondrite [@McD95; @McD03], to more similar to enstatite chondrites [@Javo10], to mixtures of carbonaceous chondrites and ordinary chondrites [e.g. @Fito12]. We instead adopt the simplest model of a composition that is identical to the Sun in refractory elements, with an oxygen abundance fixed by the fO$_{2}$ of the condensing medium. For key moderately volatile elements such as Na, we assume an ad-hoc condensation efficiency of 26% to reproduce the Earth’s bulk Na composition [@Aspl05; @McD03].
![ Predicted phases (black) and densities (red) in this model (solid) compared to Earth reference compositions [dashed, @McD03]for (a) the bulk silicate Earth and (b) melt-extracted basalt as a function of pressure. The model mantle is calculated along a 1600 K adiabat while the basalt is calculated along a 1200 K adiabat. While the simplified model presented here does not reproduce the Earth in detail, it does reproduce the correct minerals and pressures of phase transitions and magnitude of density discontinuities. Our model mantle under predicts the olivine fraction in the mantle (55% vs. 60%) while our model basalt predicts clinopyroxene at the expense of garnet relative to average basalt compositions under predicting the density by no more than 3%. Garnet is the dense phase in the basalt-eclogite transition; this model therefore represents a lower bound on the likelihood of subduction. []{data-label="fig:phases"}](phases.pdf){width="\linewidth"}
The condensation temperatures of the moderately volatile elements are highly correlated with their abundances in ordinary and CI chondrites [@Wai77], indicating a condensation process that is both composition and thermally-driven. However, because of the lower condensation temperature of these elements relative to the refractories [e.g. @Lodd03], there is likely radial mixing of material within the disk. Recent planetary formation models [e.g. @Raym06] show water ice mixing from beyond 2.5 AU into the inner Solar System, while the degree of mixing of the moderately volatile elements increases with the timescale of planet formation [@Bond10a; @Bond10b]. Other disk models addressing the mixing of moderately volatile elements in the disk systematically over-predict the Na abundance in resulting planets compared to when attempting to reproduce the Earth’s composition [@Matsu16], likely due to neglecting not accounting for volatile loss due to melting during planetary formation.
[l|cc|cc|cc|cc]{} Mg&6.1&16.6&16.9&17.4&19.9&19.7&3.8-4.6&4.3\
Si&5.8&15.8&15.3&14.4&15.9&16.4&20.0-17.7&19.1\
Fe&5.1&13.8&15.3&14.4&2.4&2.4&2.8-3.9&2.4\
Al&0.4&1.1&1.5&1.2&1.8&1.4&6.3-6.8&6.6\
Ca&0.4&1.0&1.1&1.0&1.3&1.2&4.4-4.6&3.5\
Na&0.3&0.7&0.2&0.2&0.2&0.2&1.6-2.5&3.5\
O&82.0&51.0&49.6&51.3&58.5&58.7&60.0-61.2&60.5\
\[tab:comparison\]
### Planetary Differentiation Model
We model two stages of differentiation from the condensed planet composition: separation of the iron-rich core from the bulk silicate planet (BSP) and crustal rock forming as a result of partial melting (exocrust) from adiabatic decompression of the BSP. This simplified two-stage model broadly reproduces the major element composition of the Earth, core-extracted bulk silicate Earth, and mid-ocean ridge basalt (MORB) from the Solar composition (Supplementary Table \[tab:comparison\]). Differences between our simplified Earth and the true Earth lead to comparable differences in modal abundance of minerals in both MORB and Bulk Silicate Earth (BSE, Supplementary Figure \[fig:phases\]). The resulting absolute difference in calculated densities vary by 0.6% for the bulk silicate Earth and 2.6% for the modeled MORB composition, with the depth-to-metamorphic transitions and the relative difference between the transitions for each composition indistinguishable.
We model planetary differentiation through self-consistent thermodynamic modeling [@Ghio95; @Ghio02] of cooling the bulk composition from a molten planet to calculate mineral and melt equilibria. The pressure and gas fugacity of the equilibrium calculations are chosen to best reproduce the fraction of differentiated Fe alloy and the composition of the BSE and MORB composition from the Sun’s composition. We assume Si is primary the light element in the core [@Fisch15] to be consistent with the Mg/Si ratio of the Earth’s mantle [@McD95; @McD03; @Unte16]. The melting of the mantle is calculated by assuming an adiabatic rise of material to lower pressures until the solidus temperature of the rock is reached. We calculate the primary crust composition to be that of the BSE at 5% melt fraction, which is comparable to average mid-ocean ridge basalt (Supplementary Table \[tab:comparison\]) after imposing a systematic correction to MgO to account for an over stabilization of clinopyroxene in MELTS [@Ghio02]. If no Si is present in the core, the mantle Si abundance would increase accordingly. This would enrich the melt in silica, forming a crust of buoyant andesite rather than basalt. We solve for the thermodynamically stable mineralogical host of each BSP and exocrustal composition as a function of pressure and temperature using the HeFESTo [@Stix05; @Stix11], open-source BurnMan [@Cott14; @Unte16] and the Exoplanet-Pocketknife (available at https://github.com/ScottHull/Exoplanet-Pocketknife) software packages. Temperatures are assumed to increase adiabatically, calculated self-consistently with depth. The mantle adiabat is fixed at 1600 K, approximately the potential temperature of the Earth’s mantle, while the average exocrust temperature is varied between 1000 K geotherms ($\Delta \rm{T} =600 K$) and 1600 K (thermally equilibrated, $\Delta \rm{T}=0$) From the mineralogy, compositions of the minerals, and modal abundance of each, density profiles of profile are calculated from the mineral-specific equations of state.
![Predicted mantle mineralogies along 1600 K (top) and 1000 K (bottom) geotherms of hypothetical terrestrial planets about stars HD-19994 (left) and HD-89668 (right), representing extremes in Mg/Si. cpx = clinopyroxene, opx = orthopyroxene, c2c = c2/c structured-phase of opx, mw= magnesiowustite, $\alpha$-, $\beta$-, $\gamma$-ol = olivine, wadsleyite, and ringwoodite, respectively. Densities (right axis) of the mantle along a 1600 K geotherm (solid red) are compared to the densities of the inferred slab along a 1000 K geotherm (solid blue) are a function of the density of the exocrust (dashed cyan) and cold mantle lithosphere (dashed red). []{data-label="fig:nonEarth"}](NonEarth.pdf){width="\linewidth"}
![Crustal Na$_{2}$O wt% composition as a function of molar fraction Na and chemical buoyancy force ($F_{c}$) as a function of stellar \[Na/Si\] for stars within the MELTS database. We adopt the Solar model of @Aspl05 for calculation of \[Na/Si\]. $F_{c}$ is modeled assuming Solar Na is reduced by 74% to the Earth abundance relative to Si [@Aspl05; @McD03]. []{data-label="fig:fit"}](fit.pdf){width="\linewidth"}
An additional consideration is the extent of melting for the formation of the exocrust, which is a function of both mantle composition and temperature. We find that 95% of the sample set is at 5% melt over a 65 K range, such that forming an exocrust of thickness comparable to that of the Earth occurs over a relatively narrow range. For lower mantle temperatures, little-to-no melt may be produced, while greater mantle temperatures will produce greater crustal thickness. For simplicity, we have considered a self-consistent adiabat applicable to the Earth with an Earth-like potential temperature (1600 K). Increasing the potential temperature increases the extent of melting, and consequently the thickness of melt-extracted crust. A greater extent of melting decreases Na$_2$O fraction in the melt for an approximately constant SiO$_2$ fraction when melt fraction $>$ 5% [@Ghio02], both effects increase the downward force due to chemical buoyancy for a given composition planet. The Earth’s mantle contains trace quantities of both H$_2$O and CO$_2$, which affect the melting behavior of the passive upwelling mantle. Carbon increases the depth of initial melting in the decompression melting of mantle peridotite. This melting is at significantly greater depths than assumed in the models here, but contributes very small fractions of melt [$<$0.3%; @Dasg06] and therefore unlikely to significantly affect the major element chemistry assumed here. Similarly, water decreases the solidus temperature, increasing the depth of initial melting, but not significantly altering the total melt fraction [@Asim03] and decreases the total iron in the melt.
With a simple two-stage model, we do not address the chemical secular evolution of the mantle due to the extraction of continental crust. As such, we overestimate the Na$_{2}$O in the basaltic crust due to neglecting the abundance of Na$_{2}$O sequestered in the continental crust. For a continental crust with 3.6 wt% Na$_{2}$O [@Rudn03], reducing the BSP Na$_{2}$O abundance by about 5%, reducing the Na$_{2}$O content of the resulting melt-extracted crust about 77% of our predicted Na$_{2}$O content, consistent with typical MORB compositions. This suggests that early plate tectonic processes on Earth required initial secular variation in bulk composition of incompatible elements before modern plate tectonics could commence.
The downward sinking force is quantified by the buoyancy of a potential sinking plate, made up of 10 km-thick exocrust and 50 km-thick lithospheric mantle. For simplicity, we adopt the BSP composition as the lithospheric mantle composition. The buoyancy force per unit length subducting plate, $F_{b}$, is $$F_{b} = t \int_{0}^{d}\Delta\rho\left(\Delta x,h,\Delta T\right)g\left(h\right)dh$$ where $\Delta\rho(\Delta x,h, \Delta T)$ is the density difference between the down-going plate and mantle compositions as a function of differences in the bulk compositions, $x$, depth, $h$, and the potential temperature difference between the mantle and slab, $\Delta T$, $g$ is the acceleration due to gravity, $d$ is the depth to the base of the Earth’s transition zone (where the integrated density differences are a maximum), and $t$ is the slab thickness. While $g$ is a function of the planet mass, the depth to metamorphic transitions are pressure dependent, such that for a planet of equal core to mantle proportions, the product of surface gravity and the depth at 25 GPa varies by just 1.3% over planets between 1 and 4.5 Earth masses. Therefore, we adopt Earth’s surface acceleration due to gravity and Earth’s pressure-depth relationship as a representative model valid for the calculation of net negative buoyancy in Super-Earths.
The magnitude of the chemical buoyancy force is closely tied to the abundance of Na$_{2}$O in the BSP due to its incompatibility in melting processes, $$F_{c} \approx (-9.38+1.64*X_{Na_{2}O})*10^{12} \rm{N m^{-1}}$$ where $X_{Na_{2}0}$ is the weight percent Na$_{2}$O in the planet’s mantle ($X_{Na_{2}0}$ = 4.97 wt% for our model Earth.
| {
"pile_set_name": "ArXiv"
} |
**The Increasing Rotation Period of Comet 10P/Tempel 2**\
Matthew M. Knight[^1], Tony L. Farnham[^2], David G. Schleicher$^{1}$, Edward W. Schwieterman[^3]\
Contacting author: [email protected]\
*The Astronomical Journal*\
Accepted September 13, 2010\
Manuscript pages: 22 pages text\
3 tables\
6 figures\
Abstract {#abstract .unnumbered}
========
We imaged comet 10P/Tempel 2 on 32 nights from 1999 April through 2000 March. R-band lightcurves were obtained on 11 of these nights from 1999 April through 1999 June, prior to both the onset of significant coma activity and perihelion. Phasing of the data yields a double-peaked lightcurve and indicates a nucleus rotational period of 8.941 $\pm$ 0.002 hr with a peak-to-peak amplitude of $\sim$0.75 mag. Our data are sufficient to rule out all other possible double-peaked solutions as well as the single- and triple- peaked solutions. This rotation period agrees with one of five possible solutions found in post-perihelion data from 1994 by Mueller and Ferrin (1996, Icarus, 123, 463–477), and unambiguously eliminates their remaining four solutions. We applied our same techniques to published lightcurves from 1988 which were obtained at an equivalent orbital position and viewing geometry as in 1999. We found a rotation period of 8.932 $\pm$ 0.001 hr in 1988, consistent with the findings of previous authors and incompatible with our 1999 solution. This reveals that Tempel 2 spun-down by $\sim$32 s between 1988 and 1999 (two intervening perihelion passages). If the spin-down is due to a systematic torque, then the rotation period prior to perihelion during the 2010 apparition is expected to be an additional 32 s longer than in 1999.
[**Keywords**:]{} comets: general – comets: individual (10P/Tempel 2) – methods: data analysis – techniques: photometric
Introduction {#intro}
============
Comet 10P/Tempel 2 was discovered by E.W.L. Tempel on 1873 July 4. It was recovered in 1878, but not during the 1883 or 1889 apparitions. It has been observed on every return since 1894 except three (1910, 1935, and 1941) when it was particularly poorly placed for observing [@kronk03; @kronk07; @kronk09]. While telescopic improvements now allow it to be observed at every apparition, the roughly 5.5 year period results in apparitions which alternate between favorable and unfavorable viewing geometries. Consequently, Tempel 2 was well placed for observing in 1978, 1988, and 1999, but poorly placed in 1983, 1994, and 2004 when it reached perihelion on the far side of the Sun. Tempel 2 was extensively observed during its favorable 1988 return. Because it is only weakly active until shortly before perihelion (@sekanina79 and references therein), a nucleus lightcurve can be measured. This allowed a number of investigators to measure the rotation period. The earliest results came from @jewitt88 who found likely periods of 8.9 $\pm$ 0.1 hr or 7.5 $\pm$ 0.1 hr. @jewitt89 extended these observations to near perihelion, concluding that the rotation period was 8.95 $\pm$ 0.01 hr. Comparable periods were determined by @ahearn89 (8.9 hr) and @wisniewski90 (8.93 hr). @sekanina91 combined these data to determine a sidereal period of 8.93200 $\pm$ 0.00006 hr. @mueller96 observed Tempel 2 for three nights 7–9 months after perihelion in 1994, after coma activity had subsided. Their rotation coverage was insufficient to determine a unique period solution, and they found five possible periods due to aliasing: 8.877 hr, 8.908 hr, 8.939 hr, 8.971 hr, and 9.002 hr. All of these periods were incompatible with the 1988 data, and they concluded that “the period is definitely different between the 1988 and the 1994 apparitions.” Because of aliasing, it was not known if Tempel 2 had spun-up or spun-down since 1988, merely that its rotation period had changed.
The idea of comet rotation periods changing due to outgassing dates to @whipple50 and his icy conglomerate model of the nucleus, with various authors since then arguing for spin-up or spin-down (see @samarasinha04 for a thorough review). Numerical modeling has shown that, depending on the initial conditions, either spin-up or spin-down is possible, with sustained outgassing over many orbits leading to spin-up and ultimately nucleus splitting [@samarasinha95; @neishtadt02; @gutierrez03]. Measurement of a change in the rotation period of a comet, combined with detailed knowledge of the spin axis orientation, nucleus size and shape, and the outgassing rate can constrain the bulk density and internal structure. Due to the difficulty of studying comet nuclei directly, these properties are only well known for a handful of comets, mostly the result of spacecraft visits.
While simulations have shown that changes in rotation period should be common, a strong case can be made for a changing rotation period in only a few comets: 10P/Tempel 2 (sign of the change unknown prior to this work; @mueller96), 2P/Encke (spin-up; @fernandez05), 9P/Tempel 1 (spin-up; @belton07 [@chesley10]), and Comet Levy (1990c = C/1990 K1; spin-up; @schleicher91 [@feldman92a]). The paucity of clear detections of this phenomenon is probably due to the difficulty of obtaining high quality datasets of the same comet over multiple apparitions. By combining our extensive data obtained during the 1999 apparition with those of previous authors from 1987–1988 and 1994, this work will show that 10P/Tempel 2 is the first comet known to spin-down.
We observed Tempel 2 from 1999 April until 2000 March, obtaining images in broadband and narrowband optical wavelengths. Observations prior to perihelion (1999 September 8) were primarily obtained with a broadband R filter to measure the nucleus lightcurve. Once activity began in earnest, observations were primarily obtained with narrowband comet filters [@farnham00] to study coma morphology. In this paper, we consider only the pre-perihelion nucleus lightcurve in order to resolve the ambiguity about the sign of the change in rotation period and to conclusively demonstrate that Tempel 2 has spun-down since 1988. A second paper (Paper 2) is planned and will utilize data obtained during 2010 in concert with data obtained during the active phase of the 1999 apparition to determine the pole orientation and the location of any active areas.
The layout of the paper is as follows. We summarize our observing campaign and data reductions in Section \[observations\]. In Section \[modeling\] we analyze the data, determine the rotation period, compare it with earlier datasets, and consider the effects of viewing geometry, coma contamination, and the lightcurve asymmetry. Finally, in Section \[discussion\] we discuss the implications of our results and make predictions for the 2010 apparition.
Observations and Reductions {#observations}
===========================
Observing Overview
------------------
We obtained images of Tempel 2 on a total of 32 nights between 1999 April and 2000 March, with sampling at monthly or shorter intervals. For the study presented here, we use only the data obtained from 1999 April through 1999 June, with the dates and observing circumstances listed in Table \[t:obs\_summary\]. Additional pre-perihelion data were obtained on 1999 July 10, July 16, August 4, August 5, and September 2. However, these nights all had poor weather and the data were unusable for the present study. By 1999 October, the observing window was very short and the comet had extensive coma contamination, making a period determination challenging.
The April and May observations were obtained at the Lowell Observatory Perkins 1.8-m telescope with the SITe 2K CCD. On-chip 4$\times$4 binning of the images produced a pixel scale of 0.61 arcsec. The June observations were obtained at the Hall 1.1-m telescope with the TI 800 CCD. On-chip 2$\times$2 binning produced images with a pixel scale of 0.71 arcsec. At various times during these runs, we used broadband Kron-Cousins V and R filters and the HB narrowband comet filters [@farnham00]. However, as planned, only the R filter measurements are extensive enough for the lightcurve analysis discussed here, so we limit this study to those data. Comet images were guided at the comet’s rate of motion. The comet showed evidence of a coma throughout the apparition, progressively increasing with time (see Section \[coma\_removal\]).
Reductions
----------
The data were reduced using standard bias and flat field techniques. Landolt standard stars [@landolt92] were observed to determine the instrumental magnitude and extinction coefficients on 1999 June 8–11. We observed HB narrowband standard stars [@farnham00] on all photometric nights, and used these stars as a bootstrap to estimate absolute R-band calibrations for the photometric nights on which Landolt standard stars were not observed.
Fluxes were extracted by centroiding on the nucleus and integrating inside circular apertures, with the median sky calculated in an annulus centered on the nucleus with inner and outer radii $\sim$33 and $\sim$40 arcsec, respectively. By extracting fluxes through a series of circular apertures (3, 6, 9,...30 arcsec radius), we monitored the lightcurve for incursion from passing stars which show up in the larger apertures earlier than the smaller apertures. The 6 arcsec radius aperture gave the most coherent lightcurve and was used for photometric analysis; this was selected to be large enough to include most of the light from the nucleus even when the seeing was poor and the nucleus PSF was large, while avoiding as much contamination from passing stars as possible.
In order to produce usable lightcurves from non-photometric nights, we applied extinction corrections and the absolute calibrations from photometric nights with the same telescope and instrument configuration. Thus, we applied the absolute calibrations from April 19 to April 17–18 and May 26–27, and the calibrations from June 23 to June 22. The deviation of the brightness from photometric nights helps give an estimate of how much obscuration was affecting the non-photometric nights. Field stars on each image were monitored on non-photometric nights to adjust the comet’s magnitude for varying obscuration during the night (discussed in the following subsection).
Comparison Star Corrections {#comp_stars}
---------------------------
Relative photometry of the comet with respect to field stars was carried out. However, not enough of the same field stars were available during the entire night and we therefore replaced stars which left the field of view with new ones as they entered. When possible we used only stars brighter than m$_R$ = 15, but if fewer than three were available, we used fainter stars, going as faint as m$_R$ = 16. Typically between three and six comparison stars were available at a given time.
Reliable catalog magnitudes were not available for enough comparison stars to use comparison stars for absolute calibrations. Therefore, we first used the median of the seven brightest measurements for a given star during the night as its least obscured brightness and determined the adjustment necessary to bring all fainter measurements into agreement. The comparison star correction for an image was the median offset from each star’s least obscured brightness during the night for all comparison stars on a given image. Magnitude corrections derived from comparison stars were applied on the nights specified as non-photometric in Table \[t:obs\_summary\]. No comparison star corrections were applied for the photometric nights after first using the comparison stars to confirm that the nights were indeed photometric. Nightly median comparison star corrections were 0.1–0.4 mag, although some corrections exceeded 1.0 mag on 1999 April 18.
Since the least obscured brightness of the field stars will be fainter than their ideal brightness if it had been photometric, this technique will introduce a systematic nightly magnitude offset from night to night. Therefore, after correcting the relative photometry with field stars, the entire lightcurve for each night was adjusted so that the peaks for all nights of data were at the same magnitude ($\Delta$m$_2$, discussed in Section \[mag\_adjustments\]). If the conditions varied during a night such that there was less obscuration while certain stars were observed but more obscuration while other stars were observed, the uncertainty in the comparison star correction will increase. Hence, the lightcurves on non-photometric nights may exhibit more scatter.
The calibrated, extinction corrected, and comparison star corrected R magnitudes (m$_R$) are plotted in Figure \[fig:orig\_lcs\] as a function of UT. This indicates how much nightly coverage was obtained. In April and May, Tempel 2 was observed for about five hours, while in June it was observed for about seven hours, with the observing window moving earlier each night. Because the comet brightened during the apparition and the coma increased, the median magnitude is different in each figure and later nights have smaller amplitudes (the vertical scale is held fixed in all panels but the range of magnitudes varies from panel to panel). In order to create a uniform dataset for period analysis and to study the lightcurve amplitudes, we removed the coma, corrected for changing $r$, $\Delta$, and $\phi$, and adjusted for nightly offsets. These adjustments will be discussed in the next two sections.
In Figure \[fig:orig\_lcs\_phased\] we show the same data as in Figure \[fig:orig\_lcs\] but phased to 8.941 hr (our best period which will be discussed in Section \[1999\_period\]). This figure emphasizes the increase in brightness and decrease in amplitude throughout the run, as later nights are higher in the figure and have smaller amplitudes. It also demonstrates the phase coverage obtained during each run, making it possible to determine the rotation period even without removing the coma contamination.
Removal of Coma Contamination {#coma_removal}
-----------------------------
While Tempel 2 is not strongly active, some coma was visible throughout the apparition. The coma contamination is, in general, not large and the rotation period can be determined without coma removal. However, to ensure that it was not affecting our results and to compare the amplitude of the lightcurve with those obtained by other authors, we removed the coma inside the monitoring aperture. For unchanging grains flowing radially outward from the nucleus at a constant velocity, the coma flux per pixel decreases as $\rho$$^{-1}$, where $\rho$ is the projected distance from the nucleus. Since the area of equally spaced annuli increases as $\rho$, the total coma flux in each annulus should be constant. In reality, the coma often does not fall off as $\rho$$^{-1}$ and there are factors which may cause further deviation from a constant total flux per annulus such as contamination by background stars or cosmic rays, wings of the nucleus PSF, and imperfect background removal. However, we found that a linear fit to the total annular flux as a function of annular distance from the nucleus ($\rho$) provides a reasonable first order approximation of the coma. This is illustrated in Figure \[fig:coma\_removal\] using 1999 June 9 as a representative night.
We calculated the total flux in 3 arcsec wide annuli centered on the nucleus (e.g., 0–3 arcsec, 3–6 arcsec,...27–30 arcsec) to create radial profiles (total annular flux as a function of $\rho$) for each image. These are plotted as solid light gray curves (images without significant contamination from background stars) and dotted dark gray curves (images with significant contamination from background stars) in Figure \[fig:coma\_removal\]. We then computed the median total flux in each annulus for the night, ignoring images with obvious contamination from background stars. Next, we fit a straight line to the total annular flux as a function of distance from the nucleus, $\rho$, (the heavy black line in Figure \[fig:coma\_removal\]) from $\rho$ = 7.5–19.5 arcsec for each image (where $\rho$ = 7.5 arcsec was the center of the 6–9 arcsec annulus and $\rho$ = 19.5 arcsec was the center of the 18–21 arcsec annulus). This range was chosen to exclude as much of the signal from the nucleus ($\rho <$ 6 arcsec) as possible and to minimize contamination from passing stars ($\rho >$ 21 arcsec). We extrapolated the fit in to the nucleus ($\rho$ = 0 arcsec) and out to $\rho$ = 30 arcsec. The total coma annular flux was removed from the total annular flux to give the coma corrected total annular flux (i.e. the total nucleus annular flux), which was integrated and converted back to magnitudes. Since the monitoring aperture was 6 arcsec in radius, the total coma which was removed for a night was the sum of the total coma annular flux in the 0–3 and 3–6 arcsec radius annuli.
When determining the nightly coma, we excluded images with stars which obviously altered the fit. Stars that contaminate the larger annuli tend to result in an under-removal of coma in the monitoring aperture, while stars closer to the nucleus tend to result in an over-removal of the coma in the monitoring aperture. While we excluded the images with obvious contamination from background stars, fainter stars undoubtedly remain. Since more background stars pass through the larger annuli than the smaller annuli, this results in a systematic under-removal of the coma.
As shown in Figure \[fig:coma\_removal\], typical profiles appear to have some curvature. This implies that more coma should be removed than is accomplished with a linear fit. A significant contributor to the curvature at small $\rho$ is likely the wings of the nucleus PSF, while at large $\rho$, the coma signal may simply be too small and is swamped by uncertainty in the background removal. We considered higher order fits to better match the curvature; a quadratic fit removed 20–60% more coma while an exponential fit was poor as it often implied negative nucleus counts. Although these fits removed more coma than the linear fit, they varied more from image to image, resulting in wider variance in the estimated coma. Also, small fluctuations in the larger annuli, where there is very little coma, have a large effect on the estimate of the coma contamination for these fits. Therefore, we concluded that a simple linear fit provided the best compromise for approximating the coma profile. We note that if a different fit to the coma were used, it would somewhat alter the amplitude of the lightcurve but would not change the location of its extrema. Thus, the coma removal technique does not affect the key result of this paper, the period determination (Section \[1999\_period\]).
To test that our use of a single nightly coma correction was appropriate, we determined linear coma fits for each usable image on a night. We saw no systematic variation in the estimated coma flux within the photometric aperture as a function of rotational phase throughout the entire campaign, confirming that the use of a single nightly median correction was reasonable. As another test of the coma removal technique, we investigated the total annular nucleus flux remaining after coma removal. The median fraction for all nights of the nucleus flux contained in the 3 arcsec radius aperture relative to the 9 arcsec radius aperture was 85%. The median for the 6 arcsec radius aperture relative to the 9 arcsec radius aperture was 99%. Apertures larger than 9 arcsec in radius were within $\pm$1% of the 9 arcsec radius aperture flux before deviating at $\rho$ $>$ 21 arcsec (which was beyond the range of the coma fit). These ratios were relatively constant throughout the apparition. The ratio of the 3 arcsec radius aperture relative to the 9 arcsec radius aperture varied with seeing changes, but the 6 arcsec radius aperture did not change appreciably. This confirms that the 6 arcsec radius aperture is appropriate for photometric monitoring, and that the light lost by not going to larger apertures should be minimal and roughly the same fraction in all images, causing no effect on the amplitude of the lightcurve, regardless of seeing.
Magnitude Adjustments {#mag_adjustments}
---------------------
The viewing circumstances changed significantly during our observations. Therefore, we adjusted the nucleus magnitudes using the standard asteroidal normalization $$\label{e:mag_norm}
m_{R}(1,1,0) = m_{R,cr} - 5log(r\Delta) - \beta \alpha$$ where m$_R$(1,1,0) is the normalized magnitude at $r$ = $\Delta$ = 1 AU and $\alpha$ = 0$^\circ$, m$_{R,cr}$ is the apparent magnitude, m$_R$ (which has had the absolute calibrations, extinction corrections, and comparison star corrections applied), with the coma removed, $r$ is the heliocentric distance (in AU), $\Delta$ is the geocentric distance (in AU), $\beta$ is the linear phase coefficient, and $\alpha$ is the phase angle. $\beta$ is typically 0.03–0.04 mag deg$^{-1}$ for comets, and we used $\beta$ = 0.032 mag deg$^{-1}$ since it minimized the $\Delta$m$_2$ adjustment (discussed in the following paragraph). Equation \[e:mag\_norm\] removes the secular variation in brightness and allows comparison of all lightcurves on a similar scale. The geometric corrections are given as $\Delta$m$_1$ in column (11) in Table \[t:obs\_summary\] at the midpoint of each night’s observations.
The geometric corrections do not always bring the lightcurves from different nights to the same peak brightness, with variations as high as 0.33 mag. This is due to a number of factors, including: bootstrapping Landolt standard stars from HB standards on photometric nights when Landolt stars were not observed; the application of absolute calibrations on non-photometric nights; comparison stars that are normalized to their brightest point in the night rather than a catalog value; and the shape of the coma removed. Therefore, we introduced an additional adjustment, $\Delta$m$_2$ (column (12) in Table \[t:obs\_summary\]), to adjust the individual lightcurves to a common peak brightness. It should be noted that this factor is introduced to simplify the rotation period analysis, and should not be interpreted as representing any particular physical property.
After phasing the data with preliminary $\Delta$m$_2$ values, we refined $\Delta$m$_2$ so that all lightcurves were aligned at the peak near $\sim$0.85 phase when possible. April 17 and June 22 were aligned using the peak near $\sim$0.35 phase. Since April 18 and June 9 did not have conclusive peaks, $\Delta$m$_2$ for these nights was estimated based on the cluster of points near phase 0.85.
Data
----
The photometry is given in Table \[t:photometry\]. Columns (1) and (2) are the UT date and time (at the telescope) at the midpoint of each exposure (adjustments for light travel time are given in Table \[t:obs\_summary\]). Column (3) is m$_R$, the observed R-band magnitude after photometric calibrations, extinction corrections, and comparison star corrections have been applied. Column (4) is m$_R$$^*$, the coma-removed, reduced magnitude m$_R$(1,1,0) corrected by $\Delta$m$_2$ so that all nights have a similar peak magnitude. We obtained 1016 data points, of which 124 were discarded due to contamination from background stars, tracking problems, or cosmic ray hits.
While there are several adjustments which have been applied to arrive at the m$_R$$^*$ values listed in Table \[t:photometry\], they do not affect the primary focus of the paper: the rotation period determination. The corrections for geometry ($\Delta$m$_1$) and nightly offsets ($\Delta$m$_2$) shift individual nightly lightcurves up or down, but the locations of the extrema in rotational phase do not change. While the choice of coma removal algorithm affects the amplitude of the lightcurve, it also leaves the phase of the extrema unchanged. Therefore, the same rotation period could be determined directly from the m$_R$ values prior to the corrections.
Due to the large number of sources of error for each data point (photon uncertainty in comet and background flux, extinction correction, comparison star correction, coma removal), we estimated the effective uncertainty in the magnitude by fitting a smoothed spline through each night’s lightcurve and subtracting the spline fit. The uncertainty for the night was estimated as the standard deviation of the residuals. The uncertainty was not always constant throughout a night since seeing conditions and obscuration varied on the non-photometric nights, so on some nights we estimated the uncertainty for subsections of the lightcurve. Variations in the comparison star magnitudes were used as an additional indicator of when the uncertainty changed during a night. The range of uncertainties for a night are given in column (11) in Table \[t:obs\_summary\].
Modeling and Interpretation {#modeling}
===========================
Determining the Rotation Period in 1999 {#1999_period}
---------------------------------------
We defined zero phase to be perihelion (1999 September 8.424), and accounted for the light travel time (column (8) in Table \[t:obs\_summary\]) before phasing the data. We used an interactive period search routine within the Data Desk data analysis package[^4] which updates the phased lightcurves on the fly, allowing us to easily scan through potential periods. We looked for alignment of extrema since they can be identified in phase space even if the magnitudes or amplitudes are different. As the correct solution is approached, the extrema tend to be systematically shifted in phase based on the (color-coded) date of observation, making it easy to quickly hone in on the optimal solution. However, it is difficult to quantitatively estimate the uncertainty in the period since the alignment of the features is probably well away from the optimal solution by the time the eye can distinguish it, therefore overestimating the uncertainty.
We applied this technique to the m$_R$$^*$ data given in column (4) of Table \[t:photometry\], finding an observed, i.e. synodic, period of 8.941 $\pm$ 0.002 hr for the combined dataset (April–June). We examined all period solutions between 4.47 hr (the single-peaked solution) and 13.41 hr (the triple-peaked solution) and 8.941 hr (the double-peaked solution) is the only viable solution between these extremes. The double-peaked shape is expected for a triaxial ellipsoid and was observed by previous authors in 1988 and 1994. We plot the m$_R$$^*$ data phased to 8.941 hr in the top panel of Figure \[fig:1999\_data\]. The bottom panel is phased to 8.932 hr (the period solution from 1988, discussed in the following subsection).
The uncertainty in the period was estimated from the smallest offset that produced lightcurves that were visibly out of phase. Figure \[fig:best\_period\] illustrates an exaggeration of this process using three periods: 8.937 hr (top), 8.941 (middle) and 8.945 (bottom). Note that these periods are separated by double the estimated error to emphasize the trend in the location of the extrema. We plot the m$_R$$^*$ data offset vertically by $-$0.015 mag day$^{-1}$ since 1999 April 17 to better show the chronological progression of the lightcurves. In the top panel, later lightcurves (higher on the plot) occur at a later phase, while in the bottom panel later lightcurves occur at an early phase. This clearly illustrates that the correct solution is between the two, making it easy to identify 8.941 hr as the optimal period. The inclusion of the April data is vital for this process, as the larger separation between April and the later datasets decreases the range of possible solutions.
For comparison, we also conducted period searches using Fourier [@deeming75] and phase dispersion minimization (PDM; @stellingwerf78) techniques. The Fourier routine “Period04”[^5] [@lenz05] found a single-peaked period of 4.4705 $\pm$ 0.0001 hr. Since we expect the lightcurve to be double-peaked, this implies a rotation period twice as long, or 8.9409 $\pm$ 0.0002 hr (the difference in the last digit is due to rounding). The routine determined an uncertainty in frequency and we converted this to a period uncertainty, although we note that the uncertainty of 0.0002 hr seems unrealistically small. We used our own PDM routine, which found a double-peaked period of 8.941 $\pm$ 0.005 hr. We estimated the period uncertainty by finding the range of solutions that had $\theta$ (a measure of the goodness of fit) within 50% of the minimum $\theta$. The agreement between methods confirms the robustness of our period solution.
Different Rotation Periods Due to Geometry? {#geometry}
-------------------------------------------
The default rotation period obtained by simply phasing the data is the synodic period, the length of time until the brightness appears the same from the Earth. This period changes during the apparition because the illuminated portion of the nucleus seen by the Earth varies as both the Earth and the comet move in their orbits. The sidereal rotation period is the rotation period relative to a fixed position in inertial space. A third period is the solar period, which is the time until the Sun is in the same place in the sky as seen by the comet. Because the synodic and solar rotation periods depend on the orientation of the rotational pole, converting them to a sidereal period requires a pole solution.
Since the synodic period also depends on the position of the Earth, we must consider the geometry of the Earth-comet-Sun system during our observations to confirm that the synodic period changes slowly enough that measuring a single synodic period for the whole 10-week time span is sensible. As seen from the comet, the ecliptic longitude of the Earth (column (9) of Table \[t:obs\_summary\]) and the ecliptic latitudes of the Sun and Earth all changed by less than 5$^\circ$ during our observations, resulting in minimal changes to the measured rotation period. While the ecliptic longitude of the Sun as seen from the comet (column (10) of Table \[t:obs\_summary\]) changed by $\sim$30$^\circ$, the net effect on the rotation period was still small due to the likely pole orientation (see discussion below), and obtaining a single (synodic) rotation period for the whole time span was valid.
To confirm this, we investigated the synodic rotation periods which would be observed at different times for pole solutions within 20$^\circ$ in the comet’s orbital coordinate system (obliquity and azimuthal angle) of the pole solution given by @sekanina87b. This range of solutions was chosen to include the pole solution found by @sekanina91 for the 1988 apparition and a preliminary pole solution for our 1999 coma morphology data (we will explore this in detail in Paper 2). We considered three runs discretely: April 17–19, May 26–27, and June 22–23 (the June 8–11 run gave similar results to the June 22–23 run), and determined the synodic–sidereal offset which would be observed for all pole solutions under consideration. The difference between the synodic and sidereal periods was 0.000 to $+$0.003 hr during the April run, 0.000 to $+$0.002 hr during the May run, and $-$0.001 to $+$0.001 hr during the late-June run. Considering the entire 10-week interval as a single large run (i.e. April 17–June 23), the best synodic period differed from the sidereal period by 0.000 to $+$0.002 hr. Thus, the changing geometry during our observations did not introduce an error in the period determined from the entire run larger than the estimated uncertainty in the synodic period of 0.002 hr (although if the April run was considered by itself, the maximum difference was 0.003 hr). Note that the sign of the synodic–sidereal difference reverses for a pole orientation in the opposite direction, i.e. when the comet has retrograde spin and “north” is defined in the opposite hemisphere.
The above example suggests that changes in the synodic period may be detectable during an apparition. We considered subsets of our 1999 data to look for such a change. There was no discernible difference when the rotation period was determined using all the April and May data or all the May and June data (both intervals gave acceptable solutions from 8.940–8.942 hr with a best period of 8.941 hr). We also determined the synodic period between every combination of individual nights from different runs for which a synodic period could be determined. We phased each pair of nights and defined the midpoint between the two nights as the date corresponding to the measured synodic period. This method suggests an increase in the synodic period of $\sim$0.002 hr from early-May to mid-June, although the scatter in the individual synodic measurements is very large and the range of acceptable solutions for each pair is generally at least $\pm$ 0.002 hr. If correct, this requires that the “north” pole be in the opposite hemisphere as was determined by @sekanina87b, i.e. retrograde rather than prograde rotation.
Reanalysis of the 1988 Data {#1988_data}
---------------------------
We reanalyzed the publicly available datasets from 1988 (@jewitt89, @ahearn89, and @wisniewski90), defining zero phase at perihelion (1988 September 16.738). The observing scenarios for these datasets are summarized in Table \[t:obs\_summary\_1988\]. @jewitt89 removed coma from their June data but not the February or April data. @ahearn89 obtained optical and thermal IR data simultaneously on different telescopes, and removed coma from the optical data but not from the thermal IR data, which were shown to be nearly free of coma. @wisniewski90 did not remove coma from his data. The 1987 data from @jewitt88 were not included in the 1988 period determination because they were too far away in time and were at a much different location in the orbit, making it impossible to directly compare the synodic rotation periods.
We normalized the published magnitudes to m$_R$(1,1,0) using $\beta$ = 0.032 mag deg$^{-1}$ as was used with our 1999 data. After this normalization, the optical magnitudes still differed by several tenths of a magnitude between runs and authors. We therefore adjusted all the lightcurves to peak at the same magnitude for the maximum near zero phase by applying nightly $\Delta$m$_2$ values (given in column (12) of Table \[t:obs\_summary\_1988\]). We call these data m$_R$$^*$, and used them for our period search. $\Delta$m$_2$ corrects for nightly differences between datasets due to differences in the extinction correction, absolute calibration, and coma removal (or lack thereof). We do not list a $\Delta$m$_2$ for the @ahearn89 4845 [Å]{} data, as we did not use them when phasing the data since the 4845 [Å]{} and 6840 [Å]{} data were obtained alternately throughout both nights, and their lightcurve shapes were nearly identical (although the 6840 [Å]{} data were brighter).
Visually scanning the 1988 data for a double-peaked solution, as was done for our 1999 data, yielded a period of 8.932 $\pm$ 0.001 hr. This period agrees with the synodic periods determined for the 1988 data by @sekanina91 (8.931 $\pm$ 0.001 hr, given in his Table 5) and @mueller96 (8.933 $\pm$ 0.002)[^6]. The viewing geometry in 1988 was nearly identical to the viewing geometry in 1999; the ecliptic latitudes of the Earth and Sun changed minimally and the ecliptic longitudes of the Earth and Sun (given in columns (9) and (10), respectively, in Table \[t:obs\_summary\_1988\]) mimicked the changes in 1999. As with the 1999 data (Section \[1999\_period\]), we considered discrete runs spanning the range of 1988 observations: February 24–28, April 8–12, May 18–22, and June 19–23, and determined the synodic–sidereal difference during each run for the range of pole solutions. Because of the similar viewing geometry, the synodic–sidereal difference was almost the same as in 1999: $+$0.001 to $+$0.002 hr for February 24–28, $+$0.001 to $+$0.002 hr for April 8–12, 0.000 to $+$0.001 hr for May 18–22, and $-$0.001 to $+$0.001 hr for June 19–23. Considering the entire 1988 time span from February 25 until June 30 as a single large run, the difference in the synodic and sidereal rotation periods was 0.000 to $+$0.002 hr for the range of pole solutions. As with 1999, the synodic–sidereal difference changes sign for a pole orientation in the opposite direction. Thus we can directly compare our 1988 solution (8.932 hr) with our 1999 solution (8.941 hr). In Figure \[fig:1988\_data\] we plot the 1988 m$_R$$^*$ data phased to our period solution from 1988 in the top panel and to our period solution from 1999 in the bottom panel. Similarly, we plot our 1999 m$_R$$^*$ data phased to 8.941 hr (top panel) and 8.932 hr (bottom panel) in Figure \[fig:1999\_data\]. Figures \[fig:1999\_data\] and \[fig:1988\_data\] demonstrate that the rotation periods were clearly different in 1988 and 1999. Since the synodic periods differed by $\sim$0.009 hr and synodic–sidereal offsets were similar each apparition, we can conclude that the sidereal periods also differed by $\sim$0.009 hr, and thus the rotation period increased from 1988 to 1999.
We looked for a change in the synodic period during the 1988 observations by phasing every combination of pairs of observing runs, e.g. @jewitt89’s February data with @ahearn89’s June data, @jewitt89’s February data with @wisniewski90’s June data, etc. Assigning the date of each pair as the midpoint between the two runs, there was no discernible trend in the measured synodic period over time, with acceptable values varying from 8.931–8.933 hr with a typical uncertainty of $\pm$0.001 hr. This is consistent with the conclusions of @jewitt89 who found no change in the rotation period within their uncertainty when analyzing their February, April, and June data as discrete sets. In contrast, @sekanina91 found a decrease in the synodic period during the 1988 observations of $\sim$0.003 hr (his Table 5 and Figure 7). These discrepancies and our possible finding of a slight increase in the rotation period in 1999 (Section \[1999\_period\]) suggest that @sekanina87b’s pole solution may not be definitive, with the alternate sense of rotation possible as well (resulting in a “north” pole in the exact opposite position). We will use the coma morphology visible in our post-perihelion 1999 images, combined with data we will obtain in 2010, to attempt to determine a more robust pole solution in Paper 2.
Reanalysis of the 1994 Data {#1994_data}
---------------------------
We also reanalyzed the @mueller96 data from 1994, normalizing the published magnitudes to m$_R$(1,1,0) and visually scanning for possible solutions. The data do not yield an unambiguous double-peaked lightcurve due to undersampling. @mueller96 found five possible solutions: 8.877 hr, 8.908 hr, 8.939 hr, 8.971 hr, and 9.002 hr, and we concur with their results. Only one of these, 8.939 hr, is within the uncertainty of our 1999 solution (8.941 $\pm$ 0.002 hr).
Using the same range of pole solutions as in Section \[geometry\], we find the maximum difference between the synodic and sidereal periods in October–December 1994 was smaller than $-$0.001 hr (actually about $-$1 s), in agreement with the difference found by @mueller96. Thus, the expected difference between the synodic and sidereal rotation periods is much less than the difference between the five possible solutions. Furthermore, since there were no intervening perihelion passages between the 1994 and 1999 observations, the sidereal rotation period was likely unchanged between these observations. Therefore, we rule out the remaining four possible solutions from @mueller96, and conclude that Tempel 2 had spun-down by $\sim$0.009 hr ($\sim$ 32 s) between the 1988 and 1994 determinations, which spanned two perihelion passages, and was unchanged between the 1994 post-perihelion observations and our 1999 pre-perihelion observations.
Coma Contamination and Lightcurve Amplitude {#amplitude}
-------------------------------------------
Tempel 2 was non-stellar in appearance throughout our 1999 observations. The presence of a coma suppressed the amplitude of the lightcurve, with the suppression increasing later in the apparition as activity increased. The mean coma removed from the 6 arcsec radius monitoring aperture was 0.20 mag from April 17–19, 0.27 mag from May 26–27, 0.40 mag from June 8–11, and 0.65 mag from June 22–23. We measured lightcurve amplitudes ranging from 0.2–0.5 mag prior to removal of the coma. After coma removal, the April 17–19 data exhibited amplitudes of $\sim$0.75 mag while all subsequent runs had amplitudes of 0.45–0.50 mag. If a quadratic fit to the coma was used instead of a linear fit, the amplitudes increased to $\sim$0.80 in April and 0.50–0.60 in May and June. However, note that as discussed in Section \[coma\_removal\], we opted for a more conservative linear fit rather than a higher order fit because the higher order fits varied more from image to image and produced less reliable results. Our measured amplitudes in 1999 are similar to the maximum optical amplitudes reported by other authors in 1988 when the comet was at a similar viewing geometry and activity level: 0.65 $\pm$ 0.05 mag in February and 0.60 $\pm$ 0.05 mag in April (no coma removal; @jewitt89), 0.5 mag in May (no coma removal; @wisniewski90), and 0.7 $\pm$ 0.1 mag in June (coma removed; @jewitt89).
@ahearn89 measured an amplitude of 0.8 mag in their 10.1 $\mu$m data in 1988 June, exceeding all measured optical amplitudes at similar times in both 1988 and 1999. While the amplitudes measured by ourselves and others approach the thermal IR amplitude, all are somewhat lower, and even a quadratic coma removal from our data cannot bring the amplitudes into agreement. This discrepancy raises two possible interpretations. First, the coma removal techniques employed here and by @jewitt89 and @ahearn89 are not adequately estimating the coma inside the photometric aperture. To push the amplitude up to $\sim$0.80, 50–150% more coma would need to be removed from our May and June data, requiring a significantly steeper curve than even a quadratic fit. Alternatively, the convergence of multiple authors on an optical amplitude of 0.6–0.7 mag in 1988 June and 1999 June may indicate that this is in fact the correct optical amplitude, and is different from the thermal IR amplitude. This possibility was suggested by @ahearn89 based on work by @brown85 who showed that the amplitude of the thermal IR lightcurve may be dependent on both wavelength and elongation.
The varying lightcurve amplitudes will be utilized in Paper 2 when we determine the pole solution using the coma morphology from 1999 and 2010. As the viewing geometry varies during an apparition, the amplitude of the lightcurve will vary because different cross sections of the nucleus are observed. Thus, the lightcurve amplitudes will be useful in refining the pole solution and will help in constraining the shape of the nucleus.
We looked for evidence of periodicity in the coma signal by plotting both the total flux and flux per square arcsec in each photometric annulus as a function of rotational phase. What structure is visible on individual nights does not repeat in a coherent manner from night to night or from run to run. There was also no evidence of propagation of any features (either a maximum or minimum signal) outwards during a night. This is consistent with the analyses of @jewitt89 and @ahearn89 who did not see a rotation signal in their data.
Lightcurve Asymmetry {#lc_shape}
--------------------
The rotational lightcurve in 1999 is somewhat asymmetric, allowing identification of certain features many rotations apart and eliminating $N - 0.5$, $N + 0.5$, $N + 1.5$... (where $N$ is the number of cycles) as possible solutions. This conclusively rules out the single- and triple-peaked solutions and reduces the uncertainty in the period determination since it halves the number of possible rotations. Any clear description of the lightcurve shape is made challenging because it is never observed completely on a given night. However, we note the following apparently repeated features: The minimum near 0.6 phase is deeper and steeper than the minimum near 0.1 phase when comparing lightcurves from the same observing run. There is a shoulder in the rising portion between 0.15 and 0.30 phase. Finally, while there are no nights in which both maxima were clearly observed, the maximum near 0.85 phase may be slightly higher than the maximum near 0.35 phase. We see evidence for each of these features in the 1988 data when discrete observing runs are considered[^7]. The features are also seen in the 1994 data from @mueller96, although since the viewing geometry was very different, we cannot be sure that they correspond to the same topographic features on the nucleus.
Summary and Discussion {#discussion}
======================
We imaged Tempel 2 on 32 nights from 1999 April until 2000 March. We present here R-band nucleus lightcurves obtained on 11 of these nights from April through June (prior to perihelion), a total of 892 data points. Absolute calibrations and nightly extinction corrections were determined on photometric nights, and were applied to non-photometric nights to extract usable lightcurves. Field stars in the image were monitored on all nights and were used as comparison stars to correct the comet magnitude on non-photometric nights. A median coma was determined each night and removed from all images on the night. All magnitudes were normalized to $r$ = $\Delta$ = 1 AU and $\alpha$ = 0$^\circ$. A final adjustment, which accounts for uncertainties in the calibrations that tend to create nightly magnitude offsets such as the application of absolute calibrations on non-photometric nights, was applied to make each night peak at the same magnitude.
We determined a synodic rotation period of 8.941 $\pm$ 0.002 hr; no other double-peaked solutions exist. Our dataset is sufficient to rule out all aliases other than the single- and triple-peaked solutions, and these are ruled out by the differences in shape and brightness between the two minima. Our period matches one of the possible periods obtained by @mueller96 from post-perihelion data in 1994 (during the same inter-perihelion time as our 1999 data) and rules out their other four possible solutions. We reanalyzed published data from 1988 using the same analysis techniques and found a synodic rotation period of 8.932 $\pm$ 0.001 hr. Our 1999 data cannot be fit by the rotation period from 1988, nor can the 1988 data be fit by the rotation period from 1999, even though the comet had nearly identical viewing geometries during the two sets of observations. Thus, we conclude that Tempel 2 spun-down by 0.009 hr ($\sim$32 s) from 1988 to 1999, an interval that included two perihelion passages.
This marks the first conclusive measurement of spin-down in a comet, although three comets have shown possible evidence of spin-up: 9P/Tempel 1 [@belton07; @chesley10], 2P/Encke [@fernandez05], and Comet Levy (1990c) [@schleicher91; @feldman92a]. Modeling by @samarasinha95, @neishtadt02, @gutierrez03, and others has shown that either spin-up or spin-down is possible, with spin-up more likely in the long run. These simulations have shown that under certain conditions the spin period can change by much larger amounts in a single orbit (e.g. typical changes of 0.01–10 hr per orbit according to @gutierrez03) than the measured change of 0.009 hr in two orbits for Tempel 2. Thus, the measured spin-down of 10P/Tempel 2 is probably not exceptional, and comparable changes in period could likely be measured in other comets if similarly high quality data were obtained over several epochs.
Possible causes of a change in spin-state are summarized by @samarasinha04. The spin-down of Tempel 2 may have been caused by either a one-time event or by recurring torquing each orbit. Scenarios for causing a one-time change in rotation period include a short-lived, freshly exposed active region (causing temporary torquing), fragmentation (changing the moment of inertia), collision with another object, and tidal torquing from Jupiter. The orbit of Tempel 2 is well known and tidal torquing from Jupiter should not have been significant or more severe between 1988 and 1999 than on other recent orbits. Collisions are rare in the current solar system, and any collision significant enough to alter the rotation period of Tempel 2 would likely have resulted in an increase in brightness upon subsequent orbits, which has not been seen. Fragmentation is known to be relatively common, occurring on average at least once per century per comet (c.f. @chen94 [@weissman80]). However, there is no evidence for Tempel 2 having split, such as a companion nucleus or a sustained increase in brightness as was observed in 73P/Schwassmann-Wachmann 3 [@ferrin10]. A new active region could be created by fragmention, impact, or outburst. Since outbursts are known to occur far more frequently than fragmentation or impacts, an outburst is the most plausible explanation for a one-time change in the rotation period. If the change in the period was caused by a unique event such as an outburst, it could have occurred anywhere in the orbit and we would expect no change in the rotation period on subsequent orbits.
It is challenging to devise a mechanism that is strong enough to affect the rotation period but lasts for a short time (less than one orbit). Thus, we suspect the spin-down is due to torques on the nucleus caused by asymmetric outgassing. This is believed to be the most common means of changing the spin state of a comet (c.f. @samarasinha04). Seasonal illumination of an active region could cause torquing which slows the rotation of the nucleus by a comparable amount each orbit. In the case of Tempel 2, the torquing likely occurs near or shortly after perihelion, when the comet is most active, although the length of time over which the torquing may occur is unknown. Since Tempel 2 reached perihelion twice between the 1988 and 1999 observations, in this, our preferred scenario, the rotation period must be changing by $\sim$16 s each orbit.
Tempel 2 reached perihelion on 2010 July 4, and we encourage its observation throughout the apparition to determine the rotation period. Due to a perturbation by Jupiter between the 1999 and 2005 apparitions, the viewing geometry is different in 2010 than in 1988 or 1999. The post-perihelion viewing geometry will be better this apparition and the comet will be visible longer each night. The smaller post-perihelion $\Delta$ in 2010 will provide better spatial resolution, making it easier to separate the nucleus and coma signals even though the coma is more extensive after perihelion. This may allow measurement of a nucleus lightcurve after perihelion, despite increased coma contamination relative to the pre-perihelion measurements in 1988, 1994, and 1999.
Assuming the spin-down is recurrent each orbit, we predict that the pre-perihelion rotation period should be $\sim$32 s longer than in 1999 as there have been two intervening perihelion passages (1999 September 8 and 2005 February 15) since our data were obtained in 1999. If the pole position given by @sekanina87b is correct, the sidereal rotation period should be 8.950 $\pm$ 0.003 hr. Since it is unknown if the spin-down is caused by an instantaneous impulsive event or a steady change over an extended period of time, we cannot estimate the rotation period that might be observed shortly after perihelion in 2010. Observations long enough after perihelion to allow the torquing to have taken effect should reveal an additional lengthening of the sidereal rotation period by $\sim$16 s relative to the pre-perihelion 2010 period. Using again the pole solution of @sekanina87b, this would yield a sidereal period 8.954 $\pm$ 0.004 hr.
We have begun to observe Tempel 2 in 2010 and will combine these new data with our observations obtained during its active phase in 1999 as a follow up to the current paper. As of August 2010 (when the revised manuscript was submitted), we are unaware of any period determinations from 2010 observations. In Paper 2 we will use the different viewing geometry provided by the 2010 apparition to investigate the pole orientation and the location and activity of any jets. We hope to determine the rotation period in 2010, and possibly obtain additional post-perihelion observations after activity has subsided in 2011 or later to resolve whether the comet spins down each orbit. With its relatively low inclination ($i$ $\sim$ 12$^\circ$) and perihelion close to the Earth’s orbit ($q$ $\sim$ 1.4 AU), Tempel 2 is frequently on the short list of targets for comet missions, making further study highly desirable.
Acknowledgements {#acknowledgements .unnumbered}
================
Thanks to Beatrice Mueller for a thorough and helpful review. We thank Christopher Henry, Kevin Walsh, and Wendy Williams for helping to obtain these observations, and Brian Skiff for useful discussions. We are grateful for JPL’s Horizons for generating observing geometries. The period analysis was greatly facilitated by the use of the Data Desk exploratory data analysis package from Data Descriptions, Inc. MMK and DGS were supported by NASA Planetary Astronomy grant NNX09B51G. TLF was supported by multiple NASA grants. EWS was supported by an NSF grant to Northern Arizona University for the Research Experiences for Undergraduates program.
[18]{} natexlab\#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\]\[\][[\#2](#2)]{}
, M. F., [Campins]{}, H., [Schleicher]{}, D. G., & [Millis]{}, R. L. 1989, , 347, 1155
, M. J., & [Drahus]{}, M. 2007, in Bulletin of the American Astronomical Society, vol. 38 of Bulletin of the American Astronomical Society, 498
, R. H. 1985, Icarus, 64, 53
, J., & [Jewitt]{}, D. 1994, Icarus, 108, 265
, S. R., [Belton]{}, M. J. S., [Gillam]{}, S. D., [Meech]{}, K. J., [Carcich]{}, B., & [Veverka]{}, J. 2010, in Bulletin of the American Astronomical Society, vol. 41 of Bulletin of the American Astronomical Society, 936
, T. J. 1975, , 36, 137
, T. L., [Schleicher]{}, D. G., & [A’Hearn]{}, M. F. 2000, Icarus, 147, 180
, P. D., [Budzien]{}, S. A., [Festou]{}, M. C., [A’Hearn]{}, M. F., & [Tozzi]{}, G. P. 1992, Icarus, 95, 65
, Y. R., [Lowry]{}, S. C., [Weissman]{}, P. R., [Mueller]{}, B. E. A., [Samarasinha]{}, N. H., [Belton]{}, M. J. S., & [Meech]{}, K. J. 2005, Icarus, 175, 194
, I. 2010, , 58, 365.
, P. J., [Jorda]{}, L., [Ortiz]{}, J. L., & [Rodrigo]{}, R. 2003, , 406, 1123
, D., & [Luu]{}, J. 1989, , 97, 1766
, D. C., & [Meech]{}, K. J. 1988, , 328, 974
, G. W. 2003, [Cometography: A Catalog of Comets. Volume 2, 1800-1899]{} (Cambridge, UK: Cambridge University Press)
— 2007, [Cometography: A Catalog of Comets. Volume 3, 1900-1932]{} (Cambridge, UK: Cambridge University Press)
— 2009, [Cometography: A Catalog of Comets. Volume 4, 1933-1959]{} (Cambridge, UK: Cambridge University Press)
, A. U. 1992, , 104, 340
, P., & [Breger]{}, M. 2005, Communications in Asteroseismology, 146, 53
, B. E. A., & [Ferrin]{}, I. 1996, Icarus, 123, 463
, A. I., [Scheeres]{}, D. J., [Sidorenko]{}, V. V., & [Vasiliev]{}, A. A. 2002, Icarus, 157, 205
, W. H., [A’Hearn]{}, M. F., [Carsenty]{}, U., [Millis]{}, R. L., [Schleicher]{}, D. G., [Birch]{}, P. V., [Moreno]{}, H., & [Gutierrez-Moreno]{}, A. 1990, Icarus, 88, 228
, N. H., & [Belton]{}, M. J. S. 1995, Icarus, 116, 340
, N. H., [Mueller]{}, B. E. A., [Belton]{}, M. J. S., & [Jorda]{}, L. 2004, [Rotation of Cometary Nuclei]{}, 281
, D. G., [Millis]{}, R. L., [Osip]{}, D. J., & [Birch]{}, P. V. 1991, Icarus, 94, 511
, Z. 1979, Icarus, 37, 420
— 1987, in Diversity and Similarity of Comets, edited by [E. J. Rolfe & B. Battrick]{}, vol. 278 of ESA Special Publication, 323
— 1991, , 102, 350
, R. F. 1978, , 224, 953
, P. R. 1980, , 85, 191
, F. L. 1950, , 111, 375
, W. Z. 1990, Icarus, 86, 52
[lccccccccccccc]{}\
Apr 17&7:48–11:55&1.8-m&SITe&2.025&1.275&24.0&0.177&80.1&56.3&$-$2.84&$-$0.02&0.02–0.04&thin clouds\
Apr 18&7:20–12:11&1.8-m&SITe&2.020&1.262&23.8&0.176&80.3&56.7&$-$2.81&+0.08&0.02–0.06&thin clouds\
Apr 19&6:53–12:10&1.8-m&SITe&2.014&1.248&23.7&0.174&80.5&57.1&$-$2.77&$+$0.03&0.03&photometric\
May 26&6:37–11:22&1.8-m&SITe&1.807&0.839&13.9&0.117&81.9&72.2&$-$1.36&+0.15&0.01&thin clouds\
May 27&7:43–11:30&1.8-m&SITe&1.802&0.831&13.6&0.116&81.8&72.7&$-$1.33&+0.26&0.01&thin clouds\
Jun 08&7:05–11:14&1.1-m&TI&1.741&0.748&10.7&0.104&80.1&78.4&$-$0.92&...&0.02&photometric\
Jun 09&4:45–9:36&1.1-m&TI&1.736&0.743&10.6&0.103&80.0&78.8&$-$0.90&...&0.03&photometric\
Jun 10&4:34–11:17&1.1-m&TI&1.731&0.737&10.5&0.102&79.7&79.3&$-$0.87&$-$0.02&0.02&photometric\
Jun 11&4:30–9:35&1.1-m&TI&1.726&0.732&10.4&0.102&79.6&79.8&$-$0.85&...&0.02&photometric\
Jun 22&7:29–11:08&1.1-m&TI&1.675&0.685&12.4&0.095&77.5&85.4&$-$0.69&$-$0.07&0.02–0.06&thin clouds\
Jun 23&3:57–11:04&1.1-m&TI&1.671&0.682&12.7&0.095&77.4&85.9&$-$0.69&$-$0.07&0.02–0.04&photometric\
\[t:obs\_summary\]
[lccc]{}
\
\
& & &\
\
[[** – continued from previous page**]{}]{}\
\
Apr 17&7.848&16.93&14.38\
Apr 17&7.915&16.90&14.34\
Apr 17&7.952&17.00&14.47\
Apr 17&7.989&16.90&14.34\
Apr 17&8.038&16.83&14.25\
Apr 17&8.076&16.89&14.33\
Apr 17&8.112&16.84&14.26\
Apr 17&8.201&16.76&14.16\
Apr 17&8.294&16.83&14.24\
Apr 17&8.331&16.74&14.13\
Apr 17&8.368&16.73&14.13\
Apr 17&8.421&16.73&14.13\
Apr 17&8.458&16.68&14.06\
Apr 17&8.495&16.69&14.07\
Apr 17&8.547&16.72&14.11\
Apr 17&8.637&16.71&14.10\
Apr 17&8.674&16.68&14.07\
Apr 17&8.711&16.73&14.12\
Apr 17&8.748&16.71&14.10\
Apr 17&8.785&16.67&14.04\
Apr 17&8.821&16.63&14.00\
Apr 17&8.893&16.64&14.02\
Apr 17&8.972&16.65&14.02\
Apr 17&9.067&16.62&13.99\
Apr 17&9.104&16.59&13.94\
Apr 17&9.142&16.60&13.96\
Apr 17&9.179&16.58&13.93\
Apr 17&9.216&16.61&13.98\
Apr 17&9.253&16.58&13.93\
Apr 17&9.315&16.58&13.93\
Apr 17&9.421&16.56&13.91\
Apr 17&9.486&16.55&13.91\
Apr 17&9.523&16.54&13.90\
Apr 17&9.560&16.56&13.92\
Apr 17&9.597&16.56&13.91\
Apr 17&9.634&16.59&13.96\
Apr 17&9.671&16.57&13.93\
Apr 17&9.716&16.60&13.96\
Apr 17&9.807&16.60&13.97\
Apr 17&9.844&16.63&14.00\
Apr 17&9.881&16.62&13.98\
Apr 17&9.918&16.62&13.99\
Apr 17&9.955&16.62&13.99\
Apr 17&9.992&16.62&13.99\
Apr 17&10.037&16.64&14.01\
Apr 17&10.219&16.67&14.05\
Apr 17&10.288&16.70&14.09\
Apr 17&10.325&16.68&14.07\
Apr 17&10.362&16.71&14.11\
Apr 17&10.399&16.73&14.13\
Apr 17&10.436&16.75&14.14\
Apr 17&10.473&16.75&14.16\
Apr 17&10.604&16.79&14.21\
Apr 17&11.689&16.95&14.41\
Apr 17&11.726&16.95&14.41\
Apr 17&11.763&16.93&14.38\
Apr 17&11.800&16.94&14.39\
Apr 17&11.890&16.91&14.35\
\
Apr 18&7.681&16.47&13.89\
Apr 18&7.718&16.44&13.86\
Apr 18&7.755&16.46&13.89\
Apr 18&7.792&16.48&13.91\
Apr 18&7.829&16.48&13.90\
Apr 18&7.867&16.47&13.90\
Apr 18&7.906&16.49&13.91\
Apr 18&7.943&16.51&13.94\
Apr 18&7.980&16.55&13.99\
Apr 18&8.017&16.55&13.99\
Apr 18&8.054&16.61&14.05\
Apr 18&8.091&16.62&14.06\
Apr 18&8.160&16.61&14.05\
Apr 18&8.234&16.63&14.08\
Apr 18&8.395&16.68&14.14\
Apr 18&8.432&16.69&14.15\
Apr 18&8.469&16.74&14.21\
Apr 18&8.506&16.66&14.12\
Apr 18&8.543&16.78&14.25\
Apr 18&8.580&16.73&14.20\
Apr 18&8.621&16.74&14.21\
Apr 18&8.702&16.83&14.31\
Apr 18&8.739&16.81&14.29\
Apr 18&8.777&16.84&14.33\
Apr 18&8.814&16.83&14.32\
Apr 18&8.851&16.89&14.39\
Apr 18&8.888&16.90&14.40\
Apr 18&8.962&16.93&14.44\
Apr 18&9.028&16.98&14.50\
Apr 18&9.109&17.00&14.53\
Apr 18&9.146&16.99&14.52\
Apr 18&9.183&17.04&14.57\
Apr 18&9.220&17.01&14.54\
Apr 18&9.258&17.03&14.56\
Apr 18&9.295&17.05&14.59\
Apr 18&9.332&17.07&14.61\
Apr 18&9.429&17.07&14.61\
Apr 18&9.535&17.06&14.60\
Apr 18&9.611&17.02&14.55\
Apr 18&9.649&17.03&14.56\
Apr 18&9.686&17.01&14.54\
Apr 18&9.723&17.01&14.53\
Apr 18&9.760&17.04&14.57\
Apr 18&9.797&17.07&14.61\
Apr 18&9.836&17.04&14.58\
Apr 18&9.873&17.01&14.54\
Apr 18&9.910&17.01&14.54\
Apr 18&9.947&16.98&14.51\
Apr 18&9.985&16.98&14.50\
Apr 18&10.022&16.98&14.50\
Apr 18&10.062&16.99&14.51\
Apr 18&10.391&16.73&14.20\
Apr 18&10.460&16.86&14.35\
Apr 18&10.541&16.87&14.37\
Apr 18&10.579&16.85&14.35\
Apr 18&10.616&16.83&14.32\
Apr 18&10.653&16.80&14.28\
Apr 18&10.690&16.90&14.40\
Apr 18&10.811&16.80&14.29\
Apr 18&10.917&16.75&14.22\
Apr 18&10.959&16.77&14.24\
Apr 18&10.996&16.67&14.13\
Apr 18&11.033&16.74&14.22\
Apr 18&11.070&16.74&14.21\
Apr 18&11.107&16.71&14.18\
Apr 18&11.144&16.70&14.16\
Apr 18&11.210&16.68&14.14\
Apr 18&11.277&16.67&14.13\
Apr 18&11.356&16.62&14.07\
Apr 18&11.393&16.64&14.09\
Apr 18&11.431&16.64&14.09\
Apr 18&11.468&16.66&14.12\
Apr 18&11.505&16.61&14.06\
Apr 18&11.542&16.61&14.06\
Apr 18&11.623&16.60&14.05\
Apr 18&11.664&16.57&14.01\
Apr 18&11.702&16.57&14.01\
Apr 18&11.739&16.58&14.02\
Apr 18&11.776&16.56&14.01\
Apr 18&11.813&16.55&13.99\
Apr 18&11.851&16.54&13.98\
Apr 18&11.902&16.53&13.96\
\
Apr 19&7.129&16.77&14.22\
Apr 19&7.170&16.80&14.25\
Apr 19&7.208&16.74&14.18\
Apr 19&7.824&17.07&14.57\
Apr 19&7.889&17.06&14.57\
Apr 19&7.956&17.08&14.59\
Apr 19&7.993&17.06&14.56\
Apr 19&8.031&17.09&14.61\
Apr 19&8.068&17.12&14.64\
Apr 19&8.105&17.07&14.58\
Apr 19&8.142&17.10&14.62\
Apr 19&8.185&17.07&14.59\
Apr 19&8.281&17.10&14.62\
Apr 19&8.319&17.02&14.52\
Apr 19&8.356&16.99&14.48\
Apr 19&8.393&17.02&14.52\
Apr 19&8.430&16.96&14.45\
Apr 19&8.467&17.03&14.53\
Apr 19&8.606&16.93&14.41\
Apr 19&8.692&16.82&14.27\
Apr 19&8.730&16.87&14.34\
Apr 19&8.766&16.82&14.28\
Apr 19&8.804&16.86&14.33\
Apr 19&8.841&16.83&14.29\
Apr 19&8.878&16.87&14.34\
Apr 19&9.987&16.49&13.89\
Apr 19&10.024&16.50&13.91\
Apr 19&10.062&16.48&13.88\
Apr 19&10.099&16.50&13.90\
Apr 19&10.136&16.49&13.90\
Apr 19&10.172&16.49&13.90\
Apr 19&10.211&16.49&13.90\
Apr 19&10.292&16.51&13.91\
Apr 19&10.329&16.48&13.89\
Apr 19&10.366&16.48&13.88\
Apr 19&10.384&16.48&13.88\
Apr 19&10.441&16.48&13.88\
Apr 19&10.478&16.49&13.89\
Apr 19&10.578&16.48&13.88\
Apr 19&10.658&16.49&13.90\
Apr 19&10.696&16.50&13.91\
Apr 19&10.732&16.51&13.92\
Apr 19&10.769&16.51&13.92\
Apr 19&10.806&16.50&13.91\
Apr 19&10.844&16.51&13.92\
Apr 19&10.907&16.53&13.94\
Apr 19&10.974&16.53&13.95\
Apr 19&11.889&16.89&14.36\
Apr 19&11.926&16.87&14.35\
Apr 19&11.964&16.91&14.39\
Apr 19&12.001&17.01&14.54\
Apr 19&12.038&16.92&14.44\
Apr 19&12.076&16.92&14.44\
Apr 19&12.113&16.95&14.47\
Apr 19&12.151&17.01&14.55\
\
May 26&6.640&14.91&13.99\
May 26&6.679&14.89&13.97\
May 26&6.716&14.88&13.95\
May 26&6.760&14.87&13.95\
May 26&6.801&14.88&13.95\
May 26&6.838&14.86&13.93\
May 26&6.875&14.85&13.92\
May 26&6.912&14.85&13.92\
May 26&6.992&14.84&13.91\
May 26&7.029&14.84&13.91\
May 26&7.066&14.84&13.90\
May 26&7.103&14.82&13.89\
May 26&7.140&14.83&13.90\
May 26&7.191&14.84&13.91\
May 26&7.227&14.83&13.90\
May 26&7.264&14.84&13.90\
May 26&7.302&14.83&13.89\
May 26&7.339&14.81&13.87\
May 26&7.517&14.82&13.88\
May 26&7.554&14.82&13.88\
May 26&7.591&14.81&13.88\
May 26&7.631&14.82&13.88\
May 26&7.668&14.82&13.88\
May 26&7.705&14.82&13.89\
May 26&7.742&14.82&13.88\
May 26&7.779&14.84&13.90\
May 26&7.819&14.83&13.89\
May 26&7.857&14.84&13.91\
May 26&7.894&14.85&13.92\
May 26&7.931&14.85&13.91\
May 26&7.968&14.85&13.93\
May 26&8.935&15.10&14.25\
May 26&8.974&15.13&14.30\
May 26&9.011&15.12&14.28\
May 26&9.048&15.12&14.29\
May 26&9.085&15.14&14.31\
May 26&9.122&15.13&14.30\
May 26&9.163&15.13&14.30\
May 26&9.200&15.14&14.31\
May 26&9.237&15.14&14.31\
May 26&9.274&15.13&14.30\
May 26&10.044&15.12&14.29\
May 26&10.081&15.11&14.27\
May 26&10.119&15.11&14.26\
May 26&10.156&15.10&14.26\
May 26&10.195&15.08&14.23\
May 26&10.232&15.07&14.21\
May 26&10.269&15.05&14.18\
May 26&10.307&15.04&14.18\
May 26&10.344&15.03&14.17\
May 26&10.484&15.01&14.13\
May 26&10.521&14.99&14.11\
May 26&10.558&15.00&14.11\
May 26&10.596&15.00&14.11\
May 26&10.632&14.98&14.10\
May 26&11.187&14.94&14.04\
May 26&11.225&14.94&14.04\
May 26&11.262&14.92&14.01\
May 26&11.299&14.92&14.02\
May 26&11.353&14.92&14.02\
\
May 27&7.732&15.12&14.33\
May 27&7.777&15.12&14.33\
May 27&7.820&15.13&14.34\
May 27&7.859&15.15&14.36\
May 27&7.899&15.14&14.36\
May 27&7.936&15.14&14.35\
May 27&7.974&15.15&14.36\
May 27&8.012&15.15&14.36\
May 27&8.050&15.14&14.36\
May 27&8.088&15.15&14.36\
May 27&8.129&15.15&14.37\
May 27&8.168&15.14&14.35\
May 27&8.206&15.13&14.34\
May 27&8.243&15.13&14.34\
May 27&8.280&15.13&14.34\
May 27&8.317&15.12&14.32\
May 27&8.356&15.12&14.32\
May 27&8.394&15.10&14.30\
May 27&8.431&15.10&14.30\
May 27&8.468&15.10&14.30\
May 27&8.505&15.08&14.28\
May 27&8.544&15.07&14.27\
May 27&8.581&15.06&14.25\
May 27&8.618&15.04&14.23\
May 27&8.655&15.03&14.22\
May 27&8.692&15.03&14.21\
May 27&8.734&15.02&14.20\
May 27&8.771&15.01&14.19\
May 27&8.808&15.01&14.18\
May 27&8.846&15.00&14.18\
May 27&8.883&15.00&14.18\
May 27&9.303&14.87&14.02\
May 27&9.340&14.88&14.03\
May 27&9.377&14.87&14.02\
May 27&9.414&14.87&14.02\
May 27&9.451&14.86&14.00\
May 27&9.915&14.78&13.91\
May 27&9.952&14.78&13.91\
May 27&9.989&14.77&13.90\
May 27&10.026&14.77&13.90\
May 27&10.063&14.77&13.91\
May 27&10.137&14.75&13.88\
May 27&10.175&14.75&13.88\
May 27&10.212&14.75&13.88\
May 27&10.249&14.74&13.87\
May 27&10.291&14.75&13.88\
May 27&10.366&14.75&13.88\
May 27&10.402&14.75&13.88\
May 27&10.489&14.76&13.89\
May 27&10.526&14.77&13.90\
May 27&10.563&14.76&13.90\
May 27&11.425&14.91&14.07\
\
Jun 08&7.113&14.61&14.08\
Jun 08&7.148&14.60&14.07\
Jun 08&7.184&14.59&14.06\
Jun 08&7.219&14.59&14.05\
Jun 08&7.254&14.58&14.04\
Jun 08&7.293&14.56&14.01\
Jun 08&7.328&14.55&13.99\
Jun 08&7.363&14.53&13.98\
Jun 08&7.398&14.52&13.96\
Jun 08&7.434&14.53&13.96\
Jun 08&7.568&14.50&13.93\
Jun 08&7.643&14.49&13.92\
Jun 08&7.680&14.49&13.92\
Jun 08&7.814&14.48&13.90\
Jun 08&7.851&14.48&13.90\
Jun 08&7.890&14.49&13.92\
Jun 08&8.024&14.49&13.91\
Jun 08&8.060&14.49&13.91\
Jun 08&8.095&14.49&13.92\
Jun 08&8.134&14.48&13.89\
Jun 08&8.169&14.47&13.89\
Jun 08&8.204&14.47&13.88\
Jun 08&8.239&14.47&13.89\
Jun 08&8.275&14.47&13.88\
Jun 08&8.312&14.47&13.89\
Jun 08&8.347&14.46&13.87\
Jun 08&8.382&14.47&13.88\
Jun 08&8.418&14.47&13.89\
Jun 08&8.453&14.47&13.89\
Jun 08&8.490&14.46&13.88\
Jun 08&8.525&14.46&13.87\
Jun 08&8.560&14.47&13.88\
Jun 08&8.596&14.47&13.89\
Jun 08&8.631&14.48&13.90\
Jun 08&8.676&14.48&13.91\
Jun 08&8.711&14.49&13.91\
Jun 08&8.746&14.48&13.90\
Jun 08&8.781&14.49&13.91\
Jun 08&8.816&14.50&13.93\
Jun 08&9.065&14.57&14.02\
Jun 08&9.100&14.56&14.01\
Jun 08&9.135&14.57&14.03\
Jun 08&9.170&14.57&14.02\
Jun 08&9.205&14.59&14.05\
Jun 08&9.489&14.66&14.15\
Jun 08&9.525&14.66&14.15\
Jun 08&9.560&14.67&14.16\
Jun 08&9.595&14.67&14.17\
Jun 08&9.630&14.70&14.21\
Jun 08&9.908&14.72&14.25\
Jun 08&9.943&14.72&14.24\
Jun 08&9.978&14.73&14.26\
Jun 08&10.013&14.74&14.27\
Jun 08&10.048&14.78&14.33\
Jun 08&10.194&14.76&14.31\
Jun 08&10.229&14.77&14.32\
Jun 08&10.265&14.77&14.31\
Jun 08&10.300&14.77&14.31\
Jun 08&10.335&14.77&14.32\
Jun 08&10.467&14.74&14.27\
Jun 08&10.502&14.76&14.30\
Jun 08&10.538&14.74&14.28\
Jun 08&10.573&14.76&14.30\
Jun 08&10.608&14.78&14.33\
Jun 08&10.646&14.77&14.33\
Jun 08&10.681&14.76&14.31\
Jun 08&10.716&14.74&14.27\
Jun 08&10.751&14.74&14.28\
Jun 08&10.786&14.74&14.28\
Jun 08&10.837&14.74&14.27\
Jun 08&10.872&14.74&14.27\
Jun 08&10.907&14.75&14.29\
Jun 08&10.942&14.74&14.28\
Jun 08&10.977&14.74&14.27\
Jun 08&11.160&14.70&14.21\
Jun 08&11.195&14.69&14.21\
Jun 08&11.230&14.71&14.23\
\
Jun 09&4.315&14.76&14.35\
Jun 09&4.342&14.74&14.32\
Jun 09&4.368&14.74&14.31\
Jun 09&4.395&14.70&14.25\
Jun 09&4.422&14.68&14.22\
Jun 09&4.451&14.69&14.23\
Jun 09&4.478&14.69&14.24\
Jun 09&4.505&14.69&14.23\
Jun 09&4.531&14.67&14.20\
Jun 09&4.558&14.72&14.29\
Jun 09&4.764&14.71&14.27\
Jun 09&4.790&14.72&14.28\
Jun 09&4.817&14.73&14.30\
Jun 09&4.844&14.71&14.26\
Jun 09&4.871&14.71&14.27\
Jun 09&4.900&14.71&14.27\
Jun 09&4.926&14.70&14.26\
Jun 09&4.953&14.70&14.25\
Jun 09&4.980&14.69&14.23\
Jun 09&5.007&14.69&14.24\
Jun 09&5.452&14.64&14.16\
Jun 09&5.479&14.63&14.15\
Jun 09&5.506&14.63&14.15\
Jun 09&5.533&14.62&14.14\
Jun 09&5.560&14.59&14.09\
Jun 09&5.587&14.59&14.09\
Jun 09&5.614&14.58&14.08\
Jun 09&5.641&14.58&14.07\
Jun 09&5.668&14.57&14.06\
Jun 09&5.694&14.57&14.06\
Jun 09&5.986&14.56&14.04\
Jun 09&6.012&14.55&14.03\
Jun 09&6.039&14.54&14.03\
Jun 09&6.066&14.54&14.02\
Jun 09&6.093&14.54&14.03\
Jun 09&6.120&14.55&14.03\
Jun 09&6.147&14.52&14.00\
Jun 09&6.174&14.52&14.00\
Jun 09&6.201&14.54&14.02\
Jun 09&6.457&14.50&13.97\
Jun 09&6.484&14.50&13.96\
Jun 09&6.511&14.49&13.96\
Jun 09&6.538&14.49&13.95\
Jun 09&6.565&14.49&13.95\
Jun 09&6.592&14.48&13.94\
Jun 09&7.256&14.47&13.92\
Jun 09&7.283&14.47&13.92\
Jun 09&7.310&14.48&13.94\
Jun 09&7.337&14.49&13.94\
Jun 09&7.364&14.49&13.95\
Jun 09&7.391&14.50&13.96\
Jun 09&7.418&14.49&13.96\
Jun 09&7.444&14.51&13.97\
Jun 09&7.471&14.49&13.96\
Jun 09&7.498&14.49&13.96\
Jun 09&8.011&14.62&14.14\
Jun 09&8.038&14.63&14.15\
Jun 09&8.065&14.63&14.15\
Jun 09&8.092&14.62&14.14\
Jun 09&8.119&14.63&14.15\
Jun 09&8.146&14.63&14.15\
Jun 09&8.173&14.64&14.16\
Jun 09&8.199&14.66&14.19\
Jun 09&8.226&14.65&14.17\
Jun 09&8.253&14.66&14.19\
Jun 09&8.494&14.70&14.25\
Jun 09&8.521&14.69&14.24\
Jun 09&8.548&14.72&14.28\
Jun 09&8.575&14.71&14.27\
Jun 09&8.602&14.71&14.27\
Jun 09&8.629&14.71&14.27\
Jun 09&8.656&14.72&14.28\
Jun 09&8.682&14.72&14.28\
Jun 09&8.709&14.72&14.29\
Jun 09&8.736&14.72&14.29\
Jun 09&8.768&14.74&14.31\
Jun 09&8.795&14.74&14.31\
Jun 09&8.822&14.72&14.29\
Jun 09&8.848&14.73&14.31\
Jun 09&8.875&14.74&14.31\
Jun 09&8.902&14.74&14.32\
Jun 09&8.929&14.73&14.30\
Jun 09&8.956&14.73&14.30\
Jun 09&8.983&14.73&14.30\
Jun 09&9.010&14.73&14.30\
Jun 09&9.346&14.68&14.22\
Jun 09&9.373&14.67&14.22\
Jun 09&9.400&14.71&14.27\
Jun 09&9.427&14.66&14.21\
Jun 09&9.454&14.66&14.20\
Jun 09&9.481&14.69&14.25\
Jun 09&9.507&14.69&14.25\
Jun 09&9.534&14.69&14.25\
Jun 09&9.561&14.68&14.22\
Jun 09&9.588&14.67&14.21\
Jun 09&10.269&14.47&13.93\
Jun 09&10.561&14.51&13.98\
Jun 09&10.588&14.55&14.03\
Jun 09&10.615&14.49&13.96\
Jun 09&10.642&14.48&13.94\
Jun 09&10.669&14.45&13.90\
Jun 09&10.699&14.47&13.93\
Jun 09&10.726&14.47&13.93\
Jun 09&10.753&14.46&13.91\
Jun 09&10.780&14.46&13.91\
Jun 09&10.807&14.44&13.89\
Jun 09&10.867&14.47&13.92\
Jun 09&10.925&14.44&13.89\
Jun 09&10.952&14.42&13.87\
Jun 09&10.978&14.45&13.91\
Jun 09&11.005&14.45&13.90\
Jun 09&11.032&14.45&13.90\
Jun 09&11.066&14.43&13.88\
Jun 09&11.093&14.43&13.88\
Jun 09&11.120&14.44&13.89\
\
Jun 10&4.648&14.48&13.93\
Jun 10&4.708&14.48&13.93\
Jun 10&4.735&14.47&13.92\
Jun 10&4.762&14.47&13.92\
Jun 10&4.796&14.47&13.92\
Jun 10&4.829&14.44&13.88\
Jun 10&4.856&14.44&13.88\
Jun 10&4.882&14.45&13.90\
Jun 10&4.909&14.45&13.90\
Jun 10&4.936&14.46&13.91\
Jun 10&4.963&14.42&13.85\
Jun 10&4.990&14.44&13.88\
Jun 10&5.017&14.44&13.87\
Jun 10&5.044&14.44&13.87\
Jun 10&5.070&14.45&13.89\
Jun 10&5.249&14.49&13.95\
Jun 10&5.276&14.47&13.92\
Jun 10&5.303&14.45&13.90\
Jun 10&5.329&14.47&13.92\
Jun 10&5.356&14.44&13.88\
Jun 10&5.383&14.45&13.90\
Jun 10&5.410&14.48&13.93\
Jun 10&5.437&14.47&13.92\
Jun 10&5.464&14.45&13.89\
Jun 10&5.491&14.46&13.91\
Jun 10&5.519&14.47&13.91\
Jun 10&5.573&14.48&13.93\
Jun 10&5.600&14.48&13.94\
Jun 10&5.627&14.49&13.94\
Jun 10&5.990&14.48&13.94\
Jun 10&6.017&14.52&13.99\
Jun 10&6.043&14.50&13.96\
Jun 10&6.070&14.51&13.98\
Jun 10&6.097&14.52&13.99\
Jun 10&6.124&14.54&14.02\
Jun 10&6.151&14.55&14.03\
Jun 10&6.178&14.56&14.05\
Jun 10&6.204&14.54&14.02\
Jun 10&6.231&14.53&14.01\
Jun 10&6.460&14.62&14.13\
Jun 10&6.487&14.65&14.18\
Jun 10&6.513&14.64&14.16\
Jun 10&6.540&14.61&14.12\
Jun 10&6.567&14.63&14.14\
Jun 10&6.594&14.65&14.18\
Jun 10&6.621&14.67&14.20\
Jun 10&6.648&14.65&14.18\
Jun 10&6.674&14.66&14.19\
Jun 10&6.701&14.65&14.18\
Jun 10&6.921&14.71&14.26\
Jun 10&6.948&14.71&14.26\
Jun 10&6.974&14.71&14.26\
Jun 10&7.001&14.72&14.28\
Jun 10&7.028&14.71&14.27\
Jun 10&7.055&14.71&14.27\
Jun 10&7.082&14.72&14.28\
Jun 10&7.109&14.74&14.31\
Jun 10&7.136&14.74&14.31\
Jun 10&7.162&14.74&14.32\
Jun 10&7.231&14.73&14.29\
Jun 10&7.258&14.73&14.30\
Jun 10&7.285&14.73&14.29\
Jun 10&7.339&14.75&14.33\
Jun 10&7.366&14.73&14.29\
Jun 10&7.392&14.72&14.28\
Jun 10&7.446&14.70&14.26\
Jun 10&7.473&14.70&14.25\
Jun 10&7.505&14.69&14.24\
Jun 10&7.532&14.70&14.24\
Jun 10&7.559&14.70&14.25\
Jun 10&7.586&14.70&14.24\
Jun 10&7.613&14.68&14.22\
Jun 10&7.640&14.68&14.23\
Jun 10&7.667&14.69&14.23\
Jun 10&7.693&14.68&14.22\
Jun 10&7.720&14.67&14.21\
Jun 10&7.747&14.68&14.21\
Jun 10&8.012&14.64&14.16\
Jun 10&8.039&14.64&14.16\
Jun 10&8.066&14.63&14.14\
Jun 10&8.092&14.64&14.16\
Jun 10&8.119&14.63&14.14\
Jun 10&8.146&14.62&14.13\
Jun 10&8.227&14.61&14.12\
Jun 10&8.254&14.61&14.12\
Jun 10&9.289&14.46&13.91\
Jun 10&9.315&14.47&13.92\
Jun 10&9.342&14.45&13.90\
Jun 10&9.369&14.45&13.90\
Jun 10&9.396&14.46&13.91\
Jun 10&9.423&14.45&13.89\
Jun 10&9.450&14.45&13.90\
Jun 10&9.477&14.46&13.91\
Jun 10&9.503&14.44&13.88\
Jun 10&9.533&14.43&13.88\
Jun 10&9.560&14.43&13.87\
Jun 10&9.586&14.43&13.86\
Jun 10&9.613&14.43&13.87\
Jun 10&9.640&14.43&13.86\
Jun 10&9.667&14.42&13.86\
Jun 10&9.694&14.44&13.88\
Jun 10&9.721&14.47&13.92\
Jun 10&9.748&14.46&13.91\
Jun 10&9.775&14.46&13.91\
Jun 10&10.024&14.43&13.86\
Jun 10&10.051&14.43&13.87\
Jun 10&10.078&14.43&13.86\
Jun 10&10.105&14.43&13.88\
Jun 10&10.132&14.43&13.87\
Jun 10&10.159&14.45&13.90\
Jun 10&10.185&14.44&13.89\
Jun 10&10.212&14.45&13.90\
Jun 10&10.239&14.44&13.89\
Jun 10&10.266&14.43&13.87\
Jun 10&10.496&14.45&13.90\
Jun 10&10.523&14.49&13.95\
Jun 10&10.550&14.50&13.97\
Jun 10&10.577&14.51&13.98\
Jun 10&10.604&14.49&13.95\
Jun 10&10.769&14.51&13.98\
Jun 10&10.796&14.51&13.98\
Jun 10&10.823&14.54&14.02\
Jun 10&10.850&14.52&13.99\
Jun 10&10.877&14.55&14.03\
Jun 10&11.001&14.57&14.06\
Jun 10&11.028&14.56&14.05\
Jun 10&11.055&14.59&14.09\
Jun 10&11.180&14.60&14.11\
Jun 10&11.207&14.60&14.10\
Jun 10&11.234&14.62&14.14\
Jun 10&11.263&14.62&14.13\
\
Jun 11&4.516&14.54&14.14\
Jun 11&4.544&14.53&14.13\
Jun 11&4.573&14.50&14.09\
Jun 11&4.616&14.54&14.14\
Jun 11&4.644&14.60&14.23\
Jun 11&5.322&14.70&14.39\
Jun 11&5.349&14.69&14.38\
Jun 11&5.376&14.68&14.36\
Jun 11&5.403&14.68&14.36\
Jun 11&5.429&14.68&14.37\
Jun 11&5.459&14.68&14.36\
Jun 11&5.486&14.69&14.38\
Jun 11&5.513&14.73&14.43\
Jun 11&5.540&14.69&14.38\
Jun 11&5.567&14.67&14.35\
Jun 11&5.596&14.72&14.42\
Jun 11&5.623&14.69&14.37\
Jun 11&5.650&14.69&14.37\
Jun 11&5.677&14.69&14.37\
Jun 11&5.704&14.69&14.38\
Jun 11&5.976&14.65&14.31\
Jun 11&6.003&14.64&14.30\
Jun 11&6.030&14.64&14.29\
Jun 11&6.057&14.57&14.19\
Jun 11&6.084&14.58&14.20\
Jun 11&6.215&14.58&14.20\
Jun 11&6.242&14.57&14.19\
Jun 11&6.269&14.55&14.16\
Jun 11&6.296&14.55&14.15\
Jun 11&6.323&14.55&14.15\
Jun 11&6.487&14.50&14.08\
Jun 11&6.514&14.50&14.08\
Jun 11&6.541&14.49&14.06\
Jun 11&6.568&14.49&14.07\
Jun 11&6.594&14.49&14.07\
Jun 11&6.657&14.47&14.04\
Jun 11&6.684&14.47&14.04\
Jun 11&6.711&14.47&14.04\
Jun 11&6.737&14.48&14.06\
Jun 11&6.764&14.46&14.03\
Jun 11&6.906&14.45&14.01\
Jun 11&6.933&14.45&14.02\
Jun 11&6.960&14.46&14.03\
Jun 11&6.986&14.44&14.00\
Jun 11&7.013&14.44&13.99\
Jun 11&7.076&14.42&13.97\
Jun 11&7.138&14.41&13.95\
Jun 11&7.165&14.40&13.94\
Jun 11&7.192&14.39&13.93\
Jun 11&7.219&14.38&13.91\
Jun 11&7.246&14.37&13.90\
Jun 11&7.515&14.37&13.90\
Jun 11&7.541&14.37&13.90\
Jun 11&7.568&14.36&13.88\
Jun 11&7.595&14.37&13.89\
Jun 11&7.622&14.35&13.87\
Jun 11&7.651&14.36&13.88\
Jun 11&7.678&14.36&13.88\
Jun 11&7.705&14.35&13.87\
Jun 11&7.732&14.35&13.87\
Jun 11&7.759&14.34&13.86\
Jun 11&7.788&14.35&13.86\
Jun 11&7.815&14.35&13.87\
Jun 11&7.842&14.35&13.87\
Jun 11&7.869&14.36&13.88\
Jun 11&7.896&14.36&13.89\
Jun 11&8.093&14.38&13.91\
Jun 11&8.120&14.38&13.91\
Jun 11&8.147&14.38&13.90\
Jun 11&8.174&14.39&13.92\
Jun 11&8.201&14.39&13.92\
Jun 11&8.355&14.39&13.92\
Jun 11&8.382&14.39&13.93\
Jun 11&8.409&14.40&13.94\
Jun 11&8.436&14.40&13.94\
Jun 11&8.463&14.39&13.92\
Jun 11&8.689&14.42&13.96\
Jun 11&8.716&14.42&13.97\
Jun 11&8.743&14.43&13.97\
Jun 11&8.770&14.42&13.97\
Jun 11&8.797&14.43&13.98\
Jun 11&8.920&14.43&13.99\
Jun 11&8.947&14.43&13.99\
Jun 11&8.974&14.44&14.01\
Jun 11&9.001&14.45&14.01\
Jun 11&9.028&14.45&14.01\
Jun 11&9.158&14.47&14.04\
Jun 11&9.184&14.48&14.06\
Jun 11&9.211&14.48&14.06\
Jun 11&9.238&14.48&14.06\
Jun 11&9.265&14.49&14.07\
Jun 11&9.329&14.54&14.14\
Jun 11&9.396&14.51&14.09\
Jun 11&9.423&14.50&14.09\
Jun 11&9.449&14.51&14.11\
Jun 11&9.476&14.52&14.11\
Jun 11&9.503&14.52&14.11\
\
Jun 22&7.910&14.15&13.97\
Jun 22&7.942&14.12&13.91\
Jun 22&7.971&14.11&13.91\
Jun 22&7.998&14.13&13.93\
Jun 22&8.027&14.11&13.89\
Jun 22&8.085&14.13&13.93\
Jun 22&8.143&14.13&13.93\
Jun 22&8.173&14.14&13.95\
Jun 22&8.200&14.14&13.95\
Jun 22&8.227&14.13&13.93\
Jun 22&8.368&14.15&13.96\
Jun 22&8.394&14.16&13.97\
Jun 22&8.421&14.15&13.96\
Jun 22&8.669&14.16&13.97\
Jun 22&8.695&14.16&13.98\
Jun 22&8.722&14.18&14.01\
Jun 22&8.946&14.23&14.10\
Jun 22&8.973&14.23&14.11\
Jun 22&9.000&14.25&14.13\
Jun 22&9.124&14.25&14.14\
Jun 22&9.151&14.27&14.17\
Jun 22&9.178&14.26&14.15\
Jun 22&9.489&14.30&14.23\
Jun 22&9.516&14.30&14.23\
Jun 22&9.543&14.30&14.24\
Jun 22&9.570&14.31&14.25\
Jun 22&9.597&14.31&14.25\
Jun 22&9.650&14.31&14.26\
Jun 22&9.707&14.33&14.29\
Jun 22&9.734&14.34&14.31\
Jun 22&9.761&14.33&14.30\
Jun 22&9.788&14.34&14.31\
Jun 22&10.047&14.36&14.34\
Jun 22&10.074&14.35&14.33\
Jun 22&10.101&14.35&14.32\
Jun 22&10.434&14.34&14.32\
Jun 22&10.461&14.33&14.30\
Jun 22&10.488&14.32&14.27\
Jun 22&10.687&14.34&14.31\
Jun 22&10.714&14.32&14.27\
Jun 22&10.741&14.31&14.26\
Jun 22&10.949&14.26&14.15\
Jun 22&10.976&14.25&14.14\
Jun 22&11.002&14.25&14.13\
Jun 22&11.059&14.25&14.15\
Jun 22&11.086&14.26&14.15\
Jun 22&11.113&14.24&14.13\
\
Jun 23&3.963&14.33&14.34\
Jun 23&3.999&14.34&14.36\
Jun 23&4.026&14.33&14.33\
Jun 23&4.052&14.34&14.35\
Jun 23&4.079&14.33&14.33\
Jun 23&4.106&14.35&14.37\
Jun 23&4.492&14.26&14.19\
Jun 23&4.519&14.24&14.17\
Jun 23&4.546&14.24&14.17\
Jun 23&4.606&14.23&14.15\
Jun 23&4.676&14.21&14.11\
Jun 23&4.703&14.21&14.11\
Jun 23&4.730&14.20&14.10\
Jun 23&4.963&14.17&14.04\
Jun 23&4.990&14.17&14.04\
Jun 23&5.017&14.18&14.05\
Jun 23&5.202&14.12&13.95\
Jun 23&5.229&14.14&13.97\
Jun 23&5.256&14.13&13.97\
Jun 23&5.444&14.11&13.93\
Jun 23&5.471&14.11&13.92\
Jun 23&5.498&14.10&13.92\
Jun 23&5.628&14.09&13.90\
Jun 23&5.655&14.10&13.90\
Jun 23&5.682&14.11&13.92\
Jun 23&5.710&14.11&13.92\
Jun 23&5.737&14.11&13.93\
Jun 23&5.764&14.10&13.92\
Jun 23&5.808&14.08&13.88\
Jun 23&5.834&14.07&13.87\
Jun 23&5.861&14.08&13.89\
Jun 23&5.894&14.07&13.87\
Jun 23&5.921&14.08&13.88\
Jun 23&5.947&14.09&13.89\
Jun 23&6.123&14.08&13.88\
Jun 23&6.150&14.09&13.90\
Jun 23&6.177&14.09&13.89\
Jun 23&6.341&14.09&13.89\
Jun 23&6.368&14.10&13.91\
Jun 23&6.395&14.11&13.93\
Jun 23&6.563&14.11&13.93\
Jun 23&6.590&14.13&13.96\
Jun 23&6.617&14.12&13.94\
Jun 23&6.773&14.13&13.96\
Jun 23&6.800&14.14&13.98\
Jun 23&6.827&14.14&13.99\
Jun 23&7.473&14.22&14.12\
Jun 23&7.500&14.23&14.15\
Jun 23&7.527&14.23&14.15\
Jun 23&7.557&14.24&14.17\
Jun 23&7.584&14.24&14.16\
Jun 23&7.611&14.24&14.17\
Jun 23&7.759&14.34&14.35\
Jun 23&7.786&14.32&14.32\
Jun 23&7.812&14.28&14.24\
Jun 23&7.843&14.28&14.24\
Jun 23&7.870&14.30&14.27\
Jun 23&7.897&14.30&14.28\
Jun 23&7.929&14.29&14.26\
Jun 23&7.956&14.29&14.26\
Jun 23&7.982&14.30&14.28\
Jun 23&8.012&14.30&14.29\
Jun 23&8.039&14.29&14.27\
Jun 23&8.066&14.30&14.27\
Jun 23&8.097&14.30&14.28\
Jun 23&8.124&14.30&14.27\
Jun 23&8.151&14.31&14.30\
Jun 23&8.315&14.28&14.24\
Jun 23&8.342&14.29&14.26\
Jun 23&8.368&14.30&14.27\
Jun 23&8.496&14.26&14.20\
Jun 23&8.523&14.26&14.21\
Jun 23&8.549&14.26&14.21\
Jun 23&8.761&14.24&14.16\
Jun 23&8.788&14.23&14.14\
Jun 23&8.815&14.24&14.16\
Jun 23&8.944&14.21&14.11\
Jun 23&8.971&14.21&14.11\
Jun 23&8.998&14.20&14.10\
Jun 23&9.149&14.20&14.08\
Jun 23&9.176&14.20&14.10\
Jun 23&9.203&14.20&14.09\
Jun 23&9.407&14.19&14.07\
Jun 23&9.434&14.18&14.05\
Jun 23&9.461&14.17&14.04\
Jun 23&9.607&14.17&14.03\
Jun 23&9.634&14.16&14.02\
Jun 23&9.661&14.17&14.03\
Jun 23&9.868&14.14&13.99\
Jun 23&9.895&14.13&13.97\
Jun 23&9.922&14.15&14.00\
Jun 23&10.021&14.15&14.00\
Jun 23&10.048&14.14&13.99\
Jun 23&10.075&14.15&14.00\
Jun 23&10.103&14.13&13.97\
Jun 23&10.130&14.13&13.97\
Jun 23&10.157&14.11&13.94\
Jun 23&10.186&14.11&13.93\
Jun 23&10.213&14.12&13.95\
Jun 23&10.240&14.12&13.94\
Jun 23&10.269&14.12&13.95\
Jun 23&10.296&14.09&13.90\
Jun 23&10.323&14.09&13.90\
Jun 23&10.357&14.11&13.92\
Jun 23&10.384&14.09&13.89\
Jun 23&10.411&14.10&13.91\
Jun 23&10.524&14.11&13.92\
Jun 23&10.551&14.11&13.92\
Jun 23&10.578&14.13&13.96\
Jun 23&10.709&14.11&13.92\
Jun 23&10.742&14.12&13.94\
[lcccccccccccc]{}
\
\
& & & & & & & & & & & &\
& & & & & & & & & & & &\
\
[[** – continued from previous page**]{}]{}\
\
\
\
\
1987 Mar 31&KP2.1&TI 2&R&3.991&3.162&9.0&0.438&332.6&341.6&$-$5.79&$+$0.45&JM\
1987 Apr 01&KP2.1&TI 2&R&3.987&3.168&9.2&0.439&332.5&341.7&$-$5.80&$+$0.45&JM\
1987 Apr 02&KP2.1&TI 2&R&3.984&3.174&9.4&0.440&332.3&341.8&$-$5.81&$+$0.45&JM\
1987 Apr 03&KP2.1&TI 2&R&3.981&3.181&9.7&0.441&332.2&341.9&$-$5.82&$+$0.45&JM\
1988 Feb 25&MH2.4&MASCOT&R&2.384&1.985&24.0&0.275&54.3&29.9&$-$4.14&$-$0.05&JL\
1988 Feb 27&MH2.4&MASCOT&R&2.371&1.949&24.1&0.270&54.9&30.4&$-$4.09&$-$0.05&JL\
1988 Feb 28&MH2.4&MASCOT&R&2.365&1.932&24.1&0.268&55.1&30.7&$-$4.07&$-$0.05&JL\
1988 Feb 29&MH2.4&MASCOT&R&2.359&1.914&24.1&0.265&55.4&30.9&$-$4.04&$-$0.05&JL\
1988 Apr 09&MH1.3&MASCOT&R&2.103&1.279&20.0&0.177&61.7&42.8&$-$2.79&$+$0.30&JL\
1988 Apr 10&MH1.3&MASCOT&R&2.097&1.265&19.8&0.175&61.7&43.1&$-$2.75&$+$0.27&JL\
1988 Apr 12&MH1.3&MASCOT&R&2.084&1.239&19.3&0.172&61.8&43.8&$-$2.68&$+$0.30&JL\
1988 Apr 15&MH1.3&MASCOT&R&2.065&1.201&18.6&0.166&61.8&44.8&$-$2.57&$+$0.30&JL\
1988 May 08&SO2.3&TI&VR&1.919&0.959&13.2&0.133&59.3&53.4&$-$1.75&...&WI\
1988 May 09&SO2.3&TI&VR&1.913&0.951&13.1&0.132&59.1&53.8&$-$1.72&...&WI\
1988 May 20&ML1.5&photometer&VR&1.845&0.873&12.8&0.121&56.5&58.3&$-$1.44&$-$0.10&WI\
1988 May 21&ML1.5&photometer&VR&1.839&0.867&12.9&0.120&56.2&58.7&$-$1.43&$-$0.10&WI\
1988 May 22&ML1.5&photometer&VR&1.832&0.861&13.1&0.119&56.0&59.2&$-$1.41&$-$0.10&WI\
1988 Jun 10&IRTF&bolometer&N&1.720&0.788&19.9&0.109&51.3&68.0&$-$1.30&$+$11.80&AH\
1988 Jun 10&UH2.2&photometer&6840Å&1.720&0.788&19.9&0.109&51.3&68.0&$-$1.30&$+$0.73&AH\
1988 Jun 11&IRTF&bolometer&N&1.714&0.785&20.4&0.109&51.1&68.4&$-$1.30&$+$11.70&AH\
1988 Jun 11&UH2.2&photometer&6840Å&1.714&0.785&20.4&0.109&51.1&68.4&$-$1.30&$\phantom{0}+$0.73&AH\
1988 Jun 12&ML1.5&photometer&VR&1.708&0.783&20.9&0.108&50.9&68.9&$-$1.30&$+$0.25&WI\
1988 Jun 14&ML1.5&photometer&VR&1.697&0.780&21.9&0.108&50.5&69.9&$-$1.31&$+$0.25&WI\
1988 Jun 22&MH2.4&BRICC&R&1.653&0.771&25.9&0.107&49.5&74.1&$-$1.36&$+$0.23&JL\
1988 Jun 23&MH2.4&BRICC&R&1.647&0.770&26.3&0.107&49.4&74.6&$-$1.36&$+$0.28&JL\
1988 Jun 30&MH2.4&BRICC&R&1.611&0.769&29.7&0.107&49.2&78.4&$-$1.42&$+$0.45&JL\
1994 Oct 26&KP0.9&T2KA&R&2.498&1.640&14.2&0.227&247.5&234.2&$-$3.52&...&MF\
1994 Dec 18&KP0.9&T2KA&R&2.802&1.932&11.3&0.268&234.8&245.5&$-$4.03&...&MF\
1994 Dec 19&KP0.9&T2KA&R&2.807&1.944&11.6&0.269&234.6&245.7&$-$4.06&...&MF\
\
\
[^1]: Lowell Observatory, 1400 W. Mars Hill Rd, Flagstaff, AZ 86001
[^2]: Department of Astronomy, University of Maryland, College Park, MD 20742-2421
[^3]: Physics and Space Sciences, Florida Institute of Technology, 150 W. University Blvd, Melbourne, Florida 32901
[^4]: http://www.datadesk.com.
[^5]: http://www.univie.ac.at/tops/Period04/.
[^6]: We did not see evidence of a 13 min offset of the 1988 April data from @jewitt89 as suggested by @sekanina91, but note that $\Delta$m$_2$ could mask this effect. @sekanina91 adjusted these data by 13 minutes to derive the quoted synodic period, while @mueller96 omitted these data entirely from their period analysis. We include them with no time adjustment.
[^7]: The 1988 April data from @jewitt89 showed the deeper minimum to be at $\sim$0.7 phase, in contradiction to all other 1988 datasets (see Figure \[fig:1988\_data\]). This dataset was the most difficult to normalize since neither April 9 or 12 contained a clear extremum, and the April 10 and 15 data required very different $\Delta$m$_2$ adjustments to bring the maxima in alignment near zero phase. Given that the minima were observed only five nights apart and no lightcurve change in shape is expected during this short a time interval, we are skeptical of this anomalous minimum.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The purpose of gesture recognition is to recognize meaningful movements of human bodies, and gesture recognition is an important issue in computer vision. In this paper, we present a multimodal gesture recognition method based on 3D densely convolutional networks (3D-DenseNets) and improved temporal convolutional networks (TCNs). The key idea of our approach is to find a compact and effective representation of spatial and temporal features, which orderly and separately divide task of gesture video analysis into two parts: spatial analysis and temporal analysis. In spatial analysis, we adopt 3D-DenseNets to learn short-term spatio-temporal features effectively. Subsequently, in temporal analysis, we use TCNs to extract temporal features and employ improved Squeeze-and-Excitation Networks (SENets) to strengthen the representational power of temporal features from each TCNs’ layers. The method has been evaluated on the VIVA and the NVIDIA Gesture Dynamic Hand Gesture Datasets. Our approach obtains very competitive performance on VIVA benchmarks with the classification accuracies of $91.54\%$, and achieve state-of-the art performance with $86.37\%$ accuracy on NVIDIA benchmark.'
author:
- 'Yi Zhang[^1], Chong Wang[^2], Ye Zheng, Jieyu Zhao, Yuqi Li and Xijiong Xie [^3]'
bibliography:
- 'reff.bib'
title: 'Short-Term Temporal Convolutional Networks for Dynamic Hand Gesture Recognition'
---
[***Index terms—*** Gesture Recognition, 3D-DenseNets, TCNs, multimodal.]{}
Introduction {#sec:introduction}
============
Gesture recognition is a fast expanding field with applications in human-computer interaction[@rautaray2015vision], sign language recognition[@camgoz2017subunets] and etc.. Due to subtle differences among similar gestures, complex scene background, different observation conditions, and noises in acquisition, robust gesture recognition is very challenging.\
The main task of gesture recognition is to extract features from an image or a video and then classify or determine each sample to a certain label. Gesture recognition aims to recognize and understand meaningful movement of human bodies in which arms and hands play crucial roles. Only few gestures can be identified from their spatial or structure information in an image or a single frame. In fact, motion cues and structure information simultaneously characterize a unique gesture. How to learn spatiotemporal features effectively is always the key in gesture recognition. Although in the past decades, many methods have been proposed for this issue, ranging from static to dynamic gestures, and from motion silhouettes-based to the convolutional neural network-based, there are still many challenges associated with the recognition accuracy.\
At present, although most existing models have reached a high performance for isolated gesture recognition, most methods have been developed based on Convolutional Neural Networks (CNNs)[@simonyan2014two][@neverova2014multi] or Recurrent Neural Networks (RNNs)[@molchanov2016online]. With the development of deep learning, more and more new architectures of CNNs have been proposed, especially DenseNets [@huang2017densely] what have powerful feature extraction ability. Meanwhile, a new architecture to solve sequence problem named TCNs [@bai2018empirical] have been proposed. Compared to RNNs and their canonical recurrent architectures such as LSTMs and GRUs, TCNs have comparable clarity and simplicty. In our approach, we adopt 3D-DenseNet to extract short-term stapio-temporal features, then these features are input into the TCNs to finish the task of Classification.\
However, recently, for extracting more complete temporal features, a few methods have been proposed based on attention mechanism. The research prove that there are various relationships between features’ interior in neural networks. SENets [@hu2018squeeze] are new architectural unit with the goal of improving the quality of representations produced by a network by explicitly modelling the interdependencies between the channels of its convolutional features. And in our approach, we reform SENets and combine them into TCNs to strengthen capacity of TCNs in temporal features extracting.\
The pipline of our method is depicted in Figure \[fig:pipline\], and the main contribution can be summarized as following:
- **Spatial analysis.** We design a multi-stream truncated 3D-DenseNet, which extracts spatio-temporal features from a video, and through local temporal pooling, obtain the decomposed short-term spatio-temporal features, to solve problem that single frame image can not carry enough spatial or structure information of gesture and reduce repetitive training for video clips.
- **Temporal analysis.** We employ TCN to replace RNN as the main model of sequence information feature analysis. In addition, we improve SENets and apply them in temporal domain to rescale the weights between temporal features and extract more effective temporal features to achieve higher classification accuracy.
![An overview of the proposed method. The proposed deep architecture is composed of two main steps: (a) Multimodal short-term spatio-temporal feature sequence extracting by truncated 3D-Densenet (T3D-Dense), local temporal average pooling (LTAP) and multimodal features concatenation. (b)Long-term feature sequence recognizing via TCN and TSE.[]{data-label="fig:pipline"}](overview.png){width="45.00000%"}
Related Work {#sec:relatedwork}
============
Gesture taxonomies and representations have been studied for decades.The vision based gesture recognition techniques include the static gesture oriented and the dynamic gesture oriented methods [@rautaray2015vision].\
Recently, convolution neural networks (CNNs) [@krizhevsky2012imagenet] have made a great breakthrough on computer vision related tasks by their powerful feature extraction ability, thus the features extracted by CNNs are widely used in many action classification tasks instead of hand-crafted features for better performance. Features are extracted by 2D-CNN from the starts. bi-directional rank pooling [@wang2016large][@fernando2016rank] was used to encode the spatial and temporal information of videos. Temporal convolutions for gesture recognition in videos Beyond temporal pooling [@pigou2018beyond] was proposed to solve gesture recognition problem in videos by a new temporal pooling method. On the other hand, C3D[@tran2015learning] model is developed and provides a better performance and main contribution in this research is proposed an architecture to extract spatio-temporal features from a video clip. Concurrently, a multi-stream 3D-CNN[@molchanov2015hand] was designed for hand gesture recognition and the classifier consisted of two subnetworks: a high-resolution network (HRN) and a low-resolution network (LRN) in this model.\
Meanwhile, with the development of convolutional neural networks, more and more architectures of CNNs were proposed, like AlexNet [@krizhevsky2012imagenet], VGGNet [@simonyan2014very], GoogleNet [@szegedy2015going] [@ioffe2015batch] [@szegedy2016rethinking] [@szegedy2017inception], ResNet [@he2016deep] and DenseNet [@huang2017densely]. All of these models have one target that is building a higher architectures of CNNs to dig deeper and more complete statial features from low-level image frames, and then classify. In the area of isolated gesture recognition, Res-C3D model[@miao2017multimodal] was used and won the first place twice in ChaLearn LAP Multi-modal Isolated Gesture Recognition Challenges 2016 [@escalante2016chalearn] and 2017 [@wan2017results]. Whatmore, DenseNets as one of the latest convolutional architectures, was adopted in action recognitions especially face recognitions and gesture recognitions gradually. A face recognition model named Dense Face[@zhang2018face] was proposed to explore the performance of densely connected network in face recognition. DenseNets[@hao2019spatiotemporal] also was used to classifier the different actions in recent researches.\
Regarding the temporal information of the video sequences, Long Short Term Memory(LSTM) networks is a common choice to gesture recognition. For instance, convolutional LSTM[@zhang2017learning] was introduced for spatio-temporal feature maps. 2S-RNN(RGB and Depth)[@chai2016two] was used for continuous gesture recognition. However, RNNs including LSTMs and GRUs have some weaknesses on temporal domain like short-range information learning, oversized memory capacity. To make these weaknesses up, TCNs is proposed and applied in the gesture reconition. Res-TCN[@hou2018spatial] was proposed for skeleton-based dynamic hand gesture recognition. Whatmore, a model based on TCN[@tsinganos2019improved] was proposed for gesture recognition.\
Other important works based on attention mechanism. Attention mechanism or attention model firstly was applied to neural networks by Vaswani et al[@vaswani2017attention]. After that, more and more researches are proposed based on attention mechanism, so as SENets [@hu2018squeeze] that improve ResNets to win first place of ILSVRC 2017 classification.
{width="95.00000%"} \[fig:dense\]
Our Approach {#sec:approach}
============
In the video recognition, both of the spatial and temporal information are important. Although there have been impressive progress in spatial feature extraction using 2D-CNNs based networks[@molchanov2015hand][@simonyan2014two], how to effectively learn the temporal features is still a very challenging problem. Unlike the 2D-CNNs focusing on the single image, various 3D-CNN based networks[@molchanov2016online][@wang2016robust][@carreira2017quo][@abavisani2019improving] have been proposed to process the successive frames simultaneously. For the video of dynamic hand gestures, adjacent frames are usually similar and containing the same static gesture, while the static gestures change several times during the whole video. Thus, in this paper we decompose the video to two different parts. One is the short-term spatio-temporal information in the adjacent frames, and the other is the long-term temporal information analysed by a sequential model. Based on this consideration, we raised two major questions,
- how to learn short-term spatio-temporal features effectively from video clips in the same video.
- how to reasonably classify a sequence which is combined from these consecutive features.
In order to address these issues, we designed a novel architecture to extract a sequence of short spatio-temporal features in order to recognize dynamic gestures.
As depicted in Figure \[fig:pipline\], the overall process can be divided into two parts: 1) multi-modal short-term spatio-temporal feature extraction based on 3D-DenseNets and 2) spatio-temporal sequence classify with and temporal SENets embedded TCNs. To be specific, the details of the proposed network structure is presented in Figure \[fig:dense\] and Figure \[fig:TCN\].
temporal local pooling to extract short-term features
-----------------------------------------------------
Due to the availability of various data types and the nature of signing videos, a more robust feature representation can acquired from the incorporation of multi-modal hand gesture information. To effectively present the the location, shape and sequential information in the adjacent gesture frames, we design a multi-stream DenseNet based on the C3D[@tran2015learning] to extracts short-term spatio-temporal features. Assume a given video $V$ with $n$ frames, it is firstly re-sampled to $k$ frames. Thus, the input video $V_{S}$ is denoted as, $$V_{S}=[v_{1},v_{2},...,v_{k}]\label{eq}$$ where $v_{k}$ is the $k$-th frame image of video sequence in the input.\
As aforementioned, we consider multiple modalities of gesture video data as the input. Each type of the data is set as one data stream and fed to the same network structure. The outputs of them will be fused together later as shown in Figure \[fig:pipline\]. The proposed model contains 4 dense blocks, containing 6, 12, 24, 16 layers respectively. Following the basic design in DenseNet[@huang2017densely] and C3D[@tran2015learning], the detailed network configurations are shown in Table 1. It is worth noting that most of the convolution layers are with $3\times3\times3$ filters, which limits the process only on the local spatial and temporal domain. Moreover, the temporal pooling size and stride in all the transition layers are set as 1 to avoid the fusion of the short-term temporal information, which is one major difference from the other conventional 3D-CNNs[@tran2015learning].
---------------------------------------------------------------------------------------
Layers Filter Size
---------------------------- ----------------------------------------------------------
Convolution $5\times5\times5$ conv, stride $2\times2\times1$
Pooling $3\times3\times1$ max pool, stride $2\times2\times1$
Dense Block 1 $\begin{bmatrix}
1 \times 1 \times 1\; conv\\
3 \times 3 \times 3\; conv
\end{bmatrix} \times 6 $
$1\times1\times1$ conv
$2\times2\times1$ average pool, stride $2\times2\times1$
Dense Block 2 $\begin{bmatrix}
1 \times 1 \times 1\; conv\\
3 \times 3 \times 3\; conv
\end{bmatrix} \times 12 $
\*[Transition Layer 2]{} $1\times1\times1$ conv
$2\times2\times1$ average pool, stride $2\times2\times1$
Dense Block 3 $\begin{bmatrix}
1 \times 1 \times 1\; conv\\
3 \times 3 \times 3\; conv
\end{bmatrix} \times 24 $
\*[Transition Layer 3]{} $1\times1\times1$ conv
$2\times2\times1$ average pool, stride $2\times2\times1$
Dense Block final $\begin{bmatrix}
1 \times 1 \times 1\; conv\\
3 \times 3 \times 3\; conv
\end{bmatrix} \times 16 $
\*[Classification Layer]{} global spatial average pool
global temporal average pool
fully-connected, softmax
---------------------------------------------------------------------------------------
: **3D-DenseNet architectures. The growth rate of network is k = 12. Note that each “conv” layer shown in the corresponds the sequence BN-ReLU-Conv.**[]{data-label="table:dense"}
\[tab1\]
Since the 3D-Densenet is served as a short-term spatio-temporal features extractor, we truncate it to obtain the features only. To be specific, the global temporal average pooling layer, last softmax and fully-connected layers are discarded, after the model is first pre-trained with isolated gesture data.
Therefore, we can get the global spatio-temporal feature $F_{k}$ after the global spatial average pool layer, $$F_{k}=[f_{1},f_{2},…,f_{k}]$$ where the temporal length is $k$ and represent respective spatial feature of $k$ frames. Then $T$ short-term spatio-temporal features are cut and pooled from the global feature $F_{k}$. The $t$-th short-term spatio-temporal feature $x_{t}$ is constructed as, $$x_{t}=ltap[f_{t-\frac{k}{T}},f_{t-\frac{k}{T}+1},…,f_{t+\frac{k}{T}-1}]$$ where $ltap$ is local temporal average pool layer in truncated 3D-Densenet, $\frac{k}{T}$ is half of temporal feature interval. In this way, the adjacent $ltap$ windows also overlapping that assure the relevance and completeness of the front and back frame information.\
After local temporal average pooling, we can get a sequence of short-term features in single modality. Multimodal feature sequences are fused into one sequence before input into TCN. In this paper, all feature sequences of different modality are concated in channel dimension.
{width="48.00000%"} \[fig:TCN\]
{width="96.00000%"} \[fig:instance\]
TSENet + TCN for long-term prediction
-------------------------------------
Based on the short-term spatio-temporal features extracted from all kinds of data modalities (RGB, optic flow, depth, etc.), the long-term temporal features of the whole video is considered to classify the category of the given hand gesture. In this work, a sequence recognition model named TCNs is employed and modified to process the long-term temporal information. The main characteristics of TCNs are the use of causal convolutions and the mapping of an input sequence to an output sequence of the same length. In addition, accounting for sequences with long history, this model uses dilated convolutions that enable a large receptive field as well as residual connections that allow training deeper networks. Considering that our task is to classify the category of hand gesture videos, the output layer of TCN is further processed by one fully connection layer to obtain a single class label for each gesture sequence. The structure of the proposed modified version of the TCN model is depicted in Figure \[fig:TCN\].
The short-term temporal features $X=[x_{1},…,x_{T}]$ are utilized as the input sequence of the proposed modified TCN with the outputs $Y=[y_{1},…,y_{T}]$, while the calculation of $y_{t}$, $t<T$ depends only on [$X=[x_{1},…,x_{T}]$]{}. The reason is that the dilated convolutions are calculated as, $$y_{t}=(x*_{d}h)_{t}=\sum x_{t-d_{m}}h_{m}$$ where $*_{d}$ is the operator for dilated convolutions, $d$ is the dilation factor and $h$ is the filter’s impulse response. For a TCN with $L$ layers, the output of the last layer $y^L$ is used for the sequence classification. The class label $\widehat{o}$ attributed to the sequence is found through a fully connected layer with a softmax activation function, $$\widehat{o}=softmax(W_{o}\cdot y^L + b_{o})
\label{eqn::softmax}$$ where $W_{o}, b_{o}$ are trainable parameters.\
It is noting that the short-term spatio-temporal features $x_{1},…,x_{T}$ actually have different contributions to the recognition in the long-term temporal information processing. For instance, the gesture “swipe +” in Figure 4(b) contains three paths. The first path is extremely similar to the gesture “swipe left” (Figure 4(a)) when $t<9$. The same phenomenon occurs between the third path of the gestures “swipe +” and “swipe down” (Figure 4(c)) when $t>23$. In order to assign different temporal weight to $X=[x_{1},…,x_{T}]$ , a temporal Squeeze-and-Excitation network (TSENet) block is inserted between each temporal convolution layers.
As shown in Figure \[fig:TCN\], the average pooling is applied on the channel dimensions $C$ of $X=[x_{1},…,x_{T}]$ to squeeze channel-wise information. Such obtained temporal descriptor $z=[z_{1},…,z_{T}]$ is a $T\times1$ vector, while the $t$-th element of $z$ is calculated as, $$z_{t}=F_{sq}(x_{t})=\frac{1}{C}\sum_{i=1}^{C}x_{t}(i)$$
Then another excitation operation is followed to capture the temporal dependencies, i.e. the temporal weights. To fulfil this objective, we opt to employ a simple gating mechanism with the activations:
$$s=F_{ex}(z,W)=\sigma (g(z,W))=\sigma (W_{2}\delta (W_{1}z)))$$
where $\sigma$ refers to the sigmoid fuction, $\delta$ refers to the ReLU fuction, $W_{1}\in \mathbb{R}^{\frac{T}{r}\times T}$ and $W_{2}\in \mathbb{R}^{T\times \frac{T}{r}}$, and $r$ is the size of squeeze channel.The final output of the block is obtained by rescaling the transformation output $U$ with the activations: $$\widetilde{x}=F_{scale}(u_{t},s_{t})=s_{t}\cdot u_{t}$$ where $\widetilde{X}=[\widetilde{x}_{1},\widetilde{x}_{2},...,\widetilde{x}_{T}]$ and $F_{scale}(u_{t},s_{t})$ refers to temporal-wise multiplication between the scalar $s_{t}$ and the feature map $u_{t}\in \mathbb{R}^{T}$.
An example of the weights on different TSENet layers is illustrated in Figure 5. It can be seen that the values of the weights changes corresponding to the input gesture sequence as desired.
Experiment
==========
The proposed network architecture is implemented by tensorflow, and trained using one NVIDIA Quadro GP100 GPU. Multimodal 3D-DenseNet models have same structures and are pre-trained using RGB and optic flow(if optic flow existed or can be calculated) data respectively. Adam optimizer is used for training 3D-DenseNet and the learning rate is initialized to $6.4e-4$ and decayed by 10 every 25 epochs. The weight decay is set to $1e-4$. And the dropout rate is set to 0.2. The compression rate and the growth $k$ in the DenseNet block are set as 0.5 and 12, respectively. For the TCN model, we use Adam optimizer for training, and the learning rate is initialized to $1e-4$, epsilon is $1e-8$.
Dataset
-------
In this section, we compare our method with the other state-of-the-art dynamic hand gesture methods. Two publicly available multi-modal dynamic hand gesture datasets (VIVA[@molchanov2015hand] and NVGesture[@molchanov2016online]) are used to evaluate our proposed model in the experiments.
**VIVA**[@molchanov2015hand] The VIVA challenge’s dataset is a multimodal dynamic hand gesture dataset specifically designed with difficult settings of cluttered background, volatile illumination, and frequent occlusion for studying natural human activities in real-world driving settings. This dataset was captured using a Microsoft Kinect device, and contains 885 intensity and depth video sequences of 19 different dynamic hand gestures performed by 8 subjects inside a vehicle. Figure 4 shows some gesture sequences.
**NVGesture**[@molchanov2016online] The NVGesture dataset has been captured with multiple sensors and from multiple viewpoints for studying human-computer interfaces. It contains 1532 dynamic hand gestures recorded from 20 subjects inside a car simulator with artificial lighting conditions. This dataset includes 25 classes of hand gestures. The gestures were recorded with SoftKinetic DS325 device as the RGB-D sensor and DUO-3D for the infrared streams. In the experiments, we use RGB, depth and optical flow modalities, while the optical flow is calculated from the RGB stream using the method presented in [@farneback2003two].
Data Preprocessing
------------------
In VIVA dataset, data augmentation is comprised of three other operations: reverse ordering of frames, horizontal mirroring, and applying both operations together. With these operations we generated additional samples for training. For example, applying both operations transforms the original gesture “Swipe Left” with the right hand to a new gesture “Swipe Left” with the left hand . In NVGesture dataset, for special augmentation, videos are resized to have the smaller video size of 256 pixels, and then randomly cropped with a 224x224 patch.
Data normalization is also applied on both datasets, since a fixed dimension of input data is required in the C3D model and TCN model. For the videos with different temporal lengths, uniform normalization with temporal upsampling and downsampling is used. To compress or extend a given video $V$ with $n$ frames to $k$ frames, 1) If $n>k$, we split the video $V$ into a $k$ section video set $V_{S}$ averagely, where $V_{S}=[V_{1},V_{2},...,V_{k}]$. For each piece in the video set $V_{S}$, we randomly choose one frame as the representation of the sub-video fragment. Finally we concatenate all the represent frames and make them as the result of the normalization. 2) If $n<k$, we randomly choose $k-n$ frame in the video, then repeat them follow by themselves.
In our experiments, the average number of frames $k$ is set as 32 for VIVA dataset and 64 for NVGesture dataset. Due to the high complexity of 3D convolutional calculating, the spatial size of the inputs is restricted to $112\times112$.\
Method Fused modalities Accuracy
--------------------------------- ------------------ -----------
HOG+HOG2[@ohn2014hand] RGB+Depth 64.5
CNN:LRN[@molchanov2015hand] RGB+Depth 74.4
CNN:LRN:HRN[@molchanov2015hand] RGB+Depth 77.5
C3D[@tran2015learning] RGB+Depth 77.4
I3D[@carreira2017quo] RGB+Depth 83.10
MTUT[@abavisani2019improving] RGB+Depth 86.08
3D-Dense RGB+Depth 88.21
Res3D+TCN RGB+Depth 85.97
3D-Dense+TCN RGB+Depth 90.73
3D-Dense+TCN$_{tse}(proposed)$ RGB+Depth **91.54**
: **Accuracies of different multimodal fusion-based hand gesture methods on the VIVA dataset. The top performer is denoted by boldface.**[]{data-label="table:VIVA"}
\[tab1\]
Evaluation on VIVA Dataset.
---------------------------
Table \[table:VIVA\] shows the performance of the dynamic hand gestures tested on the RGB and depth modalities of the VIVA dataset. The compared methods include the hand-crafted approach HOG+HOG2[@ohn2014hand], the recurrent CNN-based method(CNN:LRN)[@molchanov2015hand], the C3D model which were pretrained on Sport-1M dataset, the I3D method[@carreira2017quo] that performs very well in action recognition, and the Multimodal Training / Unimodal Testing (MTUT) model[@abavisani2019improving] which shows promising performance in dynamic hand gesture recognition. All the results are reported by averaging the classification accuracies. It can be seen that the proposed model achieves the highest accuracy, which is 5.46% higher than the state-of-the-art method MTUT. This experiment shows that our model is effective to extract both short-term and long-term spatio-temporal information for dynamic hand gesture recognition.
![Examples of temporal weights.[]{data-label="fig:example Excitation"}](exampleExcitation.png){width="45.00000%"}
To validate the effect of the proposed TSENet layers, the accuracy obtained by vanilla TCN is also shown in Table \[table:VIVA\]. It can be seen that the presence of the TSENet layers in the TCN can improve the recognition rate by around 0.8%. Three examples of the temporal weights produced by TSENet layers are shown in Figure 6. It is interesting to see that the weights in the third layer contain obvious large and small values, which means it does select the important ones from the short-term features. Moreover, if we change the 3D-Dense networks to Res3D which is used for extracting the short-term features. The accuracy will further drop about 4.8%. It proves the effectiveness of the structure of the proposed model.
Figure \[fig:cm1a\] shows the confusion matrix as well for the experiment. It can be seen that the proposed model confused between the Swipe and Scroll gestures performed along the same direction. Many gestures were mis-classified as the Swipe down gesture, the Rotate CW/CCW gestures were difficult for the proposed model. In some case, the propose model may have difficulties with distinguishing between the Swipe + and the Swipe X gestures.
\[tp\]\
EVALUATION ON NVgesture.
------------------------
The NVGesture dataset, containing RGB, depth and optical flow modalities, is also used to test the proposed model. Table \[table:nvGesture\] tabulates the results of our method in comparison with the recent state-of-the-art methods: HOG+HOG2[@ohn2014hand], improved dense trajectories(iDT)[@wang2016robust], R3DCNN[@molchanov2016online], two-stream CNNs[@simonyan2014two], and C3D as well as human labeling accuracy. The iDT method is often recognized as the best performing hand-crafted method. However, we observe that similar to the pervious experiments the 3D-CNN-based methods outperform the other hand gesture recognition methods, and among them, our method provides the better performance in all the modalities. Nonetheless, compare to the latest method MTUT, our method accuacies are close to the MTUT. Our method has the better performance in both of RGB and optical flow modalities, it improve accuracy by 0.73%. But in RGB+Depth modalities and in RGB+Depth+Opt.flow modalities, our method is not performing good enough. This is in part due to the knowledge that gestures in NVGesture are more complex and have more invalid information. Although through TCN and TSE, our method can key information in the frames and weaken the influence of irrelevant information, the redundant non gesture information, especially in temporal, always affects the final results of the experiment.
Method Fused modalities Accuracy
----------------------------------- --------------------- -----------
HOG+HOG2[@ohn2014hand] RGB+Depth 36.9
I3D[@carreira2017quo] RGB+Depth 83.82
MTUT[@abavisani2019improving] RGB+Depth **86.10**
Ours RGB+Depth 84.87
Two Stream CNNs[@simonyan2014two] RGB+Opt. flow 65.6
iDT[@wang2016robust] RGB+Opt. flow 73.4
I3D[@carreira2017quo] RGB+Opt. flow 84.43
MTUT[@abavisani2019improving] RGB+Opt. flow 85.48
Ours RGB+Opt. flow **86.21**
R3DCNN[@molchanov2016online] RGB+Depth+Opt. flow 83.8
I3D[@carreira2017quo] RGB+Depth+Opt. flow 85.68
MTUT[@abavisani2019improving] RGB+Depth+Opt. flow **86.93**
Ours RGB+Depth+Opt. flow 86.37
88.4
: **Accuracies of different multimodal fusion-based hand gesture methods on the NVGesture dataset. The top performer is denoted by boldface.**[]{data-label="table:nvGesture"}
\[tab1\]
Conclusion {#sec:conclusion}
==========
We developed an effective method for multi-modal (RGB, depth and optic flow data) dynamic hand gesture recognition with 3D-DenseNets and TCNs. And in TCNs, we improved and applied an attention model named SENets to learn and extract deeper temporal features. The experiments show that the proposed model achieved the highest accuracy in VIVA dataset, as well as competitive results in NVGesture dataset.
However, our model is still not an end-to-end model and has to be trained step by step. Meanwhile, NVGesture still have a large room for improvement, we still have a lot of work to enhance the accuracy of the model.
[^1]: Faculty of Electrical Engineering and Computer Science, Ningbo University, P. R. China
[^2]: Corresponding author (e-mail: [email protected])
[^3]: This work was supported in part by the National Natural Science Foundation of China (61603202, 61571247), Zhejiang Provincial Natural Science Foundation of China (LY20F030005) and K.C. Wong Magna Fund in Ningbo University.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We define a certain class $\Upsilon$ of proper posets with the $\al_2$–chain condition. The corresponding forcing axiom is a generalization of Martin’s Axiom; in fact, $\textsc{MA}(\Upsilon)_{<\k}$ implies $\textsc{MA}_{<\k}$. Also, $\textsc{MA}(\Upsilon)_{<\k}$ implies certain uniform failures of club–guessing on $\o_1$ that don’t seem to have been considered in the literature before. We show, assuming $\textsc{CH}$ and given any regular cardinal $\k\geq\o_3$ such that $\m^{\al_1}< \kappa$ for all $\m < \k$ and $\diamondsuit(\{\a<\k\,:\,cf(\a)\geq\o_2\})$ holds, that there is a proper partial order $\mtcl P$ of size $\k$ with the $\al_2$–chain condition and producing a generic extension satisfying $2^{\al_0}=\k$ together with $\textsc{MA}(\Upsilon)_{<\k}$.'
address:
- 'David Asperó, Institute of Discrete Mathematics and Geometry, Technische Universität Wien, Wiedner Hauptstrasse 8-10/104, 1040 Wien, Austria'
- 'Miguel Angel Mota, Kurt Gödel Research Center for Mathematical Logic, Währinger Straße 25, 1090 Wien, Austria'
author:
- David Asperó
- Miguel Angel Mota
title: 'A Generalization of Martin’s Axiom'
---
A generalization of Martin’s Axiom. And some of its applications.
=================================================================
Martin’s Axiom, often denoted by $\textsc{MA}$, is the following very well–known and very classical forcing axiom: If $\mtbb P$ is a partial order (poset, for short) with the countable chain condition[^1] and $\mtcl D$ is a collection of size less than $2^{\al_0}$ consisting of dense subsets of $\mtbb P$, then there is a filter $G\sub\mtbb P$ such that $G\cap D\neq\emptyset$ for every $D\in\mtcl D$.
Martin’s Axiom is obviously a weakening of the Continuum Hypothesis. Given a cardinal $\l$, $\textsc{MA}_{\l}$ is obtained from considering, in the above formulation of $\textsc{MA}$, collections $\mtcl D$ of size at most $\l$ rather than of size less than $2^{\al_0}$. Martin’s Axiom becomes interesting when $2^{\al_0}>\al_1$.
$\textsc{MA}_{\o_1}$ was the first forcing axiom ever considered ([@Martin-Solovay]). As observed by D. Martin, the consistency of $\textsc{MA}$ together with $2^{\al_0}>\al_1$ follows from generalizing the Solovay–Tennenbaum construction of a model of Suslin’s Hypothesis by iterated forcing using finite supports ([@Solovay-Tennenbaum]). Since then, a huge number of applications of $\textsc{MA}$ (+ $2^{\al_0}> \al_1$) have been discovered in set theory, topology, measure theory, group theory, and so on ([@Fremlin] is a classical reference for this).
In the present paper we generalize Martin’s Axiom to a certain class of posets $\Upsilon$ with the $\al_2$–chain condition. In fact, every poset with the countable chain condition will be in $\Upsilon$, so that for every cardinal $\l$, the forcing axiom $\textsc{MA}(\Upsilon)_\l$ for $\Upsilon$ relative to collections of size $\l$ of dense sets will imply $\textsc{MA}_\l$. Furthermore, there will be no restriction on $\l$ other than $\l < 2^{\al_0}$. More precisely, the same construction will show that $\textsc{MA}(\Upsilon)_{\al_1}$, $\textsc{MA}(\Upsilon)_{\al_{727}}$, $\textsc{MA}(\Upsilon)_{\al_{\o_2 + \o + 3}}$, and so on are all consistent.[^2] This construction will take the form of a forcing iteration, in a broad sense of the expression, involving certain symmetric systems of countable structures as side conditions as in our previous work in [@ASP] and [@ASPMEA].
That $\Upsilon$ cannot possibly consist of all posets with the $\al_2$–c.c. is clear simply by considering the collapse of $\o_1$ to $\o$ with finite conditions.[^3] On the other hand, $\Upsilon$ will be general enough to make the corresponding forcing axiom $\textsc{MA}(\Upsilon)_\l$ strictly stronger than $\textsc{MA}_\l$. In fact, we will show that $\textsc{MA}(\Upsilon)_\l$ implies certain ‘uniform’ failures of Club Guessing on $\o_1$ that don’t seem to have been considered before in the literature, and which don’t follow from $\textsc{MA}_\l$. As a matter of fact, we don’t know how to show the consistency of these statements by any method other than ours. To be a little more precise, we don’t know how to prove their consistency by means of a forcing iteration in the ‘conventional’ sense.
Let $\mtbb P$ be a poset and let $N$ be a sufficiently correct structure such that $\mtbb P\in N$. Recall that a $\mtbb P$–condition $p$ is $(N, \mtbb P)$–generic if for every extension $p'$ of $p$ and every dense subset $D$ of $\mtbb P$ belonging to $N$ (equivalently, every maximal antichain $D$ of $\mtbb P$ belonging to $N$) there is some condition in $D \cap N$ compatible with $p'$. Also, $\mtbb P$ is proper ([@Shelah]) if for every cardinal $\t \geq \av TC(\mtbb P)\av^+$, it holds that for every (equivalently, for club–many) countable $N\elsub H(\t)$ and every $p\in N\cap\mtbb P$ there is a condition $q$ in $\mtbb P$ extending $p$ and such that $q$ is $(N, \mtbb P)$–generic. Every poset $\mtbb P$ with the countable chain condition is proper as every condition is $(N, \mtbb P)$–generic for every $N$ as above.
Now we may proceed to the definition of $\Upsilon$.
Given a poset $\mtbb P$, we will say that $\mtbb P$ is *regular* if and only if the following holds.
*\[(a)\] All its elements are ordered pairs whose first component is a countable ordinal. *\[(b)\] For every regular cardinal $\lambda\geq\ |TC(\mtbb P)|^+$ there is a club $D\sub[H(\l)]^{\al_0}$ such that for every finite subset $\{N_i\,:\, i \in m\}$ of $D$ and every condition $(\nu, X)$ such that $\nu < min \{N_i\cap \o_1\, :\, i < m\}$ there is a condition extending $(\nu, X)$ and $(N_i,\, P)$–generic for all $i$.[^4]**
We will say that a poset *admits a regular representation* if it is isomorphic to a regular poset.
Note that every regular poset is proper.
$\Upsilon$ is the class of all regular posets with the $\al_2$–chain condition.
The notation $\textsc{MA}(\Upsilon)_\l$ has already shown up.
\[def00\] Given a cardinal $\l$, let $\textsc{MA}(\Upsilon)_{\l}$ be the following statement: For every $\mtbb P$ in $\Upsilon$ and for every collection $\mtcl D$ of size $\l$ consisting of dense subsets of $\mtbb P$, there is a filter $G\sub\mtbb P$ such that $G\cap D\neq\emptyset$ for all $D\in \mtcl D$.
Some consequences of $\textsc{MA}(\Upsilon)_\l$
-----------------------------------------------
Note that every poset $\mtbb P$ with the countable chain condition admits a regular representation $\p:\mtbb P\into \{0\}\times \mtbb P$ given by simply setting $\p(p) = (0, p)$. In particular, for every $\l$, $\textsc{MA}(\Upsilon)_\l$ implies $\textsc{MA}_\l$.
Also, the following can be proved by arguing very much as in the standard proof that $\textsc{MA}_\l$ implies the productiveness of c.c.c. (see e.g. [@KUNEN], Lemma 2.23 and Theorem 2.24).
$\textsc{MA}(\Upsilon)_{\al_2}$ implies that if $\mtbb P\in\Upsilon$ and $X\in[\mtbb P]^{\al_2}$, then there is $Y\in[X]^{\al_2}$ such that every nonempty $\s\in[Y]^{<\o}$ has a lower bound in $\mtbb P$. In particular, any finite support product of members of $\Upsilon$ has the $\al_2$–chain condition.
Let us see that $\textsc{MA}(\Upsilon)_\l$ implies certain uniform failures of Club Guessing on $\o_1$. It will be convenient to consider the following natural notion of rank.
Given an ordinal $\a$ and a set $X$ of ordinals,
*\[(i)\] $rank(X, \a)> 0$ if and only if $\a$ is a limit point of $X$, and*
*\[(ii)\] for every ordinal $\eta>0$, $rank(X, \a)>\eta$ if and only if $\a$ if there is $Y\sub\a$ cofinal in $\a$ such that $rank(X, \b)\geq\b$ for all $\b\in Y$.*
Given a cardinal $\l$ and a countable ordinal $\tau$, let $(\ast)^\tau_\l$ be the following statement:
For every $\l'\leq\l$ and every sequence $(A_i)_{i<\l'}$, if each $A_i$ is a subset of $\o_1$ of order type at most $\tau$, then there is a club $C\sub\o_1$ such that $C\cap A_i$ is finite for every $i<\l'$.
Given a cardinal $\l$, $(\ast)^+_\l$ is the following statement:
For every $\l'\leq\l$ and every sequence $(A_i)_{i<\l'}$ of infinite subsets of $\o_1$ there is a club $C\sub\o_1$ such that for every $\d\in C$ and every $i<\l'$, if $rank(A_i, \d)<\d$, then $C\cap A_i\cap\d$ is finite.
Clearly, for every $\l$, $(\ast)^+_\l$ implies $(\ast)^\tau_\l$ for every $\tau < \o_1$. Also, if $(\tau, \l)$, $(\tau', \l')$ are such that $\tau\leq\tau'$ and $\l\leq\l'$, then $(\ast)^{\tau'}_{\l'}$ implies $(\ast)^\tau_\l$. In particular, for every infinite $\tau$ and every $\l\geq \o_1$, $(\ast)^\tau_\l$ implies the negation of Weak Club Guessing on $\o_1$.
For every cardinal $\l\geq\o_1$, the following weakening of $(\ast)^\o_\l$ implies $2^{\al_0}>\l$: For every $\l'\leq\l$ and every sequence $(A^i_\d)_{i<\l',\,\d\in Lim(\o_1)}$, if each $A^i_\d$ is a cofinal subset of $\d$ of order type $\o$, then there is a club $C\sub\o_1$ such that $A^i_\d \nsubseteq C$ for all $i<\l'$ and $\d\in Lim(\o_1)$.
Suppose $2^{\al_0}\leq \l$ and let $(A^i_\d)_{i < \l,\,\d\in Lim(\o_1)}$ be such that for each $\d$, $\{A^i_\d\,:\,i<\l\}$ contains all cofinal subsets of $\d$. If $C\sub\o_1$ is a club and $\d\in C$ is a limit point of $C$, then there is $i<\l$ such that $A^i_\d\sub C$.
\[unifnoclub-guessing\] For every cardinal $\l\geq\o_1$, $\textsc{MA}(\Upsilon)_\l$ implies $(\ast)^+_\l$.
Let $\l'$ and $(A_i)_{i<\l'}$ be as in the definition of $(\ast)^+_\l$. Let $\mtbb P$ consist of all pairs $(f, X)$ such that
*\[(a)\] $f\sub\o_1$ is a finite function such that $rank(f(\n), f(\n))\geq\n$ for every $\n\in dom(f)$,*
*\[(b)\] $X$ is finite set of triples $(i, \n, a)$ such that $i < \l'$, $\n\in dom(f)$, $rank(A_i, f(\n))<f(\n)$, and $a$ is a finite subset of $f(\n)$, and*
*\[(c)\] for every $(i, \n, a)\in X$, $range(f\restr \n)\cap A_i=a$.*
Given $\mtbb P$–conditions $(f_0, X_0)$ and $(f_1, X_1)$, $(f_1, X_1)$ extends $(f_0, X_0)$ if $f_0\sub f_1$ and $X_0 \sub X_1$.
It is easy to check that $\mtbb P$ admits a regular representation and that it is $\al_2$–Knaster[^5] (for example by arguments as in [@ASP] for similar forcings). Also, there is a collection $\mtcl D$ of $max\{\l', \o_1\}$–many dense subsets of $\o_1$ such that if $G$ is a filter of $\mtbb P$ meeting all members of $\mtcl D$, then $range(\bigcup\{f\,:\,(f, X)\in G\textrm{ for some }X\}$ is a club witnessing $(\ast)^+_\l$ for $(A_i)_{i<\l'}$.
On the other hand, no forcing axiom $\textsc{MA}_\l$ implies $(\ast)^\tau_{\l'}$ for any infinite $\tau<\o_1$ and any $\l'\geq \o_1$. The reason is simply that $\textsc{MA}_\l$ can always be forced by a c.c.c. forcing and c.c.c. forcing preserves Weak Club Guessing.
Given a cardinal $\l$, $(\triangleleft)_\l$ is the following statement:
Let $\l'\leq\l$, and suppose $(f_i)_{i<\l'}$ is a sequence of functions such that for each $i$ there is some $\a_i<\o_1$ such that $f_i:\a_i\into \o$ is a continuous function with respect to the order topology. Then there is a club $C\sub\o_1$ such that for all $i<\l'$, $range(f_i\restr C)\neq\o$.
$(\triangleleft)_\l$ clearly implies $\lnot\mho$ in J. Moore’s terminology ([@MOORE2]) as well as $2^{\al_0}>\l$. Also, by the same argument as before, no forcing axiom of the form $\textsc{MA}_\l$ implies $(\triangleleft)_{\l'}$ for any $\l'\geq \o_1$.
The proof of the following result is as in the proof of Fact \[unifnoclub-guessing\]. The proof is similar to the proof that $\textsc{PFA}^\star(\o_1)_{\o_1}$ (see below) implies $\lnot\mho$ (cf. [@ASP]).
For every cardinal $\l$, $\textsc{MA}(\Upsilon)_\l$ implies $(\triangleleft)_\l$.
([@ASP]) Given a partial order $\mtbb P$, *$\mtbb P$ is finitely proper* if for every cardinal $\t\geq\av \mtbb P\av^+$, every finite sequence $\{N_0, \ldots N_n\}$ of countable elementary substructures of $H(\t)$ containing $\mtbb P$, and every $p\in N\cap\mtbb P$ there is a condition in $\mtbb P$ extending $p$ and $(N_i, \mtbb P)$–generic for every $i< n+1$.
([@ASP], essentially) Given a cardinal $\l$, $\textsc{PFA}^\star(\o_1)_\l$ is the forcing axiom for the class of finitely proper posets of size $\al_1$ and for collections of $\l$–many dense sets.
For every cardinal $\l\geq\o_1$, $\textsc{PFA}^\star(\o_1)_\l$ implies the following.
*\[(1)\] $\textsc{MA}_{\o_1}$*
*\[(2)\] $(\ast)^+_{\o_1}$*
*\[(3)\] $(\triangleleft)_{\o_1}$*
*\[(4)\] For every set $\mtcl F$ of size $\l$ consisting of functions from $\o_1$ into $\o_1$ there is a normal function $g:\o_1\into \o_1$ such that $\{\n<\o_1\,:\,f(\n)<g(\n)\}$ is unbounded for every $f\in\mtcl F$.*
The proofs of (1)–(3) are either immediate or as the corresponding proofs from $\textsc{MA}(\Upsilon)_\l$. (4) follows from considering Baumgartner’s forcing for adding a club $C\sub\o_1$ by finite approximations.
The following result is straightforward.
\[fin\_proper\] Every finitely proper poset of size $\al_1$ is in $\Upsilon$. In particular, for every cardinal $\l$, $\textsc{MA}(\Upsilon)_\l$ implies $\textsc{PFA}^\star(\o_1)_\l$.
The proof of the main theorem in [@ASP] essentially shows the consistency of $\textsc{PFA}^\star(\o_1)_\l$ for arbitrary $\l$.
The consistency of $\textsc{MA}(\Upsilon)_\l$ {#the_forcing_construction}
=============================================
Our main theorem is the following:
\[mainthm\] ($\textsc{CH}$) Let $\k\geq\o_3$ be a regular cardinal such that $\m^{\al_1}< \kappa$ for all $\m < \k$ and $\diamondsuit(\{\a<\k\,:\,cf(\a)\geq\o_2\})$ holds. Then there exists a proper forcing notion $\mtcl P$ of size $\k$ with the $\al_2$–chain condition such that the following statements hold in the generic extension by $\mtcl P$:
*\[(1)\] $2^{\al_0}=\k$*
*\[(2)\] $\textsc{MA}(\Upsilon)_{<2^{\al_0}}$*
The proof of Theorem \[mainthm\] is an elaboration of the proof of the main theorem in [@ASP]. Our approach in that paper consisted in building a certain type of finite support forcing iteration $\la\mtcl P_\a\,:\,\a\leq\k\ra$ of length $\k$ (in a broad sense of ‘forcing iteration’)[^6] using what one may describe as finite “symmetric” systems of countable elementary substructures of a fixed $H(\k)$[^7] as side conditions. These systems of structures were added at the first stage $\mtcl P_0$ of the iteration. Roughly speaking, the fact that the supports of the conditions in the iteration was finite ensured that the inductive proofs of the relevant facts – mainly the $\al_2$–c.c. of all $\mtcl P_\a$ and their properness – went through. The use of the sets of structures as side conditions was crucial in the proof of properness.[^8] Here we change the set–up from [@ASP] in various ways. One of the changes is the presence of a diamond–sequence which ensures that all proper posets with the $\al_2$–c.c. (with no restrictions on their size) occurring in the final extension have been dealt with at $\k$–many stages during the iteration. Of course, Theorem \[mainthm\] shows also that all forcing axioms of the form $\textsc{MA}(\Upsilon)_{<\k}$, for a fixed reasonably defined cardinal $\k$, are consistent (relative to the consistency of $\textsc{ZFC}$). As far as we know, these axioms have not been considered in the literature before for $\k\geq\al_3$.
Also, Theorem \[mainthm\] shows that no axiom of the form $\textsc{MA}(\Upsilon)_\l$ decides the size of the continuum and thus, by Fact \[fin\_proper\], fits nicely within the ongoing project of showing whether or not weak fragments of $\textsc{BPFA}$[^9] imply $2^{\al_0} = \al_2$. The problem whether (consequences of) forcing axioms for classes of posets with small chain condition decide the size of the continuum does not seem to have received much attention in the literature so far.[^10] One place where the problem has been addressed is of course our [@ASP]. Before that, M. Foreman and P. Larson showed in an unpublished note ([@FOREMAN-LARSON]) that $\textsc{FA}(\Gamma)$, for $\Gamma$ being the class of posets of size $\al_2$ preserving stationary subsets of $\o_1$, implies $2^{\al_0}=\al_2$. Several natural problems in this area remain open. For example it is not known whether the forcing axiom for the class of semi-proper posets of size $\al_2$ implies $2^{\al_0}=\al_2$, and the same is open for the forcing axiom for the class of all posets of size $\al_1$ preserving stationary subsets of $\o_1$, and even for the forcing axiom for the class of all proper posets of size $\al_1$ (let us denote these two forcing axioms by, respectively, $\textsc{MM}(\o_1)$ and $\textsc{PFA}(\o_1)$). It is open whether or not $\textsc{MM}(\o_1)$ is equivalent to $\textsc{PFA}(\o_1)$,[^11] and even whether $\textsc{MM}(\o_1)$ has consistency strength above $\textsc{ZFC}$.[^12]
As we mentioned before, Theorem \[mainthm\] shows in particular that the forcing axiom $\textsc{MA}(\Upsilon)_{<\al_2}$ has the same consistency strength as $\textsc{ZFC}$. Other articles dealing with the consistency strength of other (related) fragments of $\textsc{PFA}$ are [@Miyamoto], [@HamkinsJohnstone], [@NeemanSchimmerlingI], and [@Neeman].
For the most part our notation follows set–theoretic standards as set forth for example in [@JECH] and in [@KUNEN], but we will also make use of certain *ad hoc* pieces of notation that we introduce now. If $N$ is a set whose intersection with $\omega_1$ is an ordinal, then $\delta_N$ will denote this intersection. Throughout this paper, if $N$ and $N'$ are such that there is a (unique) isomorphism from $(N, \in)$ into $(N', \in)$, then we denote this isomorphism by $\Psi_{N, N'}$. If $\la\mtcl P_\a\,:\,\a\leq\k\ra$ is a forcing iteration and $\a\leq\k$, $\dot G_\a$ denotes the canonical $\mtcl P_\a$–name for the generic filter added by $\mtcl P_\a$. Also, $\leq_\a$ typically denotes the extension relation on $\mtcl P_\a$. We will make use of the following general fact:
\[factproper2\] For every $\mtcl Q\in \Upsilon$ and every $X\sub\mtcl Q$ there is $\mtcl R$ such that
*\[(i)\] $\mtcl R$ is a complete suborder of $\mtcl Q$,*
*\[(ii)\] $\mtcl R\in\Upsilon$,*
*\[(ii)\] $X\sub\mtcl R$, and*
*\[(iii)\] $\av\mtcl R\av = \av X\av^{\al_1}$*
Let $M$ be an elementary substructure of some large enough $H(\t)$ containing everything relevant and closed under $\o_1$–sequences. Set $\mtcl R=\mtcl Q\cap M$. Since $M$ is closed under $\o_1$–sequences, $\mtcl R$ has the $\al_2$–chain condition and is a complete suborder of $\mtcl Q$.
To see that $\mtcl R$ is regular, let $\chi\geq\av TC(\mtcl Q)\av^+$ be a cardinal in $M$ and $W$ a well–order of $H(\chi)$ also in $M$, let $D\sub[H(\chi)]^{\al_0}$ be a club witnessing the regularity of $\mtcl Q$, let $(\n, x)\in \mtcl R$, and let $N_0,\ldots N_m$ be countable elementary substructures of $(H(\chi), \in, W)$ in $D$ containing $\mtcl Q$ and such that $\n< N_i\cap\o_1$ for all $i$. In $M$, there are countable elementary substructures $M_0, \ldots M_m$ of $(H(\chi), \in, W)$ such that for all $i$,
*\[(a)\] $M_i\cap N_i=N_i\cap M$,[^13] and*
*\[(b)\] there is an isomorphism $\varphi_i: (N_i, \in, W)\into (M_i, \in, W)$ fixing $N_i\cap M_i$,*
and there is a condition $(\ov\n, \ov x)\in\mtcl R$ such that $(\ov\n, \ov x)\leq_{\mtcl Q}(\n, x)$ and such that $(\ov\n, \ov x)$ is $(M_i, \mtcl Q)$–generic for every $i$. Suppose towards a contradiction that there is $(\n', x')\leq_{\mtcl R}(\ov\n, \ov x)$, $i_0< m+1$ and some maximal antichain $A$ of $\mtcl R$ in $N_i$ such that no condition in $A\cap N_i$ is compatible with $(\n', x')$. Let $(\n^\ast, x^\ast)\in\mtcl R$ be a common extension of $(\n', x')$ and of some $(\n'', x'')\in \varphi_i(A)\cap M_i$. To see that there are such $(\n^\ast, x^\ast)$ and $(\n'', x'')$, note that $\varphi_i(A)\in M_i$ is a maximal antichain of $\mtcl Q$. This is true since $\mtcl Q\in N_i\cap M_i$.
Now note that $A\in M$ since $M$ is closed under $\o_1$–sequences and $\av A\av\leq\al_1$. It follows that $\varphi_i(A)=A$ since $\varphi_i$ is the identity on $M\cap N_i$. Also, $(\n'', x'')\in N_i$. To see this, take a surjection $f:\o_1\into A$ in $N_i\cap M$ (take for example the $W$–first surjection $f:\o_1\into A$). Then $\varphi_i(f) = f\in M_i$ is a surjection from $\o_1$ onto $\varphi_i(A)=A$. Let $\z\in M_i \cap\o_1$ such that $f(\z)=(\n'', x'')$. Then $\z\in N_i\cap\o_1$ and so $(\n'', x'')= f(\z)\in N_i$. This contradiction finishes the proof.
If $q =(p, \D_q)$, where $p$ is a function , we will use $supp(q)$ to denote $dom(p)$.[^14] If $q =(p, \D_q)$ and $r =(s, \D_r)$ are ordered pairs, $p$ and $s$ are functions, and $\a$ is an ordinal such that $dom(s)\sub\a$, then we denote by $q\wedge_\a r$ the ordered pair $$(s\cup (p \restr_{dom(p) \setminus \alpha}), \D_q\cup\D_r)\footnote{Note that the first component of $q\wedge_\a r$ is also a function.}$$
The rest of the paper is organized as follows: Section \[the\_forcing\_construction\] contains the construction of our iteration $\la\mtcl P_\a\,:\,\a\leq\k\ra$ ($\mtcl P_\k$ will witness Theorem \[mainthm\]). Section \[the\_main\_facts\] contains proofs of the main facts about $\la\mtcl P_\a\,:\,\a\leq\k\ra$. Theorem \[mainthm\] follows then easily from these facts.
The forcing construction {#the_forcing_construction}
========================
The proof of Theorem \[mainthm\] will be given in a sequence of lemmas.
Let $\vec X=\la X_\a\,:\,\a\in \k,\,cf(\a)\geq\o_2\ra$ be a $\diamondsuit(\{\a\in \k,\,cf(\a)\geq\o_2\})$–sequence. Note that $2^{<\k} = \k$. Let also $\Psi$ be a well–order of $H(\k^+)$ in order type $2^\k$. We will define an iteration $\la\mtcl P_\a\,:\,\a\leq\k\ra$, together with a function $\Phi:\k\into H(\k)$ such that each $\Phi(\a)$ is a $\mtcl P_\a$–name in $H(\k)$.
Let $\la \t_\a\,:\,\a\leq \k\ra$ be the strictly increasing sequence of regular cardinals defined as $\t_0 =|2^{\k}|^{+}$ and $\t_\a=|2^{sup\{\t_\beta\,:\,\b < \a \}}|^{+}$ if $\a>0$. For each $\a\leq\k$ let $\mtcl M^\ast_\a$ be the collection of all countable elementary substructures of $H(\t_\a)$ containing $\vec X$, $\Psi$, and $\la \t_\b\,:\,\b<\a\ra$. Let also $\mtcl M_\a=\{N^\ast\cap H(\k)\,:\,N^\ast\in\mtcl M_\a\}$ and note that if $\alpha < \beta$, then $\mtcl M^\ast_\a$ belongs to all members of $\mtcl M^\ast_\b$ containing the ordinal $\alpha$.
Our forcing $\mtcl P$ will be the direct limit $\mtcl P_{\k}$ of a certain sequence $\la\mtcl P_\a\,:\, \a < \k\ra$ of forcings. The properness of each $\mtcl P_\a$ will be witnessed by the club $\mtcl M^\ast_\a$. The main idea here is to use the elements of $\mtcl M_\a$ as side conditions to ensure properness. The actual proof of properness (Lemma \[horribilis\]) will be by induction on $\a$. For technical reasons involving the limit case of that proof we need that our side conditions satisfy certain symmetry requirements. These requirements are encapsulated in the following notion of *symmetric system of structures*.
Let $\mtcl M$ be a club of $[H(\k)]^{\al_0}$ and let $\{N_i\,:\,i<m\}$ be a finite set of members of $\mtcl M$. We will say that *$\{N_i\,:\,i<m\}$ is a symmetric system of members of $\mtcl M$* if
*\[$(A)$\] For every $i<m$, $N_i\in\mtcl M$.*
*\[$(B)$\] Given distinct $i$, $i'$ in $m$, if $\d_{N_i}=\d_{N_{i'}}$, then there is a (unique) isomorphism $$\Psi_{N_i, N_{i'}}:(N_i, \in)\into (N_{i'}, \in)$$ Furthermore, we ask that $\Psi_{N_i, N_{i'}}$ be the identity on $\k\cap N_i\cap N_{i'}$.*
*\[$(C)$\] For all $i$, $j$ in $m$, if $\d_{N_j}<\d_{N_i}$, then there is some $i'<m$ such that $\d_{N_{i'}}=\d_{N_i}$ and $N_j\in N_{i'}$.*
*\[$(D)$\] For all $i$, $i'$, $j$ in $m$, if $N_j\in N_i$ and $\d_{N_i}=\d_{N_{i'}}$, then there is some $j'<m$ such that $\Psi_{N_i, N_{i'}}(N_j)=N_{j'}$.*
If $\mtcl M$ is the club of countable $N\elsub H(\k)$, we call $\{N_i\,:\,i<m\}$ a *symmetric system of elementary substructures of $H(\k)$*.
Let us proceed to the definition of $\la\mtcl P_\a\,:\,\a\leq \k\ra$ now.
Conditions in $\mtcl P_0$ are pairs of the form $(\emptyset, \D)$, where
*\[$(A)$\] $\D$ is a finite set of ordered pairs of the form $(N, 0)$ such that $dom(\D)$ is symmetric system of members of $\mtcl M_0$.*
Given $\mtcl P_0$–conditions $q_\e=(\emptyset, \D_\e)$ for $\e\in\{0, 1\}$, $q_1$ extends $q_0$ if and only if
*\[$(B)$\] $dom(\D_0)\sub dom(\D_1)$*
Now suppose $\b \leq \k$, $\b>0$, and suppose that for each $\a<\b$, $\mtcl P_\a$ has been defined and is a partial order consisting of pairs of the form $r=(s, \D_r)$, where $s$ is a finite function with domain included in $\a$ and $\D_r$ is a set of pairs $(N, \g)$ with $N\in[H(\k)]^{\al_0}$ and $\g\in \a+1$. For each $\a<\b$ let $\mtcl N_{\dot G_\a}$ be a canonical $\mtcl P_\a$–name for $\bigcup\{\D_r^{-1}(\a)\,:\,r\in\dot G_\a\}$.
(In $V[G]$ for a $\mtcl P_\a$–generic filter $G$, for a given $\a \leq \k$ such that $\mtcl P_\a$ has been defined) A poset $\mtcl Q\sub H(\k)$ is *$H(\k^+)^V$–regular relative to $G$* if the following holds:
*\[(a)\] All its elements are ordered pairs whose first component is a countable ordinal. *\[(b)\] There is a club $D$ of $[H(\k)^V]^{\al_0}$ in $V$ such that for every finite set $\{N_i\,:\, i \in m\} \subseteq D \cap \mtcl N_{\dot G_\a}$ and every condition $(\nu, X) \in \mtcl Q$ with $\nu < min \{N_i\cap \o_1\, :\, i < m\}$ there is a condition extending $(\nu, X)$ and $(N_i[G],\, \mtcl Q)$–generic for all $i$.[^15]**
If $\b=\a_0+1$, then let $\Phi(\a_0)$ be a $\mtcl P_{\a_0}$–name in $H(\k)$ for an $H(\k^+)^V$–regular poset relative to $\dot G_{\a_0}$ with the $\al_2$–chain condition such that $\Phi(\a_0)$ is (say) the canonical $\mtcl P_{\a_0}$–name for trivial forcing on $\{(0,0)\}$ unless $X_{\a_0}$ is defined and codes (via some fixed reasonable translating function $\p$)[^16] a $\mtcl P_{\a_0}$–name $\dot X$. In that case, $\Phi(\a_0)$ is a $\mtcl P_{\a_0}$–name in $H(\k)$ for an $H(\k^+)^V$–regular poset relative to $\dot G_{\a_0}$ with the $\al_2$–chain condition such that $\Phi(\a_0)$ is $\dot X$ if $\dot X$ is such a poset and such that $\Phi(\a_0)$ is (say) trivial forcing on $\{(0,0)\}$ if $\dot X$ is not such a poset.
Let $0< \beta \leq \kappa$. Conditions in $\mtcl P_\b$ are pairs of the form $q= (p, \D)$ with the following properties.
*\[$(C0)$\] $p$ is a finite function and $dom(p)\sub\b$.*
*\[$(C1)$\] $\D$ is a finite set of pairs $(N, \g)$ such that $\g\leq \b \cap sup(N\cap \kappa)$.*
*\[$(C2)$\] For every $\a<\b$, the restriction $q\av_\a$ of $q$ to $\a$ is a condition in $\mtcl P_\a$, where $$q\av_\a:=(p\restr\a, \{(N, min\{\a, \g\})\,:\,(N, \g)\in \D\})$$*
*\[$(C3)$\] If $\a\in dom(p)$, then $p(\a) \in H(\k)$ and $q\av_\a\Vdash_{\mtcl P_\a} p(\a)\in\Phi(\a)$.*
*\[$(C4)$\] If $\a\in dom(p)$, $N \in \mtcl M_{\a+ 1}$ and $N\in\D_{q\av_{\a+1}}^{-1}(\a+1)$, then $q\av_\a$ forces in $\mtcl P_\a$ that $p(\a)$ is $(N[\dot G_\a], \Phi(\a))$–generic.*
Given conditions $q_\e=(p_\e, \D_e)$ (for $\e\in\{0, 1\}$) in $\mtcl P_\b$, $q_1$ extends $q_0$ if and only if the following holds:
*\[$(D1)$\] $q_1\av_\a \leq_\a q_0\av_\a$ for all $\a<\b$.*
*\[$(D2)$\] $dom(p_0)\subseteq dom(p_1)$ and if $\a \in dom(p_0)$, then $q_1\av_{\a}$ forces in $\mtcl P_{\a}$ that $p_1(\a)$ $\Phi(\a)$–extends $p_0(\a)$.*
*\[$(D3)$\] $\D_0^{-1}(\b)\sub \D_1^{-1}(\b)$ if $\b<\k$.*
The Main Facts {#the_main_facts}
==============
We are going to prove the relevant properties of the forcings $\mtcl P_\a$. Theorem \[mainthm\] will follow immediately from them.
Our first lemma is immediate from the definitions.
\[suborder\] $\mtcl P_\k=\bigcup_{\b<\k}\mtcl P_\b$, and $\emptyset\neq\mtcl P_\a\sub \mtcl P_\b$ for all $\a\leq\b\leq\k$.
Lemma \[compll\] shows in particular that $\la\mtcl P_\a\,:\,\a\leq\k\ra$ is a forcing iteration in a broad sense.
\[compll\] Let $\b\leq\k$ be an ordinal. If $\a < \b \leq \k$, $r\in \mtcl P_\a$, $q \in \mtcl P_\b$, and $r \leq_\a
q\av_\a$, then $q\wedge_\a r$ is a condition in $\mtcl P_\b$ extending $q$. In particular, $\mtcl P_\a$ is a complete suborder of $\mtcl P_\b$.
This proof makes use of the fact that models in $dom(\D_{q\wedge_\a r})$ come always equipped with suitable markers $\g$. New side conditions $(N, \g)$ appearing in $\D_r$ may well have the property that $[\a,\,\b)\cap N\neq\emptyset$, but they will not impose any problematic restraints – coming from clause $(C4)$ in the definition – on any $p(\x)$ for $\x\in[\a,\,\b)$. The reason is simply that $\g\leq\a$.
The following lemma gives a representation of $\mtcl P_{\a+1}$ as a certain dense subset of an iteration of the form $\mtcl P_\a\ast\dot{\mtcl Q}_\a$.
\[repr\] For all $\a<\k$, $\mtcl P_{\a+1}$ is isomorphic to a dense suborder of $\mtcl P_\a\ast\dot{\mtcl Q}_\a$, where $\dot{\mtcl Q}_\a$ is, in $V^{\mtcl P_\a}$, the collection of all pairs $(v, \mtcl Q)$ such that
*\[$(\circ)$\] there is some $r = (p, \D)\in\dot G_\a$ such that $$(p\cup\{\la\a, v\ra\}, \D\cup \{(N, \a+1)\,:\, N\in\mtcl Q\})\in\mtcl P_{\a+1},\footnote{Note that $v$ is an ordered pair whose first component is a countable ordinal.}$$*
ordered by $(v_1, \mtcl Q_1)\leq_{\dot{\mtcl Q}_\a}(v_0, \mtcl Q_0)$ if and only if
*\[(i)\] $\mtcl Q_0\sub\mtcl Q_1$, and*
*\[(ii)\] $v_1$ $\Phi(\a)$–extends $v_0$.*
Let $\wdtld{\mtcl P}_{\a+1}$ consist of all $(r, \check x)$, where $r\in\mtcl P_\a$ and $r\Vdash_{\mtcl P_\a}\check x\in\dot{\mtcl Q}_\a$. Then $\psi:\mtcl P_{\a+1}\into\wdtld{\mtcl P}_{\a+1}$ is an isomorphism, where $\psi(q)=(q\av_\a,\,\check x)$ for $x=(v, \D^{-1}(\a+1))$ if $q=(p\cup\{\la\a, v\ra\}, \D)$.
The next step in the proof of Theorem \[mainthm\] will be to show that all $\mtcl P_\a$ (for $\a\leq \k$) have the $\al_2$–chain condition.
\[cc\] For every ordinal $\b\leq \k$, $\mtcl P_\b$ has the $\al_2$–chain condition.
The proof is by induction on $\b$. The conclusion for $\b=0$ holds by a simplified version of the $\D$–system argument (using $\textsc{CH}$) we will see in a moment in the limit case.
For $\b=\a+1$ the conclusion follows immediately from Lemma \[repr\] together with the induction hypothesis for $\a$, the fact that the poset $\dot{\mtcl Q}_\a$ in Lemma \[repr\] is forced by $\mtcl P_\a$ to have the $\al_2$–c.c. (since this is true for $\Phi(\a)$), and the fact that the $\al_2$–c.c. is preserved under forcing iterations of length $2$.
Now suppose $\b\leq\k$ is a nonzero limit ordinal. Let $q_\x$ be $\mtcl P_\b$–conditions for $\x<\o_2$. By a $\D$–system argument using $\textsc{CH}$ we may assume that $dom(\D_{q_\x} \cup \D_{p_{\x'}})$ is a symmetric system of structures, that $\{supp(q_\x)\,:\,\x<\o_2\}$ forms a $\D$–system with root $R$ and, furthermore, that for all distinct $\x$, $\x'$ in $\o_2$, $supp(q_\x)\setminus R$ has empty intersection with $\bigcup dom(\D_{q_{\x'}})$. Let $\a<\b$ be a bound for $R$. By induction hypothesis we may find distinct $\x$, $\x'$ such that $q_\x\av_\a$ and $q_{\x'}\av_\a$ are compatible $\mtcl P_\a$–conditions. Let $r$ be a common extension of $q_\x\av_\a$ and $q_{\x'}\av_\a$. It follows now that the natural amalgamation of $r$, $q_\x$ and $q_{\x'}$ is a common extension of $q_\x$ and $q_{\x'}$. The case $\b=\k$ follows also from $\mtcl P_\k=\bigcup_{\b<\k}\mtcl P_\b$ together with $cf(\k)\geq\o_3$.
\[cor00\] For every $\b\leq\k$, $\mtcl P_\b$ forces $H(\k)^{V[\dot G_\b]}=H(\k)^V[\dot G_\b]$ and forces $N^\ast[\dot G_\b]\cap H(\k)= (N^\ast\cap H(\k))[\dot G_\b]$ whenever $\t\geq\k^+$ is regular and $N^\ast$ is a countable elementary substructure of $H(\t)$ such that $\mtcl P_\b\in N^\ast$.
Given $\a\leq \kappa$, a condition $q\in\mtcl P_\a$, and a countable elementary substructure $N \prec H(\kappa)$, we will say that $q$ is $(N,\, \mtcl P_\a)$–pre-generic in case
*\[$(\circ)$\] $\a < \k$ and the pair $(N, \a)$ is in $\Delta_q$, or else*
*\[$(\circ)$\] $\a = \k$ and the pair $(N, sup(N\cap \k))$ is in $\Delta_q$.*
The properness of all $\mtcl P_\a$ is an immediate consequence of the following lemma.
\[horribilis\]
Suppose $\a\leq \kappa$ and $N^\ast\in\mtcl M^\ast_\a$. Let $N=N^\ast\cap H(\k)$. Then the following conditions hold.
*\[$(1)_\a$\] For every $q\in N$ there is $q'\leq_\a q$ such that $q'$ is $(N,\, \mtcl P_\a)$–pre-generic.*
*\[$(2)_\a$\] If $\mtcl P_\a\in N^\ast$ and $q\in\mtcl P_\a$ is $(N,\, \mtcl P_\a)$–pre-generic, then $q$ is $(N^\ast,\, \mtcl P_\a)$–generic.*
The proof will be by induction on $\a$. We start with the case $\a =0$. For simplicity we are going to identify a $\mtcl P_0$–condition $q=(\emptyset, \D_q)$ with $dom(\D_q)$. The proof of $(1)_0$ is trivial: It suffices to set $q'= q\cup\{(N, 0)\}$.
The proof of $(2)_0$ is also easy: Let $E$ be a dense subset of $\mtcl P_0$ in $N^\ast$. It suffices to show that there is some condition in $E\cap N^\ast$ compatible with $q$. Notice that $q\cap N^\ast\in\mtcl P_0$. Hence, we may find a condition $q^\circ\in E\cap N^\ast$ extending $q\cap N^\ast$. Now let $$q^\ast=q\cup\{\Psi_{N, \ov N}(M) \,:\, M \in q^\circ,\,\ov N\in dom(\D_q),\,\d_{\ov N}=\d_N\}$$ It takes a routine verification to check that $q^\ast$ is a condition in $\mtcl P_0$ extending both $q$ and $q^\circ$.
Let us proceed to the more substantial case $\a=\s+1$. We start by proving $(1)_\a$. Without loss of generality we may assume that $\s$ is in the support of $q$ (otherwise the proof is easier). So, $q$ is of the form $q=(p\cup\{(\sigma, v)\}, \D_q)$, where $v$ is an ordered pair whose first component is a countable ordinal less than $\d_N$. By $(1)_\s$ we may assume that there is a condition $t\in \mtcl P_\s$ extending $q\av_\s$ and $(N, \mtcl P_\s)$–pre-generic. In $V^{\mtcl P_\sigma \restr t}$ let $\dot D$ be the $\Psi$–first club in $V\cap N^\ast[\dot G_\sigma]$ witnessing the $H(\k^+)^V$–regularity of $\Phi(\s)$ relative to $\dot G_{\s}$,[^17] and note that $N\in \dot D$ since $\dot D$ is closed in $V$ and $N[\dot G_\s]\cap V = N$. There is then some $v^{\ast}\in H(\k)$ which is a $(N[\dot G_{\s}], \Phi(\s))$–generic condition $\Phi(\s)$–extending $v$. Now it suffices to pick $a=(b, \D_a)$ $\leq_{\s}$–extending $t$ and deciding $v^\ast$ and set $$q^\dag=(b\cup\{\la \s, v^\ast\ra\}, \D_a\cup\D_q\cup\{(N, \a)\})$$
\[quecosa\] Starting from $\sigma \notin supp(q)$, we can also run the same argument and find a condition $q'$ extending $q$ and such that $\sigma \in supp(q)$.
This is true since $(2)_\sigma$ guarantees that $q\av_\s$ is also $(M^\ast,\, \mtcl P_\s)$–generic for all $M \in\D^{-1}_q(\s+1) \cap \mtcl M_{\s+1}$ (which implies that the above $t$ forces that all these $M$ are in $\dot D$). Also note that, by its being definable, the first component $\n$ of the weakest condition of $\Phi(\s)$ is such that $\n < \d_M = \d_{M^\ast[\dot G_\s]}$ for all these $M$’s.
Now let us prove $(2)_\a$. Let $A$ be a maximal antichain of $\mtcl P_\a$ in $N^\ast$, and assume without loss of generality that $q=(p, \D_q)$ extends some condition $q^\ast$ in $A$. We must show $q^\ast\in N$. Note that $A$ is in $N$ by the $\al_2$–c.c. of $\mtcl P_\a$. Let us work in $V^{\mtcl P_\s\restr(q\av_\s)}$. Let $E$ be the set of $\Phi(\s)$–conditions $v$ such that either
*\[(i)\] there exists some $a =(b, \D_a)\in \mtcl P_\a$ extending some member of $A$ such that $a\av_\s\in \dot G_\s$, $\s\in dom(b)$, and such that $b(\s)=v$, or else*
*\[(ii)\] there is no $a=(b, \D_a)\in \mtcl P_\a$ extending any member of $A$ such that $a\av_\s\in \dot G_\s$, $\s\in dom(b)$, and such that $b(\s)\leq_{\Phi(\s)} v$.*
$E$ is a dense subset of $\Phi(\s)$, and $E\in N^\ast[\dot G_\s]$ since $N^\ast[\dot G_\s]\elsub H(\k^+)^V[\dot G_\s]$ and $N^\ast$ contains $\mtcl P_\a$. Note that $E$ is in fact in $N[\dot G_\s]$ by Corollary \[cor00\]. Suppose $\s\in dom(p)$ and suppose $p(\s)=\ov v$ (the case $\s\notin dom(p)$ is slightly simpler). Since $\ov v$ is $(N[\dot G_\s],\,\Phi(\s))$–generic, we may find some $v'\in E\cap N[\dot G_\s]$ and some $v^\ast$ $\Phi(\s)$–extending both $v'$ and $\ov v$.
\[cl\] Condition (i) above holds for $v'$.
Let $r=(s, \D_r)$ be a condition in $\dot G_\s$ extending $q\av_\s$ and deciding $v^\ast$, and let $u = (s\cup\{\la\s, v^\ast\ra\}, \D_r\cup \D_q)$. Note that $u$ is a $\mtcl P_\a$–condition extending $q$. In particular, $u$ extends a condition in $A$, and therefore it witnesses the negation of condition (ii) for $\n'$, so condition (i) must hold for $\n'$.
By the above claim and by $N^\ast[G_\s]\elsub H(\k^+)^V[G_\s]$ there is $a=(b, \D_a)$ in $N^\ast[G_\s]$ witnessing that condition (i) holds for $v'$, and actually $a\in N$ since $N^\ast[G_\s]\cap V = N^\ast$ by induction hypothesis. Now we extend $q\av_\s$ to a condition $r=(s, \D_r)$ deciding $a$, and deciding also some common extension $v^\ast\in\Phi(\s)$ of $\ov v$ and $v'$. We may also assume that $r$ extends $a\av_\s$. Now it is straightforward to verify, by the usual arguments, that $(s\cup\{\la\s, v^\ast\ra\}, \D_r\cup \D_a\cup\D_q)$ is a $\mtcl P_\a$–condition extending $a$ and $q$. It follows that $q^\ast=a$.
It remains to prove the lemma for the case when $\a$ is a nonzero limit ordinal. The proof of $(1)_\a$ is easy. Let $\s\in N\cap\a$ be above $supp(q)$. By induction hypothesis we may find $r\in\mtcl P_\s$ extending $q\av_\s$ and such that $(N, \s)\in\D_r$. Check that the result of stretching the marker $\s$ in $(N, \s)$ up to $\a$ if $\a < \k$ and up to $sup(N\cap\k)$ if $\a = \k$ is a condition in $\mtcl P_\a$ extending $q$ with the desired property.
For $(2)_\a$, let $A$ be a maximal antichain of $\mtcl P_\a$ in $N^\ast$, and assume without loss of generality that $q=(p, \D_q)$ extends some condition $q^\ast$ in $A$. We must show $q^\ast\in N$. Suppose first that $cf(\a)=\o$. In this case we may take $\s\in N^\ast\cap\a$ above $supp(q)$. Let $G_\s$ be $\mtcl P_\s$–generic with $q\av_\s\in G_\s$. In $N^\ast[G_\s]$ it is true that there is a condition $q^\circ\in\mtcl P_\a$ such that
*\[(a)\] $q^\circ\in A$ and $q^\circ\av_\s\in G_\s$, and*
*\[(b)\] $supp(q^\circ)\sub \s$.*
(the existence of such a $q^\circ$ is witnessed in $V[G_\s]$ by $q^\ast$.)
Since $q\av_\s$ is $(N^\ast, \mtcl P_\s)$–generic by induction hypothesis, $q^\circ\in N^\ast$. By extending $q$ below $\s$ if necessary, we may assume that $q\av_\s$ decides $q^\circ$ and extends $q^\circ\av_\s$. But now, if $q=(p, \D_q)$, the natural amalgamation $(p, \D_q\cup \D_{q^\circ})$ of $q$ and $q^\circ$ is a $\mtcl P_\a$–condition extending them. It follows that $q^\ast = q^\circ$.
Finally, suppose $cf(\a)\geq\o_1$. This will be the only place where we use the symmetry of $dom(\D)$ for every $\mtcl P_\a$ condition $(p', \D)$. Notice that if $N'\in dom(\D_q)$ and $\d_{N'}<\d_N$, then $sup(N'\cap N\cap \a)\leq sup(\Psi_{\ov N, N}(N')\cap\a)\in N\cap\a$[^18] whenever $\ov N\in dom(\D_q)$ is such that $\d_{\ov N}=\d_N$ and $N'\in \ov N$. Hence we may fix $\s\in N\cap \a$ above $supp(q)\cap N$ and above $sup(N'\cap N\cap \a)$ for all $N'\in dom(\D_q)$ with $\d_{N'}<\d_N$.
As in the above case, if $G_\s$ is $\mtcl P_\s$–generic with $q\av_\s\in G_\s$, then in $N^\ast[G_\s]$ we can find a condition $q^\circ=(p^\circ, \Delta_{\circ}) \in\mtcl P_\a$ such that $q^\circ\in A$ and $q^\circ\av_\s\in G_\s$ (again, the existence of such a condition is witnessed in $V[G_\s]$ by $q$), and such a $q^\circ$ will necessarily be in $N^\ast$. By extending $q$ below $\s$ we may assume that $q\av_\s$ decides $q^\circ$ and extends $q^\circ\av_\s$. The proof of $(2)_\a$ in this case will be finished if we can show that there is a condition $q^\dag$ extending $q$ and $q^\circ$. The condition $q^\dag$ can be built by recursion on $supp(q)\cup supp(q^\circ)$. This finite construction mimics the proof of $(1)_\a$ for successor $\a$. Note for instance that if $\eta$ is in the support of $q^\circ$ and $\s \leq \eta < \alpha$, then $p^{\circ}(\eta)=v$ is an ordered pair whose first component is a countable ordinal less than $\delta_N$. Such an $\eta$ satisfies that if $\eta \in N'$ for some $N'\in dom(\D_{q})$, then $\d_{N'}\geq\d_{N}$. Hence, there must exist and ordered pair $v^\ast$ and a common extension of $q\av_\eta$ and $q^\circ\av_\eta$ forcing that $v^{\ast}$ $\Phi(\eta)$–extends $v$ and is $(N'[\dot G_\eta],\,\Phi(\eta))$–generic for all relevant $N'$. The reason is that there is, in $V^{\mtcl P_\eta\restr (q^\dag\av_\eta)}$, a club $\dot D$ witnessing the $H(\k^+)^V$–regularity of $\Phi(\eta)$ relative to $\dot G_\eta$, and such that every relevant $N'$ is in $\dot D$. This $\dot D$ can be taken to be the first club, in the well–order of $H(\k^+)[\dot G_\eta]$ induced by $\Psi$, witnessing the $H(\k^+)^V$–regularity of $\Phi(\eta)$ relative to $\dot G_\eta$ (it is clear that, since all relevant $N'$ contain $\Psi$, they contain a name for $\dot D$, and therefore are in $\dot D$ by $(2)_\eta$). This finishes the proof of $(2)_\a$ for limit $\a$ and the proof of the lemma.
For all $\a\leq\k$, $\mtcl P_\a$ is proper.
The following two lemmas are trivial.
\[genh\] For every $\a<\k$ and every condition $q\in\mtcl P_\k$, $q$ forces that the collection of all $v$ such that there is some $(p, \D)\in\dot G_\k$ with $p(\a)= v$ generates a $V[\dot G_\a]$–generic filter on $\Phi(\a)$.
See Remark \[quecosa\].
\[v-proper\] $\mtcl P_\k$ forces that every regular poset $\mtcl R\sub H(\k)$ is $H(\k^+)^V$–regular relative to $\dot G_\k$.
Lemma \[bound\] follows from the usual counting of nice names for subsets of $\k$ using $(\k^{<\k})^V=\k$ and Lemma \[cc\].
\[bound\] $\mtcl P_\k$ forces $\k^{<\k}=\k$.
Lemma \[fa\] will make use of the following result.
\[refl\] Let $Q$ be an elementary substructure of $H(\t)$, for some large enough $\t$, and suppose $Q$ is closed under $\o_1$–sequences and contains $\Psi$ and $\vec X$. Suppose $Q\cap\k$ is an ordinal $\d$ in $\k$. Then for every $\mtcl P_\d$–condition $q$ there is a $\mtcl P_\d$–condition $q^\ast\in Q$ such that every condition $\leq_\d$–extending $q$ is compatible with $q^\ast$ and every condition $\leq_\d$–extending $q^\ast$ is compatible with $q$.
Suppose $q=(p, \D_q)$ and $\D_q=\{(N_i, \g_i)\,:\,i<n\}$. For all $i$ let $\tld\g_i=\g_i$ if $\g_i<\d$, and let $\tld\g_i=sup(N_i\cap\d)$ if $\g_i=\d$. Note that $p$ is in $Q$ since $X_\a\in Q$ for each $\a<\d$. Since $Q\elsub H(\t)$ contains all reals and $\{(N_i\cap Q, \tld\g_i)\,:\,i<n\}\in Q$, we may find in $Q$ a set $\{M_i\,:\,i<n\}$ with the property that for all $i \in n$ and $\alpha \in N_i \cap Q$, $\tld\g_i \cap N_i= \tld\g_i\cap M_i$, $M_i\cap N_i = N_i\cap Q$, $M_i \in \mtcl M_{\alpha+1}$ iff $N_i \in \mtcl M_{\alpha+1}$, and there is an isomorphism $\varphi_i:(M_i, \in)\into (N_i, \in)$ fixing $N_i\cap Q$, and such that $q^\ast = (p, \{(M_i, \tld\g_i)\,:\,i\in n\})$ is in $\mtcl P_\k$.
Let $q'=(p', \D')$ be a condition in $\mtcl P_\d$ extending $q$. We want to show that $\tld q = (p', \D'\cup\D_{q^\ast})$ is a $\mtcl P_\d$–condition (the proof that every condition in $\mtcl P_\d$ extending $q^\ast$ is compatible with $q$ is similar). We prove by induction on $\a\leq\d$ that $\tld q\av_\a$ is a condition in $\mtcl P_\a$. Let $\a<\d$ and suppose $\tld q\av_\a\in\mtcl P_\a$. It suffices to show that if $\a\in dom(p')$, then $\tld q\av_\a$ forces that $p'(\a)$ is $(M_i[\dot G_\a],\,\Phi(\a))$–generic for every $i<n$ such that $\tld\g_i>\a$ and such that $M_i\in\mtcl M_{\a+1}$.
Note that $N_i\in\mtcl M_{\a+1}$ and that $\g_i>\a$. Hence, $\tld q\av_\a$ forces that $p'(\a)$ is $(N_i[\dot G_\a],\,\Phi(\a))$–generic. Work now in $V^{\mtcl P_\a\restr (\tld q\av_\a)}$ and suppose towards a contradiction that there is a condition $y\leq_{\Phi(\a)} p'(\a)$ and a maximal antichain $A$ of $\Phi(\a)$ in $M_i[\dot G_\a]$ such that no condition in $A\cap M_i[\dot G_\a]$ is compatible with $y$. Let $\dot A\in M_i$ be a $\mtcl P_\a$–name for $A$. Then $\varphi_i(\dot A)\in N_i$ is a $\mtcl P_\a$–name for a maximal antichain of $\Phi(\a)$ (note that both $\mtcl P_\a$ and $\Phi(\a)$ are fixed by the isomorphism $\varphi_i$ since these objects are in $N_i\cap Q$) and, by the $\al_2$–c.c. of $\mtcl P_\a$ together with the $\al_2$–c.c.of $\Phi(\a)$ in $V^{\mtcl P_\a}$ and the closure of $Q$ under $\o_1$–sequences, $\varphi_i(\dot A)\in Q$. It follows that $\varphi_i(\dot A)=\dot A$. The rest of the argument is as in the proof of Fact \[factproper2\], using the fact that there is a surjection from $\o_1$ onto $\dot A$ in $N_i$ fixed by $\varphi_i^{-1}$ and the fact that $N_i[\dot G_\a]\cap\o_1 = N_i\cap\o_1= M_i\cap\o_1=M_i[\dot G_\a]\cap\o_1$ is forced by $\tld q\av_\a$ thanks to Lemma \[horribilis\].
Given $Q$ and $q$ as in the hypothesis of Lemma \[refl\], we will say that the condition $q^\ast$ given by its conclusion is a *projection of $q$ to $Q$*.
\[fa\] $\mtcl P_\k$ forces $\textsc{MA}(\Upsilon)_{<\kappa}$.
Let $q$ be a $\mtcl P_\k$–condition, let $\chi < \k$, and let $\dot{\mtcl R}$ and $\dot A_i$ ($i<\chi$) be $\mtcl P_\k$–names such that $q$ forces that $\dot{\mtcl R}$ is an $\al_2$–c.c. $H(\k^+)^V$–regular poset relative to $\dot G_\k$ defined on $\k$ and that each $\dot A_i$ is a maximal antichain of $\dot{\mtcl R}$. By Lemma \[bound\], Fact \[factproper2\], and Lemma \[v-proper\] it suffices to show that there is some condition extending $q$ and forcing that there is a filter on $\dot{\mtcl R}$ intersecting all $\dot A_i$. Let $X$ be a subset of $\k$ coding the $\mtcl P_\k$–name $\dot{\mtcl R}$ via our fixed translating function $\p$.
Now, using the fact that $\vec X$ is a $\diamondsuit(\{\a<\k\,:\,cf(\a)\geq\o_2 \})$–sequence we may fix an elementary substructure $Q$ of some large enough $H(\t)$ containing $q$, $\mtcl P_\k$, $\dot R$, $(\dot A_i)_{i\in\chi}$, $\vec X$ and $X$, closed under $\o_1$–sequences, and such that $\d= Q\cap\k$ is an ordinal such that $X_\d=X\cap\d$ (since $\m^{\al_1} < \k$ for all $\m < \k$, the set of $\d\in\k$ for which there is a $Q$ as above contains a $\l$–club for every regular cardinal $\l<\k$, $\l\geq\o_2$). Furthermore we may assume that $q$ forces for all $\x$, $\x'$ in $\d$ that if $\p(\x)$ and $\p(\x')$ are compatible conditions in $\dot{\mtcl R}$, then there is an ordinal below $\d$ coding a common extension in $\dot{\mtcl R}$ of $\p(\x)$ and $\p(\x')$.
The following claim follows from the closure of $Q$ under $\o_1$–sequences together with the above choice of $\d$.
Letting $\dot{\mtcl R}_0$ be the $\mtcl P_\d$–name coded by $X_\d$, $\mtcl P_\d\restr q$ forces that $\dot{\mtcl R}_0$ is an $H(\k^+)^V$–regular poset relative to $\dot G_\d$ with the $\al_2$–c.c.
The $\al_2$–c.c. of $\dot{\mtcl R}_0$ in $V^{\mtcl P_\d\restr q}$ follows from Lemma \[refl\] together with the fact that $q$ forces for all $\x$, $\x'$ in $\d$ that if $\x$ and $\x'$ code compatible conditions $\dot c$, $\dot c'$ in $\dot{\mtcl R}$, then there is an ordinal below $\d$ coding a common extension in $\dot{\mtcl R}$ of $\dot c$ and $\dot c'$. In fact, since $Q$ is closed under $\o_1$–sequences, $q$ forces $\dot{\mtcl R}_0$ to be a complete suborder of $\dot{\mtcl R}$. The proof of the $H(\k^+)^V$–regularity of $\dot{\mtcl R}_0$ in $V^{\mtcl P_\d\restr q}$ relative to $\dot G_\d$ is essentially as in the proof of Lemma \[refl\].
It follows from the above claim that $q$ forces $\Phi(\d)=\dot{\mtcl R}_0$. Finally, we may extend $q$ to a condition $q'$ such that $\d\in supp(q')$. Then, by Lemma \[genh\], $q'$ forces that there is a filter $H$ on $\dot{\mtcl R}_0$ meeting all $\dot A_i$, and of course $H$ generates a filter on $\dot{\mtcl R}$.
\[kappa\] $\mtcl P_{\k}$ forces $2^{\al_0}=\k$.
$V^{\mtcl P_\k}\models 2^{\al_0}\geq\k$ follows for example from the fact that $\mtcl P_\k$ forces $\textsc{MA}(\Upsilon)_{<\kappa}$. $V^{\mtcl P_\k}\models 2^{\al_0}\leq\k$ follows from Lemma \[bound\].
Lemma \[kappa\] finishes the proof of Theorem \[mainthm\].
D. Asperó, *Generic absoluteness for $\Sigma_1$ formulas and the continuum problem*, in Logic Colloquium 2002, Lecture Notes in Logic, vol. 27, Association for Symbolic Logic, Wellesley, Massachusetts (2006), 1–27.
D. Asperó and M. A. Mota, *Forcing consequences of $\textsc{PFA}$ together with the continuum large*, submitted (2010). Available at [arXiv:1203.1235v1 \[math.LO\]]{}
D. Asperó and M. A. Mota, *Measuring club–sequences with a large continuum*, submitted (2012). Available at [arXiv:1203.1238v1 \[math.LO\]]{}
J. Bagaria, *Bounded forcing axioms as principles of generic absoluteness*, Archive for Mathematical Logic, vol. 39, pp. 393–401, 2000.
D. H. Fremlin, *Consequences of Martin’s Axiom*, Cambridge Tracts in Mathematics, no. 84, Cambridge (1984).
M. Foreman and P. B. Larson, *Small posets and the continuum*, unpublished note (2004).
J. D. Hamkins and T. A. Johnstone, *The proper and semi-proper forcing axioms for forcing notions that preserve $\al_2$ or $\al_3$*, Proc. American Mathematical Society, 137 (5), pp. 1823–1833, 2009.
T. Jech, *Set Theory: The Third Millenium Edition, Revised and Expanded*, Springer, Berlin (2002).
K. Kunen, *Set Theory, An introduction to independence proofs*, North-Holland Publ. Co., Amsterdam (1980).
D. Martin and R. M. Solovay, *Internal Cohen extensions*, Annals of Mathematics, vol. 2, pp. 143–178, 1970.
T. Miyamoto, *A note on weak segments of $\textsc{PFA}$*. In *Proceedings of the Sixth Asian Logic Conference (Beijing, 1996)*, pp. 175–197, World Sci. Publ., River Edge, NJ, 1998.
J. T. Moore, *Set Mapping Reflection*, Journal of Mathematical Logic, vol. 5, 1, pp. 87–98, 2005.
J. T. Moore, *Aronszajn lines and the club filter*, Journal of Symbolic Logic, vol. 73, 3, pp. 1029–1035, 2008.
I. Neeman, *Hierarchies of forcing axioms, II*, Journal of Symbolic Logic, 73, pp. 522–542, 2008.
I. Neeman and E. Schimmerling, *Hierarchies of forcing axioms, I*, Journal of Symbolic Logic, 73, pp. 343–362, 2008.
R. M. Solovay and S. Tennenbaum, *Iterated Cohen extensions and Souslin’s problem*, Annals of Mathematics, 94, pp. 201–245, 1971.
S. Shelah, *Proper and improper forcing*, Springer, Berlin (1998).
[^1]: A partial order has the countable chain condition if and only if it has no uncountable antichains. More generally, given a cardinal $\k$, a partial oder has the $\k$–chain condition if it has no antichains of size $\k$.
[^2]: The same is true for the Solovay-Tenennbaum construction, i.e., the same construction shows the consistency of Martin’s Axiom together with $2^{\al_0}$ being $\al_2$, $\al_{728}$, $\al_{\o_2 + \o + 4}$, and so on.
[^3]: This poset $\mtbb P$ has size $\al_1$ and therefore has the $\al_2$–c.c. On the other hand, the forcing axiom for collections of $\al_1$–many dense subsets of $\mtbb P$ is obviously false.
[^4]: Note that we are not assuming that $(\nu, X) \in \bigcap\{N_i\, :\, i < m\}$.
[^5]: Given a cardinal $\m$, a partial order $\mtbb P$ is $\m$–Knaster if for every $X\in[\mtbb P]^\m$ there is $Y\in [X]^\m$ consisting of pairwise compatible conditions.
[^6]: In the sense that $\mtcl P_\b$ is a regular extension of $\mtcl P_\a$ whenever $\a<\b\leq\k$. It follows of course that $\la\mtcl P_\a\,:\,\a\leq\k\ra$ is forcing–equivalent to a forcing iteration $\la\mtbb P_\a\,:\,\a\leq\k\ra$ in the ordinary sense (that is, such that $\mtbb P_{\a+1}\cong\mtbb P_\a\ast\dot{\mtbb Q}_\a$ for all $\a$, where $\dot{\mtbb Q}_\a$ is a $\mtbb P_\a$–name for a poset), but such a presentation of $\la\mtcl P_\a\,:\,\a\leq\k\ra$ is not really ‘natural’.
[^7]: This $\k$ is exactly the value that $2^{\al_0}$ attains at the end of the construction.
[^8]: For more on the motivation of this type of construction see [@ASP].
[^9]: $\textsc{BPFA}$ is the assertion that $H(\o_2)^V$ is a $\S_1$–elementary substructure of $H(\o_2)^{V^{\mtbb P}}$ for every proper poset $\mtbb P$ ([@Bagaria]).
[^10]: Usually the focus has been on deriving $2^{\al_0}=\al_2$ from bounded forms of forcing axioms, that is forms of forcing axioms in which one considers only small maximal antichains but where the posets are allowed to have large antichains as well.
[^11]: On the other hand, $\textsc{PFA}(\o_1)$ is trivially equivalent to the foxing axiom for the class of semi-proper posets of size $\al_1$.
[^12]: The existence of a non–proper poset of size $\al_1$ preserving stationary subsets of $\o_1$ is consistent. In fact, Hiroshi Sakai has recently constructed such a poset assuming a suitably strong version of $\diamondsuit_{\o_1}$ holding in $\textsc{L}$ and which can always be forced.
[^13]: In particular, $\mtcl Q\in M_i$.
[^14]: $supp(q)$ stands for the *support* of $q$.
[^15]: Note that every such club $D$ is in $H(\k^+)^V$.
[^16]: $\p$ can be taken to be for example the following surjection $\p:\mtcl P(\k)\into H(\k^+)$: if $a\in H(\k^+)$, then $\p(X) = a$ if and only if $X\sub\k$ codes a structure $(\k', E)$ isomorphic to $(TC(\{a\}), \in)$ (for some unique cardinal $\k'\leq\k$).
[^17]: We can find such a club in $N^\ast[\dot G_{\s}]$ since $\Psi\in N^\ast$, $N^\ast[\dot G_{s}]\elsub H((2^\k)^+)[\dot G_{\s}]$, and $\mtcl P^V(H(\k)^V)\in N^\ast[\dot G_{\s}]$.
[^18]: Recall that $\Psi_{\ov N, N}$ fixes $\ov N \cap N \cap \kappa$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Spatially-sparse predictors are good models for brain decoding: they give accurate predictions and their weight maps are interpretable as they focus on a small number of regions. However, the state of the art, based on total variation or graph-net, is computationally costly. Here we introduce sparsity in the local neighborhood of each voxel with social-sparsity, a structured shrinkage operator. We find that, on brain imaging classification problems, social-sparsity performs almost as well as total-variation models and better than graph-net, for a fraction of the computational cost. It also very clearly outlines predictive regions. We give details of the model and the algorithm.'
author:
-
bibliography:
- 'biblio.bib'
title: 'Social-sparsity brain decoders: faster spatial sparsity'
---
brain decoding, sparsity, spatial regularization
Introduction: Spatial sparsity in decoding
==========================================
Machine learning can predict behavior or phenotype from brain images. Across subjects, it can give indications on a pathology or its progression, for instance capturing atrophy patterns of Alzheimer-related cognitive decline from anatomical imaging [@fan2008spatial]. To study brain function, it has been used extensively to predict some cognitive parameters associated with the presented stimuli from functional brain images such as functional Magnetic Resonance Imaging (fMRI) [@haxby2001]. In these applications, the number of samples is small (hundreds or less), while the number of features is typical the number of voxels in the brain, 50000 or more. The estimation is ill-posed, as there are much more parameters to estimate than the number of samples, which calls for regularization. In brain decoding, good prediction is important, but also retrieving and understanding the aspects of brain images that drive this prediction. Linear models, as linear support vector machines (SVM), are often used because they work well in the low-sample regime. An additional benefit of these models is that their weights form brain maps [@mouraomiranda2005]. However, suitable regularization is necessary for these weights to retrieve well the important regions [@varoquaux2012small; @gramfort2013]. Sparse penalties select voxels [@yamashita2008; @carroll2009], but only a subset of the important ones [@varoquaux2012small]. Spatial and sparse penalties, using total variation (TV) [@michel2011] or graph-net [@grosenick2013interpretable], help the decoder capture full regions [@gramfort2013]. TV and its variants are state-of-the-art regularizers for brain images. In addition to decoding [@michel2011; @baldassarre2012; @gramfort2013], they have been used with great success for applications as diverse as MR image reconstruction [@huang2011efficient], super-resolution [@tourbier2015efficient], prediction from anatomical images [@dubois2014predictive], and even regularization of spatial registration [@fiot2014longitudinal]. The main drawback of spatial sparsity as in TV and related penalties is its computational cost.
Here we introduce spatial sparsity based on “social sparsity”, a relaxed structured penalty with a simple closed-form shrinkage operator. We first detail the mathematical underpinnings of the model and then perform decoding experiments that show that the model is almost as accurate in prediction as the TV-based state of the art and much faster.
#### Notations {#notations .unnumbered}
Vectors are written in bold: $\B{x}$. Indexing vectors is written as sub-scripts: $\B{x}_i$. In an iterative scheme, the iteration number is a sub-script in parenthesis: $\B{x}_{(k)}$.
Methods: Introducing social sparsity
====================================
Spatial penalties for neuroimaging
----------------------------------
A sparsity assumption on the decoders maps has been identified early as a means to select voxels relevant for decoding [@yamashita2008; @carroll2009]. Sparsity has been used with great success in statistics, signal processing, and machine learning. Indeed, it combines good properties for prediction and denoising when the ground truth is actually sparse [@hastie2015sparsity]. Sparse models can be estimated by adding to the data-fit term an $\ell_1$ penalty for sparsity, as in the famed Lasso. The Iterative Shrinkage/Thresholding Algorithm (ISTA) is a common algorithm to solve the corresponding minimization problem [@daubechies2004iterative]. It alternates a gradient-descent step for the data-fit energy and, for sparsity, a *soft-thresholding*: a proximal operator which corresponds to the Euclidean projection on some $\ell_1$ ball:$$\ell_1 \text{ prox of } \B{w} \text{, scale }\lambda:
\quad\forall i
\quad\B{w}_i \leftarrow
\B{w}_i \, \biggl(1 -
\frac{\lambda}{|\B{w}_i|}\biggr)^{\!+}$$ where the operation is applied element-wise –for every coordinate $\B{w}_i$ of $\B{w}$– and $(.)^{+}$ is the positive part. This operation is reminiscent of element-wise thresholding: it sets to zero all the entries of $\B{w}$ that are smaller than $\lambda$ in absolute value, and shrinks by $\lambda$ the remaining.
However, on brain images, simple sparse models as with the $\ell_1$ penalty select a subset of the important voxels [@varoquaux2012small; @rish2012; @rondina2014scors]. Indeed as the information in a voxel is very correlated to its neighbors, sparsity focuses on one representative in a local neighborhood [@varoquaux2012small]. Improved penalties add a spatial term to couple neighbors and create structured sparsity. Total-variation (TV) imposes sparsity in the image gradients [@michel2011]. Coupled with sparsity in a TV-$\ell_1$ penalty it recovers very well the predictive brain regions [@baldassarre2012; @gramfort2013]. It is the state of the art for brain decoding in terms of predictive power and of interpretability of the decoder maps. The main drawback of total variation is that it leads to very slow solvers. Indeed, it cannot be formulated in terms of a thresholding-like operator. Its proximal operator is computational costly as it couples all the voxels. Another spatial penalty, related but faster, Graph-net, imposes smoothness, rather than sparsity, on the image gradients [@grosenick2013interpretable]. Because it is a differentiable penalty, it leads to faster optimization, though it imposes less spatial structure.
The challenges of solving TV and related penalties arise from their very strength: they impose sparsity not on the voxels of the image, but on a representation capturing differences between voxels. This concept is an instance of *analysis sparsity*, which leads to other penalties successful for predicting from brain images [@eickenberg2015total]. Overlapping group sparsity is another analysis penalty which has been heavily used in image processing [@mairal2014sparsemodelling]: rather than putting voxels to zeros, it penalizes a full local neighborhood, using a penalty on blocks. As a voxel is in several neighborhoods, these blocks are overlapping. This overlap leads to an optimization problem that has the same structure –and same cost– as that of TV-$\ell_1$.
All these estimators, and related optimization problems, can be solved with an ISTA algorithm, or its accelerated variant the FISTA. In fact, these approaches are the best option in the case of brain decoding [@dohmatob2014benchmarking] although the TV-$\ell_1$ proximal operator involves a costly sub-iteration. In the case of non-overlapping blocks, group sparsity is simply $\ell_{21}$ penalty applied on each group, as in the famed group-lasso. The proximal operator is closed-form and performs a soft-thresholding on each group based on its Euclidean norm: $$\parbox{.18\linewidth}{$\ell_{21}$ prox\\
on group $\mathcal{G}$}:
\quad\forall i \in \mathcal{G}
\quad\B{w}_i \leftarrow
\B{w}_i \, \biggl(1 -
\frac{\lambda}{\sqrt{\sum_{j \in \mathcal{G}} \B{w}_j^2}}\biggr)^{\!+}$$ for a coordinate $\B{w}_i$ in the group $\mathcal{G}$. As with soft-thresholding for the $\ell_1$ penalty, the operation is applied element-wise for all the groups and leads to fast optimizations.
Fast spatial structure with social sparsity
-------------------------------------------
There is a large gap in computational cost between sparsity imposed on separate items, coordinates or groups, and coupling the sparsity of a voxel to that of its neighbor, as in overlapping group sparsity or penalties related to TV. We introduce social sparsity [@kowalski2013social] which is a tradeoff between the two scenarios. It applies a soft-thresholding similar to group-lasso, using the norm of a local neighborhood, but modifies only the coordinate $\B{w}_i$ at the center of this neighborhood. In a sense, social sparsity “forgets” overlaps across neighborhoods. This makes solvers much faster, as we show in the following.
#### The social-sparsity operator
We focus here on the most popular “social-sparse” operator: the windowed group-lasso. For each element $\B{w}_i$, we associate a neighborhood $\mathcal{N}(i)$ of coordinates. The shrinkage operator reads $$\parbox{.16\linewidth}{$S(\B{w}, \lambda)$ social shrinkage}:
\quad\forall i
\quad\B{w}_i \leftarrow
\B{w}_i \, \biggl(1 -
\frac{\lambda}{\sqrt{\sum_{j \in \mathcal{N}(i)} \alpha_j^{i} \B{w}_j^2}}\biggr)^{\!+}
\label{eq:social_shrinkage}$$ where the $\alpha_j^{i}$ are weights representing the shape of the neighborhood (the simplest choice being the rectangular window: all $\alpha_j^{i}$ are equal).
An interesting interpretation of this operator, is that it can be seen as a fast approximation of the group-lasso with overlaps as presented in [@jacob2009group]. Indeed, it can be shown that the social-sparse operator is equivalent to applying a regular group-lasso operator in a high dimensional space where all the variables are duplicated in order to form independent groups (there are as many groups as variables). Then the result is projected back in the original space following an oblique projection [@kowalski2013social]. In practice, similar performances are obtained without the cost encountered by the group-lasso with overlaps.
#### Choice of the neighborhood
Coupling neighboring voxels is crucial, as demonstrated by the success of spatial penalties in brain imaging. However the spatial extent of that coupling should be small in brain imaging. Indeed, the typical scale used to smooth fMRI data is of 6mm, for 3mm voxels. Similarly, anatomical images, with voxels of 1mm, are often smoothed with a kernel of 2mm. We use as a neighborhood $\mathcal{N}(i)$ of a voxel $i$ its 6 immediate neighbors. To let the behavior of a group be driven more by its central voxel, we set the relative weights $\alpha$ of the 6 other voxels to .7.
#### A FISTA solver
The social-sparsity shrinkage $S$ is not the proximal operator of a known penalty. Yet, it has been shown to yield good estimations in proximal optimization schemes [@kowalski2013social]. We use a FISTA, an accelerated variant of ISTA, as for TV-$\ell_1$ [@dohmatob2014benchmarking; @eickenberg2015total]. The brain-imaging data fit appears via a loss $L: \B{w}\in \mathbb{R}^p \rightarrow
\mathbb{R}$, the gradient of which should be Lipschitz-continuous. Typically, the logistic loss is used for classification problems and the squared loss for regression. For completeness we detail the scheme in Algorithm \[algo:fista\].
$\B{v}_{(1)}\gets \B{w}_{(0)},\qquad$ $k\gets 1,\qquad$ $t_{(0)} \gets 0$
#### Parameter selection
We set the regularization parameter $\lambda$ by nested cross-validation, using the same strategies as can be used in Graph-net or TV-$\ell_1$ to speed up computation [@dohmatob2015speeding]. We do 8 folds. For each fold, we scan the $\lambda$ parameter from large values to small values, with warm start of the solver. As the model is similar to an $\ell_1$ penalty but with additional constraints, values of $\lambda$ that lead pure $\ell_1$ models to be fully sparse will also give fully sparse social-sparsity models. Hence we start our path at $\lambda_\text{max}$, the largest $\lambda$ giving non-empty $\ell_1$ model[^1]. We visit 5 values on a logarithmic scale from $\lambda_\text{max}$ to $\frac{1}{20} \lambda_\text{max}$. As in Graph-net or TV-$\ell_1$ solvers [@dohmatob2015speeding], we do early stopping of the optimizer on the left-out prediction error during parameter selection. The final coefficients are the average of the coefficient for the optimal $\lambda$ on each of the 8 folds.
Finally, as it can be done with graph-net and TV-$\ell_1$, we use univariate feature screening, retaining 20% of the features. The motivation for this screening is that sparse models are highly likely to put the corresponding features to zero [@dohmatob2015speeding].
Experiments: Accuracy and run time
==================================
#### Datasets {#datasets .unnumbered}
We study social sparsity on publicly-available datasets for two applications: intra-subject brain-decoding from fMRI, and inter-subject prediction from voxel-based morphometry (VBM).
For fMRI brain decoding, we use a standard visual-object recognition dataset [@haxby2001]. We perform on all 5 subjects intra-subject 2-class decoding of 14 pairs of stimuli of varying difficulties (listed on Figure \[fig:times\_and\_scores\]). To measure prediction accuracy, we perform cross-validation, leaving out 2 of the 12 sessions.
For VBM inter-subject prediction, we use the OASIS anatomical imaging dataset [@marcus2007]. We predict gender and use a cross-validation of random splits of 20% of the subjects.
#### Experimental settings {#experimental-settings .unnumbered}
On all classification tasks, we fit a model with a logistic loss and measure prediction accuracy as well as wall time in ten iterations of cross-validation.
We compare social sparsity to graph-net and TV-$\ell_1$, as implemented in the Nilearn library, as well as the most commonly-used decoder, a linear SVM with 20% univariate feature selection. To fully replicate common practice we use the default $C=1$ for the SVM, rather than cross-validation. All models, including social-sparsity, are implement in Python, using scikit-learn for the SVM [@abraham2014machine]. The graph-net and TV-$\ell_1$ implementation use the same feature-selection and path strategies as our social-sparsity solver to speed up computation. An important detail for computation time is the stopping criteria of the algorithms. We use the same for Graph-net, TV-$\ell_1$, and social sparsity: a $10^{-4}$ cutoff on the relative maximum change on the weights $\B{w}$ (see Alg. \[algo:fista\]).
Results: Striking a good tradeoff
=================================
![**Prediction accuracy and computation time** for different classification tasks from brain images: 14 on the Haxby dataset, and gender prediction on OASIS. Values are displayed relative to the mean over 4 classifiers: TV-$\ell_1$, graph-net, social sparsity, and a linear SVM with 20% univariate feature selection. Each subject of the Haxby study gives one data point. \[fig:times\_and\_scores\]](times_and_scores.pdf){width="\linewidth"}
Experiments outline a tradeoff between prediction accuracy and computation time. Fig. \[fig:times\_and\_scores\] displays the relative prediction accuracy and run time. TV-$\ell_1$ predicts best on average over the various classification tasks. However, it is followed closely by social sparsity which outperforms graph-net[^2]. The SVM performs much worse than the spatial sparsity, aside from the VBM data where we find that all models perform similarly.
In terms of run time, we find that graph-net is on average 4 times faster than TV-$\ell_1$, but social sparsity is 3 times faster than graph-net. The SVM-based decoder is 20 times faster than social sparsity, *ie* 240 times faster than TV-$\ell_1$.
Finally, an important aspect of the brain decoders is whether they segment well the brain regions that support the decoding. Such a question is hard to validate, yet there is evidence that TV-$\ell_1$ is a good approach [@gramfort2013]. Fig. \[fig:maps\] displays the decoder maps for the object-recognition tasks. For these tasks, we expect prediction to be driven by the functional areas of the visual cortex [@grill2004human]. Indeed, the maps outline regions in known visual areas. The graph-net maps are much more scattered and less structured than the others. Conversely, the social sparsity maps are sparser and outline a smaller number of clusters.
Conclusion: Be social
=====================
Brain decoders benefit strongly from spatial sparsity that helps them narrow on regions important for prediction. Total variation is a powerful and principled solution, but it comes with a hefty computational cost. Social sparsity can be used to introduce penalties on local neighborhoods in the image. It is more heuristic, as it does not minimize a known convex cost. We find empirically that it performs very slightly less well that TV-$\ell_1$ in terms of prediction accuracy, but is more than ten times faster. The corresponding decoder maps are more sparse and focus on a smaller number of regions than TV-$\ell_1$. Social sparsity outperforms graph-net, the faster contender to TV-$\ell_1$, on speed, accuracy, and interpretability. We have found that it strikes a very interesting balance for brain decoding: much faster and almost as good for prediction as TV-$\ell_1$. A full social-sparsity model-fit with hyper-parameter selection takes only 20 times longer than a simple SVM with default hyper-parameters. With social-sparsity, spatially-structured brain decoders are fast: typically 30s with parameter selection on a 2GHz i7 CPU.
#### Acknowledgment {#acknowledgment .unnumbered}
[OASIS was supported by grants P50 AG05681, P01 AG03991, R01 AG021910, P50 MH071616, U24 RR021382, R01 MH56584. The authors acknowledge funding from the EU FP7/2007-2013 under grant agreement 604102 (HBP). ]{}
[^1]: $\lambda_\text{max} = \|\B{X}^\mathsf{T} \B{y}\|_\infty$ for lasso, and $\lambda_\text{max} = \frac{1}{n} \|\B{X}^\mathsf{T} \tilde{\B{y}}\|_\infty$ for $\ell_1$ logistic, where $\tilde{\B{y}}$ is the weighted output vector: $\tilde{\B{y}}_i = \frac{n^+}{n}$ for samples in the positive class, and $\tilde{\B{y}}_i = \frac{n^-}{n}$ for samples in the negative class, with $n$ the number of samples, $n^+$ (resp. $n^-$) the number in the positive (resp. negative) class.
[^2]: All differences are significant in a Wilcoxon rank test.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present an `A`daptively `B`iased `M`olecular `D`ynamics (`ABMD`) method for the computation of the free energy surface of a reaction coordinate using non-equilibrium dynamics. The `ABMD` method belongs to the general category of umbrella sampling methods with an evolving biasing potential, and is inspired by the metadynamics method. The `ABMD` method has several useful features, including a small number of control parameters, and an $O(t)$ numerical cost with molecular dynamics time $t$. The `ABMD` method naturally allows for extensions based on *multiple walkers* and *replica exchange*, where different replicas can have different temperatures and/or collective variables. This is beneficial not only in terms of the speed and accuracy of a calculation, but also in terms of the amount of useful information that may be obtained from a given simulation. The workings of the `ABMD` method are illustrated via a study of the folding of the peptide in a gaseous and solvated environment.'
author:
- Volodymyr Babin
- Christopher Roland
- Celeste Sagui
date: 'January 16, 2008'
title: Adaptively Biased Molecular Dynamics for Free Energy Calculations
---
\[sec:introduction\] Introduction
==================================
When investigating the equilibrium properties of a complex polyatomic system, it is customary to identify a suitable *reaction coordinate* $\sigma(\rr_1,\dots,\rr_N):\mathbb{R}^{3N}\mapsto\mathbb{Q}$ that maps atomic positions $\rr_1,\dots,\rr_N$ onto the points of some manifold $\mathbb{Q}$, and then to study its equilibrium probability density: $$\nonumber
p(\xi) = \big<\delta\left[\xi - \sigma(\rr_1,\dots,\rr_N)\right]\big>,\;
\xi\in\mathbb{Q}$$ (angular brackets denote an ensemble average). The density $p(\xi)$ provides information about the relative stability of states corresponding to different values of $\xi$ along with useful insights into the transitional kinetics between various stable states. In practice, the Landau free energy[@Frenkel] $$\nonumber
f(\xi) = -k_BT\ln p(\xi),$$ is typically preferred over $p(\xi)$, because it tends to be more intuitive. Either $p(\xi)$ or $f(\xi)$ is said to provide a coarse-grained description of the system — in terms of $\xi$ alone — with the rest of the degrees of freedom of the original system integrated out. Quite naturally, the reaction coordinate (often also referred to as *collective variable* or *order parameter*) is typically chosen to represent the slowest degrees of freedom of the original system, although this is not formally required.
In the past few years, several methods targeting the computation of $f(\xi)$ using non-equilibrium dynamics have become popular. First methods that introduced a time evolving potential to bias the original potential energy were the the `L`ocal `E`levation `M`ethod (`LEM`)[@Huber_T_94], by Huber, Torda and van Gunsteren in the `MD` context and the Wang-Landau approach in `MC` one[@Wang_F_01]. More recent approaches include the adaptive-force bias method[@Darve_E_01], and the non-equilibrium metadynamics[@Laio_A_02; @Iannuzzi_M_03] method. These methods all estimate the free energy of the reaction coordinate from an “evolving” ensemble of realizations[@Lelievre_T_2007; @Bussi_G_06], and use that estimate to bias the system dynamics, so as to flatten the effective free energy surface. Collectively, they can all be considered as umbrella sampling methods, with an evolving potential. In the long time limit, the biasing force is expected to compensate for the free energy gradient, so that the biasing potential eventually reproduces the free energy surface.
In this work, we present an `A`daptively `B`iased `M`olecular `D`ynamics (`ABMD`) method whose implementation is particularly efficient and suited for free energy calculations. The method has an $O(t)$ scaling with molecular dynamics time t and is characterized by only two control parameters. In addition, the method allows for extensions based on [*multiple walkers*]{} and [*replica exchange*]{} for both temperature and/or the collective variables. The `ABMD` method has been implemented in the `AMBER` software package[@Amber9], and is to be distributed freely.
Before discussing `ABMD`, it is helpful to review the salient features of the metadynamics (`MTD`) method. Essentially, the `MTD` method is built upon the `LEM` method by exploiting Car-Parrinello dynamics: the phase space of the system is extended to include additional dynamical degrees of freedom harmonically coupled to the collective variable. These additional degrees of freedom are assumed to have masses associated with them, and evolve in time according to Newton’s laws. The masses are supposed to be large enough, so that the dynamics of these extra-variables is driven by the free energy gradient. Their trajectory is then used to construct a history-dependent biasing potential by means of placing many small Gaussians along the trajectory. When combined with Car-Parrinello *ab-initio* dynamics, `MTD` has been successfully used to explore complex reaction pathways involving several energy barriers.
While `MTD` continues to be used successfully, there are several known limitations associated with the initial implementation of the method, which provided the motivation for the development of the `ABMD` method. First, in order to calculate reliable free energies with a *controllable accuracy*, long runs are needed, especially for the “corrective” follow-up at equilibrium[@Babin_V_2006]. This is especially true for biomolecular systems, which typically are characterized by many degrees of freedom and non-negligible entropy contributions to the free energies. Long runs, however, may be precluded by the `MTD` method, because of its unfavorable scaling with molecular dynamics `MD` time $t$. While one can readily speed up the original `MTD` method using such tricks as truncated Gaussians and kd-trees[@Babin_V_2006], the bottleneck there is the explicit calculation of the history-dependent potential. Since at every `MD` step, Gaussians from all previous time steps need to be added, the number of Gaussians grows linearly with $t$. The numerical cost of `MTD` therefore grows as $O(t^2)$ which, in some cases, may prove itself to be prohibitively expensive, especially when long runs are needed. Another undesirable feature is that the `MTD` method (at least in its original implementation) is characterized by a relatively large number of parameters ([*e.g.*]{}, the masses and spring constants associated with the collective variable, the characteristics of the Gaussians to be added, multiple control parameters, etc), all of which influence the dynamics in an entangled and non-transparent way. A successful `MTD` run often requires a careful balancing of these parameters, which is especially nontrivial for multidimensional collective variables. More recent implementations of `MTD` [@Laio_A_05] have reduced the number of parameters. As will be discussed, the `ABMD` method is characterized by only [*two*]{} control parameters and scales as $O(t)$ with simulation time.
\[sec:method\] The Adaptively Biased Molecular Dynamics Method.
================================================================
The `ABMD` method is formulated in terms of the following set of equations: $$\nonumber
m_a\frac{\mathrm{d}^2\rr_a}{\mathrm{d}t^2} = \mathbf{F}_a
- \frac{\partial}{\partial\rr_a}
U\big[t|\sigma\left(\rr_1,\dots,\rr_N\right)\big],$$ $$\nonumber
\frac{\partial U(t|\xi)}{\partial t} = \frac{k_B T}{\tau_F}
G\big[\xi - \sigma\left(\rr_1,\dots,\rr_N\right)\big],$$ where the first set represents Newton’s equations that govern ordinary `MD` (temperature and pressure regulation terms are not shown) augmented with the additional force coming from the time-dependent biasing potential $U(t|\xi)$ (with $U(t=0|\xi) = 0$), whose time evolution is given by the second equation. In the following, we refer to $\tau_F$ as *flooding timescale* and to $G(\xi)$ as *kernel* (in analogy to the *kernel density estimator* widely used in statistics[@Silverman_Density_Estimation]). The kernel is supposed to be positive definite ($G(\xi)>0$) and symmetric ($G(-\xi)=G(\xi)$). It can be perceived as a smoothed Dirac delta function. For large enough $\tau_F$ and small enough width of the kernel, the biasing potential $U(t|\xi)$ converges towards $-f(\xi)$ as $t\to\infty$[@Lelievre_T_2007; @Bussi_G_06].
Our numerical implementation of the `ABMD` method involves the following. We stick with $\mathbb{Q} = \mathbb{Q}_1\times\dots\times\mathbb{Q}_D$ where $\mathbb{Q}_k$ is either $[a,b]\in\mathbb{R}^1$ or a one-dimensional torus, and use cubic B-splines (or products of thereof for $D>1$) to discretize $U(t|\xi)$ in $\mathbb{Q}$: $$\nonumber
U(t|\xi) = \sum\limits_{m\in\mathbb{Z}^D} U_m(t)B(\xi/\Delta\xi - m),$$ $$\nonumber
B(\xi) = \left\{
\begin{array}{ll}
(2 - |\xi|)^3/6, & 1\leq |\xi| < 2, \\
\xi^2(|\xi| - 2)/2 + 2/3, & 0\leq |\xi| < 1, \\
0, & \mathrm{otherwise}.
\end{array}
\right.$$ We use the biweight kernel[@Silverman_Density_Estimation] for $G(\xi)$: $$\nonumber
G(\xi) = \frac{48}{41}\left\{
\begin{array}{ll}
\left(1 - \left.\xi^2\right/4\right)^2, & -2 \leq \xi \leq 2 \\
0, & \mathrm{otherwise}
\end{array}
\right.,$$ and an Euler-like discretization scheme for the time evolution of the biasing potential: $$\nonumber
U_m(t + \Delta t) = U_m(t)
+ \Delta t\frac{k_BT}{\tau_F}G\big[\sigma/\Delta\xi - m\big],$$ where $\sigma = \sigma(\rr_1,\dots,\rr_N)$ is at time $t$. Note that this time discretization may be readily improved. This, however, is not really important here, since the goal is not to recover the solution of the `ABMD` equations *per se*, but rather to flatten $U(t|\xi) + f(\xi)$ in the $t\to\infty$ limit. Note also, that the numerical cost of evaluation of the time-dependent potential is constant over time, and so `ABMD` scales trivially as $O(t)$, which is computationally quite favorable. The storage requirements of the `ABMD` are also quite reasonable, especially if sparse arrays are used for $U_m$. In addition, it is characterized by only [*two*]{} control parameters: the flooding timescale $\tau_F$ and the kernel width $4\Delta\xi$.
`ABMD` admits two important extensions. The first is identical in spirit to the *multiple walkers metadynamics*[@Lelievre_T_2007; @Raiteri_P_2006]. It amounts to carrying out several different `MD` simulations biased by the same $U(t|\xi)$, which evolves via: $$\nonumber
\frac{\partial U(t|\xi)}{\partial t} = \frac{k_B T}{\tau_F}
\sum\limits_{\alpha}%
G\big[\xi - \sigma\left(\rr_1^{\alpha},\dots,\rr_N^\alpha\right)\big]\;,$$ where $\alpha$ labels different `MD` trajectories. A second extension is to gather several different `MD` trajectories, each bearing its own biasing potential and, if desired, its own distinct collective variable, into a generalized ensemble for “replica exchange” with modified “exchange” rules[@Sugita_Y_2000; @PTMetaD; @BiasExchange]. Both extensions are advantageous and lead to a more uniform flattening of $U(t|\xi) + f(\xi)$ in $\mathbb{Q}$. This enhanced convergence to $f(\xi)$ is due to the improved sampling of the “evolving” canonical distribution.
![\[fig:1\]The peptide in a $\beta$-hairpin conformation (sketch). ](fig1){width="\linewidth"}
We have implemented the `ABMD` method in the `AMBER` package[@Amber9], with support for both replica exchange and multiple-walkers. In pure “parallel tempering” replica exchange (same collective variable in all replicas), $N_r$ replicas are simulated at different temperatures $T_n$, $n=1,\dots,N_r$. Each replica has its own biasing potential $U^n(t|\xi)$, $n=1,\dots,N_r$, that evolves according to its dynamical equation. Exchanges between neighboring replicas are attempted at a prescribed rate, with an exchange probability given by[@Sugita_Y_2000]: $$\label{eq:exchange1}
w(m|n) = \left\{%
\begin{array}{ll}
1, & \Delta\leq 0,\\
\exp (-\Delta), & \Delta > 0,
\end{array}
\right.$$ $$\begin{aligned}
\label{eq:exchange2}
\Delta &=& \left(\frac{1}{k_BT_n} - \frac{1}{k_BT_m}\right)
\Big(E_p^m - E_p^n\Big)\\
\nonumber
&+& \frac{1}{k_BT_m}\Big[U^m(\xi^n) - U^m(\xi^m)\Big]\\
\nonumber
&-& \frac{1}{k_BT_n}\Big[U^n(\xi^n) - U^n(\xi^m)\Big],\end{aligned}$$ where $E_p$ denotes the atomic potential energy. The biasing potentials are temperature-bound and converge in the $t\to\infty$ limit to the free energies at their respective temperatures.
We have also implemented a more *general* replica exchange scheme, where different replicas can have different collective variables and/or temperatures, and can experience either an evolving or a static biasing potential (the latter obviously includes the case of $U^n(t|\xi)=0$). Exchanges between random pairs of replicas are then tried at a prescribed rate. This method is simply a generalization [@BiasExchange] of the “Hamiltonian replica exchange” method described in Ref., and reduces to it when all biasing potentials are static. The big advantage here is that, by using replicas with different collective variables, it is possible to obtain several *one-* or *two-*dimensional projections of the free energy surface for the corresponding variables. This is very useful because it not only increases the amount of information that can be gathered from a given simulation, but it also allows for previously obtained information for a collective variable to be used to compute the free energy associated with different variables. For instance, suppose that in the course of a simulation it becomes apparent that one wishes to address additional questions involving different collective variables. Instead of starting from “scratch”, one can re-use the already obtained biasing potentials and thereby greatly accelerate the free energy calculation for the new variables. It is also worth noting that for replicas running at the same temperature, the exchange probability does not depend on the atomic potential energies (Eqs. (\[eq:exchange1\])-(\[eq:exchange2\]) above). This implies that the number of replicas needed to maintain acceptable exchange rates can be made independent of the solvent degrees of freedom, provided that one is interested in the properties of the solute only (so that the collective variables do not depend explicitly on the atomic coordinates of the solvent) and that the structure is adequately solvated. This can be exploited to sample a solute with a minimum amount of solvent, and to accelerate the averaging over the solvent degrees of freedom. These last two applications of the *general* replica exchange method are illustrated in the next section.
![\[fig:2\]RMS error of the free energy over $3.3\text{\AA}<R_g<6.3\text{\AA}$ at $T=300K$ for the `ABMD` (blue) and reference `MTD` (red) simulations. ](fig2){width="\linewidth"}
\[sec:output\] Case study: a short peptide.
============================================
To illustrate the method, we have simulated the hydrophobic peptide (sketched in Fig.\[fig:1\]) in the gas phase and in a solvated environment, using cyclohexane as the explicit solvent. The free energy of this peptide at $T=300\,\mathrm{K}$ in gas phase has previously been investigated with the `MTD` method[@Babin_V_2006], and is characterized by a “double-well” structure (see Fig.\[fig:8\]), with the wells corresponding to the peptide in a “globular” (left minimum in the Fig.\[fig:8\]) and a $\beta$-hairpin (right minimum in the Fig.\[fig:8\]) folded conformation, respectively. While simple enough, the molecule possesses all the typical features of larger peptide systems usually studied with biomolecular simulations. Simulation parameters are as in a previous study[@Babin_V_2006]: the atoms are described by the 1999 version of the Cornell *et al.* force field[@Cornell_W_95], with no cutoff for the non-bonded interactions. The Berendsen thermostat is chosen with $\tau_{tp}=1\mathrm{fs}$ for temperature control. The `MD` time step ($\Delta t$) is 1 fs for the parallel tempering simulations, and 2 fs otherwise.
![\[fig:3\]RMS error of the free energy over $3.3\text{\AA}<R_g<6.3\text{\AA}$ for `ABMD` simulations at $T=300K$ with $\tau_F = 180\;\mathrm{ps}$ (1), $360\;\mathrm{ps}$ (2), $720\;\mathrm{ps}$ (3) and $1440\;\mathrm{ps}$ (4). ](fig3){width="\linewidth"}
The radius of gyration of the heavy atoms was chosen to be the collective variable: $$\label{eq:rg}
R_g = \sum_{a}\frac{m_a}{m_{\Sigma}}\left(\rr_a
- \mathbf{R}_{\Sigma}\right)^2\,.$$ Here, $\mathbf{R}_{\Sigma} = \sum_{a}\left(m_a/m_{\Sigma}\right)\rr_a$ is the center of mass, with $m_{\Sigma} = \sum_{a}m_a$, and the sum runs over all atoms except hydrogen. The initial configuration is the fully unfolded peptide. A reference free energy profile, whose error[@Babin_V_2006] in the region of interest is less than 0.15 kcal/mol, was computed for benchmarking purposes (see Appendix \[ap:reference\] for details). As a measure of the RMS free energy error, the following construction was used: $$\nonumber
E_{RMS} = \sqrt{\frac{1}{b - a}\int\limits_{a}^{b}%
\ud\xi\Big(f_1(\xi) - f_2(\xi) - \Delta\Big)^2},$$ where $$\nonumber
\Delta = \frac{1}{b - a}\int\limits_{a}^{b}%
\ud\xi\Big(f_1(\xi) - f_2(\xi)\Big),$$ accounts for the arbitrary additive constants in the free energies $f_{1,2}(\xi)$. Here $a=3.3$[Å]{} and $b=6.3$[Å]{}, which correspond to the physical region of interest (see Fig.\[fig:8\]).
![\[fig:4\]RMS error of the free energy over 3.3[Å]{}$<R_g<$6.3[Å]{} for multiple walkers `ABMD` simulations with $\tau_F = 1\,\mathrm{ns}$ using one (1), two (2), four (4) and eight (8) trajectories at $T = 300$K. ](fig4){width="\linewidth"}
Figure \[fig:2\] presents the time dependence of the RMS free energy error for the `ABMD` and reference `MTD` run[@Babin_V_2006]. Both simulations have exactly the same kernel width $4\Delta\xi = 0.25\;$[Å]{}, and flooding timescale $\tau_F=90\;\mathrm{ps}$ that corresponds to the *a posteriori* hills acceptance rate reported in Ref.. It is evident that the `AMBD` run is more accurate than the corresponding `MTD` run. The `ABMD` method owes its better convergence to the smoother time evolution of the biasing potential and accurate discretization in $\mathbb{Q}$. The amount of memory used by the `ABMD` simulation to store $U_m$ values was only $\approx 200\times 8$ bytes (considering double precision) for roughly $10^8$ tiny “hills” that were accumulated by the end of the run. The reference `MTD` simulation with merely $5\times 10^3$ Gaussians required roughly 25 times more memory for the biasing potential (with only the positions of the Gaussians stored explicitly). One can expect, that `ABMD` will be even more economical when it comes to dealing with multidimensional collective variables, provided that sparse arrays are used for $U_m$ with only non-zero elements being stored explicitly.
Although an *a priori* error estimate for this type of non-equilibrium simulation is really not feasible, it is expected that the error should decrease for increased $\tau_F$. This point is illustrated in Fig.\[fig:3\], which shows the error for increasing values of $\tau_F$.
In order to decrease the simulation time required for accurate free energy estimates even further, the multiple-walker variation of `ABMD` proves to be useful. For a moderate number of walkers, the speedup is nearly linear, with an additional increase in accuracy coming from the better sampling of the “evolving” canonical distribution (see Fig.\[fig:4\]). Parallel tempering improves both the speed and the accuracy even more. To this end, we first ran `ABMD` with $\tau_F = 90\,\mathrm{ps}$ using 2, 4, 6 and 8 replicas at $T=300\,\mathrm{K}$, $331\,\mathrm{K}$, $365\,\mathrm{K}$, $403\,\mathrm{K}$, $445\,\mathrm{K}$, $492\,\mathrm{K}$, $543\,\mathrm{K}$ and $600\,\mathrm{K}$ (during equilibrium `MD` runs, the peptide configuration jumps between the two minima on a picosecond timescale at $T=600\mathrm{K}$). In all cases, the $E_{RMS}$ was found to be $\sim 1$ kcal/mol, or less as $t\to\infty$ (data not shown). Again, the improvement in accuracy stems from the better sampling of the “evolving” canonical distribution. Then, we ran 8 replicas with smaller values $\tau_F$ and were surprised that the accuracy does not degrade, even for $\tau_F = 11.2\,\mathrm{ps}$ as shown in Fig.\[fig:5\].
![\[fig:5\]RMS error of the free energy over $3.3\text{\AA}<R_g<6.3\text{\AA}$ at $T=300K$ for parallel tempering `ABMD` simulations using eight replicas running at $T=300\,\mathrm{K}$, $331\,\mathrm{K}$, $365\,\mathrm{K}$, $403\,\mathrm{K}$, $445\,\mathrm{K}$, $492\,\mathrm{K}$, $543\,\mathrm{K}$ and $600\,\mathrm{K}$ with $\tau_F = 90\,\mathrm{ps}$ (1), $45\,\mathrm{ps}$ (2), $22.5\,\mathrm{ps}$ (3) and $11.25\,\mathrm{ps}$ (4). ](fig5){width="\linewidth"}
Finally, we turn to aspects related to the *general* replica exchange method and illustrate its potential. As already noted, by using replicas with different collective variables and swapping these at prescribed rates, it is possible to obtain projections of the free energy surface for the corresponding variables. It is also possible to use previously obtained information with respect to one collective variable to compute the free energy associated with a different variable. For example, suppose that instead of the [*one-dimensional*]{} free energy profile as as a function of $R_g$ already discussed, one realizes that what is actually needed is a *two-dimensional* profile that includes information with respect to the number of O-H bonds along the backbone. The two-dimensional free energy map is computationally quite expensive, but the calculation can be greatly accelerated with the help of the general replica exchange method. We therefore simulated $8+1=9$ replicas. The eight replicas were run at the previously stated temperatures, with each replica biased by a *static* (not evolving) biasing potential corresponding to the negated free energy associated with the radius of gyration $R_g$ at the corresponding temperature, as shown in Fig.\[fig:8\]. The additional ninth replica was run at $T=300\,\mathrm{K}$ with `ABMD` flooding in the two collective variables, [*i.e.,*]{} $R_g$ and the number of O-H bonds along the backbone as given by $$\nonumber
N_{\mathrm{OH}} = \sum\limits_{\mathrm{O}, \mathrm{H}}
\frac{1 - \big(\left.r_{\mathrm{OH}}\right/r_0\big)^{6\hphantom{1}}}
{1 - \big(\left.r_{\mathrm{OH}}\right/r_0\big)^{12}}\,,$$ where $r_{\mathrm{OH}}$ is the distance between a pair of hydrogen and oxygen atoms and $r_0=2.5\text{\AA}$. The sum runs over the unique O-H pairs (*i.e.*, each O-H pair is counted only once), with O and H separated by one or more amino-bases along the backbone (27 pairs in total). In other words, we “re-use” the previously computed free energies for $R_g$ to get the free energy in the two-dimensional space $(R_g, N_{\mathrm{OH}})$: the eight first replicas serve as a “sampling enhancement device” for the ninth replica. The calculation is carried out in two stages: a “coarse” stage ($15\,\mathrm{ns}$ with $\tau_F = 10\,\mathrm{ps}$ and $4\Delta R_g = 0.25\text{\AA}$, $4\Delta N_{\mathrm{OH}} = 0.5$) followed by a “fine” stage ($50\,\mathrm{ns}$ with $\tau_F = 100\,\mathrm{ps}$ and $4\Delta R_g = 0.1\text{\AA}$, $4\Delta N_{\mathrm{OH}} = 0.25$). In both runs exchanges between four randomly chosen pairs of replicas were attempted every $100\,\mathrm{fs}$. The final free energy map is shown in Fig.\[fig:6\]. It is clear that this two-dimensional free energy landscape conveys additional information not contained in the one-dimensional free energy plots already discussed. In particular, it allows for a better characterization of the “globular” states of the : specifically, it is apparent from the Fig.\[fig:6\] that there are at least two such states with different values of $N_{\mathrm{OH}}$ (both correspond to the left minimum in Fig.\[fig:8\]). Of course, one could have re-used the information in the one-dimensional $R_g$ profiles to include other collective variables, in addition to $N_{\mathrm{OH}}$.
![\[fig:6\]Free energy map for peptide in the gas phase as a function of the collective variables $R_g$ and $N_{\mathrm{OH}}$. (see Ref. for details regarding the algorithm used to make this plot). ](fig6){width="\linewidth"}
The general replica exchange `ABMD` may also be advantageous for explicit solvent simulations, which are often notoriously lengthy. Specifically, if one is interested in the solute and the collective variables do not depend on the solvent degrees of freedom, then the number of replicas required to maintain an adequate exchange rate depends only very weakly on the amount of solvent (which must of course be sufficient as to adequately solvate the structure), provided that all the replicas are simulated at the same temperature. This is because the exchange probability does not explicitly depend on the potential energy difference when the temperature of the replicas is the same. While not every choice of collective variables for different replicas will lead to decent exchange rates, one can nevertheless take advantage of this property and use general replica exchange to enhance the sampling in a solvated environment.
In order to demonstrate the method in this regime, we simulated the peptide at $T=300\,\mathrm{K}$ solvated by 171 cyclohexane ($\mathrm{C}_6\mathrm{H}_{12}$) molecules (the total number of atoms was 3139) under periodic boundary conditions using the General[@GAFF] `AMBER` Force-Field (`GAFF`) for the solvent. We used a *truncated octahedron* cell of fixed size (constant volume) that corresponds to the equilibrium density at $T=300\,\mathrm{K}$ (the equilibrium density value was obtained from a $10\,\mathrm{ns}$ simulation under constant pressure at $T=300\,\mathrm{K}$). The Particle-Mesh Ewald (`PME`[@Darden_T_93]) method was used for the electrostatic forces, with a $36\times 36\times 36$ `FFT` grid and an $8\,\text{\AA}$ cutoff for the direct sum (same cutoff was used for van der Waals interactions). First, we ran 10 replicas in the “flooding” mode ([*i.e.*]{}, under evolving biasing potentials) using as collective variables the distances $r_{\mathrm{CC}}$ between the backbone carbons separated by at least 2 amino-acids (there are 10 such distances for ). We ran for $5\,\mathrm{ns}$ with $\tau_F=30\,\mathrm{ps}$ and $4\Delta r_{\mathrm{CC}} = 1\text{\AA}$, and then for another $5\,\mathrm{ns}$ with $\tau_F=150\,\mathrm{ps}$ and $4\Delta r_{\mathrm{CC}} = 0.5\text{\AA}$, attempting exchanges between five randomly selected pairs every $0.5\,\mathrm{ps}$. As expected, at the start of the simulation, the exchange rate was nearly 100% decreasing later as the biasing potentials were built up. By the end of the simulation, when all possible values of the distances had been covered, the exchange rate was very disparate between different pairs of replicas. However, for every replica there was at least one other replica such that the exchange rate between them was reasonable ([*i.e.,*]{} the whole simulation did not degenerate into ten different non-interacting trajectories). We then set $\tau_F=\infty$ in these 10 replicas, and added an eleventh replica whose collective variable was chosen as the radius of gyration of the heavy atoms. In other words, as before, we use the ten replicas as a “sampling enhancement device” for the last one. We then ran a two-stage flooding scheme: a $5\,\mathrm{ns}$ coarser stage, with $\tau_F=25\,\mathrm{ps}$ and $4\Delta R_g = 0.25\text{\AA}$; followed by a $10\,\mathrm{ns}$ finer stage, with $\tau_F=180\,\mathrm{ps}$ and $4\Delta R_g = 0.2\text{\AA}$. As before, the exchange attempts between 5 randomly selected pairs of replicas were performed every $0.5\,\mathrm{ps}$. The “raw” `ABMD`-computed free energy associated with $R_g$ after that stage is shown in Fig.\[fig:7\]. In a next step, we ran 64 biased simulations (each comprising of 11 replicas) for $7\,\mathrm{ns}$ (first $2\,\mathrm{ns}$ for equilibration followed by $5\,\mathrm{ns}$ of “production” runs) starting from different initial configurations. We set $\tau_F=\infty$ in all replicas (static biasing potentials) and recorded the values of $R_g$ in the eleventh replica every 10 picoseconds. We then used the log-spline algorithm of Stone[@LogSpline] *at. al.* to estimate the (biased) log-density of the $R_g$ values at *equilibrium*. This led us to the final shape of the free energy curve shown in Fig.\[fig:7\]. Compared to the gas phase, the folded $\beta$-turn in the cyclohexane solvated peptide is clearly favored over the globular structure.
![\[fig:7\]Free energy for peptide solvated in cyclohexane at $T=300\,\mathrm{K}$ as obtained via: coarse (non-equilibrium) replica-exchange `ABMD` (1); finer (non-equilibrium) replica-exchange `ABMD` (2); including the correction coming from equilibrium biased replica exchange (3). ](fig7){width="\linewidth"}
\[sec:outro\] Conclusions and outlook.
=======================================
In summary, we have presented an `ABMD` method that computes the free energy surface of a reaction coordinate using non-equilibrium dynamics. The method belongs to the general category of umbrella sampling methods with an evolving potential, and is characterized by only two control parameters (the flooding timescale and the kernel width) and a favorable $O(t)$ scaling with molecular dynamics time t. This scaling can be very important for large-scale classical `MD` biomolecular simulations when long simulation times are required (see, for example Ref., and references therein).
`ABMD` has also been extended to support multiple walkers and replica exchange. Both variations improve speed and accuracy of the method due to the better sampling of the “evolving” canonical distribution. The replica exchange `ABMD` has been generalized to include different temperatures and/or collective variables, that move under either an evolving or a static biasing potential. Aside from enhancing the sampling, this swapping of replicas has several important practical advantages. Most importantly, it enables one to obtain projections of the free energy surface for [*any number of collective variables*]{} one might wish to investigate. In addition, one can re-use previously obtained results in order to enhance the sampling of new collective variables. It is also possible to exploit the fact that exchange rates at the same temperature are independent of the potential energy to enhance sampling of a solute in a minimum amount of solvent (for collective variables independent of solvent atom coordinates). We have implemented the `ABMD` method in the `AMBER` package[@Amber9], and plan to distribute it freely. Here, we have demonstrated the workings of the `ABMD` method with a study of the folding of the peptide, The application of `ABMD` to more complicated biomolecular systems is reserved for future publications. This research was partly supported by NSF under grants ITR-0121361 and CAREER DMR-0348039. In addition we thank NC State HPC for computational resources.
\[ap:reference\] Reference free energy curve.
==============================================
Here, we provide simulation details with regards to the reference free energy curve. We first ran short (5 ns, eight walkers with $\tau_F=60\mathrm{ps}$ and $4\Delta\xi=0.2$[Å]{}) multiple-walkers `ABMD` at $T=600\,\mathrm{K}$ to reconstruct the global well. This was followed by parallel tempering `ABMD` runs, using the biasing potential obtained from the multiple-walkers simulation as the zero-time value for the biasing potentials at different temperatures. We used eight replicas at $T=300\,\mathrm{K}$, $331\,\mathrm{K}$, $365\,\mathrm{K}$, $403\,\mathrm{K}$, $445\,\mathrm{K}$, $492\,\mathrm{K}$, $543\,\mathrm{K}$ and $600\,\mathrm{K}$ and attempted exchanges every 100 `MD` steps ($0.1\,\mathrm{ps}$). The simulation started with $\tau_F=60\mathrm{ps}$ and $4\Delta\xi=0.2$[Å]{}, and ran for $10^5$ exchanges. We then set $\tau_F$ to $600\mathrm{ps}$, $4\Delta\xi$ to $0.1$[Å]{} and ran for $5\times 10^5$ more exchanges. Finally, this was followed up with $1\times 10^6$ more exchanges with $\tau_F=6\mathrm{ns}$ and $4\Delta\xi = 0.1$[Å]{}.
![\[fig:8\]The accurate free energies of peptide in gas phase as function of $R_g$ at $T=300\,\mathrm{K}$, $331\,\mathrm{K}$, $365\,\mathrm{K}$, $403\,\mathrm{K}$, $445\,\mathrm{K}$, $492\,\mathrm{K}$, $543\,\mathrm{K}$ and $600\,\mathrm{K}$ (from bottom to top). ](fig8){width="\linewidth"}
We then ran a very long ($3\times 10^7$ exchanges, $0.1\,\mathrm{ps}$ between exchanges) biased parallel tempering simulation in the spirit of Ref., in order to get an *a posteriori* error estimate. From the resulting histogram it follows that the error does not exceed $\approx 0.15\,\mathrm{kcal/mol}$ for 3.3[Å]{}$<R_g<$6.3[Å]{}. The RMS error is probably much smaller, since 0.15 corresponds to the absolute non-uniformity of the histogram, [*i.e.*]{}, the maximum error, over 3.3[Å]{}$<R_g<$6.3[Å]{}. The accurate free energy curves as a function of temperature are shown in Fig.\[fig:8\].
[31]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, **, Computational Science Series (, ).
, , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, , , ****, ().
, , , ****, ().
, , , , , , , , , , ****, ().
, , , , , ****, ().
, , , ****, ().
, , , ****, ().
, , , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , , , , ****, ().
, , , ****, ().
, , , , ****, ().
, , , , , ****, ().
, **, Monographs on statistics and applied probability (, ).
, , , , , ****, ().
, , , ****, ().
, , , , ****, ().
, ****, ().
, , , , , , , , , , ****, ().
, ****, ().
, , , , , ****, ().
, , , ****, ().
, , , , ****, ().
, , , , ****, ().
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show that the torus knot topology is inherent in electromagnetic and gravitational radiation by constructing spin-$N$ fields based on this topology from the elementary states of twistor theory. The twistor functions corresponding to the elementary states admit a parameterization in terms of the poloidal and toroidal winding numbers of the torus knots, allowing one to choose the degree of linking or knotting of the associated field configuration. Using the gravito-electromagnetic formalism, we show that the torus knot structure is exhibited in the tendex and vortex lines for the analogous linearized gravitational solutions. We describe the topology of the gravitational fields and its physical interpretation in terms of the tidal and frame drag forces of the gravitational field.'
author:
- 'Amy Thompson[^1]'
- 'Joe Swearngin $^*$'
- Dirk Bouwmeester
title: Linked and Knotted Gravitational Radiation
---
Introduction
============
Knots and links are quite remarkable given that they are as old and ubiquitous as ropes and thread and yet have only relatively recently seen a rigorous formulation within mathematics. The study of knots and links has enjoyed a close relationship with physics since its inception by Gauss [@Przytycki2007history]. Today the application of these topological structures in theoretical physics is more widespread than it has ever been, from fault resistant quantum computing [@Nayak2008], hadron models [@Skyrme1962; @Faddeev1997], topological MHD and fluid mechanics [@Kamchatnov1982; @Thompson2014plasma], classical field theories [@Ranada2002; @Irvine2008; @Swearngin2013], quantum field theory [@Witten1989; @Robertson1989], DNA topology [@Arsuaga2005], to nematic liquid crystals [@Machon2013] just to name a few. In this article we shall focus on the application of an important class of knots, torus knots, to classical electromagnetic and gravitational radiation.
A *hopfion* is a field configuration based on a topology derived from the Hopf fibration. The electromagnetic hopfion (EM hopfion) is a null solution to the source free Maxwell equations such that any two field lines associated to either the electric, magnetic, or Poynting vector fields (EBS fields) are closed and linked exactly once [@Ranada2002]. When an EM hopfion is decomposed onto hyperplanes of constant time there always exists a hyperplane wherein the EBS fields are tangent to the fibers of three orthogonal Hopf fibrations. If one extends the Poynting vector to be a future pointing light-like 4-vector its integral curves comprise a space filling shear-free null geodesic congruence dubbed the Robinson congruence by Roger Penrose.
Ra[ñ]{}ada [@Ranada1989] rediscovered the EM hopfion solution and noted that its topology was invariant under time evolution. The search for generalizations of the hopfion solution led to the introduction of a set of non-null EM solutions based on torus knots [@Trueba2011], but the topology was not preserved during time evolution. Around the same time, the Kerr-Robinson theorem was used to derive the hopfion from the Robinson congruence itself using 2-spinor methods [@Dalhuisen2012]. Inspired by the role of the Robinson congruence in the evolution of the hopfion, the conservation of field line topology was tied to the shear-free property of the Robinson congruence [@irvine2010linked].
Complex Analytic and Twistor Methods
====================================
The fundamental role of the Robinson congruence and its connection to the topology of physical systems leads one to consider its relationship to twistor theory. It has long been known that complex contour integral transforms can be used to find the solution to real PDEs. In 1903, Whittaker used this technique to construct the general solution to Laplace’s equation [@Whittaker1904]. In 1915, Bateman extended this method to give solutions to the vacuum Maxwell equations [@Bateman1915]. Twistor theory was developed by Roger Penrose in the late 1960’s as an extension of the $sl(2,\mathbb{C})$ spinor algebra. From this perspective, the complex analytic structure of Bateman can be related to the geometry of spinor fields on space-time [@Penrose1975], which encode linear and angular momentum and are represented by two $SL(2,\mathbb{C})$ spinors $\pi_{A'}$ and $\omega ^{A}$. The linear momentum is the flagpole of $\pi_{A'}$ so that $p^a = \pi^{A'}\bar{\pi}^{A}$ and the angular momentum bivector is related to $\omega ^{A}$ by $$M^{ab} = i\omega ^{(A}\overline{\pi }^{B)} \epsilon^{A'B'} + c.c.$$ These spinors are combined into a single object $Z^{\alpha }=\left( \omega ^{A},\pi _{A'}\right)$ called a twistor. In this formalism, massless linear relativistic fields are expressed in the form of symmetric spinor fields. In 1969, the general solution to the massless spin-$N$ field equations was given as a complex contour integral transform now called the Penrose transform [@Penrose1969solutions].
Within the twistor framework the solutions of the massless spin-$N$ equations are represented by the Penrose transform of homogeneous twistor functions (see Appendix). The elementary states are a canonical example of such functions whose singularities define Robinson congruences on Minkowski space $\mathbb{M}$. The space-time fields corresponding to the elementary states are finite-energy, and in the null case are everywhere non-singular [@Penrose1987origins]. For integer spin fields, the expansion of a solution over the elementary states in twistor space $\mathbb{T}$ is related to the expansion over spherical harmonics in $\mathbb{M}$ through the Penrose transform [@Grgin1966thesis]. These properties have made the elementary states the topic of many studies [@Penrose1972; @Hughston1979advances; @Hodges1982diagrams], and for many problems it is assumed that considering the elementary states is sufficient to describe any solution [@Eastwood1991density].
While investigating these twistor functions and their connection to field topology in $\mathbb{M}$, we have previously shown that the EM hopfion and the analogous gravitational hopfion are elementary states of twistor theory [@Swearngin2013]. Using the earlier construction for EM fields by Bateman [@Bateman1915], Kedia, *et al.* have shown that the EM hopfion is the simplest case in a set of null EM fields based on torus knots [@Irvine2013]. Here we show that field configurations based on all the torus knots are contained within the elementary states of twistor theory. The Hopf fibration appears as the degenerate case whereby the linked and knotted toroidal structure degenerates down to the linked hopfion configuration. This generalization leads to a construction for spin-$N$ fields based on torus knots. We will focus our analysis on the spin-1 and spin-2 fields, where the topology is physically manifest in the field lines.[^2]
The concept of tendex and vortex lines, gravitational lines of force for a particular observer, was developed by Nichols, *et al.* [@Nichols2011], who were motivated by the desire to understand the non-linear dynamics of curved space-time in a more intuitive, directly physical way than previous approaches. The physical understanding of the electromagnetic field is based upon the decomposition of the Faraday field strength tensor onto hyperplanes of constant time yielding two spacial vector fields interpreted as the electric and magnetic fields. Analogously, the Weyl curvature tensor admits a decomposition onto constant time hyperplanes yielding two spacial tensors called the gravito-electric (GE) and gravito-magnetic (GM) tensors. The integral curves of the eigenvector fields of these tensors are called tendex and vortex lines respectively and represent the gravitational analog of electromagnetic field lines. This method was elucidated through a series of papers where it was applied to I) weak field solutions [@Nichols2011], II) stationary black holes [@Zhang2012], and III) weak perturbations of stationary black holes [@Nichols2012]. This method of GEM decomposition is well-suited to studying linked and knotted fields because, as we have shown previously [@Swearngin2013], the field topology is manifest in the lines of force for both the electromagnetic and analogous gravitational solutions.
Parameterization of the Elementary States
=========================================
We will now relate the torus knot topology to the twistor elementary states to obtain solutions to the EM and gravitational spinor field equations.
Torus knots are closed curves on the surface of a torus which wind an integer number of times about the toroidal direction $n_{t}$ and poloidal direction $n_{p}$ as in Fig. \[fig:core\_lines\], where $n_t$ and $n_p$ are coprime and both greater than one. If $n_t$ and $n_p$ are not coprime, then there are $n_g=gcd(n_t,n_p)$ linked curves, each corresponding to a $(n_t,n_p)\bmod{n_g}$ torus knot. If either $n_t$ or $n_p$ is equal to one, then the knot is trivial with $n_g$ linked curves.
![\[fig:core\_lines\]Torus knots (green) wind $(n_t,n_p)$ times around a torus (purple) in the toroidal and poloidal directions, respectively. Shown here are the cases of a) trefoil (2,3) knot, b) cinquefoil (2,5) knot, c) septafoil (2,7) knot, and d) nonafoil (2,9) knot.](core_lines.png){width="100.00000%"}
Following the twistor program, we represent solutions to the massless spin-$N$ equations in $\mathbb{M}$ as the Penrose transform of functions $f(Z)$ in twistor space [@PenroseSpinors2]. The details of the Penrose transform calculation for these fields are given in the Appendix.
Consider the twistor functions corresponding to the elementary states [@Penrose1972] $$\label{eqn:twsitor_function}
f(Z)= \frac{(\bar C_{\gamma} Z^{\gamma})^{c}(\bar D_{\delta} Z^{\delta})^{d}}{(\bar A_{\alpha} Z^{\alpha})^a(\bar B_{\beta} Z^{\beta})^b}$$ where $(\bar A_{\alpha} Z^{\alpha})$ is the $SU(2,2)$ twistor inner product. Choosing $a=1$ yields null/Type N solutions and we must have $b=2h+1+c+d$ to give the correct homogeneity $\mathpzc{h}=-2h-2$ for a solution with helicity $h$. We will show that the class of generating functions of the form[^3] $$\label{eqn:generating_function}
f(Z)= \frac{(\bar C_{\gamma} Z^{\gamma})^{h(n_{p}-1)}(\bar D_{\delta} Z^{\delta})^{h(n_{t}-1)}}{(\bar A_{\alpha} Z^{\alpha})(\bar B_{\beta} Z^{\beta})^{h(n_{p}+n_{t})+1}},$$ lead to field configurations with a torus knot topology where $n_{p}$ and $n_{t}$ correspond to the poloidal and toroidal winding numbers.
We choose the dual twistors $$\begin{aligned}
\bar{A}_\alpha &= \imath (0,\sqrt{2},0,1)\notag\\
\bar{B}_\beta &= \imath (-\sqrt{2},0,-1,0)\notag\\
\bar{C}_\gamma &= (0,-\sqrt{2},0,1)\notag\\
\bar{D}_\delta &= \imath (-\sqrt{2},0,1,0). \label{eqn:dual_twistors}\end{aligned}$$ $\bar{A}_\alpha$ and $\bar{C}_\gamma$ correspond to Robinson congruences with opposite twist, both with central axes aligned along the $+\hat z$-direction. $\bar{B}_\beta$ and $\bar{D}_\delta$ correspond to Robinson congruences with opposite twist, but in the $-\hat z$-direction. This choice leads to spin-$N$ fields which propagate $+\hat z$-direction with field line configurations that are based on a torus knot structure.
Electromagnetic Torus Knots
===========================
After applying the spin-1 Penrose transform to Eqn. (\[eqn:generating\_function\]), the resulting spinor field is $$\label{spin1solution}
\phi_{A'B'}(x) = \frac{(\mathcal{A}_{C'}\mathcal{C}^{C'})^{n_{p}-1}(\mathcal{A}_{D'}\mathcal{D}^{D'})^{n_{t}-1}}{(\mathcal{A}_{E'}\mathcal{B}^{E'})^{n_{p}+n_{t}+1}}\mathcal{A}_{A'}\mathcal{A}_{B'}$$ Note that the Latin script spinor variables are the spinors associated to the Latin twistor variable. Ergo, $\mathcal{A}_{A'}$ is defined implicitly by $\bar A_{\alpha}Z^{\alpha} = \mathcal{A}_{A'} \pi^{A'}$ (see Appendix). The solution in Eqn. (\[spin1solution\]) satisfies the source-free spinor field equation by construction and yields the field strength spinor $$\begin{aligned}
\nabla^{AA'}\varphi_{A'B'} &= 0, \\
F_{A'B'AB} &= \varphi_{A'B'}\epsilon_{AB} + c.c. \end{aligned}$$
The spin-1 fields are null torus knots with a Poynting vector that is everywhere tangent to a Hopf fibration and propagates in the $\hat z$-direction without deformation. The solutions have the same topology[^4] as the electromagnetic fields in Ref. [@Irvine2013]. The electric and magnetic vector fields each have the following topological structure as shown in Fig. \[fig:knot23surfaces\]. There are $2n_g$ core field lines, where $n_g=gcd(n_t,n_p)$, which are linked (and knotted if $n_t,n_p>1$). Each core line has the same configuration as the corresponding torus knot with $(n_t,n_p)$ shown in Fig. \[fig:core\_lines\]. With the choice for $C$ and $D$ given in Eqn. (\[eqn:dual\_twistors\]), the poloidal and toroidal winding numbers for the EM case are related to the exponents in Eqn. (\[eqn:twsitor\_function\]) by $c=n_p-1$ and $d=n_t-1$. A single core field line is surrounded by nested, toroidal surfaces, each filled by one field line. A second core field line, also surrounded by nested surfaces, is linked with the first so that there are $2n_g$ sets of linked nested surfaces which fill all of space. A complete solution to Maxwell’s equations consists of an electric and a magnetic field orthogonal to each other, both with this field line structure. The (1,1) case corresponds to the electromagnetic hopfion.
At $t=0$ the electric and magnetic fields are tangent to orthogonal torus knots, as shown in Fig. \[fig:knot23\] (first row). The fields will deform under time evolution, but the topology will be conserved since $\vec E \cdot \vec B=0$ [@irvine2010linked; @Arrayas2011exchange].
![\[fig:knot23surfaces\]The field line structure based on (2,3) trefoil knot. a) The core field line is a torus knot (green). b) Each field line except the core lies on the surface of a nested, deformed torus. c) One field line fills a complete surface (red). d) Another field line fills a second surface (blue) linked with the first. The two linked core field lines and the nested surfaces around them fill all of space.](knot23surfaces.png){width="100.00000%"}
![A comparison of the spin-1 (EM) and spin-2 (gravity) trefoil knots at $t=0$. The first row is the EM trefoil knot: **a** the electric field, **b** the magnetic field, and **c** the Poynting vector field. The second row is the gravito-electric trefoil knot: **d** the negative eigenvalue field $\vec e_-$, **e** the positive eigenvalue field $\vec e_+$, and **f** the zero eigenvalue field $\vec e_0$.The third row is the gravito-magnetic trefoil knot: **g** the negative eigenvalue field $\vec b_-$, **h** the positive eigenvalue field $\vec b_+$, and **i** the zero eigenvalue field $\vec b_0$. The color scale indicates magnitude of the eigenvalue, with lighter colors indicating a higher magnitude.[]{data-label="fig:knot23"}](knot23.png){width="80.00000%"}
Gravito-electromagnetic Torus Knots
===================================
The spin-2 solutions will be analyzed in terms of the gravito-electromagnetic tidal tensors. The Weyl tensor $C_{abcd}$ can be decomposed into an even-parity “electric” part $E_{ij}$ corresponding to the tidal field and an odd-parity “magnetic” part $B_{ij}$ for the frame-drag field, in direct analogy with the decomposition of the electromagnetic field strength tensor into an electric field and a magnetic field. For an observer at rest, this gives $$\begin{aligned}
E_{ij} &= C_{i0j0} \\
B_{ij} &= - \ast C_{i0j0}.\end{aligned}$$ These tensors are symmetric and traceless, and are thus characterized entirely by their eigenvalues and eigenvectors. One may then study the eigensystem associated with these matrix-valued fields instead of the Weyl tensor itself, allowing for a more intuitive approach to understanding the gravitational field. The integral curves of the eigenvectors of the tidal tensor are called tendex lines and their eigenvalues define the tidal acceleration along these lines. The integral curves of the eigenvectors of the frame-drag tensor are called vortex lines and their eigenvalues define the gyroscope precession about the vortex lines. Together, the tendex and vortex lines are the analog of electromagnetic field lines.[@Nichols2011]
After applying the spin-2 Penrose transform to Eqn. (\[eqn:generating\_function\]), the resulting spinor field is $$\phi_{A'B'C'D'}(x) = \frac{(\mathcal{A}_{F'}\mathcal{C}^{F'})^{2(n_{p}-1)}(\mathcal{A}_{G'}\mathcal{D}^{G'})^{2(n_{t}-1)}}{(\mathcal{A}_{E'}\mathcal{B}^{E'})^{2(n_{p}+n_{t})+1}}\mathcal{A}_{A'}\mathcal{A}_{B'}\mathcal{A}_{C'}\mathcal{A}_{D'}.$$ The source-free field equation and Weyl field strength spinor are $$\begin{aligned}
\nabla^{AA'}\varphi_{A'B'C'D'} &= 0, \\
C_{A'B'C'D'ABCD} &= \varphi_{A'B'C'D'}\epsilon_{AB}\epsilon_{CD} + c.c. \end{aligned}$$
The Weyl tensor can then be decomposed into the GEM components. For Type N, the eigenvalues for both the GE and GM tensors take the form $\{ -\Lambda, 0, +\Lambda \}$, with $-\Lambda(x) \leq 0 \leq +\Lambda(x)$ for all points $x$ in space-time The magnitude of the eigenvalues is $$\label{eqn:eigenvalue_Lambda}
\Lambda=\frac{2^{2n_p-3}(1 + r^2 + t^2 -2tz)^2(r^2 - z^2)^{n_p - 1}(r^4 - 2 r^2 (1 + t^2) + (1 + t^2)^2 + 4 z^2)^{n_t - 1}}{(r^4 - 2 r^2 (-1 + t^2) + (1 + t^2)^2)^{\frac{5}{2} + n_t + n_p}}$$
We label the eigenvectors $\{\vec{e}_-, \vec{e}_0, \vec{e}_+\}$ and $\{\vec{b}_-, \vec{b}_0, \vec{b}_+\}$ corresponding to the eigenvalues for the tidal and frame-drag fields respectively. For the zero eigenvalue, the eigenvectors $\vec{e}_0$ and $\vec{b}_0$ are both aligned with the Poynting vector of the null EM torus knots. For the remaining eigenvectors, we can construct Riemann-Silberstein (RS) vectors $\vec{f}_e = \vec{e}_- + i\vec{e}_+$ and $\vec{f}_b = \vec{b}_- + i\vec{b}_+$ which are related to each other by $$\label{eqn:RS_relation}
\vec{f}_e = e^{i\pi/4} \vec{f}_b.$$ At $t=0$, the eigenvectors of the GE fields have precisely the same structure as the EM fields, and the GM eigenvector fields have the same structure but rotated by $45^\circ$. For the spin-$2$ case, the poloidal and toroidal winding numbers are related to the exponents in Eqn. (\[eqn:twsitor\_function\]) by $c=2(n_p-1)$ and $d=2(n_t-1)$. The surfaces of the $\vec{e}_-$ eigenvector, color-scaled according to the magnitude of the eigenvalue, for different values of $(n_t,n_p)$ are shown in Fig. \[fig:Eminus\_t0\]. The other GEM fields can be constructed by rotating $\vec{e}_-$ according to Eqn. (\[eqn:RS\_relation\]): $\vec{e}_+$ is found by rotating $\vec{e}_-$ by $90^\circ$ about the Poynting vector. $\vec{b}_-$ and $\vec{b}_+$ are found by rotating $\vec{e}_-$ and $\vec{e}_+$ by $45^\circ$, respectively. The eigenvalues of the GEM fields for a given $(n_t,n_p)$ have the same magnitude (color-scaling) given by $|\Lambda(x)|$ in Eqn. (\[eqn:eigenvalue\_Lambda\]).
![The eigenvector field $\vec e_-$ for the gravitational field based on the a) trefoil (2,3) knot, b) cinquefoil (2,5) knot, c) septafoil (2,7) knot, and d) nonafoil (2,9) knot. The color scaling is the same as in Fig. \[fig:knot23\].[]{data-label="fig:Eminus_t0"}](Eminus_t0.png){width="100.00000%"}
Conclusion
==========
Here we have shown that the null EM torus knot solutions correspond to a class of elementary states characterized the poloidal and toroidal winding numbers of the associated field configuration. Using the relationship between fields of different spin in the twistor formalism, we constructed the analogous gravitational radiation configuration that possesses tendex and vortex lines based on a torus knot structure. Since the topology is manifest in the tendex and vortex lines, the gravito-electromagnetic tidal tensor decomposition is a straightforward method for characterizing these field configurations.
The elementary states were known as early as the 1970’s [@Penrose1972], however the explicit forms of their associated spinor and tensor representations on $\mathbb{M}$ were never published.[^5] The modern rediscovery of these solutions has raised interest in obtaining a more complete physical understanding of the topological properties of these fields. The parameterization of the twistor functions corresponding to the elementary states in terms of the poloidal and toroidal winding indicates that the torus knot structure is indeed inherent in the elementary states. For both electromagnetism and gravity, the topology appears in the configuration of the lines of force.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors would like to thank J.W. Dalhuisen for discussions and A. Wickes for help with the figures. This work is supported by NWO VICI 680-47-604 and NSF Award PHY-1206118.
Appendix: The Penrose Transform for Spin-$N$ Torus Knots {#sec:transform .unnumbered}
========================================================
To obtain a spinor field $\varphi _{A_{1}'\cdots A_{2h}'}\left( x\right)$ with helicity $h$ which satisfies the spin-$N$ massless field equation $$\nabla ^{AA_{1}'}\varphi _{A_{1}'\cdots A_{2h}'}\left( x\right) = 0$$ we will calculate the Penrose transform $$\label{penroseTransform}
\varphi_{A'_1\cdots A'_{2h}}(x) = \frac{1}{2\pi i} \oint_\Gamma\pi_{A'_1}\cdots\pi_{A'_{2h}}f(Z)\pi_{B'}d\pi^{B'}$$ where $\Gamma$ is a contour on the Celestial sphere of $x$ that separates the poles of $f(Z)$. Consider the twistor function given by Eqn. (\[eqn:generating\_function\]) $$f(Z)= \frac{(\bar C_{\gamma} Z^{\gamma})^{h(n_{p}-1)}(\bar D_{\delta} Z^{\delta})^{h(n_{t}-1)}}{(\bar A_{\alpha} Z^{\alpha})(\bar B_{\beta} Z^{\beta})^{h(n_{p}+n_{t})+1}}. \notag$$ (For further review of the background material on twistors and the Penrose transform see Ref. [@Swearngin2013]). Let $\overline{A}_\alpha = (\mu_A,\lambda^{A'})$ be a dual twistor such that $$\begin{aligned}
\overline{A}_\alpha Z^\alpha &= i\mu_A x^{AA'} \pi_{A'} + \lambda^{A'} \pi_{A'} \notag \\
&\equiv \mathcal{A}^{A'} \pi_{A'}\end{aligned}$$ where $Z^\alpha = (ix^{AA'} \pi_{A'}, \pi_{A'})$. Similar relations hold for the other dual twistors $\overline{B}_\beta Z^\beta \equiv \mathcal{B}^{B'} \pi_{B'}$, $\overline{C}_\gamma Z^\gamma \equiv \mathcal{C}^{C'} \pi_{C'}$, and $\overline{D}_\delta Z^\delta \equiv \mathcal{D}^{D'} \pi_{D'}$. We want to write the Penrose transform as an integral over the $\mathbb{CP}^1$ coordinate $\zeta = \pi_{1'} / \pi_{0'}$, so we have $$\begin{aligned}
\pi_{C'} d\pi^{C'} &= \pi_{C'} d\pi_{D'} \epsilon^{D'C'} \notag \\
&= \pi_{0'} d\pi_{1'} - \pi_{1'} d\pi_{0'} \notag \\
&= (\pi_{0'})^2 d(\frac{\pi_{1'}}{\pi_{0'}}).\end{aligned}$$ Adopting the canonical spin bases $\{o_{A'}, \iota_{A'}\}$ we have that $$\begin{aligned}
\pi_{A'} &= \pi_{0'} o_{A'} + \pi_{1'} \iota_{A'} \notag \\
&= \pi_{0'} (o_{A'} + (\frac{\pi_{1'}}{\pi_{0'}}) \iota_{A'}).\end{aligned}$$ Observing that $$\begin{aligned}
\frac{1}{\pi_{0'}} \mathcal{A}^{A'} \pi_{A'} &= \mathcal{A}^{0'} + \mathcal{A}^{1'} (\frac{\pi_{1'}}{\pi_{0'}})\end{aligned}$$ and similarly for $\mathcal{B}$, $\mathcal{C}$, and $\mathcal{D}$, we see that the Penrose transform becomes an integral manifestly over $\mathbb{CP}^1$. Thus $$\begin{aligned}
\label{eqn:penroseTransformResult}
\varphi_{A'_1\cdots A'_{2h}}(x) &= \frac{1}{2\pi i} \oint_\Gamma f(Z)\pi_{A'_1} \cdots \pi_{A'_{2h}} \pi_{B'}d\pi^{B'} \\
&= \frac{1}{2\pi i} \oint_\Gamma \frac {(\mathcal{C}^{0'}+\mathcal{C}^{1'}(\frac{\pi_{1'}}{\pi_{0'}}))^{h(n_p-1)}(\mathcal{D}^{0'}+\mathcal{D}^{1'}(\frac{\pi_{1'}}{\pi_{0'}}))^{h(n_t-1)} } {(\mathcal{A}^{0'}+\mathcal{A}^{1'}(\frac{\pi_{1'}}{\pi_{0'}}))(\mathcal{B}^{0'}+\mathcal{B}^{1'}(\frac{\pi_{1'}}{\pi_{0'}}))^{h(n_p+n_t)+1} } (o_{A_1}+(\frac{\pi_{1'}}{\pi_{0'}}) \iota_{A'_1}) \cdots (o_{A_{2h}}+(\frac{\pi_{1'}}{\pi_{0'}}) \iota_{A'_{2h}}) d(\frac{\pi_{1'}}{\pi_{0'}}) \notag \\
&= \frac{(\mathcal{C}^{1'})^{h(n_p-1)}(\mathcal{D}^{1'})^{h(n_t-1)}}{2\pi i \mathcal{A}^{1'} (\mathcal{B}^{1'})^{h(n_p+n_t)+1}} \oint_\Gamma \frac{(\rho+\zeta)^{h(n_p-1)}(\tau+\zeta)^{h(n_t-1)}} {(\mu+\zeta)(\nu+\zeta)^{h(n_p+n_t)+1}} (o_{A_1}+\zeta\iota_{A'_1})\cdots(o_{A_{2h}}+\zeta\iota_{A'_{2h}}) d\zeta \notag\end{aligned}$$ where $\mu = \mathcal{A}^{0'} / \mathcal{A}^{1'}$, $\nu = \mathcal{B}^{0'} / \mathcal{B}^{1'}$, $\rho = \mathcal{C}^{0'} / \mathcal{C}^{1'}$, and $\tau = \mathcal{D}^{0'} / \mathcal{D}^{1'}$.
The above change of variables leaves the contour integral in a form that is straightforward to calculate. The contour $\Gamma$ is taken to enclose the pole $-\mu$ giving the result $$\begin{aligned}
\varphi_{A'_1 \cdots A'_{2h}}(x) &= \frac{(\mathcal{C}^{1'})^{h(n_p-1)}(\mathcal{D}^{1'})^{h(n_t-1)}}{\mathcal{A}^{1'} (\mathcal{B}^{1'})^{h(n_p+n_t)+1}} \underset{\zeta=-\mu}{\text{Res}} \frac{(\rho+\zeta)^{h(n_p-1)}(\tau+\zeta)^{h(n_t-1)} } {(\mu+\zeta)(\nu+\zeta)^{h(n_p+n_t)+1}} (o_{A_1}+\zeta\iota_{A'_1})\cdots(o_{A_{2h}}+\zeta\iota_{A'_{2h}}) \notag \\
&= \frac{(\mathcal{C}^{1'})^{h(n_p-1)}(\mathcal{D}^{1'})^{h(n_t-1)}}{\mathcal{A}^{1'} (\mathcal{B}^{1'})^{h(n_p+n_t)+1}} \frac{(\rho -\mu)^{h(n_p-1)}(\tau-\mu)^{h(n_t-1)}} {(\nu-\mu)^{h(n_p+n_t)+1}} (o_{A'_1}-\mu\iota_{A'_1})\cdots(o_{A'_{2h}}-\mu\iota_{A'_{2h}}) \notag \\
&= \frac{(\mathcal{A}^{1'}\mathcal{C}^{0'} - \mathcal{A}^{0'}\mathcal{C}^{1'})^{h(n_p-1)}(\mathcal{A}^{1'}\mathcal{D}^{0'} - \mathcal{A}^{0'}\mathcal{D}^{1'})^{h(n_t-1)}} {(\mathcal{A}^{1'}\mathcal{B}^{0'} - \mathcal{A}^{0'}\mathcal{B}^{1'})^{h(n_p+n_t)+1}} (\mathcal{A}^{1'} o_{A'_1} - \mathcal{A}^{0'} \iota_{A'_1}) \cdots (\mathcal{A}^{1'} o_{A'_{2h}} - \mathcal{A}^{0'} \iota_{A'_{2h}}) \notag \\
&= \frac{(\epsilon_{C'D'}\mathcal{A}^{C'}\mathcal{C}^{D'})^{h(n_p-1)}(\epsilon_{E'F'}\mathcal{A}^{E'}\mathcal{D}^{F'})^{h(n_t-1)}}{(\epsilon_{A'B'}\mathcal{A}^{A'}\mathcal{B}^{B'})^{h(n_p+n_t)+1}}\mathcal{A}_{A'_1} \cdots \mathcal{A}_{A'_{2h}} \notag \\
&=
\frac{(\mathcal{A}_{C'}\mathcal{C}^{C'})^{h(n_p-1)}(\mathcal{A}_{D'}\mathcal{D}^{D'})^{h(n_t-1)}}{(\mathcal{A}_{B'}\mathcal{B}^{B'})^{h(n_p+n_t)+1}}\mathcal{A}_{A'_1} \cdots \mathcal{A}_{A'_{2h}}
\label{eqn:knotField}.\end{aligned}$$ In the case of spin-1 and spin-2, the classical field strength spinors are given by $$\begin{aligned}
F_{A_{1}'A_{2}'\ A_{1}A_{2}} &=&\varphi _{A_{1}'A_{2}'}\epsilon _{A_{1}A_{2}} + \overline{\varphi} _{A_{1}A_{2}}\epsilon _{A_{1}'A_{2}'} \notag \\
C_{A_{1}'\cdots A_{4}'\ A_{1}\cdots A_{4}} &=& \varphi
_{A_{1}'\cdots A_{4}'}\ \epsilon _{A_{1}A_{2}}\ \epsilon
_{A_{3}A_{4}}+\overline{\varphi}_{A_{1}\cdots A_{4}}\ \epsilon _{A_{1}'A_{2}'}\ \epsilon_{A_{3}'A_{4}'}.
\label{classicalFieldStrength}\end{aligned}$$ Choosing the standard basis to be the extended Pauli matrices we define the following symbols, referred to as the *Infeld-van der Waerden symbols*
$$\begin{array}{cccc}
\sigma _{0}^{AA'}\equiv \frac{1}{\sqrt{2}}
\begin{pmatrix}
1 & 0 \\
0 & 1
\end{pmatrix}
& \sigma _{1}^{AA'}\equiv \frac{1}{\sqrt{2}}
\begin{pmatrix}
0 & 1 \\
1 & 0
\end{pmatrix}
& \sigma _{2}^{AA'}\equiv \frac{1}{\sqrt{2}}
\begin{pmatrix}
0 & i \\
-i & 0
\end{pmatrix}
& \sigma _{3}^{AA'}\equiv \frac{1}{\sqrt{2}}
\begin{pmatrix}
1 & 0 \\
0 & -1
\end{pmatrix}
.
\end{array}
\label{eqn:Pauli_matrices}$$
The spinor fields are then related to the world tensor description by $$\begin{aligned}
F_{ab} &=& F_{A_{1}'A_{2}'\ A_{1}A_{2}}\sigma _{a}^{A_1A_1'}\sigma _{b}^{A_2A_2'} \notag \\
C_{abcd} &=& C_{A_{1}'\cdots A_{4}'\ A_{1}\cdots A_{4}}\sigma _{a}^{A_1A_1'} \cdots \sigma _{d}^{A_4A_4'}.\end{aligned}$$
[10]{}
J. H. Przytycki. Knots: From combinatorics of knot diagrams to combinatorial topology based on knots. 2007, arXiv:math/0703096.
C. Nayak, S. H. Simon, A. Stern, M. Freedman, and S. Das Sarma. Non-abelian anyons and topological quantum computation. , 80(3):1083-1159, 2008.
T. H. R. Skyrme. A unified field theory of mesons and baryons. , 31:556–569, 1962.
L. Faddeev and Antti J. Niemi. Stable knot-like structures in classical field theory. , 387:58–61, 1997.
A. M. Kamchatnov. Topological solitons in magnetohydrodynamics. , 82:117–124, 1982.
Amy Thompson, Joe Swearngin, Alexander Wickes, and Dirk Bouwmeester. Constructing a class of topological solitons in magnetohydrodynamics. , 89(4):043104–1–5, 2014.
Antonio F. Rañada and Jos[é]{} L. Trueba. Topological electromagnetism with hidden nonlinearity. , 119:197–253, 2002.
William T. M. Irvine and Dirk Bouwmeester. Linked and knotted beams of light. , 4:716–720, 2008.
Joe Swearngin, Amy Thompson, Alexander Wickes, Jan Willem Dalhuisen, and Dirk Bouwmeester. Gravitational hopfions. 2014, arXiv:gr-qc/1302.1431.
Edward Witten. Quantum field theory and the Jones polynomial. , 121(3):351–399, 1989.
G. D. Robertson. Torus knots are rigid string instantons. , 226(3-4):244–250, 1989.
Javier Arsuaga, Mariel Vazquez, Paul McGuirk, Sonia Trigueros, De Witt Sumners, and Joaquim Roca. DNA knots reveal a chiral organization of DNA in phage capsids. , 102(26):9165-9169, 2005.
T. Machon and G. Alexander. Knots and nonorientable surfaces in chiral nematics. , 110(35):14174-14179, 2013.
Antonio F. Ra[ñ]{}ada. A topological theory of the electromagnetic field. , 18(2):97–106, 1989.
Manuel Array[á]{}s and Jos[é]{} L. Trueba. Electromagnetic torus knots. 2011, arXiv:hep-th/1106.1122.
J. W. Dalhuisen and D. Bouwmeester. Twistors and electromagnetic knots. , 45(13):135201, 2012.
William T. M. Irvine. Linked and knotted beams of light, conservation of helicity and the flow of null electromagnetic fields. , 43(38):385203, 2010.
E.T. Whittaker. On the partial differential equations of the mathematical physics. , 57:333–355, 1903.
Harry Bateman. . Cambridge University Press, Cambridge, 1915.
Roger Penrose. Twistor theory: its aims and achievements. In C. J. Isham, R. Penrose, and D. W. Sciama, editors, [*Quantum Gravity, an Oxford Symposium*]{}. Oxford University Press, Oxford, 1975.
Roger Penrose. Solutions of the zero-rest-mass equations. , 10(38):38–39, 1969.
Roger Penrose. On the origins of twistor theory. In W. Rindler and A. Trautman, editors, [*Gravitation and geometry, a Volume in Honour of I. Robinson*]{}, pages 341–361. Bibliopolis, Naples, 1987.
Emil Grgin. . Ph.D. Thesis, Syracuse University, 1966.
Roger Penrose and M. A. H. MacCallum. Twistor theory: An approach to the quantisation of fields and space-time. , 6(4):241–316, 1972.
Lane P. Hughston and Richard S. Ward. An introduction to twistor theory. In [*Advances in twistor theory*]{}. Pitman Publishing Limited, London, 1979.
Andrew Hodges. Twistor diagrams. , 114(1):157–175, 1982.
M. G. Eastwood and A. M. Pilato. On the density of twistor elementary states. , 151(2):201–215, 1991.
Hridesh Kedia, Iwo Bialynicki-Birula, Daniel Peralta-Salas, and William M. T. Irvine. Tying knots in light fields. , 111(15):150404-1–4, 2013.
David A. Nichols, Robert Owen, Fan Zhang, Aaron Zimmerman, Jeandrew Brink, Yanbei Chen, Jeffrey D. Kaplan, Geoffrey Lovelace, Keith D. Matthews, Mark A. Scheel, and Kip S. Thorne. Visualizing spacetime curvature via frame-drag vortexes and tidal tendexes: General theory and weak-gravity applications. , 84:124014, Dec 2011.
Fan Zhang, Aaron Zimmerman, David A. Nichols, Yanbei Chen, Geoffrey Lovelace, Keith D. Matthews, Robert Owen, and Kip S. Thorne. Visualizing spacetime curvature via frame-drag vortexes and tidal tendexes II. Stationary black holes. , 86:084049, 2012.
David A. Nichols, Aaron Zimmerman, Yanbei Chen, Geoffrey Lovelace, Keith D. Matthews, Robert Owen, Fan Zhang, and Kip S. Thorne. Visualizing spacetime curvature via frame-drag vortexes and tidal tendexes III. Quasinormal pulsations of schwarzschild and Kerr black holes. , 86:104028, 2012.
R. Penrose and W. Rindler. . Cambridge monographs on mathematical physics. Cambridge University Press, Cambridge, 1988.
Manuel Array[á]{}s and Jos[é]{} L. Trueba. Exchange of helicity in a knotted electromagnetic field. 2011, arXiv:hep-th/1105.6285.
[^1]: email: [email protected]
[^2]: For the Weyl fields, the linked and knotted topology appears in the current.
[^3]: We use the conventions given in Eqn. of the Appendix.
[^4]: There is an overall constant factor of $4 n_t n_p$ in Ref. [@Irvine2013] that does not appear in our construction, but it does not affect the topology.
[^5]: from private discussions with Roger Penrose
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We obtain a discrete time analog of E. Noether’s theorem in Optimal Control, asserting that integrals of motion associated to the discrete time Pontryagin Maximum Principle can be computed from the quasi-invariance properties of the discrete time Lagrangian and discrete time control system. As corollaries, results for first-order and higher-order discrete problems of the calculus of variations are obtained.'
author:
- |
Delfim F. M. Torres\
`[email protected]`
date: |
Department of Mathematics\
University of Aveiro\
3810-193 Aveiro, Portugal\
`http://www.mat.ua.pt/delfim`
title: |
Integrals of Motion for Discrete-Time\
Optimal Control Problems[^1]
---
**Keywords:** discrete time optimal control, discrete time calculus of variations, discrete time mechanics, discrete time Pontryagin extremals, quasi-invariance up to difference gauge terms, discrete version of Noether’s theorem.
**Mathematics Subject Classification 2000:** 49-99, 39A12.
Introduction
============
Most physical systems encountered in nature exhibit symmetries: there exists appropriate infinitesimal-parameter family of transformations which keep the system invariant. From the well-known theorem of Emmy Noether [@JFM46.0770.01; @MR53:10538], one can discover the integrals of motion from those invariance transformations. Noether’s theorem plays a fundamental role in modern physics, and is usually formulated in the context of the calculus of variations: from the invariance properties of the variational integrals, the integrals of motion of the respective Euler-Lagrange differential equations, that is, expressions which are preserved along the extremals, are obtained. The result is, however, much more than a theorem. It is an universal principle, which can be formalized in a precise statement, as a theorem, on very different contexts and, for each such context, under very different assumptions. Let us consider, for example, classical mechanics or, more generally, the calculus of variations. Typically, Noether transformations are considered to be point-transformations (they are considered to be functions of coordinates and time), but one can consider more general transformations depending also on velocities and higher derivatives [@ZBL0964.49001] or within the broader context of dynamical symmetries [@MR81c:58054]. For an example of an integral of motion which comes from an invariance transformation depending on velocities, see [@MoyoLeach2002]. In most formulations of Noether’s principle, the Noether transformations keep the integral functional invariant ( [@MR2000m:49002 §1.5]). It is possible, however, to consider transformations of the problem up to an exact differential ( [@MR37:5752 p. 73]), called a gauge-term [@MR83c:70020]. Once strictly-invariance of the integral functional is no more imposed, one can think considering additional terms in the variation of the Lagrangian – see the quasi-invariance and semi-invariance notions introduced by the author respectively in [@torresMED2002] and [@torresControlo2002]. Formulations of Noether’s principle are possible for problems of the calculus of variations: on Euclidean spaces ( [@MR58:18024]) or on manifolds ( [@MR57:13703]); with single or multiple integrals ( [@MR95b:58049]); with higher-order derivatives ( [@MR54:6786]); with holonomic or nonholonomic constraints ( [@MR97a:49001 Ch. 7], [@MR2000a:37060]); and so on. Other contexts for which Noether’s theorems are available include supermechanics [@MR96g:58011], control systems [@MR83k:49054; @MR83k:93011], and optimal control (see [@MR49:5979; @MR1806135; @delfim3ncnw; @delfimEJC]). For a survey see [@MR96i:49037; @torresSpecialJMS]. Here we are interested in providing a formulation of the Noether’s principle in the discrete time setting. For a description of discrete time mechanics, discrete time calculus of variations, and discrete optimal control see, , [@MR98k:81076; @MR98k:81077; @MR99m:81275a; @MR99m:81275b], [@ZBL0193.07601], and [@MR2001k:49058]. Illustrative examples of real-life problems which can be modeled in such framework can be found in [@SegundaEdicaoMR84g:49002 Ch. 8]. Versions of the Noether’s principle for the discrete calculus of variations, and applicable to discrete analogues of classical mechanics, appeared earlier in [@MR48:6739; @MR48:6741; @MR82g:70041; @MR95i:58098; @discreteNoetherPhD; @MR99d:70004], motivated by the advances of numerical and computational methods. There, the discrete analog of Noether’s theorem is obtained from the discrete analog of the Euler-Lagrange equations. To the best of our knowledge, no Noether type theorem is available for the discrete time optimal control setting. One such formulation is our concern here. The result is obtained from the discrete time version of the Pontryagin maximum principle. As corollaries, we obtain generalizations of the previous results for first-order and higher-order discrete problems of the calculus of variations which are quasi-invariant and not necessarily invariant.
Discrete-Time Optimal Control
=============================
Without loss of generality ( [@MR86c:49028 §2]), we consider the discrete optimal control problem in Lagrange form. The time $k$ is a discrete variable: $k \in \mathbb{Z}$. The horizon consists of $N$ periods, $k = M, M + 1, \ldots, M + N - 1$, where $M$ and $N$ are fixed integers, instead of a continuous interval. We look for a finite control sequence $u(k) \in \mathbb{R}^r$, $k = M, \ldots, M + N - 1$, and the corresponding state sequence $x(k) \in \mathbb{R}^n$, $k = M, \ldots, M + N$, which minimizes or maximizes the sum $$J\left[x(\cdot),u(\cdot)\right] = \sum_{k = M}^{M+N-1}
L\left(k,x(k),u(k)\right) \, ,$$ subject to the discrete time control system $$\label{eq:DCS}
x(k+1) = \varphi\left(k,x(k),u(k)\right) \, , \quad
k = M, \ldots, M + N - 1 \, ,$$ the boundary conditions $$\label{eq:BC}
x(M) = x_{M} \, , \quad x(M+N) = x_{M+N} \, ,$$ and the control constraint $$u(k) \in \Omega \subseteq \mathbb{R}^r \, , \quad
k = M, \ldots, M + N - 1 \, .$$ A sequence-pair $\left(x(k),u(k)\right)$, $k = M, \ldots, M + N - 1$, satisfying the recurrence relation and conditions , is said to be admissible: $x(k)$ is an admissible state sequence and $u(k)$ an admissible control sequence. Functions $L(k,x,u) : \left\{M, \ldots, M + N - 1\right\}
\times \mathbb{R}^n \times \mathbb{R}^r \rightarrow \mathbb{R}$ and $\varphi(k,x,u) : \left\{M, \ldots, M + N - 1\right\}
\times \mathbb{R}^n \times \mathbb{R}^r \rightarrow \mathbb{R}^n$ are assumed to be continuously differentiable with respect to $x$ and $u$ for all fixed $k = M, \ldots, M + N - 1$, and convex in $u$ for any fixed $k$ and $x$. They are in general nonlinear. The control constraint set $\Omega$ is assumed to be convex. The problem is denoted by $(P)$.
For continuous optimal control problems, the convexity assumptions we are imposing are not needed in order to derive the Pontryagin maximum principle [@MR29:3316b]. This differs from the discrete time optimal control setting. Our hypothesis can be, however, weakened to directional convexity or even more weak conditions (see [@MR86c:49028], [@SegundaEdicaoMR84g:49002 §8.3] and references in [@ZBL0388.49002 Ch. 6] and [@MR86c:49028]).
It is possible to formulate problem $(P)$ with the first-order difference equations in terms of the forward or backward difference operators $\Delta$ or $\nabla$, defined by $\Delta x(k) = x(k+1) - x(k)$, $\nabla x(k) = x(k) - x(k-1)$. The results of the paper are written in those terms in a straightforward way.
The following theorem provide a first-order necessary optimality condition ( [@BertsekasVol1 §3.3.3], [@MR81i:90157], [@ZBL0388.49002 Ch. 6]) in the form of Pontryagin’s maximum principle [@MR29:3316b]. For a good survey on the history of the development of maximum principle to the optimization of discrete time systems, we refer the reader to [@MR86c:49028].
\[th:DTMP\] If $\left(x(k),u(k)\right)$ is a minimizer or a maximizer of the problem $(P)$, then there exists a nonzero sequence-pair $(\psi_0,\psi(k))$, $k = M+1, \ldots, M+N$, where $\psi_0$ is a constant less or equal than zero and $\psi(k) \in \mathbb{R}^n$, such that the sequence-quadruple $$\left(x(k),u(k),\psi_0,\psi(k+1)\right) \, , \quad
k = M, \ldots, M+N-1 \, ,$$ satisfies:
(i)
: the Hamiltonian system $$\label{eq:HamSyst}
\begin{cases}
x(k+1) = \frac{\partial H}{\partial \psi}\left(k,x(k),u(k),\psi_0,\psi(k+1)\right) \, ,
& k = M, \ldots, M+N-1 \, ,\\
\psi(k) = \frac{\partial H}{\partial x}\left(k,x(k),u(k),\psi_0,\psi(k+1)\right) \, ,
& k = M+1, \ldots, M+N-1 \, ;
\end{cases}$$
(ii)
: the maximality condition $$\label{eq:MaxCond}
H\left(k,x(k),u(k),\psi_0,\psi(k+1)\right) = \max_{u \in \Omega}
H\left(k,x(k),u,\psi_0,\psi(k+1)\right) \, ,$$ $k = M, \ldots, M+N-1$;
with the Hamiltonian $$H\left(k,x,u,\psi_0,\psi\right) = \psi_0 L(k,x,u) + \psi \cdot \varphi(k,x,u) \, .$$
The first equation in the Hamiltonian system is just the control system . The second equation in the Hamiltonian system is known as the adjoint system. The multipliers $\psi(\cdot)$ are called adjoint multipliers or co-state variables.
In the absence of the initial conditions $x(M) = x_{M}$ and/or terminal conditions $x(M+N) = x_{M+N}$, there corresponds additional conditions in the Discrete-Time Maximum Principle called transversality conditions. Our version of Noether’s theorem only require the use of the adjoint system and maximality condition. Therefore, the result is valid under all types of boundary conditions under consideration.
A sequence-quadruple $\left(x(k),u(k),\psi_0,\psi(k+1)\right)$, $k = M,\ldots,M+N-1$, $\psi_0 \le 0$, satisfying the Hamiltonian system and the maximality condition, is called an extremal for problem $(P)$. An extremal is said to be normal if $\psi_0 \ne 0$ and abnormal if $\psi_0 = 0$.
As we will see on Section \[sec:DCV\], there are no abnormal extremals both for first-order and higher-order discrete problems of the calculus of variations. In particular, there are no abnormal extremals for problems of discrete time mechanics. For our general problem $(P)$, however, abnormal extremals do exist. In fact, they happen to occur frequently. For a throughout study of abnormal extremals see [@MR1845332].
Integrals of Motion
===================
We obtain a systematic procedure to establish integrals of motion, , to establish expressions which are preserved on the extremals of the discrete optimal control problem $(P)$, from the (quasi-)invariance properties of the discrete Lagrangian $L\left(k,x(k),u(k)\right)$ and discrete control system $x(k+1) = \varphi\left(k,x(k),u(k)\right)$.
\[def:INV\] Let $X : \left\{M,\ldots,M+N-1\right\} \times \mathbb{R}^n
\times \Omega \times \mathcal{B}(0;\varepsilon) \rightarrow \mathbb{R}^n$, $\varepsilon > 0$, $\mathcal{B}(0;\varepsilon) =
\left\{s=(s_1,\ldots,s_\rho)| \left\|s\right\| = \sqrt{\sum_{i=1}^{\rho} (s_i)^2}
< \varepsilon \right\}$, be an infinitesimal $\rho$-parameter transformation such that for each $k$, $k = M,\ldots,M+N-1$, $X(k,\cdot,\cdot,\cdot)$ is continuously differentiable with respect to all arguments, and such that $X(k,x,u,0) = x$ for all $k = M,\ldots,M+N-1$, $x \in \mathbb{R}^n$, and $u \in \Omega$. If there exists a real function $\Phi\left(k,x,u,s\right)$ and for all $s \in \mathcal{B}(0;\varepsilon)$ and admissible $\left(x(k),u(k)\right)$ there exists a control sequence $u(k,s)$, $u(k,0) = u(k)$, such that: $$\begin{gathered}
\label{eq:INVi}
L\left(k,x(k),u(k)\right) + \Delta \Phi\left(k,x(k),u(k),s\right)
+ \delta\left(k,x(k),u(k),s\right) \\
= L\left(k,X\left(k,x(k),u(k),s\right),u(k,s)\right) \, ,\end{gathered}$$ $$\begin{gathered}
\label{eq:INVii}
X\left(k+1,x(k+1),u(k+1),s\right) + \delta\left(k,x(k),u(k),s\right) \\
= \varphi\left(k,X\left(k,x(k),u(k),s\right),u(k,s)\right) \, ,\end{gathered}$$ for each $k = M,\ldots,M+N-1$ and where $\delta(k,x,u,s)$ is an arbitrary function satisfying $$\label{eq:DdeltasEQ0}
\left.\frac{\partial \delta(k,x,u,s)}{\partial s_i}\right|_{s = 0} = 0 \, ,
\quad i = 1,\ldots,\rho \, ,$$ for each $k$, $x$, $u$, then the problem $(P)$ is said to be quasi-invariant with respect to the transformation $X(k,x,u,s)$ up to the difference gauge term $\Phi\left(k,x(k),u(k),s\right)$.
In the relation , $\Delta$ is the forward difference operator: $$\Delta \Phi\left(k,x(k),u(k),s\right) =
\Phi\left(k+1,x(k+1),u(k+1),s\right) - \Phi\left(k,x(k),u(k),s\right) \, .$$
When $\delta \equiv 0$ and $\Phi \equiv 0$, we have (strict-)invariance. The term *quasi-invariant* refers to the possibility of $\delta$ to be different from zero.
\[Th:MainResult\] If $(P)$ is quasi-invariant with respect to the $\rho$-parameter transformation $X$ up to the difference gauge term $\Phi$, in the sense of Definition \[def:INV\], then all its extremals $\left(x(k),u(k),\psi_0,\psi(k)\right)$, $k = M,\ldots,M+N-1$, satisfy the following $\rho$ expressions ($i = 1,\ldots,\rho$): $$\psi_0 \left.\frac{\partial}{\partial s_i}
\Phi\left(k,x(k),u(k),s\right)\right|_{s = 0}
+ \psi(k) \cdot \left.\frac{\partial}{\partial s_i}
X\left(k,x(k),u(k),s\right)\right|_{s = 0}
= \text{constant} \, .$$
The integrals of motion obtained by Theorem \[Th:MainResult\] are “momentum” integrals. Due to the fact that time $k$ is discrete, one can not vary $k$ continuously and, for that reason, one can not obtain the “energy” integrals as in the continuous optimal control case ( [@delfim3ncnw; @delfimEJC]). To address the problem another method needs to be developed. This will be addressed in a forthcoming paper.
Together with the continuous results in [@delfim3ncnw; @delfimEJC], Theorem \[Th:MainResult\] provides a framework to obtain a generalization of Noether’s theorem for hybrid-systems. This and related questions are under study and will be addressed elsewhere.
Let $\left(x(k),u(k),\psi_0,\psi(k)\right)$ be an extremal for problem $(P)$. Differentiating and with respect to the parameter $s_i$, $i = 1,\ldots,\rho$, and setting $s = \left(s_1,\ldots,s_{\rho}\right) = 0$, we get (recall and that $X\left(k,x(k),u(k),0\right)
= x(k)$, $u(k,0) = u(k)$): $$\begin{gathered}
\label{eq:DsL}
\Delta \left.\frac{\partial}{\partial s_i}
\Phi\left(k,x(k),u(k),s\right)\right|_{s=0}
= \frac{\partial L}{\partial x}\left(k,x(k),u(k)\right)
\cdot \left.\frac{\partial}{\partial s_i} X\left(k,x(k),u(k),s\right)\right|_{s=0} \\
+ \frac{\partial L}{\partial u}\left(k,x(k),u(k)\right)
\cdot \left.\frac{\partial}{\partial s_i} u\left(k,s\right)\right|_{s=0} \, ,\end{gathered}$$ $$\begin{gathered}
\label{eq:DsFi}
\left.\frac{\partial}{\partial s_i}
X\left(k+1,x(k+1),u(k+1),s\right)\right|_{s=0} \\
= \frac{\partial \varphi}{\partial x}\left(k,x(k),u(k)\right)
\cdot \left.\frac{\partial}{\partial s_i} X\left(k,x(k),u(k),s\right)\right|_{s=0} \\
+ \frac{\partial \varphi}{\partial u}\left(k,x(k),u(k)\right)
\cdot \left.\frac{\partial}{\partial s_i} u\left(k,s\right)\right|_{s=0} \, .\end{gathered}$$ From the adjoint system $\psi(k) = \frac{\partial H}{\partial x}\left(k,x(k),u(k),\psi_0,\psi(k+1)\right)$, we know that $$-\psi_0 \frac{\partial L}{\partial x}\left(k,x(k),u(k)\right)
= \psi(k+1) \cdot \frac{\partial \varphi}{\partial x}\left(k,x(k),u(k)\right)
- \psi(k) \, ,$$ and multiplying by $-\psi_0$ one obtains: $$\begin{gathered}
\label{eq:DepMultpsi0}
\psi_0 \left(\Delta \left.\frac{\partial}{\partial s_i}
\Phi\left(k,x(k),u(k),s\right)\right|_{s=0}
- \frac{\partial L}{\partial u}\left(k,x(k),u(k)\right)
\cdot \left.\frac{\partial}{\partial s_i} u\left(k,s\right)\right|_{s=0}\right) \\
+ \left(\psi(k+1) \cdot \frac{\partial \varphi}{\partial x}\left(k,x(k),u(k)\right)
- \psi(k)\right)
\cdot \left.\frac{\partial}{\partial s_i} X\left(k,x(k),u(k),s\right)\right|_{s=0}
= 0 \, .\end{gathered}$$ As far as $u(k,0) = u(k)$, according to the maximality condition of the Discrete-Time Maximum Principle the function $$s \mapsto \psi_0 L\left(k,x(k),u(k,s)\right)
+ \psi(k+1) \cdot \varphi\left(k,x(k),u(k,s)\right)$$ attains its maximum for $s = 0$. Therefore, $$\left.\frac{\partial}{\partial s_i} \left( \psi_0 L\left(k,x(k),u(k,s)\right)
+ \psi(k+1) \cdot \varphi\left(k,x(k),u(k,s)\right)\right)\right|_{s = 0} = 0 \, ,$$ that is, $$\begin{gathered}
\label{eq:DaCondMax}
\psi_0 \frac{\partial L}{\partial u}\left(k,x(k),u(k)\right)
\cdot \left.\frac{\partial}{\partial s_i} u(k,s)\right|_{s = 0} \\
+ \psi(k+1) \cdot \frac{\partial \varphi}{\partial u}\left(k,x(k),u(k)\right)
\cdot \left.\frac{\partial}{\partial s_i} u(k,s)\right|_{s = 0} = 0 \, .\end{gathered}$$ From and it comes $$\begin{gathered}
\psi_0 \Delta \left.\frac{\partial}{\partial s_i}
\Phi\left(k,x(k),u(k),s\right)\right|_{s=0} \\
+ \left(\psi(k+1) \cdot \frac{\partial \varphi}{\partial x}\left(k,x(k),u(k)\right)
- \psi(k)\right) \cdot \left.\frac{\partial}{\partial s_i}
X\left(k,x(k),u(k),s\right)\right|_{s=0} \\
+ \psi(k+1) \cdot \frac{\partial \varphi}{\partial u}\left(k,x(k),u(k)\right)
\cdot \left.\frac{\partial}{\partial s_i} u(k,s)\right|_{s = 0} = 0 \, .\end{gathered}$$ Using , this last equality is equivalent to $$\Delta\left(\psi_0 \left.\frac{\partial}{\partial s_i}
\Phi\left(k,x(k),u(k),s\right)\right|_{s = 0}
+ \psi(k) \cdot \left.\frac{\partial}{\partial s_i}
X\left(k,x(k),u(k),s\right)\right|_{s = 0}\right) = 0 \, .$$ The proof is complete.
An Example
==========
We now illustrate the use of Theorem \[Th:MainResult\] by the following example ($n = 3$, $r = 2$, $\Omega = \mathbb{R}^2$): $$\begin{gathered}
\sum_{k} \left(u_1(k)\right)^2 - \left(u_2(k)\right)^2 \longrightarrow \textrm{extr} \, , \\
\begin{cases}
x_1(k+1) = x_2(k) + u_1(k) \, , \\
x_2(k+1) = x_1(k) + u_2(k) \, , \\
x_3(k+1) = x_2(k) u_1(k) \, ,
\end{cases}\end{gathered}$$ subject to fixed endpoints. In this case the Hamiltonian is given by $$\begin{gathered}
H(x_1,x_2,u_1,u_2,\psi_0,\psi_1,\psi_2,\psi_3) \\
= \psi_0 L(u_1,u_2) + \psi_1 \varphi_1(x_2,u_1)
+ \psi_2 \varphi_2(x_1,u_2) + \psi_3 \varphi_3(x_2,u_1) \, ,\end{gathered}$$ with $L(u_1,u_2) = u_1^2 - u_2^2$, $\varphi_1(x_2,u_1) = x_2 + u_1$, $\varphi_2(x_1,u_2) = x_1 + u_2$, and $\varphi_3(x_2,u_1) = x_2 u_1$. From the adjoint system we get the evolution equations $$\begin{split}
\psi_1(k) &= \psi_2(k+1) \, ,\\
\psi_2(k) &= \psi_1(k+1) + \psi_3(k+1) u_1(k) \, , \\
\psi_3(k) &= 0 \, ,
\end{split}$$ while from the maximality conditions we get ($\psi_3 = 0$) $$\begin{split}
\psi_1(k+1) &= -2 \psi_0 u_1(k) \, ,\\
\psi_2(k+1) &= 2 \psi_0 u_2(k) \, .
\end{split}$$ There are no abnormal extremals for the problem, and one can fix $\psi_0 = -\frac{1}{2}$. The extremals are obtained solving five difference-equations of order one, $$\begin{cases}
x_1(k+1) = x_2(k) + \psi_1(k+1) \, , \\
x_2(k+1) = x_1(k) - \psi_2(k+1) \, , \\
x_3(k+1) = x_2(k) \psi_1(k+1) \, , \\
\psi_1(k+1) = \psi_2(k) \, , \\
\psi_2(k+1) = \psi_1(k) \, ,
\end{cases}$$ together with the boundary conditions (or the transversality conditions), by standard techniques. On the other hand, the problem is quasi-invariant with respect to the one-parameter ($\rho = 1$) transformations $$\begin{split}
X_1\left(x_1(k),s\right) &= x_1(k)+2s \, , \\
X_2\left(x_2(k),s\right) &= x_2(k)+s \, , \\
X_3\left(x_1(k),x_3(k),s\right) &= x_3(k) + s x_1(k) \, ,
\end{split}$$ up to the difference gauge term $\Phi\left(x_1(k),x_2(k),s\right) = 2\left(x_1(k)+x_2(k)\right) s$. To see that we choose $$u_1(k,s) = u_1(k) + s \, , \quad
u_2(k,s) = u_2(k) - s \, ,$$ in the Definition \[def:INV\]. We notice that $X_1\left(x_1(k),0\right) = x_1(k)$, $X_2\left(x_2(k),0\right) = x_2(k)$, $X_3\left(x_1(k),x_3(k),0\right) = x_3(k)$, $u_1(k,0) = u_1(k)$, and $u_2(k,0) = u_2(k)$. Direct verifications show that the quasi-invariance conditions are satisfied: $$\begin{split}
L(u_1(k,s),&u_2(k,s))
= \left(u_1(k)\right)^2 - \left(u_2(k)\right)^2 + 2\left(u_1(k)+u_2(k)\right) s \\
&= L\left(u_1(k),u_2(k)\right) + 2 \left(x_1(k+1)-x_2(k)+x_2(k+1)-x_1(k)\right) s \\
&= L\left(u_1(k),u_2(k)\right) + \Delta \Phi\left(x_1(k),x_2(k),s\right) \, ,
\end{split}$$ $$\begin{gathered}
\begin{split}
\varphi_1\left(X_2\left(x_2(k),s\right),u_1(k,s)\right)
&= x_2(k) + u_1(k) + 2s = x_1(k+1) + 2s \\
&= X_1\left(x_1(k+1),s\right) \, , \\
\end{split}\\
\begin{split}
\varphi_2\left(X_1\left(x_1(k),s\right),u_2(k,s)\right)
&= x_1(k) + u_2(k) + s = x_2(k+1) + s \\
&= X_2\left(x_2(k+1),s\right) \, , \\
\end{split}\\
\begin{split}
\varphi_3\left(X_2\left(x_2(k),s\right),u_1(k,s)\right)
&= \left(x_2(k)+s\right) \left(u_1(k)+s\right) \\
&= x_2(k) u_1(k)
+ s \left(x_2(k) + u_1(k)\right) + s^2 \\
&= x_3(k+1) + s x_1(k+1) + \delta(s) = X_3(k+1) + \delta(s) \, .
\end{split}\end{gathered}$$ By Theorem \[Th:MainResult\] we obtain the following conservation law for the problem: $$\label{eq:FIMEx}
2 \psi_0 \left(x_1(k)+x_2(k)\right) + 2 \psi_1(k)
+ \psi_2(k) + \psi_3(k) x_1(k) = \text{constant} \, .$$ Using the information from the discrete time maximum principle, condition is equivalent to $$\label{eq:FIMExCondNec}
\left(x_1(k)+x_2(k)\right) + 2 u_2(k) - u_1(k) = \text{constant} \, .$$ The conservation law is a necessary optimality condition. It is trivially satisfied choosing the control variables according to: $$\begin{gathered}
u_1(k) = x_1(k) \, ,\\
u_2(k) = - \frac{1}{2} x_2(k) \, .\end{gathered}$$ An extremal is then obtained with the co-state variables given by $$\begin{gathered}
\psi_1(k) = - u_2(k) = \frac{1}{2} x_2(k) \, , \\
\psi_2(k) = u_1(k) = x_1(k) \, , \\
\psi_3(k) = 0 \, .\end{gathered}$$
Discrete Calculus of Variations {#sec:DCV}
===============================
We now obtain a discrete Noether’s theorem for the problems of the discrete time calculus of variations which are quasi-invariant with respect to infinitesimal transformations having $\rho$ parameters, $\rho \ge 1$, up to a difference gauge term.
The fundamental Problem
-----------------------
The fundamental problem in the discrete calculus of variations is a special case of our problem $(P)$: $r = n$; no restrictions on the controls ($\Omega = \mathbb{R}^n$); $\varphi(k,x,u) = u$. The problem is then to determine a finite sequence $x(k) \in \mathbb{R}^n$, $k = M,\ldots,M+N$, $x(M) = x_{M}$, $x(M+N) = x_{M+N}$, for which the discrete cost function $$J\left[x(\cdot)\right] = \sum_{k = M}^{M+N-1}
L\left(k,x(k),x(k+1)\right)$$ is extremized. The maximality condition in the Theorem \[th:DTMP\] implies in this case the conditions $$\frac{\partial H}{\partial u}\left(k,x(k),u(k),\psi_0,\psi(k+1)\right) = 0 \, ,
\quad k = M,\ldots,M+N-1 \, ,$$ that is, $$\label{eq:NoAbnPB}
\psi(k+1) = -\psi_0 \frac{\partial L}{\partial u}\left(k,x(k),x(k+1)\right) \, ,$$ while from the adjoint system one gets $$\label{eq:SAPBCV}
\psi(k) = \psi_0 \frac{\partial L}{\partial x}\left(k,x(k),x(k+1)\right) \, .$$ We note that no abnormal extremals exist for the fundamental problem of the discrete calculus of variations: $\psi_0 = 0$ implies that $\psi(k+1)$ is zero for all $k = M,\ldots,M+N-1$, a possibility excluded by the discrete time maximum principle. So it must be the case that $\psi_0 \ne 0$. From and , a necessary condition to have an extremum is that $x(k)$, $k = M,\ldots,M+N-2$, must satisfy the second-order difference equation $$\label{eq:DiscELeq}
\frac{\partial L}{\partial x}\left(k+1,x(k+1),x(k+2)\right)
+ \frac{\partial L}{\partial u}\left(k,x(k),x(k+1)\right) = 0 \, .$$ Equations share resemblances with the continuous Euler-Lagrange equations, and are called the discrete Euler-Lagrange equations.
\[def:INVPBCV\] The discrete Lagrangian $L\left(k,x(k),x(k+1)\right)$ is said to be quasi-invariant with respect to the infinitesimal $\rho$-parameter transformation $X(k,x,u,s)$, $s = \left(s_1,\ldots,s_{\rho}\right)$, $\left\|s\right\| < \varepsilon$, $X(k,x,u,0) = x$ for all $k = M,\ldots,M+N-1$, $x, u \in \mathbb{R}^n$, up to the difference gauge term $\Phi\left(k,x(k),x(k+1),s\right)$, if for each $k = M,\ldots,M+N-2$ $$\begin{gathered}
\label{eq:INViPBCV}
L\left(k,x(k),x(k+1)\right) + \Delta \Phi\left(k,x(k),x(k+1),s\right)
+ \delta\left(k,x(k),x(k+1),s\right)\\
= L\left(k,X\left(k,x(k),x(k+1),s\right),X\left(k+1,x(k+1),x(k+2),s\right)\right) \, ,\end{gathered}$$ where $\delta(\cdot,\cdot,\cdot,\cdot)$ is a function satisfying .
\[CorMainResultPBCV\] If $L\left(k,x(k),x(k+1)\right)$ is quasi-invariant with respect to the $\rho$-parameter transformation $X$ up to the difference gauge term $\Phi$, in the sense of Definition \[def:INVPBCV\], then all solutions $x(k)$, $k = M,\ldots,M+N-2$, of the discrete Euler-Lagrange difference equation satisfy $$\begin{gathered}
\frac{\partial L}{\partial u}\left(k-1,x(k-1),x(k)\right)
\cdot \left.\frac{\partial}{\partial s_i}
X\left(k,x(k),x(k+1),s\right)\right|_{s = 0} \\
- \left.\frac{\partial}{\partial s_i}
\Phi\left(k,x(k),x(k+1),s\right)\right|_{s = 0} = \text{constant} \, ,\end{gathered}$$ $i = 1,\ldots,\rho$.
Higher Order Discrete Problems
------------------------------
Let us now consider the problem of optimizing $$\label{eq:Pr:HODP}
\sum_{k} L\left(k,x(k),x(k+1),\ldots,x(k+m)\right) \, ,$$ where $L(k,x^0,x^1,\ldots,x^{m})$ is continuously differentiable with respect to all variables. This problem is analogous to the continuous problems of the calculus of variations for which the Lagrangian $L$ depends on higher-order derivatives. It is easily written in the optimal control form $(P)$. Introducing the notation $$\begin{gathered}
x^0(k) = x(k) \, , \\
x^1(k) = x(k+1) \, , \\
\vdots \\
x^{m-1}(k) = x(k+m-1) \, , \\
u(k) = x(k+m) \, ,\end{gathered}$$ one gets: $$\begin{gathered}
\sum_{k} L\left(k,x^0(k),\ldots,x^{m-1}(k),u(k)\right)
\longrightarrow \textrm{extr} \, , \\
\begin{cases}
x^0(k+1) = x^1(k) \, , \\
x^1(k+1) = x^2(k) \, , \\
\vdots \\
x^{m-2}(k+1) = x^{m-1}(k) \, , \\
x^{m-1}(k+1) = u(k) \, .
\end{cases}\end{gathered}$$ The Hamiltonian is given by $$H = \psi_0 L\left(k,x^0,\ldots,x^{m-1},u\right)
+ \left(\sum_{j=0}^{m-2} \psi^j \cdot x^{j+1}\right) + \psi^{m-1} u \, .$$ From the maximality condition $$\label{eq:MCHighOrderPr}
\psi^{m-1}(k+1) = - \psi_0
\frac{\partial L}{\partial x^m}\left(k,x^0(k),\ldots,x^{m-1}(k),u(k)\right) \, ,$$ while from the adjoint system $$\begin{gathered}
\psi^0(k) = \psi_0
\frac{\partial L}{\partial x^0}\left(k,x^0(k),\ldots,x^{m-1}(k),u(k)\right)
\, , \label{eq:ASHighOrderPr1} \\
\psi^{j}(k) = \psi_0
\frac{\partial L}{\partial x^j}\left(k,x^0(k),\ldots,x^{m-1}(k),u(k)\right)
+ \psi^{j-1}(k+1) \, , \label{eq:ASHighOrderPr2}\end{gathered}$$ $j = 1,\ldots,m-1$. From , , and , we conclude that: similarly to the fundamental problem of the calculus of variations, no abnormal extremals exist in the higher order case; the equation $$\label{eq:EulerPoissonOCnotation}
\sum_{j=0}^{m} \frac{\partial L}{\partial
x^j}\left(k+m-j,x^0(k+m-j),\ldots,x^{m-1}(k+m-j),u(k+m-j)\right) = 0$$ holds. Going back to the initial notation, is nothing more than the discrete Euler-Poisson equation of order $2m$ for the $m$-th order discrete problem of the calculus of variations : $$\label{eq:EulerPoisson}
\sum_{j=0}^{m} \frac{\partial L}{\partial
x^j}\left(k+m-j,x(k+m-j),\ldots,x(k+2m-1-j),x(k+2m-j)\right) = 0 \, .$$
\[def:INVPHOCV\] The discrete Lagrangian $L\left(k,x(k),\ldots,x(k+m)\right)$ is said to be quasi-invariant with respect to the infinitesimal $\rho$-parameter transformation $X(k,x^0,\ldots,x^m,s)$, $s = \left(s_1,\ldots,s_{\rho}\right)$, $\left\|s\right\| < \varepsilon$, $X(k,x^0,\ldots,x^m,0) = x^0$ for all $k$, and $x^j$, $j = 0,\ldots,m$, up to the difference gauge term $\Phi\left(k,x(k),\ldots,x(k+m),s\right)$, if for each $k$ $$\begin{gathered}
\label{eq:INViPHOCV}
L\left(k,x(k),\ldots,x(k+m)\right)
+ \Delta \Phi\left(k,x(k),\ldots,x(k+m),s\right)
+ \delta\left(k,x(k),\ldots,x(k+m),s\right) \\
= L\left(k,X\left(k,x(k),\ldots,x(k+m),s\right),\ldots,
X\left(k+m,x(k+m),\ldots,x(k+2m),s\right)\right) \, ,\end{gathered}$$ where $\frac{\partial \delta}{\partial s_i} = 0$, $i = 1,\ldots,\rho$.
\[CorMainResultPHOCV\] If $L\left(k,x(k),\ldots,x(k+m)\right)$ is quasi-invariant with respect to the $\rho$-parameter transformation $X$ up to the difference gauge term $\Phi$, in the sense of Definition \[def:INVPHOCV\], then all solutions $x(k)$ of the discrete Euler-Poisson difference equation satisfy $$\begin{gathered}
\label{eq:LeiConsPHOCV}
\left.\frac{\partial}{\partial s_i}
\Phi\left(k,x(k),\ldots,x(k+m),s\right)\right|_{s = 0} \\
+ \sum_{j=0}^{m-1} \sum_{l=0}^{j}
\frac{\partial L}{\partial x^l}\left(k+j-l,x(k+j-l),\ldots,x(k+j-l+m)\right) \\
\cdot \left.\frac{\partial}{\partial s_i}
X\left(k+j,x(k+j),\ldots,x(k+j+m),s\right)\right|_{s = 0}
= \text{constant} \, ,\end{gathered}$$ $i = 1,\ldots,\rho$.
In the case $m=1$ the discrete Euler-Poisson equation reduces to the discrete Euler-Lagrange equation , and the conservation law reduces to $$\begin{gathered}
\left.\frac{\partial}{\partial s_i}
\Phi\left(k,x(k),x(k+1),s\right)\right|_{s = 0} \\
+ \frac{\partial L}{\partial x^0}\left(k,x(k),x(k+1)\right)
\cdot \left.\frac{\partial}{\partial s_i}
X\left(k,x(k),x(k+1),s\right)\right|_{s = 0}
= \text{constant} \, ,\end{gathered}$$ or, which is the same, $$\begin{gathered}
\left.\frac{\partial}{\partial s_i}
\Phi\left(k,x(k),x(k+1),s\right)\right|_{s = 0} \\
- \frac{\partial L}{\partial x^1}\left(k-1,x(k-1),x(k)\right)
\cdot \left.\frac{\partial}{\partial s_i}
X\left(k,x(k),x(k+1),s\right)\right|_{s = 0}
= \text{constant} \, .\end{gathered}$$ This is precisely the conservation law given by Corollary \[CorMainResultPBCV\].
[10]{}
D. Anderson. Noether’s theorem in generalized mechanics. [*J. Phys. A*]{}, 6:299–305, 1973.
A. V. Arutyunov. [*Optimality conditions*]{}. Kluwer Academic Publishers, Dordrecht, 2000.
J. C. Baez and J. W. Gilliam. An algebraic approach to discrete mechanics. [*Lett. Math. Phys.*]{}, 31(3):205–212, 1994.
D. P. Bertsekas. [*Dynamic programming and optimal control. [I]{}*]{}. Athena Scientific, Belmont, MA, 2 edition, 2000.
P. Blanchard and E. Br[ü]{}ning. [*Variational methods in mathematical physics*]{}. Springer-Verlag, Berlin, 1992.
G. Blankenstein and A. van der Schaft. Optimal control and implicit [H]{}amiltonian systems. In [*Nonlinear control in the year 2000, Vol. 1 (Paris)*]{}, pages 185–205. Springer, London, 2001.
J. A. Cadzow. Discrete calculus of variations. [*Int. J. Control, I. Ser.*]{}, 11:393–407, 1970.
J. F. Cariñena and H. Figueroa. A geometrical version of [N]{}oether’s theorem in supermechanics. [*Rep. Math. Phys.*]{}, 34(3):277–303, 1994.
D. S. Djukic. Noether’s theorem for optimum control systems. [*Internat. J. Control (1)*]{}, 18:667–672, 1973.
I. M. Gelfand and S. V. Fomin. [*Calculus of variations*]{}. Dover Publications, Mineola, NY, 2000.
J. W. Gilliam. [*Lagrangian and Symplectic Techniques in Discrete Mechanics*]{}. Ph.[D]{}. dissertation, University of California Riverside, August 1996.
R. Hilscher and V. Zeidan. Discrete optimal control: the accessory problem and necessary optimality conditions. [*J. Math. Anal. Appl.*]{}, 243(2):429–452, 2000.
G. Jaroszkiewicz and K. Norton. Principles of discrete time mechanics. [I]{}. [P]{}article systems. [*J. Phys. A*]{}, 30(9):3115–3144, 1997.
G. Jaroszkiewicz and K. Norton. Principles of discrete time mechanics. [I]{}[I]{}. [C]{}lassical field theory. [*J. Phys. A*]{}, 30(9):3145–3163, 1997.
J. Jost and X. Li-Jost. [*Calculus of variations*]{}. Cambridge University Press, Cambridge, 1998.
J. D. Logan. First integrals in the discrete variational calculus. [*Aequationes Math.*]{}, 9:210–220, 1973.
J. D. Logan. Higher dimensional problems in the discrete calculus of variations. [*Internat. J. Control (1)*]{}, 17:315–320, 1973.
J. D. Logan. [*Invariant variational principles*]{}. Academic Press \[Harcourt Brace Jovanovich Publishers\], New York, 1977.
D. Lovelock and H. Rund. [*Tensors, differential forms, and variational principles*]{}. Wiley-Interscience \[John Wiley & Sons\], New York, 1975.
M. Lutzky. Dynamical symmetries and conserved quantities. [*J. Phys. A: Math. Gen.*]{}, 12(7):973–981, 1979.
S. Maeda. Extension of discrete [N]{}oether theorem. [*Math. Japon.*]{}, 26(1):85–90, 1981.
P. Michel. Programmes mathématiques mixtes. [A]{}pplication au principe du maximum en temps discret dans le cas déterministe et dans le cas stochastique. [*RAIRO Rech. Opér.*]{}, 14(1):1–19, 1980.
S. Moyo and P. G. L. Leach. A note on the construction of the [E]{}rmakov-[L]{}ewis invariant. [*J. Phys. A: Math. Gen.*]{}, 35:5333–5345, 2002.
Z. Nahorski, H. F. Ravn, and R. V. V. Vidal. The discrete-time maximum principle: a survey and some new results. [*Internat. J. Control*]{}, 40(3):533–554, 1984.
H. Nijmeijer and A. van der Schaft. Controlled invariance for nonlinear systems. [*IEEE Trans. Automat. Control*]{}, 27(4):904–914, 1982.
E. Noether. Invariante variationsprobleme. [*Gött. Nachr.*]{}, pages 235–257, 1918.
E. Noether. Invariant variation problems. [*Transport Theory Statist. Phys.*]{}, 1(3):186–207, 1971. English translation of the original paper [@JFM46.0770.01].
K. Norton and G. Jaroszkiewicz. Principles of discrete time mechanics. [I]{}[I]{}[I]{}. [Q]{}uantum field theory. [*J. Phys. A*]{}, 31(3):977–1000, 1998.
K. Norton and G. Jaroszkiewicz. Principles of discrete time mechanics. [I]{}[V]{}. [T]{}he [D]{}irac equation, particles and oscillons. [*J. Phys. A*]{}, 31(3):1001–1023, 1998.
L. S. Pontryagin, V. G. Boltyanskii, R. V. Gamkrelidze, and E. F. Mishchenko. [*The mathematical theory of optimal processes*]{}. Interscience Publishers John Wiley & Sons, Inc. New York-London, 1962.
H. Rund. [*The [H]{}amilton-[J]{}acobi theory in the calculus of variations: [I]{}ts role in mathematics and physics*]{}. D. Van Nostrand Co., Ltd., London-Toronto, Ont.-New York, 1966.
A. P. Sage and C. C. I. White. [*Optimum systems control*]{}. Prentice-Hall Inc., Englewood Cliffs, N.J., 2nd edition edition, 1977.
W. Sarlet and F. Cantrijn. Generalizations of [N]{}oether’s theorem in classical mechanics. [*SIAM Rev.*]{}, 23(4):467–494, 1981.
S. P. Sethi and G. L. Thompson. [*Optimal control theory*]{}. Kluwer Academic Publishers, Boston, MA, second edition, 2000. Applications to management science and economics.
J. Sniatycki. Nonholonomic [N]{}oether theorem and reduction of symmetries. [*Rep. Math. Phys.*]{}, 42(1-2):5–23, 1998.
H. J. Sussmann. Symmetries and integrals of motion in optimal control. In [*Geometry in nonlinear control and differential inclusions (Warsaw, 1993)*]{}, pages 379–393. Polish Acad. Sci., Warsaw, 1995.
D. F. M. Torres. Conservation laws in optimal control. In [*Dynamics, Bifurcations and Control*]{}, volume 273 of [*Lecture Notes in Control and Information Sciences*]{}, pages 287–296. Springer-Verlag, Berlin, Heidelberg, 2002.
D. F. M. Torres. Conserved quantities along the [P]{}ontryagin extremals of quasi-invariant optimal control problems. In [*Proc. 10th Mediterranean Conference on Control and Automation, MED2002, Lisbon, Portugal, 10 pp. (electronic)*]{}, 2002.
D. F. M. Torres. On optimal control problems which admit an infinite continuous group of transformations. In [*Proc. 5th Portuguese Conference on Automatic Control, Controlo 2002, Aveiro, Portugal*]{}, pages 247–251, 2002.
D. F. M. Torres. On the [N]{}oether theorem for optimal control. [*European Journal of Control*]{}, 8(1):56–63, 2002.
D. F. M. Torres. [C]{}arathéodory-equivalence, [N]{}oether theorems, and [T]{}onelli full-regularity in the calculus of variations and optimal control. [*Special Issue of the *J. of Mathematical Sciences**]{}, Submitted for publication.
A. van der Schaft. Symmetries and conservation laws for [H]{}amiltonian systems with inputs and outputs: a generalization of [N]{}oether’s theorem. [*Systems Control Lett.*]{}, 1(2):108–115, 1981/82.
F. Y. M. Wan. [*Introduction to the calculus of variations and its applications*]{}. Chapman & Hall, New York, 1995.
J. M. Wendlandt and J. E. Marsden. Mechanical integrators derived from a discrete variational principle. [*Phys. D*]{}, 106(3-4):223–246, 1997.
[^1]: [Accepted to be presented at the IFAC Workshop on Control Applications of Optimization – *CAO’2003* – to be held in Visegrád, Hungary, 30 June – 2 July 2003, and to appear in the respective conference proceedings. The date of this version is January 24, 2003.]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We solve two variants of the Reifenberg problem (minimizing or not the free boundary) for all coefficient groups. We carry out the direct method of the calculus of variation and search a solution as a “weak limit” of a minimizing sequences. This strategy has been introduced by De Lellis, De Philippis, De Rosa, Ghiraldin and Maggi in [@Ita1],[@Ita2],[@Ita35],[@Ita4] and allowed them to solve the Reifenberg problem. We use an analogous strategy proved in [@Lab] which has the advantage to take into account the free boundary. Moreover, we show that the Reifenberg class is closed under weak limits without restriction on the coefficient group.'
author:
- Camille Labourie
title: 'Solutions of the (free boundary) Reifenberg Plateau problem'
---
Definitions and main results
============================
Introduction
------------
We present the Reifenberg approach to the Plateau problem ([@Rei], 1960). Reifenberg works with *sets* of the Euclidean space which span a boundary in the sense of algebraic topology and minimizes their (spherical) *Hausdorff measure*. A $d$-dimensional set $E$ spans a boundary $\Gamma$ if $E$ contains $\Gamma$ and cancel the $(d-1)$-cycles of $\Gamma$ (or a subgroup of them).
Fix $\Gamma$ a compact subset of $\mathbf{R}^n$ and let $L$ be a subgroup of the homology group $H_{d-1}(\Gamma)$. A Reifenberg competitor is a compact subset $E \subset \mathbf{R}^n$ such that $E$ contains $\Gamma$ and the morphism induced by inclusion, $$\begin{tikzcd}
H_{d-1}(\Gamma) \arrow[r] & H_{d-1}(E),
\end{tikzcd}$$ is zero on $L$.
Reifenberg proved the existence of a solution to his problem in 1960 for the Čech homology theory and compact Abelian coefficient groups ([@Rei]). The continuity property of the Čech theory implies that a Hausdorff limit of competitors is a competitor. Thus, a solution can be searched as a Hausdorff limit of a minimizing sequence. However, the area is not lower semicontinuous with respect to Hausdorff limits. For instance, one can imagine a minimizing sequence which has more and more dense tentacles so that the limit set is too large. Reifenberg worked with compact coefficient groups to have the Excision Axiom and thus to be able to cut out the tentacles and patch the holes. His construction leads to an alternative minimizing sequence for which the area is lower semicontinuous.
Nakauchi ([@Nakauchi]) stated and solved a free boundary variant in 1984 (for compact Abelian coefficient groups as well). We mean by this that the intersection $E \cap \Gamma$ varies among the competitors $E$.
{width=".3\linewidth"} {width=".3\linewidth"}
Fix a compact subset $\Gamma$ of $\mathbf{R}^n$ and a subgroup $L$ of $H_{d-1}(\Gamma)$. The Nakauchi competitors are the compacts sets $E \subset \mathbf{R}^n$ such that for all $v \in L$, there exists $u \in H_{d-1}(E \cap \Gamma)$ such that $i_*(u) = v$ and $i'_*(u) = 0$ where $i_*$ and $i'_*$ are the morphisms induced by inclusions: $$\begin{tikzcd}
& H_{d-1}(\Gamma)\\
H_{d-1}(E \cap \Gamma) \arrow[ur,"i_*"] \arrow [dr,"i'_*"]\\
& H_{d-1}(E).
\end{tikzcd}$$
Nakauchi minimizes $H^d(E \setminus \Gamma)$ but it would be also interesting to minimize $H^d(E)$ to take into account the free boundary.
In 2015, Fang ([@Fang]) gave a new proof of the Reifenberg and Nakauchi problems for all coefficient groups and minimizing $H^d(E \setminus \Gamma$). His idea is to take advantage of the lower semicontinuity of the area on quasiminimal sets (we present them in the next paragraph). Thanks to a construction of Feuvrier ([@Feuv]), he obtains an alternative minimizing sequence composed of quasiminimal sets. Fang also replaces the Hausdorff measures with more general functionals, called *elliptic integrands*.
In [@Ita1] and [@Ita2], De Lellis, De Philippis, De Rosa, Ghiraldin and Maggi introduced a new type of direct method. As Reifenberg, they work with sets and Hausdorff measures of the Euclidean space but they replace local Hausdorff limits by *weak limits* of minimizing sequences in the ambiant space $\mathbf{R}^n \setminus \Gamma$. In [@Ita3], [@Ita4], the authors replace the Hausdorff measure by elliptic integrands and solve the problem of Reifenberg. However, they do not minimize the free boundary and they require a compact coefficient group.
In [@Lab], we adapted this notion of weak limit to quasiminimizing sequences and we were able to take into account the free boundary. Thus, we can follow the strategy of De Lellis, De Philippis, De Rosa, Ghiraldin and Maggi and minimize the part of the competitor which lies on the boundary.
Reifenberg competitors
----------------------
*Given a topological space $X$ and an integer $k$ , $H_k(X;G)$ is the $k$^th^ Čech homology group of $X$ over an Abelian coefficient group $G$. We abbreviate this notation as $H_k(X)$ since the coefficient group is not significant for us. Throughout this section, we fix a closed set $\Gamma$ of $\mathbf{R}^n$ and a subgroup $L$ of $H_{d-1}(\Gamma)$.*
\[rei\_definition\]A Reifenberg competitor is a compact subset $E \subset \mathbf{R}^n$ such that the morphism induced by inclusion, $$\begin{tikzcd}
H_{d-1}(\Gamma) \arrow[r] & H_{d-1}(E \cup \Gamma),
\end{tikzcd}$$ is zero on $L$.
This definition is different from the original definition of Reifenberg because the competitor $E$ may not contain $\Gamma$. In the remainder of this paragraph, we compare this definition to the definition of Nakauchi ([@Nakauchi]). Let $E$ be a compact subset of $\mathbf{R}^n$ and consider the following commutative diagram induced by the inclusions: $$\begin{tikzcd}
& H_{d-1}(\Gamma) \arrow[dr,"j_*"] &\\
H_{d-1}(E \cap \Gamma) \arrow[ur,"i_*"] \arrow [dr,"i'_*"] & & H_{d-1}(E \cup \Gamma).\\
& H_{d-1}(E) \arrow[ur,"j'_*"] &
\end{tikzcd}$$ The set $E$ is a Nakauchi competitor provided that for all $v \in L$, there exists $u \in H_{d-1}(E \cap \Gamma)$ such that $i_*(u) = v$ and $i'_*(u) = 0$. Assuming that the Mayer Vietoris sequence holds for the sets $\Gamma$, $E$ in $E \cup \Gamma$, the following sequence is exact: $$\begin{tikzcd}
H_{d-1}(E \cap \Gamma) \arrow[r,"{(i_*,i'_*)}"] & H_{d-1}(\Gamma) \otimes H_{d-1}(E) \arrow[r,"j_* - j'_*"] & H_{d-1}(E \cup \Gamma).
\end{tikzcd}$$ Observe that $E$ satisfies Definition \[rei\_definition\] if and only if all elements of the form $(v,0) \in L \otimes H_{d-1}(E)$ are in the kernel of $j_* - j'_*$. And $E$ is a Nakauchi competitor if and only if all elements of the form $(v,0) \in L \otimes H_{d-1}(E)$ are in the image of $(i_*,i'_*)$. Thus, the Mayer Vietoris sequence implies that Definition \[rei\_definition\] is equivalent to the definition of Nakauchi. In that sense, we consider these definitions to be essentially equivalent. We favor Definition \[rei\_definition\] because we are able to prove that it is stable under weak limits (see Lemma \[rei\_limit\]).
Our goal is to prove the two following results (we omit the regularity of the boundary here, see Theorems \[sol1\] and \[sol2\] for the full statement).
We assume that $$m = \inf \set{H^d(E) | E \ \text{Reifenberg competitor}} < \infty$$ and that there exists a compact set $C \subset \mathbf{R}^n$ such that $$m = \inf \set{H^d(E) | E \ \text{Reifenberg competitor},\ E \subset C}.$$ Then there exists a Reifenberg competitor $E \subset C$ such that $H^d(E) = m$.
The next theorem is similar to Theorem $1.3$ of [@Fang] (which is based on Feuvrier’s construction) and Theorem $3.4$ of [@Ita4] (which is based on weak limits of minimizing sequences). However, we have not yet dealt with elliptic integrands.
We assume that $$m = \inf \set{H^d(E \setminus \Gamma) | E \ \text{Reifenberg competitor}} < \infty$$ and that there exists a compact set $C \subset \mathbf{R}^n$ such that $$m = \inf \set{H^d(E \setminus \Gamma) | E \ \text{Reifenberg competitor},\ E \subset C}.$$ Then there exists a Reifenberg competitor $E \subset C$ such that $H^d(E \setminus \Gamma) = m$.
If $\Gamma$ is compact and $H^d(\Gamma) < \infty$, this amounts to minimizing $H^d(E)$ among Reifenberg competitors containing $\Gamma$.
Sliding deformations
--------------------
Our ambiant space is an open set $X$ of $\mathbf{R}^n$. We fix an integer $1 \leq d \leq n$. The term *a closed set $S \subset X$* means that $S$ is relatively closed in $X$. The interval $[0,1]$ is denoted by the capital letter $I$. Given a set $E \subset X$ and a function $F\colon I \times E \to X$, the notation $F_t$ means $F(t,\cdot)$. Given two sets $A, B \subset \mathbf{R}^n$, the notation $A \subset \subset B$ means that there exists a compact set $K \subset \mathbf{R}^n$ such that $A \subset K \subset B$. For a ball $U$ of center $x$ and radius $r$, for $h \geq 0$, the symbol $hU$ denotes the ball of center $x$ and radius $hr$. We fix a closed subset $\Gamma$ of $X$ (the boundary).
\[sliding\_definition\]Let $E$ be a closed, $H^d$ locally finite subset of $X$. A *sliding deformation* of $E$ in an open set $U \subset X$ is a Lipschitz map $f\colon E \to X$ such that there exists a continuous homotopy $F\colon I \times E \to X$ satisfying the following conditions:
$$\begin{aligned}
&F_0 = \mathrm{id}\\
&F_1 = f\\
&\forall t \in I,\ F_t(E \cap \Gamma) \subset \Gamma\\
&\forall t \in I,\ F_t(E \cap U) \subset U\\
&\forall t \in I,\ F_t = \mathrm{id} \ \text{in} \ E \setminus K,\end{aligned}$$
where $K$ is some compact subset of $E \cap U$. Alternatively, the last axiom can be stated as $$\Set{x \in E | \exists \, t \in I, \ F_t(x) \ne x} \subset \subset E \cap U.$$
Operations on the competitors
=============================
We present three operations that preserve the Reifenberg competitors: supsets, continuous image by sliding deformations and weak limits.
Supset
------
\[rei\_supset\]Let $E$ be a Reifenberg competitor. Let $F$ be a compact subset of $\mathbf{R}^n$ containing $E$. Then $F$ is a Reifenberg competitor.
This follows from the following commutative diagram $$\begin{tikzcd}
H_{d-1}(\Gamma) \arrow[r] \arrow [rd] & H_{d-1}(E \cup \Gamma) \arrow [d]\\
& H_{d-1}(F \cup \Gamma)
\end{tikzcd}$$ where the arrows are the morphisms induced by inclusion.
Continuous image
----------------
\[rei\_image\]Let $E$ be a Reifenberg competitor. Let $f\colon E \cup \Gamma \to \mathbf{R}^n$ be a continuous map such that there exists a continuous map $F\colon I \times \Gamma \to \Gamma$ satisfying $F_0 = \mathrm{id}$ and $F_1 = f$. Then $f(E)$ is a Reifenberg competitor.
Consider the following commutative diagram $$\begin{tikzcd}
H_{d-1}(\Gamma) \arrow[r] \arrow[d,"f_*"] & H_{d-1}(E \cup \Gamma) \arrow[d,"f_*"]\\
H_{d-1}(\Gamma) \arrow[r] & H_{d-1}(f(E) \cup \Gamma)
\end{tikzcd}$$ where the unlabeled arrows are the morphisms induced by inclusion. As $f\colon \Gamma \to \Gamma$ is homotopic to $\mathrm{id}$, $f_* = \mathrm{id}$ on $H_{d-1}(\Gamma)$.
The lemma assumed $f$ to be defined on $E \cup \Gamma$ but the image $f(E)$ depends only on the values of $f$ on $E$. In the two following remarks, we are going to see that it is generally enough for $f$ to be defined on $E$. In particular, the second remark applies to sliding deformations.
Let $f\colon E \to \mathbf{R}^n$ be a continuous map such that $f = \mathrm{id}$ on $E \cap \Gamma$. As $E$ and $\Gamma$ are closed sets of $\mathbf{R}^n$, the gluing $$g =
\begin{cases}
f &\text{in} \ E\\
\mathrm{id} &\text{in} \ \Gamma
\end{cases}$$ is continuous. Then $G_t = (1-t) \mathrm{id} + tg$ is a continuous homotopy from $\mathrm{id}$ to $g$ and $G_t = \mathrm{id}$ on $\Gamma$. We deduce that $f(E)$ is a Reifenberg competitor.
Let $f\colon E \to \mathbf{R}^n$ be a continuous map such that there exists a continuous map $F\colon I \times (E \cap \Gamma) \to \Gamma$ satisfying $F_0 = \mathrm{id}$ and $F_1 = f$. Let us assume that $\Gamma$ is a neighborhood retract i.e. there exists an open set $O \subset \mathbf{R}^n$ and a continuous map $r\colon O \to \Gamma$ such that $r = \mathrm{id}$ on $\Gamma$. According to the Homotopy Extension Lemma, $F$ extends as a continuous map $F\colon I \times \Gamma \to \Gamma$. Moreover, the gluing $$g =
\begin{cases}
f &\text{in} \ E\\
F_1 &\text{in} \ \Gamma
\end{cases}$$ is continous because $E$ and $\Gamma$ are closed sets of $\mathbf{R}^n$. We deduce that $f(E)$ is a Reifenberg competitor.
Weak limit
----------
We finally present our lemma about weak limits of Reifenberg competitors.
\[rei\_limit\]Let $(E_k) \subset \mathbf{R}^n$ be a sequence of Reifenberg competitors. Let $E$ be a compact subset of $\mathbf{R}^n$. We assume that
1. there exists a compact set $C \subset \mathbf{R}^n$ such that for all $k$, $E_k \subset C$;
2. for all open sets $V$ containing $E \cup \Gamma$, $$\lim_k H^d(E_k \setminus V) = 0.$$
Then $E$ is a Reifenberg competitor.
The proof requires a preliminary lemma about the general position of spheres. For $x \in \mathbf{R}^n$ and $r > 0$, let $S(x,r)$ denote the Euclidean sphere of center $x$ and radius $r$ of $\mathbf{R}^n$. Given an integer $k$, a *$k$-sphere* is an Euclidean sphere of positive radius relative to a $(k+1)$-affine plane. We extend this definition to $k < 0$, by calling *$k$-sphere* the empty set.
\[sphere-intersection\] Let $S^k$ be a $k$-sphere in $\mathbf{R}^n$ and let $x$ be a point in $\mathbf{R}^n$. Then for all $r > 0$ (except for at most one value), $S^k \cap S(x,r)$ is a subset of a $(k-1)$-sphere.
We assume $k \geq 1$. The proof is based on the observation that the intersection of a sphere with a $k$-affine plane is either empty, a point, or a $(k-1)$-sphere. In all cases, this intersection is part of a $(k-1)$-sphere. Let $P_0$ be the $(k+1)$-affine plane associated to $S^k$, let $x_0 \in P_0$ be the center of $S^k$ and $r_0 > 0$ be its radius. For $r > 0$, a point $y \in S^k \cap S(x,r)$ is characterized by the system
$$\begin{aligned}
&y \in P_0\\
&\abs{y - x} = r\\
&\abs{y - x_0} = r_0\end{aligned}$$
or equivalently
$$\begin{aligned}
&y \in P_0\label{sphere1}\\
&\abs{y - x} = r\label{sphere3}\\
&\abs{y - x}^2 - \abs{y - x_0}^2 = r^2 - r_0^2\label{sphere2}.\end{aligned}$$
Assume $x = x_0$. If $r \ne r_0$ (this removes one value of $r$), equation (\[sphere2\]) has no solutions. Then, $S^k \cap S(x,r)$ is empty and it is part of a $(k-1)$-sphere. Assume $x \ne x_0$. Equation (\[sphere2\]) defines an hyperplane and, if $\abs{x - x_0}^2 \ne r^2 - r_0^2$ (this removes at most one value of $r$), this hyperplane does not contain $x_0$. Then, the intersection of the two planes (\[sphere1\]) and (\[sphere2\]) is included in a $k$-affine plane. The intersection of this plane with the sphere (\[sphere3\]) is part of a $(k-1)$-sphere as seen in introduction.
The following proof makes use of the notion of complex (Subsection $3.1$ of [@Lab]) and of the Federer-Fleming projection (Proposition $3.1$ of [@DS] or Proposition $3.3.1$ of [@Lab]).
Observe that the sequence $(E_k \cup E)_k$ also satisfies the Lemma assumptions. So without loss of generality, we assume that for all $k$, $E \subset E_k$. We define a *general covering* as an open family $\gamma = (\gamma_j)_{j \in V_\gamma}$ of $\mathbf{R}^n$ satisfying the following properties:
1. there exists $k$ such that $E_k \cup \Gamma \subset \bigcup_{j \in V_\gamma} \gamma_j$;
2. for every subset $S \subset V_\gamma$ of cardinal $d+1$, $$\bigcap_S \gamma_j \ne \emptyset \implies (E \cup \Gamma) \cap \bigcap_S \gamma_j \ne \emptyset.$$
The main goal of the proof is to show that for any open covering $\alpha = (\alpha_i)_i$ of $E \cup \Gamma$, there exists a general covering $\gamma = (\gamma_j)_{j \in V_\gamma}$ such that $((E \cup \Gamma)\cap \gamma_j)_j$ is a refinement of $\alpha$. Let us explain how to conclude from there. A general covering $\gamma$ induces simplicial complexes: $$\begin{aligned}
K(\Gamma) &= \set{S \subset V_\gamma \ \text{finite} | \Gamma \cap \bigcap_S \gamma_j \ne \emptyset},\\
K(E \cup \Gamma) &= \set{S \subset V_\gamma \ \text{finite} | (E \cup \Gamma) \cap \bigcap_S \gamma_j \ne \emptyset},\\
K(E_k \cup \Gamma) &= \set{S \subset V_\gamma \ \text{finite} | (E_k \cup \Gamma) \cap \bigcap_S \gamma_j \ne \emptyset}.\end{aligned}$$ The inclusions $K(\Gamma) \subset K(E \cup \Gamma) \subset K(E_k \cup \Gamma)$ induce morphisms $i_*$ and $j_*$: $$\begin{tikzcd}
H_{d-1}(K(\Gamma)) \arrow[r,"i"] & H_{d-1}(K(E \cup \Gamma)) \arrow[r,"j"] & H_{d-1}(K(E_k \cup \Gamma)).
\end{tikzcd}$$ As $E_k$ is a Reifenberg set, we have $j_* \circ i_* = 0$ on $L$. However, the second axiom of general coverings says that the simplicial complexes $K(E \cup \Gamma)$ and $K(E_k \cup \Gamma)$ have the same $d$-simplexes. Hence the $d$-chains of $K(E \cup \Gamma)$ and $K(E_k \cup \Gamma)$ are identical and they induce the same boundaries. We deduce that $j_*$ is injective and then, $i_* = 0$ on $L$. Since every open covering $\alpha$ of $E \cup \Gamma$ is refined by such general covering $\gamma$, we conclude that the morphism induced by inclusion $H_{d-1}(\Gamma) \to H_{d-1}(E \cup \Gamma)$ is nul on $L$. *Step 1.* We fix a relative open covering $\alpha = (\alpha_i)_i$ of $E \cup \Gamma$ and we build a locally finite open sequence $\beta = (\beta_j)_{j \in \mathbf{N}}$ in $\mathbf{R}^n$ such that
1. $\beta$ cover $E \cup \Gamma$ and $((E \cup \Gamma) \cap \beta_j)_j$ is a refinement of $\alpha$;
2. for every finite subset $S \subset \mathbf{N}$, the intersection of boundaries $\bigcap_S \partial \beta_i$ is included in a finite union of $(n-m)$-spheres, where $m$ is the cardinal of $S$;
3. for every finite subset $S \subset \mathbf{N}$, $$\bigcap_S \beta_j \ne \emptyset \implies (E \cup \Gamma) \cap \bigcap_S \beta_j \ne \emptyset.$$
We work with the closed set $F := E \cup \Gamma$. For all $x \in F$, there exists $i$ such that $x \in \alpha_i$ so there exists an open ball $B$ centred at $x$ such that $F \cap 2B \subset \alpha_i$. We extract a sequence of open ball $(B_j)_{j \in \mathbf{N}}$ covering $F$ such that $(2B_j)_j$ is locally finite in $\mathbf{R}^n$ and $(F \cap 2B_j)_j$ is a refinement of $\alpha$. Next, we build by induction an open sequence $(\beta_j)_{j \in \mathbf{N}}$ such that for all $j$,
1. $F \cap \overline{B_j} \subset \beta_j$ and there exists $i$ such that $F \cap \beta_j \subset \alpha_i$.
2. for every subset $S \subset \set{1,\ldots,j}$, the intersection of boundaries $\bigcap_S \partial \beta_i$ is included in a finite union of $(n-m)$-spheres, where $m$ is the cardinal of $S$;
3. for every subset $S \subset \set{1,\ldots,j}$, $$\bigcap_S \beta_i \ne \emptyset \implies F \cap \bigcap_S \beta_i \ne \emptyset.$$
Assume that $\beta_0,\ldots,\beta_{j-1}$ has been built and let us built $\beta_j$. For all $x \in F \cap \overline{B_j}$, there exists an open ball $B$ centered at $x$ such that
1. $B \subset 2B_j$;
2. for all finite subset $S \subset \set{1,\ldots,j-1}$, the intersection of boundaries $\partial B \cap \bigcap_S \partial \beta_i$ is included in a finite union of $(n-m-1)$-spheres, where $m$ is the cardinal of $S$;
3. for all finite subset $S \subset \set{1,\ldots,j-1}$, $$(F \cap \overline{B_j}) \subset \mathbf{R}^n \setminus \bigcap_S \overline{\beta_j} \implies \overline{B} \subset \mathbf{R}^n \setminus \bigcap_S \overline{\beta_j}$$ or, équivalently $$\overline{B} \cap \bigcap_S \overline{\beta_j} \ne \emptyset \implies (F \cap \overline{B_j}) \cap \bigcap_S \overline{\beta_j} \ne \emptyset.$$
Extract a finite covering of $F \cap \overline{B_j}$ by such balls $B$ and denote $\beta_j$ their union. Then $\beta_j$ solves the next step of the induction.
*Step 2. We complete the family $\beta$ with an open set $\beta_\infty$ to obtain a covering of one of the $E_k$. We take care not to introduce new $d$-simplexes on $E \cup \Gamma$.* We want to reduce the problem to the case where for some $k$, $E_k \setminus \bigcup_j \beta_j$ is a $(d-1)$-dimensional grid. Using a Federer-Fleming projection, we are going to project $E_k$ in a $(d-1)$-dimensional grid away from $E \cup \Gamma$. Let $\ell > 0$ and consider a complex $K$ describing a uniform grid of sidelength $\ell$ in $\mathbf{R}^n$. In particular, $\mathbf{R}^n = \abs{K} = U(K)$ and the cells of $K$ have a diameter $\leq \sqrt{n}\ell$. We select the cells in which we want to perform the Federer-Fleming projection. Let $B_0$ be an open ball such that for all $k$, $E_k \subset \overline{B_0}$. Let $L$ be the subcomplex of $K$ defined by $$L = \set{A \in K | \exists x \in A,\ x \in \overline{2B_0} \ \text{and} \ \mathrm{d}(x,E \cup \Gamma) \geq 2\sqrt{n}\ell}.$$ Consider $x \in \overline{2B_0}$ such that $\mathrm{d}(x,E \cup \Gamma) \geq 2\sqrt{n}\ell$. As $\mathbf{R}^n = U(K)$, there exists a cell $A \in K$ such that $x \in \mathrm{int}(A)$ and, in particular, $A \in L$. We deduce that $$\label{grille1}
\set{x \in \overline{2B_0} | \mathrm{d}(x,E \cup \Gamma) \geq 2\sqrt{n}\ell} \subset U(L)$$ As $E \cup \Gamma$ is a closed set included in $\bigcup_j \beta_j$, the function $x \mapsto \mathrm{d}(x,E \cup \Gamma)$ is positive on $\mathbf{R}^n \setminus \bigcup_j \beta_j$. Moreover, $\overline{2B_0} \setminus \bigcup_j \beta_j$ is compact so the function $x \mapsto \mathrm{d}(x,E \cup \Gamma)$ has a positive minimum on $\overline{2B_0} \setminus \bigcup_j \beta_j$. This minimum does not depend on $\ell$ so we can assume $\ell$ small enough so that for all $x \in \overline{2B_0} \setminus \bigcup_j \beta_j$, $\mathrm{d}(x,E \cup \Gamma) > 4\sqrt{n}\ell$. By contraposition, $$\label{grille2}
\set{x \in \overline{2B_0} | \mathrm{d}(x,E \cup \Gamma) \leq 4\sqrt{n}\ell} \subset \bigcup_j \beta_j.$$ Next, we introduce the Federer-Fleming projection of $E_k \cap \abs{L}$ in $L$. First, we justify that $H^d(E_k \cap \abs{L}) < \infty$. In fact, we are going to have much better. By local finitness of $K$, $\abs{L}$ is a closed subset of $\mathbf{R}^n$. Since the cells of $K$ have a diameter $\leq \sqrt{n}\ell$, the definition of $L$ implies that the cells of $L$ cannot meet $E \cup \Gamma$. Thus, the set $V = \mathbf{R}^n \setminus \abs{L}$ is open and contains $E \cup \Gamma$. According to the Lemma assumptions, $$\lim_k H^d(E_k \cap \abs{L}) = 0.$$ Now, we apply Proposition $3.3.1$ of [@Lab] and we obtain a continuous map $\phi\colon \abs{L} \to \abs{L}$ such that
1. for all $A \in L$, $\phi(A) \subset A$;
2. $\phi(E_k \cap \abs{L}) \subset \abs{L} \setminus \bigcup \set{\mathrm{int}(A) | A \in L,\ \mathrm{dim}(A) > d}$;
3. for all $A \in L^d$, $$\label{third_axiom}
H^d(\phi(E_k \cap \abs{L}) \cap A) \leq C H^d(E_k \cap \abs{L})$$
where $C$ is a positive constant that depends only on $n$. When $k$ is big enough (depending on $\ell$), $H^d(E_k \cap \abs{L})$ becomes sufficiently small so that one can perform additional projections in the $d$-dimensional cells of $L$. Thus, the second axiom becomes $$\phi(E_k \cap \abs{L}) \subset \abs{L} \setminus \bigcup \set{\mathrm{int}(A) | A \in L,\ \mathrm{dim}(A) \geq d};$$ In particular, $$\label{phi_image}
\phi(E_k \cap \abs{L}) \cap U(L) \subset \bigcup \set{\mathrm{int}(A) | A \in L,\ \mathrm{dim}(A) \leq d-1}.$$ The sets $E \cup \Gamma$ and $\abs{L}$ are disjoint and closed so we can extend $\phi$ continuously on $E \cup \Gamma$ by $\phi = \mathrm{id}$. Observe that $\abs{\phi - \mathrm{id}} \leq \sqrt{n}\ell$ because $\phi$ preserves the cells of $L$. We can extend $\phi$ continuously on $\mathbf{R}^n$ in such that way that $\abs{\phi - \mathrm{id}} \leq \sqrt{n}\ell$. Let us show that $$\label{E_k_image}
\phi(E_k) \subset \abs*{L^{d-1}} \cup \bigcup_{j \in \mathbf{N}} \beta_j.$$ Remember that $E_k \subset \overline{B_0}$. We assume $\ell$ small enough so that $\sqrt{n}\ell \leq 1$ whence $\phi(E_k) \subset \overline{2B_0}$. For $x \in E_k$, we distinguish two cases. If $\mathrm{d}(x,E \cup \Gamma) \leq 3\sqrt{n}\ell$, then $\mathrm{d}(\phi(x),E \cup \Gamma) \leq 4\sqrt{n}\ell$ so $\phi(x) \in \bigcup_j \beta_j$ by (\[grille2\]). If $\mathrm{d}(x,E \cup \Gamma) \geq 3\sqrt{n}\ell$, then we have both $\mathrm{d}(x,E \cup \Gamma) \geq 2\sqrt{n}\ell$ and $\mathrm{d}(\phi(x),E \cup \Gamma) \geq 2\sqrt{n}\ell$ so (\[grille2\]) shows that $x \in U(L)$ and $\phi(x) \in U(L)$. By (\[phi\_image\]), we have thus $\phi(x) \in \abs{L^{d-1}}$.
Now, we are all set to introduce $$\beta_\infty = \mathbf{R}^n \setminus (E \cup \Gamma \cup \bigcup_{\abs{S} = d} \bigcap_S \overline{\beta_j}).$$ First, we justify that $\beta_\infty$ is open. It suffices to show that the family $$\left(\bigcap_S \overline{\beta_j}\right)_{\abs{S}=d}$$ is locally finite in $\mathbf{R}^n$. In step 1, we have built the family $(\beta_j)_{j \in \mathbf{N}}$ such that it is locally finite: for all $x \in \mathbf{R}^n$, there exists an open set $U$ containing $x$ such that the set $$S_0 = \set{j \in \mathbf{N} | U \cap \beta_j \ne \emptyset}$$ is finite. Let $S$ be a subset of $\mathbf{N}$ with cardinal $d$ such that $U$ meets $\bigcap_S \overline{\beta_j}$. Then for all $j \in S$, $U \cap \overline{\beta_j} \ne \emptyset$ and thus $U \cap \beta_j \ne \emptyset$ because $U$ is open. This means that $S \subset S_0$. We deduce that there exists only a finite number of subsets $S \subset \mathbf{N}$ of cardinal $d$ such that $U$ meets $\bigcap_S \overline{\beta_j} \ne \emptyset$. We conclude that $\beta_\infty$ is open. Observe that $\beta_\infty$ is disjoint from $E \cup \Gamma$ and that for all $S \subset \mathbf{N}$ of cardinal $d$, $$\label{gamma_simplex}
\beta_\infty \cap \bigcap_S \overline{\beta_j} = \emptyset.$$ In other words, for all $S \subset \mathbf{N}$ of cardinal $d+1$, the condition $\bigcap_S \beta_j \ne \emptyset$ implies $S \subset \mathbf{N}$. This means that the family $(\beta_j)_{j \in \mathbf{N} \cup \set{\infty}}$ does not induce additional $d$-simplexes. Finally, we would like $$\label{E_k_image2}
\phi(E_k) \subset \beta_{\infty} \cup \bigcup_{j \in \mathbf{N}} \beta_j.$$ This is where $(d-1)$-dimensional grid helps us a lot. According to (\[E\_k\_image\]), the condition (\[E\_k\_image2\]) holds if $\abs*{L^{d-1}} \setminus \beta_\infty \subset \bigcup_j \beta_j$, that is, for all $S \subset \mathbf{N}$ of cardinal $d$, $$\label{condition_grille}
\abs*{L^{d-1}} \cap \bigcap_S \overline{\beta_j} \subset \bigcup_j \beta_j.$$ We are going to see that a suitable translation of $K$ allows to assume that $\abs*{K^{d-1}}$ is disjoint from the intersection of boundaries $\bigcap_S \partial \beta_j$. Fix $S \subset \mathbf{N}$ of cardinal $d$. As $\bigcap_S \partial \beta_j$ is included in a finite union of $(n-d)$-spheres, we deduce that for all $(d-1)$-linear plane $P$, $$H^n(\bigcap_S \partial \beta_j + P) = 0.$$ In particular, $$H^n(\bigcap_S \partial \beta_j + (-\abs*{K^{d-1}})) = 0.$$ This means that for almost every $x \in \mathbf{R}^n$, $x + \abs*{K^{d-1}}$ is disjoint from $\bigcap_S \partial \beta_j$. There are only a countable number of subsets $S \subset \mathbf{N}$ of cardinal $d$ so we can find $x$ such that this is true for all of them. To simplify the notation, we assume that $x = 0$ and that (\[condition\_grille\] holds).
We are about to finish the proof. We define the domain $V_\gamma = \mathbf{N} \cup \set{\infty}$ and for $j \in V_\gamma$, we define the open set $\gamma_j = \phi^{-1}(\beta_j)$. Remember that $\phi = \mathrm{id}$ on $E \cup \Gamma$ so for all $j \in \mathbf{N} \cup \set{\infty}$, $$(E \cup \Gamma) \cap \gamma_j = (E \cup \Gamma) \cap \beta_j.$$ The family $\gamma$ covers $E_k \cup \Gamma$ because $(\beta_j)_{j \in \mathbf{N} \cup \set{\infty}}$ covers $E \cup \Gamma$ and $\phi(E_k)$. The family $((E \cup \Gamma) \cap \gamma_j)_{j \in V_\gamma}$ is a refinement of $\alpha$ because $(E \cup \Gamma) \cap \gamma_\infty = \emptyset$ and because for all $j \in \mathbf{N}$, $\gamma_j$ coincides with $\beta_j$ on $E \cup \Gamma$. Finally, for all $S \subset V_\gamma$ of cardinal $d+1$, the condition $$\bigcap_S \gamma_j \ne \emptyset$$ implies $\bigcap_S \beta_j \ne \emptyset$ and then by (\[gamma\_simplex\]), $S \subset \mathbf{N}$. By construction of $(\beta_j)_{j \in \mathbf{N}}$, we have $(E \cup \Gamma) \cap \bigcap_S \beta_j \ne \emptyset$ or equivalently, $$(E \cup \Gamma) \cap \bigcap_S \gamma_j \ne \emptyset$$ since $\gamma_j$ coincides with $\beta_j$ on $E \cup \Gamma$.
Existence of Plateau solutions
==============================
We solve two formulations of the Reifenberg Plateau problem. In the first one, we work in $X = \mathbf{R}^n$ and minimize $H^d(E)$ among Reifenberg competitors $E$. In the second one, we work in $X = \mathbf{R}^n \setminus \Gamma$ (that is, away from the boundary) and minimize $H^d(E \setminus \Gamma)$ among Reifenberg competitors $E$. In this second case, we do not require regularity on the boundary.
Direct method
-------------
We recall Corollary $3.4.1$ of [@Lab]. This is the same direct method as [@Ita1], [@Ita2] but this version allows to minimize the competitors on the boundary. Our working space is an open set $X$ of $\mathbf{R}^n$. We need a few assumptions on the boundary; see Definition $2.2.2$ (*$H^d$ regular sets*) and $3.1.9$ (*Lipschitz subset* in [@Lab].
\[direct\_method\]Fix a Lipschitz subset $\Gamma$ of $X$ which is $H^d$ regular. Let $\mathcal{C}$ be a class of closed subsets of $X$ such that $$m = \inf \set{H^d(E) | E \in \mathcal{C}} < \infty$$ and assume that for all $E \in \mathcal{C}$, for all sliding deformations $f$ of $E$ in $X$, $$m \leq H^d(f(E)).$$ Let $(E_k)$ be a minimizing sequence for $H^d$ in $\mathcal{C}$. Up to a subsequence, there exists a coral[^1] minimal set $E_\infty$ in $X$ such that $$H^d {\mathbin{\vrule height 1.6ex depth 0pt width 0.13ex\vrule height 0.13ex depth 0pt width 0.8ex}}E_k \rightharpoonup H^d {\mathbin{\vrule height 1.6ex depth 0pt width 0.13ex\vrule height 0.13ex depth 0pt width 0.8ex}}E_\infty.$$ where the arrow $\rightharpoonup$ denotes the weak convergence of Radon measures in $X$. In particular, $H^d(E_\infty) \leq m$.
In the works of Reifenberg, the Hausdorff limit of a minimizing sequence is a competitor but the area is not lower semicontinuous. Reifenberg worked with a compact coefficient group to build an alternative minimizing sequence. With weak limits, the lower semicontinuity follows from the previous proposition. Moreover, we proved in the last section that the limit is a competitor. Thus, there not much work to do.
Applications
------------
\[sol1\]Fix a Lipschitz subset $\Gamma$ of $\mathbf{R}^n$ which is $H^d$ regular and fix a subgroup $L$ of $H_{d-1}(\Gamma)$. We assume that $$m = \inf \set{H^d(E) | E \ \text{Reifenberg competitor}} < \infty$$ and that there exists a compact set $C \subset \mathbf{R}^n$ such that $$m = \inf \set{H^d(E) | E \ \text{Reifenberg competitor},\ E \subset C}.$$ Then there exists a Reifenberg competitor $E \subset C$ such that $H^d(E) = m$.
We work in $X = \mathbf{R}^n$ and we consider the class $$\mathcal{C} = \set{E | E \ \text{is a Reifenberg competitor}}.$$ By Lemma \[rei\_image\], the class $\mathcal{C}$ is preserved by sliding deformations in $\mathbf{R}^n$ so it satisfies the requirement of Corollary \[direct\_method\]. Let $(E_k)$ be a minimizing sequence of $\mathcal{C}$ such that for all $k$, $E_k \subset C$. According to Corollary \[direct\_method\], there exists a coral set $E_\infty$ of $\mathbf{R}^n$ such that $$H^d {\mathbin{\vrule height 1.6ex depth 0pt width 0.13ex\vrule height 0.13ex depth 0pt width 0.8ex}}E_k \rightharpoonup H^d {\mathbin{\vrule height 1.6ex depth 0pt width 0.13ex\vrule height 0.13ex depth 0pt width 0.8ex}}E_\infty.$$ We prove that $E_\infty$ is a Reifenberg competitor. First, we show that $E_\infty$ is a compact subset of $C$. Observe that $\mathbf{R}^n \setminus C$ is an open set and that by lower semicontinuity, $$H^d(E_\infty \setminus C) \leq \liminf_k H^d(E_k \setminus C) = 0.$$ This proves that the support of $H^d {\mathbin{\vrule height 1.6ex depth 0pt width 0.13ex\vrule height 0.13ex depth 0pt width 0.8ex}}E_\infty$ is included in $C$. As $E$ is coral, $E$ is a subset of $C$ and therefore compact. Next, we appy Lemma \[rei\_limit\] to the set $E_\infty$. For all open set $V$ containing $E_\infty \cup \Gamma$, $$\begin{aligned}
\limsup_k H^d(E_k \setminus V) &= \limsup_k H^d(E_k \cap C \setminus V)\\
&\leq H^d(E_\infty \cap C \setminus V)\\
&\leq 0.\end{aligned}$$ We conclude that $E_\infty$ is a Reifenberg competitor. Finally, we show that $H^d(E_\infty) = m$. As $E_\infty$ is a Reifenberg competitor, we have of course $H^d(E_\infty) \geq m$. The fact that $H^d(E_\infty) \leq m$ was already observed in Corollary \[direct\_method\].
The next theorem is similar to Theorem $1.3$ of [@Fang] (which is based on Feuvrier’s construction) and Theorem $3.4$ of [@Ita4] (which is based on weak limits of minimizing sequences). However, we have not dealt with elliptic integrands yet.
\[sol2\]Fix a closed set $\Gamma$ of $\mathbf{R}^n$ and a subgroup $L$ of $H_{d-1}(\Gamma)$. We assume that $$m = \inf \set{H^d(E \setminus \Gamma) | E \ \text{Reifenberg competitor}} < \infty$$ and that there exists a compact set $C \subset \mathbf{R}^n$ such that $$m = \inf \set{H^d(E \setminus \Gamma) | E \ \text{Reifenberg competitor},\ E \subset C}.$$ Then there exists a Reifenberg competitor $E \subset C$ such that $H^d(E \setminus \Gamma) = m$.
If $\Gamma$ is compact and $H^d(\Gamma) < \infty$, this amounts to minimizing $H^d(E)$ among Reifenberg competitors containing $\Gamma$.
We work in $X = \mathbf{R}^n \setminus \Gamma$ (away from the boundary) and we consider the class $$\mathcal{C} = \set{E \setminus \Gamma | E \ \text{is a Reifenberg competitor}}.$$ By Lemma \[rei\_image\], the class $\mathcal{C}$ is preserved by sliding deformations in $X$ so it satisfies the requirement of Corollary \[direct\_method\]. Let $(E_k)$ be a sequence of Reifenberg competitor such that $(E_k \setminus \Gamma)$ is a minimizing sequence of $\mathcal{C}$ and for all $k$, $E_k \subset C$. According to Corollary \[direct\_method\], there exists a coral set $S_\infty$ of $X$ such that $$H^d {\mathbin{\vrule height 1.6ex depth 0pt width 0.13ex\vrule height 0.13ex depth 0pt width 0.8ex}}(E_k \setminus \Gamma) \rightharpoonup H^d {\mathbin{\vrule height 1.6ex depth 0pt width 0.13ex\vrule height 0.13ex depth 0pt width 0.8ex}}S_\infty \ \text{in $X$}.$$ We prove that there exists a Reifenberg competitor $E_\infty \subset C$ such that $S_\infty = E_\infty \setminus \Gamma$. First, we justify that $S_\infty \subset C$. Observe that $X \setminus C$ is an open set of $X$ and that by lower semicontinuity, $$H^d(S_\infty \setminus C) \leq \liminf_k H^d((E_k \setminus \Gamma) \setminus C) = 0.$$ As a consequence, the support of $H^d {\mathbin{\vrule height 1.6ex depth 0pt width 0.13ex\vrule height 0.13ex depth 0pt width 0.8ex}}S_\infty$ in $X$ is included in $C$. As $S_\infty$ is coral in $X$, $S_\infty$ is a subset of $C$. Now, let $$\begin{aligned}
E_\infty &= (S_\infty \cup \Gamma) \cap C\\
&= S_\infty \cup (\Gamma \cap C).\end{aligned}$$ The set $S_\infty$ is closed in $X$ so $S_\infty \cup \Gamma$ is closed in $\mathbf{R}^n$ and $E_\infty$ is compact. We appy Lemma \[rei\_limit\] to the set $E_\infty$. For all open set $V$ containing $E_\infty \cup \Gamma$, $C \setminus V$ is a compact subset of $X$ so $$\begin{aligned}
\limsup_k H^d(E_k \setminus V) &= \limsup_k H^d(E_k \cap C \setminus V)\\
&= \limsup_k H^d((E_k \setminus \Gamma) \cap C \setminus V)\\
&\leq H^d(E_\infty \cap C \setminus V)\\
&\leq 0.\end{aligned}$$ In conclusion, $E_\infty$ is a Reifenberg competitor and $S_\infty = E_\infty \setminus \Gamma \in \mathcal{C}$. Finally, we show that $H^d(S_\infty) = m$. As $S_\infty \in \mathcal{C}$, we have of course $H^d(S_\infty) \geq m$. The fact that $H^d(S) \leq m$ has already been observed in Corollary \[direct\_method\].
[AAA]{} G. David, Should we solve Plateau’s problem again? Advances in analysis: the legacy of Elias M. Stein, 108-145, Princeton Math. Ser., 50, Princeton Univ. Press, Princeton, NJ, 2014.
C. De Lellis, F. Ghiraldin, and F. Maggi, A direct approach to Plateau’s problem. J. Eur. Math. Soc. (JEMS) 19 (2017), no. 8, 2219-2240.
G. De Philippis, A. De Rosa, and F. Ghiraldin, A direct approach to Plateau’s problem in any codimension. Adv. Math. 288 (2016), 59-80.
G. De Filippis, A. De Rosa, and F. Ghiraldin, Rectifiability of varifolds with locally bounded first variation with respect to anisotropic surface energies. Comm. Pure Appl. Math. 71 (2018), no. 6, 1123-1148.
C. De Lellis, A. De Rosa and F. Ghiraldin, A direct approach to the anisotropic Plateau problem. Adv. Calc. Var. 12 (2019), no. 2, 211–223.
G. De Filippis, A. De Rosa, and F. Ghiraldin, Existence results for minimizers of parametric elliptic functionals.
G. David and S. Semmes, Uniform rectifiability and quasiminimizing sets of arbitrary codimension. Mem. Amer. Math. Soc. 144 (2000), no. 687, viii+132 pp.
S. Eilenberg and N. Steenrod, Foundations of algebraic topology. Princeton University Press, Princeton, New Jersey, 1952. xv+328 pp.
V. Feuvrier, Un résultat d’existence pour les ensembles minimaux par optimisation sur des grilles polyédrales. PhD thesis, Université Paris Sud (Orsay).
Yangqin Fang, Existence of minimizers for the Reifenberg plateau problem. Ann. Sc. Norm. Super. Pisa Cl. Sci. (5) 16 (2016), no. 3, 817-844.
C. Labourie, Weak limits of quasiminimizing sequences. 2020. Available at arXiv:2002.08876
N. Nakauchi, On free boundary Plateau problem for general-dimensional surfaces. Osaka J. Math. 21 (1984), no. 4, 831-841.
E. R. Reifenberg, Solution of the Plateau Problem for $m$-dimensional surfaces of varying topological type. Acta Math. 104 1960 1-92.
[^1]: A set $E \subset X$ is coral in $X$ if $E$ is the support of $H^d {\mathbin{\vrule height 1.6ex depth 0pt width 0.13ex\vrule height 0.13ex depth 0pt width 0.8ex}}E$ in $X$. Equivalently, $E$ is closed in $X$ and for all $x \in E$ and for all $r > 0$, $H^d(E \cap B(x,r)) > 0$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A Multi-Phase Transport (AMPT) model has been shown to describe experimental data well, such as the bulk properties of particle spectra and elliptic anisotropy ($v_{2}$) in heavy ion collisions. Recent studies have shown that AMPT describes the $v_{2}$ data in small system collisions as well. In these proceedings, we first investigate the origin of the mass ordering of identified hadrons $v_{2}$ in heavy ion as well as small system collisions. We then study the production mechanism of the charm $v_{2}$ in light of the escape mechanism for the light quark $v_{2}$.'
address:
- '$^{1}$College of Science, Wuhan University of Science and Technology, Wuhan, Hubei 430065, China '
- '$^{2}$Department of Physics and Astronomy, Purdue University, West Lafayette, IN 47907, USA'
- '$^{3}$Department of Physics, East Carolina University, Greenville, NC 27858, USA'
author:
- 'Hanlin Li$^{1,2}$, Zi-Wei Lin$^{3}$, Fuqiang Wang$^{2}$'
title: 'The physics mechanisms of light and heavy flavor $v_{2}$ and mass ordering in AMPT'
---
Introduction\[sec:intro\]
=========================
Relativistic heavy ion collisions aim to create the quark-gluon plasma (QGP) and study its properties at the extreme conditions of high temperature and energy density [@Arsene:2004fa; @Back:2004je]. The collective flow as a soft probe is often used to study the QGP properties in experimental and theoretical investigations. Both hydrodynamics [@Heinz:2013th; @SongMultistrange]and transport theory [@Lin:2004en]can describe the bulk data in heavy ion collisions. For example, the string melting version of A Multi-Phase Transport (AMPT) model [@Lin:2004en; @Lin:2001zk] reasonably reproduces particle yields, $p_{\perp}$ spectra, and $v_2$ of low-$p_{\perp}$ pions and kaons in central and mid-central Au+Au collisions at 200A GeV and Pb+Pb collisions at 2760A GeV [@Lin:2014tya]. The small system data can be also satisfactorily described by AMPT [@Bzdak:2014dia].
Recent studies have shown that light parton $v_{2}$ is mainly generated by the anisotropic parton escape from the collision zone and hydrodynamics may play only a minor role [@He:2015hfa; @Lin:2015ucn]. It suggests that the mass ordering of $v_{2}$, commonly believed as a signature of hydrodynamics, may arise from other mechanisms. In these proceedings, we first investigate the physic mechanisms of mass ordering in AMPT [@Li:2016flp; @Li:2016ubw]. We then study the production mechanism of the charm $v_{2}$ in light of the escape mechanism for the light quark $v_{2}$.
Model details and analysis method \[sec:setup1\]
================================================
We employ the string melting version of AMPT [@Lin:2004en; @Lin:2001zk] in our study. The model consists of four components: fluctuating initial conditions, parton elastic scatterings, quark coalescence for hadronization, and hadronic interactions. The partons interaction are modeled by Zhang’s parton cascade (ZPC) [@Zhang:1997ej]. We use Debye screened differential cross-section $d\sigma/dt\propto\alpha_s^2/(t-\mu_D^2)^2$ [@Lin:2004en], with strong coupling constant $\alpha_s=0.33$ and Debye screening mass $\mu_D=2.265$/fm (the total cross section is then $\sigma=3$ mb) for all AMPT simulation in our work. After partons stop interaction, a simple quark coalescence model is applied to combine two nearest partons into a meson and three nearest partons into a baryon(or antibaryon). The hadronic interactions are described by the ART model [@B:art]. We terminate the hadronic interactions at a cutoff time.
We simulate three collision systems: [Au+Au]{} collisions with [$b=6.6$-8.1 fm]{} at the nucleon-nucleon center-of-mass energy ${\sqrt{s_{_{\rm NN}}}}=200$ GeV, [$d$+Au]{} collisions with [$b=0$ fm]{} at ${\sqrt{s_{_{\rm NN}}}}=200$ GeV, and [$p$+Pb]{} collisions with [$b=0$ fm]{} at ${\sqrt{s_{_{\rm NN}}}}=5$ TeV by AMPT. We analyze the momentum-space azimuthal anisotropy of partons in the final state before hadronization, of hadrons right after hadronization before hadronic rescatterings take place, and of freeze-out hadrons in the final state. The momentum anisotropy is characterized by Fourier coefficients [@Voloshin:1994mz] $$v_2=\langle \cos 2(\phi -\psi_2^{(r)}) \rangle,$$ where $\phi$ is the azimuthal angle of the particle (parton or hadron) momentum. The $\psi_2^{(r)}$ is the harmonic plane of each event from its initial configuration of all partons. All results shown are for particles within pseudo-rapidity window $|\eta|<1$.
Mass ordering of $v_2$ \[sec:setup\]
====================================
![(Color online) Mid-rapidity ($|\eta|<1$) $v_2$ of $\pi$, K, p($\bar{p}$) and charged hadrons ($h^\pm$) at $0.7<p_{\perp}<0.8$ GeV/$c$ at different stages of system evolution in $b=6.6$-8.1 fm Au+Au (left) and $b=0$ fm [$d$+Au]{} (right) collisions at ${\sqrt{s_{_{\rm NN}}}}$=200 GeV by AMPT: right after coalescence hadronization and before hadronic rescattering (initial $v_2$ of primordial hadrons), hadron initial $v_2$ including decays, and after hadronic rescattering at freezeout (hadron final $v_2$). \[fig:fig3\_s\]](fig3){width="30pc"}
Considering that observed pions and protons are made of only light constituent quarks, the difference between their $v_2$ should come from the hadronization process and/or hadronic rescattering. We first study the primordial hadrons right after hadronization, before any decay and hadronic rescatterings take place. Figure \[fig:fig1\](a) shows the $v_2$ of primordial charged pion, kaon, and (anti)proton as a function of $p_{\perp}$ (solid curves). It demonstrates that the mass ordering of $v_{2}$ at low $p_{\perp}$ in AMPT comes from the dynamic in coalescence. The dynamical “selections” of constituent quarks into pions, kaons, and protons are manifest in the constituent quark $v_2$ distributions shown by the dashed curves in Fig. \[fig:fig1\](a), plotted at the respective [*hadron*]{} $p_{\perp}$.
Figure \[fig:fig1\](b) shows the $v_2$ results by $\phi$-randomized AMPT [@He:2015hfa] for primordial hadrons right after coalescence hadronization at the corresponding constituent quark $v_2$’s. No hydrodynamic anisotropic flow is present in the $\phi$-randomized case [@He:2015hfa], however, mass splitting is still present. It implies that the mass splitting is mainly due to kinematics in the coalescence process. It is therefore not a unique signature of collective anisotropic flow or hydrodynamics.
After hadronization, particles undergo decay and rescattering. We evaluate the $v_2$ of hadrons before hadronic rescattering but including effects of resonance decays by setting the cut-off time to be 0.6 [fm]{}/c. The results are shown in Fig. \[fig:fig2\](a) by the dashed curves. The decay product $v_2$ is usually smaller than their parent $v_2$ [@Li:2016ubw]. By including decay products, the hadron $v_2$ is reduced from that of primordial hadrons (solid curves in Fig. \[fig:fig1\](a)).
The $v_2$ values before hadronic rescattering (including resonance decay effects) can be considered as the initial $v_2$ for the hadronic evolution stage. The final-stage freezeout hadron $v_2$’s (also including decay daughters) are shown in Fig. \[fig:fig2\](a) by the solid curves. Pion $v_2$ increases while proton $v_2$ decreases from before to after hadronic rescattering. This may be due to interactions between pions and protons, after which they tend to flow together at the same velocity. Thus, the same-velocity pions and protons (i.e. small $p_{\perp}$ pions and large $p_{\perp}$ protons) will tend to have the same anisotropy. It will yield lower $v_2$ for the protons and higher $v_2$ for the pions at the same $p_{\perp}$ value. This should happen whether or not there is a net gain in the overall charged hadron $v_2$, which depends on the initial configuration geometry from which the hadronic evolution begins [@Li:2016ubw]. The similar results in $d$+Au collisions as shown in Fig. \[fig:fig2\](b) [@Li:2016ubw].
To summarize the origin of $v_2$ mass splitting, we plot the $v_2$ of pions, kaons, and protons within $0.7<p_{\perp}<0.8$ GeV/$c$, a typical region for different stages of the collision system evolution as shown in Fig. \[fig:fig3\_s\]. The evolution stages contain: (i) right after coalescence hadronization including only primordial hadrons (labeled “prim.”); (ii) right after coalescence hadronization but including decay products (labeled “w/ decay”); and (iii) at final freezeout (labeled “w/ rescatt. w/ decay). We can see the mass ordering actually comes from interplay of several physics effects; it comes from coalescence, and more importantly, from hadronic rescattering process.
Charm $v_2$ mechanism\[sec:setup\]
==================================
In this section we discuss the production mechanism of charm quark flow in pPb. We compare the light and charm quark freeze out $v_2$, integrated over all $p_{\perp}$ at the same proper time. This is shown in the left panel of Fig. \[fig:fig4\_c\]. The charm $v_{2}$ is systematically lower than light quark $v_{2}$. This may suggest that the charm quark is not as thermalized as light quarks. The right panel of Fig. \[fig:fig4\_c\] shows $\phi$-randomized results. The charm and light quark $v_2$ are similar suggesting a common escape mechanism.
Summary\[sec:summary\]
======================
In these proceedings, we investigate the origin of mass splitting of identified hadron $v_{2}$. The mass splitting is due to coalescence and, more importantly, hadronic rescattering. It is therefore not a unique signature of hydrodynamics. We also study the development of charm $v_{2}$ and found the escape mechanism to be the major contribution in pPb similar to light quarks.
This work is supported in part by US Department of Energy Grant No. DE-FG02-88ER40412 (LH,FW,WX) and No. DE-FG02-13ER16413 (DM). HL acknowledges financial support from the China Scholarship Council.
References {#references .unnumbered}
==========
[9]{}
I. Arsene et al. [BRAHMS Collaboration]{}, [*Nucl. Phys.*]{} A [**757**]{}, 1 (2005).
B. Back et al. PHOBOS Collaboration [*Nucl. Phys.*]{} A [**757**]{}, 28 (2005).
U. Heinz and R. Snellings, Ann. [*Rev. Nucl. Part. Sci.* ]{} [**63**]{}, 123 (2013).
Zhu X, Meng F, Song H, and Liu Y X [*Phys. Rev.*]{} C [**91**]{}, 034904 (2015).
Z.-W. Lin, C. M. Ko, B.-A. Li, B. Zhang and S. Pal , [*Phys.Rev.*]{} C [**72**]{}, 064901 (2005).
Z.-W. Lin and C. M. Ko, [*Phys.Rev.*]{} C [**65**]{}, 034904 (2002).
Z.-W. Lin, [*Phys.Rev.*]{} C [**90**]{}, 014904 (2014).
A. Bzdak and G.-L. Ma, [*Phys.Rev.*]{} Lett. [**113**]{}, 252301 (2014).
L. He, T. Edmonds, Z.-W. Lin, F. Liu, D. Molnar, and F. Wang, [*Phys. Lett.*]{} B [**735**]{}, 506 (2016).
Z. -W. Lin, L. He, T. Edmonds, F. Liu, D. Molnar, and F. Wang [*arXiv:1512.06465*]{} (2015).
H. Li, L. He, Z. W. Lin, D. Molnar, F. Wang and W. Xie, [*Phys.Rev.*]{} C [**93**]{}, 051901(R) (2016).
H. Li, L. He, Z. W. Lin, D. Molnar, F. Wang and W. Xie, [*arXiv:1604.07387*]{} (2016).
B. Zhang,[*Comput.Phys.Commun.*]{} [**109**]{}, 193 (1998).
B. A. Li and C. M. Ko, [*Phys. Rev.*]{} C [**52**]{}, 2037 (1995).
S. Voloshin and Y. Zhang, [*Z.Phys.*]{} [**C70**]{},[665]{} (1996).
| {
"pile_set_name": "ArXiv"
} |
INTRODUCTION
============
The experimental realization of Bose-Einstein condensates [@And; @Dav; @Brad] in dilute alkali gases has stimulated great interest in the study of vortices in these systems. In the absence of experimental detection of a condensate containing a vortex, several authors have addressed the important question of the stability of such a vortex [@Baym96; @Dalf; @Lundh; @Dodd; @Rokh; @Fetter98; @SF98; @Svid98; @Pu; @IsoJ99; @Garcia99; @Feder99; @Butts]. The simplest criterion is energetic stability, determined by comparing the total energy of a condensate with and without a vortex, based on the time-independent Gross-Pitaevskii equation. Separately, numerical and analytical studies of the Bogoliubov equations for the excitations of a condensate with a vortex found a normal mode with negative energy and positive normalization, suggesting a dynamical instability [@Dodd; @Rokh; @Fetter98; @IsoJ99]. Moreover, in the limit of large particle number, there is a metastable regime, for which the free energy of the vortex condensate develops a local minimum at the center of the trap, stabilizing the vortex against small lateral displacements [@Svid98; @Feder99]. In the Thomas-Fermi limit, the onset of metastability coincides with $|\omega_a|$, implying a close connection between the two phenomena [@Svid98].
In the present work, we consider a small cylindrically symmetric condensate with a central vortex line along the $z$ axis; this weak-coupling limit allows a perturbation expansion of the Gross-Pitaevskii (GP) equation. The additional energy of the condensate with a vortex can be compensated by rotating the whole condensate with an angular velocity $\Omega_c$ [@Dalf]. The Bogoliubov equations for the same system have an anomalous mode with negative frequency $\omega_a$. To first order in the (small) interaction parameter, $|\omega_a|=\Omega_c$ [@Fetter98; @Butts]. To understand the possible connection between these two types of instabilities, we here evaluate the second-order corrections to $\Omega_c$ and $\omega_a$. Furthermore, the metastable frequency $\Omega_m$ computed for the weak-coupling limit agrees to first order with the modulus of the anomalous mode, just as in the Thomas-Fermi limit. This result indicates that the instability with respect to microscopic oscillations and the onset metastability are closely related, for the system becomes dynamically stable when rotated faster than the metastable frequency.
The following section uses the GP equation for a condensate with a central vortex to determine the critical rotation frequency $\Omega_c$ to second order. In the third section, a similar perturbation expansion of the Bogoliubov equations for the anomalous mode yields the second-order correction to the corresponding eigenvalue of this mode $\omega_a$. The fourth section is devoted to determining the metastable frequency $\Omega_m$ variationally, and the conclusion discusses the implications of our results.
CRITICAL angular velocity
=========================
Consider a condensate containing $N$ particles in a harmonic trap with radial and axial frequencies $\omega_\perp$ and $\omega_z$. The interparticle interactions are characterized by a positive $s$-wave scattering length $a>0$. The starting point for the perturbation theory is the time-independent Gross-Pitaevskii equation [@Pit61; @Gross61] $$(\,H_0 + 4 \pi \gamma\, |\Psi|^2\,)\Psi = \mu \,\Psi \, ,$$ with $\mu$ the chemical potential. Here, $$H_0 =\case{1}{2} \bigg[-\nabla_\perp^2 + r^2+
\lambda\bigg(-{\partial^2\over
\partial z^2}+z^2\bigg)\bigg] \,$$ is the Hamiltonian for the noninteracting condensate, expressed in dimensionless units (the radial and axial coordinates are scaled with the radial and axial oscillator lengths $d_{\perp} = \sqrt{\hbar/m \omega_{\perp}}$ and $d_z =
\sqrt{\hbar/m \omega_z}$, and frequencies are scaled with the radial trap frequency $\hbar\omega_\perp$). The condensate wave function $\Psi$ is normalized to unity, $\gamma \equiv Na/d_z$ is the small interaction parameter and $\lambda=\omega_z/\omega_\perp$ represents the geometrical asymmetry of the cylindrical trap.
The normalized eigenfunctions for the noninteracting condensate are taken as a product of the one-dimensional axial oscillator state $\varphi_l(z)$ and the two-dimensional wave function for a $q$-fold quantized vortex $\chi_{n+q,n}(r,\phi)$ at the center of the trap, so that \[compare Eqs. (\[varphi\]) and (\[chi\]) in the appendix\] $$\psi_{n+q,n,l} (\vec{r}) = \chi_{n+q,n}(r,\phi)\varphi_l(z) \, ,$$ where $\psi_{q00}$ is the lowest energy state for a condensate with a given quantized circulation $q$. The noninteracting single-particle states have an energy $\epsilon_{n+q,n,l}^{(0)}
= 2n +q +1 +(l+\frac{1}{2})\lambda$.
The thermodynamic criterion for stability of a rotating condensate with a $q$-fold vortex is the vanishing of the free-energy difference $\Delta F_q$ between the free energy of the condensate with and without the vortex. The free energy for a $q$-fold vortex state in the frame rotating with angular velocity $\Omega$ is $F_q
= E_q -\Omega q N$, where $E_q$ is the energy of the condensate and $q$ is the angular-momentum quantum number. Setting $F_q-F_0
=0$ gives the critical rotation frequency
$$\Omega_c = \frac{E_q -E_0}{N q} \, .$$
To construct the perturbation theory, the chemical potential and the condensate wave function are expanded in the interaction parameter $\mu \approx
\mu^{(0)}+\gamma\mu^{(1)} + \cdots$ and $\Psi \approx
\Psi^{(0)}+\gamma\Psi^{(1)} + \cdots\,$. The first-order correction to the chemical potential of the condensate is easily seen to be $\mu^{(1)} = 4\pi\,\langle \Psi^{(0)} |\,|\Psi^{(0)}|^2
\,|\Psi^{(0)}
\rangle$, and the thermodynamic relation $\mu =
\partial E/\partial N$ then gives the first-order correction to the condensate energy (note that $\gamma \propto N$) $$E^{(1)} = \case{1}{2} N\mu^{(1)} =2 \pi N \langle \Psi^{(0)}
|\,|\Psi^{(0)}|^2
\,|\Psi^{(0)}
\rangle \, .\label{eq:E1}$$ For a $q$-fold vortex with $\Psi^{(0)}=\psi_{q00}$, direct evaluation of the matrix element gives $$\frac{E^{(1)}_q}{N} = \case{1}{2} \mu_q^{(1)}= \frac{(2q)!}{(q!)^2\,
2^{2q}\,\sqrt{2 \pi}\,} \,.\label{eq:mu1}$$ The first-order correction to the energy decreases with increasing $q$, and the corresponding critical rotation frequency
$$\Omega_c \approx 1-\frac{E_0^{(1)}-E_q^{(1)}}{Nq}\,\gamma$$
increases with increasing $q$. Figure 1 shows the first-order approximation to the thermodynamic critical angular velocity as a function of the interaction parameter $\gamma$ for the first few values of $q$. Since the resulting $\Omega_c$ is always smallest for $q=1$, a singly quantized vortex will appear first, in which case $\Omega_c \approx
1-\frac{1}{2}\gamma/\sqrt{2\pi}$. We consider only this case ($q=1$) in determining the second-order correction to the critical angular velocity.
\[q\]
The second-order correction to the energy of the condensate can be found by writing $\Psi^{(1)}$ as a sum over the unperturbed eigenfunctions of the appropriate symmetry (this form maintains the normalization to first order) $$\Psi^{(1)}={\sum_{nl}}'c_{nl}\,
\chi_{n+q,n}\,\varphi_{2l}\,,\label{eq:Psi1a}$$ where the prime on the sum indicates that it runs over all nonnegative values of $\{n,l\}$, omitting the single state $ \{0,0\}$. Substitution into the GP equation and use of the orthogonality of the unperturbed solutions determine the coefficients $$\begin{aligned}
c_{nl}&=& -\frac{2\pi}{n+\lambda l}\, \langle
\psi_{n+q,n,2l}|\,|\psi_{q00}|^2\,|\psi_{q00}\rangle \nonumber \\
&=&-\frac{2\pi}{n+\lambda l}\, I_{l}\,J^{nq}_{qqq}
\label{eq:Psi1b}\end{aligned}$$ in terms of definite integrals discussed in the appendix. For example, the axial integration over the product of four harmonic-oscillator eigenfunctions yields $$I_l=\int_{-\infty}^\infty
dz\,\varphi_{2l}(z)\,\varphi_0(z)^3=
\frac{(-1)^l}{l!\,2^{2l}}\,\sqrt{\frac{(2l)!}{2\pi}}\,,\label{eq:Il}$$ with a similar but more complicated expression $J^{nq}_{ijk}$ for the integral over the radial eigenfunctions (they involve Laguerre polynomials).
The corresponding second-order correction to the condensate energy follows immediately from the second-order chemical potential
$$\begin{aligned}
\frac{E_q^{(2)}}{N} &=& \case{1}{3} \mu_q^{(2)} = 16\pi^2 \,{\sum_{nl}}'
\,\frac{|\langle \psi_{n+q,n,2l}| \,|\psi_{q00}|^2 \,|\psi_{q00}
\rangle|^2}
{\mu^{(0)}_q -\epsilon_{n+q,n,2l}^{(0)}} \nonumber \\
&=&
-8\pi^2\,{\sum_{nl}}'\,\frac{(I_{l}\,J^{n,q}_{qqq})^2}{n+\lambda
l}\, ,\label{eq:E2}\end{aligned}$$
where $ \mu^{(0)}_q = \epsilon_{q00}^{(0)}= q + 1+
\frac{1}{2} \lambda$. The angular quantum number $q$ is one for the singly quantized vortex or zero for the nonvortex ground state.
In contrast to the first-order correction $E_q^{(1)}$ in Eq. (\[eq:E1\]), the second-order correction now depends explicitly on the asymmetry parameter $\lambda$. The appendix describes a systematic way of writing the matrix elements needed to carry out the infinite sums, and the difference between the second-order energy contributions for the vortex and nonvortex state $\Delta E^{(2)}$ yields the second-order correction to the critical rotation frequency
$$\Omega^{(2)}_c(\lambda) =\frac{1}{16 \pi}\,{\sum_{nl}}'
\frac{1}{n+\lambda
\,l}\,\frac{12 +3 n^2 -n^3}{\,2^{2n}}\,
\frac{(2l)!}{2^{4l}\, (l!)^2} \, .$$
Consequently the second-order approximation to the thermodynamic critical rotation frequency (determined from the GP equation) is
$$\Omega_c
= 1 - \frac{1}{2\,\sqrt{2 \pi}}\,\gamma
+\Omega^{(2)}_c(\lambda)\, \gamma^2 \,+\cdots .$$
Table I lists the second-order contribution $\Omega_c^{(2)}(\lambda)$ for various asymmetry parameters $\lambda$. This contribution is always positive, which counteracts the (negative) first-order contribution. For large $\lambda$ (a disk-shaped condensate), $\Omega_c^{(2)}$ becomes constant (from the terms with $l=0$), whereas for small $\lambda$ (a cigar-shaped condensate), it grows like $\Omega_c^{(2)} \propto
1/\lambda$ (from the terms with $n=0$). The resulting thermodynamic critical angular velocity for several $\lambda$ is shown in Fig. 2, illustrating the increasing importance of the second-order term for small $\lambda$. For comparison, this figure also includes the critical angular velocity for a $q=1$ vortex with only the first-order correction taken into account.
[ddddd]{} & $\lambda$ & $\Omega_c^{(2)}$ & $-\omega_a^{(2)}$&\
&$10^4$ & 0.0805 & 0.0508 &\
&$10$ & 0.0849 & 0.0546 &\
&$2 \sqrt{2}$& 0.0954 & 0.0637 &\
&$1$ & 0.1196 & 0.0851 &\
&$10^{-1}$ & 0.4227 & 0.3614 &\
&$10^{-2}$ & 3.4033 & 3.0933 &\
&$10^{-4}$ & 331.15 & 303.53 &
\[l\]
ANOMALOUS MODE
==============
The preceding section dealt solely with the GP equation for the condensate wave function $\Psi$ and the chemical potential $\mu$, comparing the energy for a condensate with and without a vortex. We now turn to the Bogoliubov equations [@Bog47] that describe the small-amplitude excitations of a condensate containing a singly quantized vortex:
$$\begin{aligned}
\left(H_0 - \mu +8\pi \gamma\,|\Psi|^2\right)u_j -4 \pi \gamma\,
(\Psi)^2 v_j &=& \omega_j u_j \, , \\
-4 \pi \gamma\,
(\Psi^*)^2 u_j +\left(H_0 - \mu +8\pi \gamma\,|\Psi|^2\right)v_j
&=& - \omega_j v_j \, ,\end{aligned}$$
where $u_j$ and $v_j$ are the normal-mode amplitudes and $\omega_j$ is the corresponding frequency. In zero order, the condensate wave function $\Psi^{(0)}$ for a singly quantized vortex is given by $\psi_{100} = \chi_{10}\varphi_0$ and the associated chemical potential is $\mu^{(0)}=2
+\frac{1}{2}\lambda$. For physical solutions, the normal-mode amplitudes have positive normalization $\int dV (|u_j|^2-|v_j|^2)
= 1$. Here, we focus on the anomalous mode, which has a negative eigenvalue $\omega_a <0$. Hence the creation of quasiparticles in this mode lowers the energy relative to that of the condensate, implying a possible instability [@Dodd; @Rokh; @Fetter98; @IsoJ99; @Fetter99].
A noninteracting condensate with a singly quantized vortex is unstable because the particles in the condensate with $\psi_{100}$ can make a transition to the nonrotating condensate with $ \psi_{000}$, giving up an energy $1$ in our dimensionless units. In the presence of interactions, however, the description becomes more complicated, and the eigenfunctions of the Bogoliubov equation include both $u_a$ and $v_a$, even in zeroth order. At this level, the Bogoliubov equations for the anomalous mode become
$$\begin{aligned}
\left(H_0 - \mu^{(0)} \right)\,u_a^{(0)} &=&
\omega_a^{(0)} u_a^{(0)} \, , \\
\left(H_0 - \mu^{(0)} \right)\,v_a^{(0)}
&=& - \omega_a^{(0)} v_a^{(0)} \, ,\end{aligned}$$
where the preceding argument suggests that $u_a^{(0)}\propto
\psi_{000}= \chi_{00}\varphi_0$ characterizes the nonrotating vortex-free condensate, so that $\omega_a^{(0)} = -1$. Since the full Bogoliubov equation for $u_a$ contains the coupling term $(\Psi)^2\,v_a\propto
e^{2i\phi}v_a$, it is natural to assume that $v_a\propto
e^{-2i\phi}$, and the unperturbed state $v_a^{(0)}\propto
\psi_{020}=\chi_{02}\varphi_0$ has the correct energy to satisfy the remaining zero-order Bogoliubov equation. To ensure the proper normalization, we merely take $$u_a^{(0)}= \cosh\theta\,\chi_{00} \varphi_0 \, \, \hbox{and} \,
\, v_a^{(0)}= \sinh\theta\,\chi_{02}\varphi_0 \, ,\label{eq:zero}$$ where the parameter $\theta$ can only be determined by including the higher-order terms in the perturbation expansion.
The first-order terms in the Bogoliubov equations become $$\begin{aligned}
\left(H_0 - \mu^{(0)}\right) u_a^{(1)} &+&\left( 8\pi
\,|\Psi^{(0)}|^2-\mu^{(1)}\right)u_a^{(0)} \nonumber \\ -4\pi\,
(\Psi^{(0)})^2 v_a^{(0)} &=& \omega_a^{(0)}
u_a^{(1)}+\omega_a^{(1)} u_a^{(0)} \, ,\label{eq:u1}\\
\noalign{\vspace{.4cm}}
\left(H_0 - \mu^{(0)}\right) v_a^{(1)} &+&\left( 8\pi
\,|\Psi^{(0)}|^2-\mu^{(1)}\right)v_a^{(0)} \nonumber \\
-4\pi\, (\Psi^{(0)*})^2 u_a^{(0)} &=& - \omega_a^{(0)}
v_a^{(1)}-\omega_a^{(1)} v_a^{(0)} \, ,\label{eq:v1}\end{aligned}$$ and we must also expand the parameter [$\theta\approx
\theta^{(0)}+\gamma \theta^{(1)} + \cdots\,$]{}. As a result, the zero-order functions in Eq. (\[eq:zero\]) are evaluated with $\theta^{(0)}$. The first-order contributions then become $$\begin{aligned}
u_a^{(1)} &=& \theta^{(1)}\sinh\theta^{(0)} \chi_{00}\,\varphi_0
+ {\sum_{nl}}' a_{nl} \,\chi_{nn}\,\varphi_{2l}\,,\label{eq:ua1}
\\
v_a^{(1)} &=& \theta^{(1)}\cosh\theta^{(0)}\chi_{02}\,\varphi_0 +
{\sum_{nl}}' b_{nl}\,
\chi_{n,n+2}\,\varphi_{2l}\, ;\label{eq:va1}\end{aligned}$$ these expansions maintain the proper normalization to first order in $\gamma$.
Multiply Eqs. (\[eq:u1\]) and (\[eq:v1\]) by $u_a^{(0)*}$ and $v_a^{(0)*}$, respectively, and integrate. Straightforward manipulations yield the pair of equations $$\begin{aligned}
\mu^{(1)}+\omega_a^{(1)}& =& 4\pi I_0\,\Big(2\int
d^2r\,|\chi_{00}|^2\,|\chi_{10}|^2 \nonumber \\
&&-\tanh\theta^{(0)}\int
d^2r\,\chi_{00}^*\,(\chi_{10})^2\,\chi_{02}\Big)\nonumber\\
&=&\frac{1}{\sqrt{2\pi}}\left(2-\frac{\tanh\theta^{(0)}}{\sqrt
2}\right)\,,\\ \mu^{(1)}-\omega_a^{(1)} &= &4\pi I_0\Big(2\int
d^2r\,|\chi_{02}|^2\,|\chi_{10}|^2 \nonumber \\ &&-
\coth\theta^{(0)}\int
d^2r\,\chi_{02}^*\,(\chi_{10}^{*})^2\,\chi_{00}\Big)\nonumber\\
&=&
\frac{1}{\sqrt{2\pi}}\left(\case{3}{2}-\frac{\coth\theta^{(0)}}
{\sqrt2}\right)\,,\end{aligned}$$ where $I_0$ is given in Eq. (\[eq:Il\]), and the radial integrals are evaluated with the results in the appendix. The sum of these equations determines $\mu^{(1)}$, and comparison with Eq. (\[eq:mu1\]) yields $$\begin{aligned}
\cosh\theta^{(0)}& = &\sqrt 2,\quad\hbox{and}\quad \nonumber \\
\sinh\theta^{(0)}&= &1.\end{aligned}$$ Correspondingly, the difference now determines the first-order correction to the anomalous frequency [@Fetter98] $$\omega_a^{(1)} = \frac{1}{2\sqrt{2\pi}}\,,$$ so that $\Omega_c = -\omega_a$ through first order.
To determine the coefficients in the first-order expansions in Eqs. (\[eq:ua1\]) and (\[eq:va1\]), project Eqs. (\[eq:u1\]) and (\[eq:v1\]) onto the appropriate unperturbed eigenfunctions. In this way, we find (see Appendix B) $$\begin{aligned}
a_{nl}& =&\frac{2\pi\,I_l}{n+\lambda\, l} \,\Big(\int
d^2r\,\chi_{nn}^*\,(\chi_{10})^2\,\chi_{02}\nonumber \\ &&- 2\sqrt
2 \int
d^2r\,\chi_{nn}^*\,|\chi_{10}|^2\,\chi_{00}\,\Big)\nonumber\\
&=&\frac{2\pi\,I_l}{n+\lambda \, l}\left(J_{112}^{n0}-2\sqrt
2\,J_{110}^{n0}\right)\, ,\label{eq:anl}\\ b_{nl}&=&
\frac{2\pi\,I_l}{n+\lambda\, l} \,\Big(\sqrt 2\,\int
d^2r\,\chi_{n,n+2}^*\,|\chi_{10}|^2\,\chi_{02}\nonumber \\ &&- 2
\int
d^2r\,\chi_{n,n+2}^*\,(\chi_{10}^{*})^2\,\chi_{00}\,\Big)\nonumber\\
&=&\frac{2\pi\,I_l}{n+ \lambda \, l}\left(\sqrt 2
J_{112}^{n2}-2J_{110}^{n2}\right)\,.\label{eq:bnl}\end{aligned}$$
The second-order Bogoliubov equations are
$$\begin{aligned}
&&(\omega_a^{(0)}u_a^{(2)} + \omega_a^{(1)} u_a^{(1)} +
\omega_a^{(2)} u_a^{(0)}) \nonumber \\ \noalign{\vspace{.1cm}}&& =
(H_0- \mu^{(0)})\, u_a^{(2)} - \mu^{(1)} u_a^{(1)} - \mu^{(2)}
u_a^{(0)} \nonumber \\ && \, \,\, +8\pi\, [\,|\Psi^{(0)}|^2
u_a^{(1)} +
\left(\Psi^{(0)*}\Psi^{(1)}+\Psi^{(0)}\Psi^{(1)*}\right)u_a^{(0)}]
\nonumber
\\ && \, \, \, - 4 \pi \,[(\Psi^{(0)})^2 \,v_a^{(1)} + 2
\Psi^{(0)}\Psi^{(1)} v_a^{(0)}] \, , \label{bog1}\\
\noalign{\vspace{.4cm}} && - (\omega_a^{(0)}v_a^{(2)} +
\omega_a^{(1)} v_a^{(1)} + \omega_a^{(2)} v_a^{(0)})\nonumber \\
\noalign{\vspace{.1cm}}&& = (H_0 -\mu^{(0)}) \,v_a^{(2)} -
\mu^{(1)} v_a^{(1)} - \mu^{(2)} v_a^{(0)} \nonumber
\\ &&\, \, \,+ 8 \pi \,[\,|\Psi^{(0)}|^2 v_a^{(1)} +
\left(\Psi^{(0)*}\Psi^{(1)}+\Psi^{(0)}\Psi^{(1)*}\right)v_a^{(0)}]
\nonumber \\
&& \, \, \,- 4 \pi \,[(\Psi^{(0)*})^2\,u_a^{(1)}
+ 2\Psi^{(0)*}\Psi^{(1)*}u_a^{(0)}] \, . \label{bog2}\end{aligned}$$
Here, Eqs. (\[eq:Psi1a\]) and (\[eq:Psi1b\]) give the condensate wave function $\Psi^{(1)}$, and Eqs. (\[eq:ua1\]), (\[eq:va1\]), (\[eq:anl\]) and (\[eq:bnl\]) give the corrections $u_a^{(1)}$ and $v_a^{(1)}$.
The remaining steps are essentially the same as in first order, yielding a pair of equations for $\mu^{(2)}\pm\omega_a^{(2)}$. The parameter $\theta^{(1)}$ must be chosen so that the sum reproduces Eq. (\[eq:E2\]) for the chemical potential $\mu^{(2)}$. The difference then gives the second-order correction to the anomalous frequency $$\begin{aligned}
\omega^{(2)}_a(\lambda)
&=&8\pi^2{\sum_{nl}}' \frac{(I_l)^2}{n + l \lambda} \nonumber
\\ &\times&\Big[13(J_{111}^{n1})^2 - 8(J_{110}^{n0})^2
-4(J_{112}^{n2})^2 - 2(J_{110}^{n2})^2 \nonumber
\\ &&- (J_{112}^{n0})^2 - 8J_{100}^{n1}J_{111}^{n1}
- 4J_{122}^{n1}J_{111}^{n1}\nonumber \\ &&+ 4\sqrt{2}
(J_{112}^{n0}J_{110}^{n0} + J_{110}^{n2}J_{112}^{n2})\,\Big] \, ,\end{aligned}$$ and the total anomalous frequency through second order becomes $$\omega_a
= -1 + {1\over 2}\frac{1}{\sqrt{2 \pi}}\,\gamma
+\omega^{(2)}_a(\lambda) \,\gamma^2 \, .$$ This is the main result of the present section, illustrated in Fig. 3.
\[a\]
For several different axial asymmetries, Table I compares the second-order correction $\Omega_c^{(2)}(\lambda)$ to the critical rotation frequency from the preceding section with the second-order correction $\omega_a^{(2)}(\lambda)$ to the anomalous frequency. It can be seen that $\omega_a^{(2)}$ is always negative and smaller in absolute value than the corresponding correction $\Omega_c^{(2)}$.
To understand the significance of this result, consider a normal mode of the $q$-fold quantized vortex with $u_j(r,\phi) = \exp[i(m_j +q)\phi]\, \tilde u_j(r)$ and $v_j(r,\phi)
=
\exp[i(m_j-q)\phi]\,\tilde v_j(r)$ [@Dodd; @SF98], where $m_j$ is the angular momentum of the normal mode relative to that of the vortex in the condensate (namely, the corresponding perturbations in the density and velocity potential have the angular dependence $\propto\exp im_j\phi$). In a frame rotating with angular velocity $\Omega$, the frequency of a given normal mode becomes $\omega_j (\Omega) =
\omega_j (0) - m_j
\Omega$. For the present case of a singly quantized vortex with $q=1$, the anomalous mode has $m_a = -1$, so that the shifted anomalous frequency in the rotating frame is $\omega_a(\Omega) = \omega_a (0) +\Omega$. With increasing external rotation $\Omega$, the condensate with a vortex becomes dynamically stable against microscopic oscillations for $\Omega\ge \Omega^* = |\omega_a|$. Our numerical results show that $\Omega^*<\Omega_c $, so the onset of dynamical stability occurs before the singly quantized vortex in the condensate becomes energetically favorable. Since the sum $\Omega_c+\omega_a$ vanishes through order $\gamma$, it is necessary to include the second-order corrections to decide the relative value of $\Omega^*$ and $\Omega_c$. The difference is illustrated in Fig. 4 for two geometries, including the corresponding results for strong coupling (in the Thomas-Fermi limit) [@Svid98], which shows that the sequence of the two types of stabilization is the same in both regimes.
\[omega\]
Metastable frequency
====================
The third important frequency in this problem is the metastable angular velocity $\Omega_m$ at which a slightly off-center vortex becomes trapped in a local minimum of the free energy at the center of the confining potential. Consider the corresponding energy functional $$\label{efunc}
E(\Psi) = \int dV \left[{\Psi^*}\left(H_0 -\Omega L_z\right)\Psi
+ 2 \pi \gamma |\Psi|^4 \right]\, \, .$$ We now use cartesian coordinates, with $$\label{H0}
H_0 = \frac{1}{2} \left[ - \partial^2_x- \partial^2_y + (x^2 +y^2)
+ \lambda \left( -\partial^2_z+z^2\right)\right]$$ and $L_z = -i(x \partial y - y \partial x)$ is the $z$-component of the angular-momentum operator.
In order to evaluate the energy functional Eq. (\[efunc\]), we construct a trial condensate wave function. This trial function is assumed to be unchanged along the axis of symmetry, which allows us to use the ground-state gaussian $\varphi_0(z)$. In the weak-coupling limit, the radius of the vortex core is comparable with the radius of the condensate, so that the displacement ${\bf r}_0=(x_0,y_0)$ of the vortex, the displacement ${\bf r}_1=(x_1,y_1)$ of the condensate, and the induced velocity of the condensate must all be included in the following trial function $$\begin{aligned}
\psi_v(x,y,z)&=&\frac{C}{\pi^{3/4}}[(x-x_0)+i(y-y_0)] \nonumber \\
&\times& e^{-\case{1}{2}{(x-x_1)^2}-\case{1}{2}{(y-y_1)^2}-
\case{1}{2}{z^2}}
e^{i(\alpha_x x+\alpha_y y)} \, ,\end{aligned}$$ where $C^{-2} = 1+|{\bf r}_1-{\bf r}_0|^2$ and the constants $\alpha_x$ and $\alpha_y$ characterize the velocity components.
In evaluating the integration in $E(\psi_v)$, we retain all terms up to second order in the small parameters $\alpha_i$ and in the displacements. For convenience, we introduce new variables $\bbox{\delta}={\bf
r}_1-{\bf r}_0 =(\delta_x,\delta_y)$ and $\bbox{\epsilon}=2{\bf
r}_1-{\bf r}_0 =(\epsilon_x,\epsilon_y)$, which will turn out to be the normal modes of the system. The variational energy is
$$\begin{aligned}
\label{Ecross}
E_{\rm var} (\bbox{\alpha},\bbox{\delta},\bbox{\epsilon} \,)
&=& 2+{\lambda \over 2} -\Omega +{\gamma\over
2\sqrt{2\pi}}\nonumber \\
&+&{\alpha^2 \over 2} +\alpha_x\,(-\delta_y+\Omega\epsilon_y)
+\alpha_y\,(\delta_x-\Omega\epsilon_x) \nonumber \\
&+&{\epsilon^2 \over 2} -\Omega \, \bbox{\delta}\cdot\bbox{\epsilon}
+\Big(2\Omega-{3\over2}+{\gamma\over \sqrt{2\pi}}\Big)\delta^2.\end{aligned}$$
The values $$\begin{aligned}
\alpha_x &=& \delta_y -\Omega \epsilon_y \\
\alpha_y &=& -\delta_x +\Omega \epsilon_x\end{aligned}$$ minimize this expression with respect to the velocity parameter $\bbox{\alpha} =(\alpha_x,\alpha_y)$, and the energy then becomes diagonal in the variables $\bbox{\delta}$ and $\bbox{\epsilon}$, which thus represent the appropriate normal modes $$\begin{aligned}
\label{Ede}
E_{\rm var}(\bbox{\delta},\bbox{\epsilon} \,) &=& 2 +\frac{\lambda}{2}+
\frac{\gamma}{2\sqrt{2 \pi}} - \Omega \nonumber \\
&+&\left( -2 +\frac{\gamma}{\sqrt{2 \pi}} + 2 \Omega \right){\delta}^2
+\frac{\left(1- \Omega^2\right)}{2}{\epsilon}^2 \, .\end{aligned}$$
The energy of the vortex-free ground state is $E_0 = 1+\lambda /2+ \gamma /(\sqrt{2
\pi})$. Thus the difference $\Delta F(0) = E_{\rm var}(0) -E_0$ vanishes at the expected critical angular velocity $\Omega_c = 1 -\gamma/(2 \sqrt{2
\pi})$, in agreement with the first-order result obtained from the Gross-Pitaevskii equation. Furthermore, if $\bbox{\delta}$ vanishes, then the vortex and condensate move rigidly, and the resulting dipole oscillation mode is stable throughout the range $|\Omega|\le 1$. For larger angular velocity, the motion becomes unstable, as is familiar from the behavior of a classical particle in a parabolic potential.
We next consider how the relative displacement $\bbox{\delta}$ affects the variational energy Eq. (\[Ede\]), which is clearly unstable if $\Omega=0$. With increasing external rotation, however, the relative displacement becomes metastable at a frequency $$\label{Om}
\Omega_m = 1 -\frac{\gamma}{2\sqrt{2 \pi}} \, ,$$ when $E_{\rm var}$ changes from a local maximum to a local minimum for small relative displacements. To this (first) order in the interaction parameter $\gamma$, the value $\Omega_m$ coincides with the modulus of the anomalous mode frequency $|\omega_a|$ and hence with the critical frequency $\Omega_c$.
In order to illustrate the importance of the combined effect of the vortex and the condensate, we can study the behavior if only the vortex is displaced (namely ${\bf r}_1=0$) or if the induced velocity is neglected ($\bbox{\alpha}=0$). In the former case, the energy (\[Ede\]) is given by $$\begin{aligned}
\label{Ex0}
E_{\rm var}({\bf r}_0) &=& 2 +\frac{\lambda}{2}+ \frac{\gamma}{2\sqrt{2 \pi}}
-
\Omega \nonumber \\
&+&\left( -\frac{3}{2} +\frac{\gamma}{\sqrt{2 \pi}} + 2 \Omega
-\frac{\Omega^2}{2}\right)
\left(x_0^2 +y_0^2\right),\end{aligned}$$ and the metastable frequency is determined by requiring a positive coefficient of the displacement contribution \[the second line in Eq. (\[Ex0\])\]. To first order in the interaction parameter, we find $$\Omega_m^* = 1 -\frac{\gamma}{\sqrt{2 \pi}},\label{Omegam*}$$ which differs from the value $\Omega_m$ found with the more general approach. In the second case, the metastable frequency is determined by the condition that the last line in Eq. (\[Ecross\]) is positive, namely the determinant of the coefficients must be positive. This gives the same metastable frequency (\[Omegam\*\]) as that found from Eq. (\[Ex0\]). Therefore, the metastable frequency has the same value as the absolute value of the anomalous mode (up to first order) only if the displacement of the condensate *and* the induced velocity are both taken into account. This indicates that the vortex becomes confined in a local central energy minimum at the same rotation frequency $|\omega_a|$ for which the instability due to the anomalous mode disappears. The same scenario had been found in the strong-coupling limit [@Svid98], and our results strengthen the idea of a common underlying phenomenon.
Note that the metastable frequency found here \[Eq. (\[Om\])\] does not coincide with the metastable rotation frequency defined by Feder, Clark, and Schneider [@Feder99]. Instead of requiring a local minimum in the free energy, they identify the onset of metastability with the frequency for which the chemical potentials for a condensate with and without a vortex are equal. In first-order perturbation theory, their criterion leads to the expression (\[Omegam\*\]) found by omitting the displacement of the condensate.
CONCLUSIONS
===========
It has been shown that the critical rotation frequency $\Omega_c$ exceeds the modulus of the anomalous Bogoliubov mode $\omega_a$ when second-order corrections in the interaction parameter are included. This result agrees with numerical results for one particular trap geometry found by Feder, Clark, and Schneider [@Feder99]. Furthermore, the sequence of stabilizing a singly quantized vortex through rotation is found to be the same as in the strong-coupling limit [@Svid98]: First, the vortex becomes stable against microscopic oscillations at $\Omega^* = |\omega_a|$, and only at the higher rotation speed $\Omega_c$ does the vortex become energetically stable. The geometry dependence introduced through the second-order terms indicates that pancake geometries are more favorable for vortex detection in rotating traps, since the stabilization frequencies are lower. This is consistent with the numerical results of García-Ripoll and Pérez-García [@Garcia99]. For extreme cigar-shaped condensates, the critical frequency can even exceed the trap frequencies [@Feder99], a regime that is experimentally inaccessible. The angular velocity $\Omega_m$ for the onset of metastability coincides with the modulus of the anomalous frequency $\omega_a$ in first order, when the combined effect of the vortex and the condensate is taken into account. For $\Omega >\Omega_m$, an energy barrier stabilizes the vortex at the center; at the same frequency, the instability due to microscopic oscillations disappears. These features in the stability scenario also agree with the behavior in the Thomas-Fermi limit [@Svid98].
We thank A. A. Svidzinsky for valuable discussions. This work was supported in part by NSF Grant No. 94-21888 and by the DAAD (German Academic Exchange Service) “Doktorandenstipendium im Rahmen des gemeinsamen Hochschulsonderprogramms III von Bund und Ländern" (M. L.).
Noninteracting eigenstates
==========================
The eigenstates for the noninteracting Bose condensate in a cylindrical trap can be classified with the quantum numbers of positive and negative circulation $n_+,n_-$ around the $z$ axis and the axial harmonic-oscillator energy quantum number $l$ in the $z$ direction [@Cohen]. In particular, the $z$-dependent parts of the eigenfunctions are simple harmonic oscillator eigenfunctions
$$\label{varphi}
\varphi_l(z) =\frac{1}{\sqrt{ \pi^{1/2}\,2^l
l!}}\,H_l(z)
\,e^{-\frac{1}{2} z^2} \, ,$$
where the $H_l(z)$ are the Hermite polynomials [@Lebedev].
In terms of the circular quanta, the normalized two-dimensional eigenfunctions are
$$\chi_{n_+,n_-}(x,y) =
\frac{1}{\sqrt{\pi\,n_+!\,n_-!}}\left(a_+^\dagger\right)^{n_+}
\left(a_-^\dagger\right)^{n_-}e^{-\frac{1}{2}(x^2 +y^2)} \, ,$$
where $a_\pm^\dagger= (a_x^\dagger \pm ia_y^\dagger )/\sqrt2=
\frac{1}{2}[x\pm iy-(\partial_x\pm i\partial_y)]$ are the creation operators for right and left circular quanta, respectively. In terms of the new variables $\zeta = x+i y$, $\zeta^* = x-i y $, these operators take the form
$$a_+^\dagger = \frac{\zeta}{2} - \frac{\partial}{\partial\zeta^*}\,
,\qquad a_-^\dagger =
\frac{\zeta^*}{2} - \frac{\partial}{\partial\zeta} \, .$$
The identity $(\frac{1}{2}\zeta^*-\partial_\zeta)\exp(-\frac{1}{2}\zeta\zeta^*)
=\exp(\frac{1}{2}\zeta\zeta^*)
(-\partial_\zeta)$ $\exp(-\zeta\zeta^*)$ and its complex conjugate readily yield $$\begin{aligned}
\chi_{n+q,n} &=&
\frac{e^{\case{1}{2}\zeta\zeta^*}}{\sqrt{\pi\,(n+q)!\,n!}}
\left(-\frac{\partial}{\partial\zeta}\right)^n\,
\left(-\frac{\partial}{\partial\zeta^*}\right)^{n+q}\,
e^{-\zeta\zeta^*}\nonumber\\
&=&\frac{(-1)^n \,e^{\case{1}{2}\zeta\zeta^*}}{\sqrt{\pi\,(n+q)!\,n!}}
\left(\frac{\partial}{\partial\zeta}\right)^n\,
\left(\zeta^{n+q}e^{-\zeta\zeta^*}\right)\,.\end{aligned}$$ Comparison with the standard formula for the associated Laguerre polynomials $L_n^q$ [@Lebedev] yields
$$\label{chi}
\chi_{n+q,n}(r,\phi) =
(-1)^{n}\sqrt{\frac{n!}{\pi\,(n+q)!}}\,e^{-\frac{1}{2}r^2}
\,e^{iq\phi}\,r^q
\,L_n^q(r^2) \, .$$
The complete normalized three-dimensional eigenfunctions of the noninteracting system are $\psi_{n+q,n,l} (\vec{r}) = \chi_{n+q,n}(r,\phi)\varphi_l(z)$.
Second-order matrix elements
============================
The above expressions allow us to evaluate the necessary matrix elements, and the $z$-dependent part factorizes out in the form $$\begin{aligned}
I_l &=& \int_{-\infty}^\infty dz\,\varphi_{2l}(z)\,\varphi_0(z)^3 \nonumber \\
&=&
\frac{1}{\pi\sqrt{2^{2l}\,(2l)!}}\int_{-\infty}^\infty dz\,H_{2l}(z)
e^{-2z^2},\end{aligned}$$ where only states with an even number of quanta contribute. The generating function for the Hermite polynomials $$e^{-t^2+2zt} = \sum_{n=0}^\infty\frac{H_n(z)}{n!}\,t^n$$ readily shows that $\exp(-\frac{1}{2}t^2)$ is the generating function for $\sqrt{2^{2l}\,(2l)!}\,I_l$, and a straightforward analysis gives the desired expression
$$\label{B}
I_l =
\frac{1}{\sqrt{2 \pi}} \frac{(-1)^l \sqrt{(2 l)!}}{2 ^{2l} l!} \,.$$
For the radial part note that $\chi_{n,n+q}(r,\phi) =
\chi_{n+q,n}^*(r,\phi)$. Since the phase factor is the only complex part, it can be separated explicitly
$$\begin{aligned}
\chi_{n+q,n}(r,\phi) &=&e^{iq\phi} \tilde{\chi}_{n+q,n}(r) \, ,\nonumber \\
\chi_{n,n+q}(r,\phi) &=& e^{-iq\phi}\tilde{\chi}_{n+q,n}(r) \, ,\end{aligned}$$
where $\tilde\chi $ is real. The matrix elements $ J^{nq}_{ijk}=\int d^2r\,
\tilde{\chi}_{n+q,n}\,\tilde{\chi}_{i,0}\,\tilde{\chi}_{j,0}\,
\tilde{\chi}_{k,0}$ involve a product of four of these eigenfunctions, where three refer to the condensate in the lowest energy state, namely $n=0$ (note that $L_0^\alpha=1$ for any $\alpha$). Due to the angular integration, only products with no net overall phase remain. Use of Eq. (\[chi\]) yields
$$J^{nq}_{ijk}= \frac{(-1)^n}{ \pi} \,\frac{1}{\sqrt{i!\,j!\,k!}}
\,
\sqrt{\frac{n!}{ (n+q)!}}\,\int_0^\infty du \,u^{p} e^{-2u}
L^q_n(u)
\, ,$$
where $p = \frac{1}{2}(q+i+j+k)$ is an integer because of the angular phase factors (note also that $J_{ijk}^{nq}$ is symmetric under interchange of its subscripts). As in the preceding example of the axial matrix elements, the generating function for the Laguerre polynomials [@Lebedev] $$\sum_{n=0}^{\infty} L_n^q (u) \,t^n =
\frac{1}{(1-t)^{1+q}}\, \exp\left(\frac{-u t}{1-t}\right)$$ facilitates the radial integration. For example the function $2(1-t)/(2-t)^3$ provides a generating function for $(-1)^n\,\pi\,\sqrt{n+1}\,J_{111}^{n1}$, which is the matrix element with four singly quantized vortex eigenfunctions, and we find
$$\label{J}
J^{n1}_{111} =
\frac{(-1)^n}{2^{n+3}\,\pi}\,\sqrt{n+1}\,(2-n) \, .$$
A similar technique leads to all the other relevant radial integrals.
The sums occurring in the second-order terms contain the square of the matrix elements $(J^{nq}_{ijk}
\,I_l)^2$ appropriately weighted with the energy denominator. For example, the second-order contribution to the chemical potential for the singly quantized vortex (obtained from the Gross-Pitaevskii equation) is
$$\begin{aligned}
\label{mu2}
\mu_1^{(2)} &=&-24\pi^2\, {\sum_{nl}}'
\frac{ (J^{n1}_{111}\, I_l)^2}{n + l\lambda}
\nonumber \\
&=&-\frac{3}{16 \pi}\,{\sum_{nl}}'
\frac{1}{n+\lambda l}\frac{(n+1)(2-n)^2}{2^{2n}}
\frac{(2l)!}{2^{4l} (l!)^2}\, .\end{aligned}$$
M. H. Anderson, J. R. Ensher, M. R. Matthews, C. E. Wieman, and E. A. Cornell, Science [**269**]{}, 198 (1995).
K. B. Davis, M.-O. Mewes, M. R. Andrews, N. J. van Druten, D. S. Durfee, D. M. Kurn, and W. Ketterle, Phys. Rev. Lett. [**75**]{}, 3969 (1995).
C. C. Bradley, C. A. Sackett, and R. G. Hulet, Phys. Rev. Lett. [**78**]{}, 985 (1997).
G. Baym and C. J. Pethick, Phys. Rev. Lett. [**76**]{}, 6 (1996).
F. Dalfovo and S. Stringari, Phys. Rev. A[**53**]{}, 2477 (1996).
E. Lundh, C. J. Pethick, and H. Smith, Phys. Rev. A [**55**]{}, 2126 (1997).
R. J. Dodd, K. Burnett, M. Edwards, and C. W. Clark, Phys. Rev. A [**56**]{}, 587 (1997).
D. S. Rokhsar, Phys. Rev. Lett. [**79**]{}, 2164 (1997).
A. L. Fetter, J. Low Temp. Phys. [**113**]{}, 198 (1998).
A. A. Svidzinsky and A. L. Fetter, Phys. Rev. A [**58**]{}, 3168 (1998).
A. A. Svidzinsky and A. L. Fetter, cond-mat/9811348.
H. Pu, C. K. Low, J. H. Eberly, and N. P. Bigelow, Phys. Rev. A. [**59**]{}, 1533 (1999).
T. Isoshima and K. Machida, J. Phys. Soc. Jpn. [**68**]{}, 487 (1999).
J. J. García-Ripoll and V. M. Pérez-García, cond-mat/9903353.
D. L. Feder, C. W. Clark, and B. I. Schneider, cond-mat/9904269.
D. A. Butts and D. S. Rokhsar, Nature [**397**]{}, 327 (1999).
L. P. Pitaevskii, Zh. Éksp. Teor. Fiz. [**40**]{}, 646 (1961) \[Sov. Phys. JETP [**13**]{}, 45 (1961)\].
E. P. Gross, Nuovo Cimento [**20**]{}, 454 (1961); J. Math. Phys. [**4**]{}, 195 (1963).
N. Bogoliubov, J. Phys. (Moscow) [**11**]{}, 23 (1947).
A. L. Fetter, in [*Proceedings of the International School of Physics “Enrico Fermi,”*]{} July, 1998, to be published.
C. Cohen-Tannoudji, B. Diu, F. Laloë, [*Quantum Mechanics*]{} (Wiley, 1977), Vol. I, pp. 500-502, 733-738.
N. N. Lebedev, [*Special Functions and Their Applications*]{} (Dover, 1972), Chap. 4.
| {
"pile_set_name": "ArXiv"
} |
---
address:
- 'S.I.S.S.A., Via Beirut 2-4, 34014, Trieste, Italy'
- 'S.I.S.S.A., Via Beirut 2-4, 34014, Trieste, Italy'
author:
- Alessandro Giacomini
- Marcello Ponsiglione
title: |
A discontinuous finite element approximation of\
quasi-static growth of brittle fractures
---
.2truecm
Introduction {#intr}
============
In this paper we formulate a discontinuous finite element approximation for a model of quasi-static growth of brittle fractures in linearly elastic bodies proposed by Francfort and Marigo [@FM]. Their model is based on the classical Griffith’s criterion which involves a competition between [*bulk*]{} and [*surface*]{} energies. To be precise, let ${\Omega}\subseteq {{\mathbb R}}^3$ be an elastic body, $\partial_D {\Omega}$ be a part of its boundary and let $g: \partial_D {\Omega}\to {{\mathbb R}}^3$ be the spatial displacement of ${\Omega}$ at the points of $\partial_D {\Omega}$. According to Griffith’s theory, given a preexisting crack $\Gamma_1 \subseteq {\overline{\Omega}}$, the new crack $\Gamma$ and the displacement $u: {\Omega}\setminus \Gamma \to {{\mathbb R}}^3$ associated to $g$ at the equilibrium minimizes the following energy $$\label{griffithenergy}
{{\mathcal E}}(v,g,\Gamma):= \int_{\Omega}\mu |Ev|^2 +\frac{\lambda}{2}\,|{\rm div}\,v|^2 \,dx+ {{\mathcal H}}^2(\Gamma),$$ among all cracks $\Gamma$ with $\Gamma_1 \subseteq \Gamma$ and all displacements $v: {\Omega}\setminus \Gamma \to {{\mathbb R}}^3$ with $v=g$ on $\partial_D {\Omega}\setminus \Gamma$. Here $Ev$ denotes the symmetric part of the gradient of $v$, and ${{\mathcal H}}^2$ denotes the two dimensional Hausdorff measure, while $\mu$ and $\lambda$ are the Lamé coefficients. The boundary condition is required only on $\partial_D {\Omega}\setminus \Gamma$ because the displacement in a fractured region is assumed not to be transmitted. Let us indicate by ${{\mathcal E}}(g,\Gamma)$ the minimum value of (\[griffithenergy\]) among all $v: {\Omega}\setminus \Gamma \to {{\mathbb R}}^3$ with $v=g$ on $\partial_D {\Omega}\setminus \Gamma$.
Supposing that the boundary displacement $g$ varies with the time $t \in [0,1]$, the quasi-static evolution $t \to \Gamma(t)$ proposed in [@FM] requires that:
- $\Gamma(t)$ is increasing in time, i.e., $\Gamma(t_1) \subseteq \Gamma(t_2)$ for all $0 \le t_1 \le t_2 \le 1$ (irreversibility of the process);
- - ${{\mathcal E}}(g(t),\Gamma(t)) \le {{\mathcal E}}(g(t),\Gamma)$ for all cracks $\Gamma$ such that $\cup_{s<t} \Gamma(s) \subseteq \Gamma$ (equilibrium condition);
- - the total energy ${{\mathcal E}}(g(t),\Gamma(t))$ is absolutely continuous in time and (conservation of energy) $$\frac{d}{dt} {{\mathcal E}}(g(t),\Gamma(t))= 2 \mu \int_{\Omega}Eu(t) E\dot{g}(t)\,dx
+\lambda \int_{\Omega}{\rm div}\,u(t) \,{\rm div}\,\dot{g}(t) \,dx.$$
A precise mathematical formulation of this model has been given by Dal Maso and Toader [@DMT] in the case of [*anti-planar shear*]{} in dimension two assuming that the fractures are compact sets with a finite number of connected components. Recently Francfort and Larsen [@FL], using the framework of $SBV$-functions, proved the existence of a quasi-static growth of brittle fractures in the case of [*generalized anti-planar shear*]{} and without assumptions on the structure of the fractures which are dealt with the set of jumps of the displacements. To be precise, they consider as elastic body an infinite cylinder whose section ${\Omega}\subseteq {{\mathbb R}}^N$ is subject to a displacement $u \in SBV({\Omega})$ in the direction orthogonal to ${\Omega}$. The crack at time $t$ on the section ${\Omega}$ is defined as $$\Gamma(t):=\bigcup_{s<t} \left[ S_{u(s)} \cup
(\partial_D {\Omega}\cap \{u(s) \not= g(s)\})
\right],$$ where $S_u$ denotes the set of jumps of $u$. Moreover the pair $(u(t), \Gamma(t))$ is such that:
- for all $v \in SBV({\Omega})$ $$\label{intrfl1}
\int_{\Omega}|\nabla u(t)|^2\,dx + {{\mathcal H}}^{N-1}(\Gamma(t)) \le
\int_{\Omega}|\nabla v|^2\,dx + {{\mathcal H}}^{N-1}(S_v \cup
(\partial_D {\Omega}\cap \{v \not= g(t)\})
\cup \Gamma(t));$$
- - the total energy ${{\mathcal E}}(t):= \int_{\Omega}|\nabla u(t)|^2\,dx+ {{\mathcal H}}^{N-1}(\Gamma(t))$ is absolutely continuous and $$\label{intrfl2}
{{\mathcal E}}(t)={{\mathcal E}}(0) + 2\int_0^t \int_{\Omega}\nabla u(\tau) \nabla \dot{g}(\tau) \,dx \,d\tau.$$
The aim of this paper is to discretize the model using a suitable finite element method and to a give a rigorous proof of its convergence to a quasi-static evolution in the sense of Francfort and Larsen. We restrict our analysis to a two dimensional setting considering only a polygonal reference configuration ${\Omega}\subseteq {{\mathbb R}}^2$.
The discretization of the domain ${\Omega}$ is carried out, following [@N], considering two parameters ${\varepsilon}>0$ and $a \in ]0,\frac{1}{2}[$ . We consider a regular triangulation ${{\bf R}}_{\varepsilon}$ of size ${\varepsilon}$ of ${\Omega}$, i.e. we assume that there exist two constants $c_1$ and $c_2$ so that every triangle $T \in {{\bf R}}_{\varepsilon}$ contains a ball of radius $c_1 {\varepsilon}$ and is contained in a ball of radius $c_2 {\varepsilon}$. In order to treat the boundary data, we assume also that $\partial_D {\Omega}$ is composed of edges of ${{\bf R}}_{\varepsilon}$. On each edge $[x,y]$ of ${{\bf R}}_{\varepsilon}$ we consider a point $z$ such that $z=tx+(1-t)y$ with $t \in [a, 1-a]$. These points are called [*adaptive vertices*]{}. Connecting together the adaptive vertices, we divide every $T \in {{\bf R}}_{\varepsilon}$ into four triangles. We take the new triangulation ${{\bf T}}$ obtained after this division as the discretization of ${\Omega}$. The family of all such triangulations is denoted by ${{\mathcal T}}_{{\varepsilon},a}({\Omega})$.
The discretization of the energy functional is obtained restricting the total energy to the family of functions $u$ which are affine on the triangles of some triangulation ${{\bf T}}(u) \in {{\mathcal T}}_{{\varepsilon},a}({\Omega})$ and are allowed to jump across the edges of ${{\bf T}}(u)$. We indicate this space by ${\mathcal A}_{{\varepsilon},a}({\Omega})$. The boundary data is assumed to belong to the space ${{{\mathcal A}}{{\mathcal F}}_{\varepsilon}({\Omega})}$ of continuous functions which are affine on every triangle $T \in
{{\bf R}}_{\varepsilon}$.
Given the boundary data $g \in W^{1,1}([0,1],H^1({\Omega}))$ with $g(t) \in {{{\mathcal A}}{{\mathcal F}}_{\varepsilon}({\Omega})}$ for all $t \in [0,1]$, we divide $[0,1]$ into subintervals $[t^\delta_i,t^\delta_{i+1}]$ of size $\delta>0$ for $i=0, \ldots, N_\delta$, we set $g^\delta_i=g(t^\delta_i)$, and for all $u \in {{{\mathcal A}}_{\varepsilon,a}({\Omega})}$ we indicate by $S_D^{g^\delta_i}(u)$ the edges of the triangulation ${{\bf T}}(u)$ contained in $\partial_D {\Omega}$ on which $u \not= g^\delta_i$. Using a variational argument we construct a [*discrete evolution*]{} $\{u^{\delta,i}_{{\varepsilon},a}\,:\,i=0, \ldots,N_\delta\}$ such that $u^{\delta,i}_{{\varepsilon},a} \in {{{\mathcal A}}_{\varepsilon,a}({\Omega})}$ for all $i=0, \ldots, N_\delta$, and such that considering the [*discrete fracture*]{} $$\Gamma^{\delta,i}_{{\varepsilon},a}:=
\bigcup_{r=0}^i \big[ S_{u^{\delta,r}_{{\varepsilon},a}} \cup
S_D^{g^\delta_r}(u^{\delta,r}_{{\varepsilon},a}) \big],$$ the following [*unilateral minimality property*]{} holds: $$\label{pieceminintr}
\int_{\Omega}|\nabla u^{\delta,i}_{{\varepsilon},a}|^2\,dx \le \int_{\Omega}|\nabla v|^2\,dx +
{{\mathcal H}}^1\left( \big( S_v \cup S_D^{g^\delta_i}(v) \big) \setminus
\Gamma^{\delta,i-1}_{{\varepsilon},a}\right).$$ Moreover we get suitable estimates for the discrete total energy $${{\mathcal E}}^{\delta,i}_{{\varepsilon},a}:=\|\nabla u^{\delta,i}_{{\varepsilon},a}\|^2_{L^2({\Omega};{{\mathbb R}}^2)}+
{{\mathcal H}}^1 \left( \Gamma^{\delta,i}_{{\varepsilon},a}\right).$$ The definition of the discrete fracture ensures that $\Gamma^{\delta,i}_{{\varepsilon},a} \subseteq \Gamma^{\delta,j}_{{\varepsilon},a}$ for all $i \le j$, recovering in this discrete setting the irreversibilty of the growth given in $(1)$. The minimality property is the reformulation in the finite element space of the equilibrium condition $(2)$.
In order to perform the asymptotic analysis of the [*discrete evolution*]{} $\{u^{\delta,i}_{{\varepsilon},a}\,:\,i=0, \ldots,N_\delta\}$, we make the piecewise constant interpolation in time $u^{\delta}_{{\varepsilon},a}(t)=u^{\delta,i}_{{\varepsilon},a}$ and $\Gamma^{\delta}_{{\varepsilon},a}(t)=\Gamma^{\delta,i}_{{\varepsilon},a}$ for all $t^\delta_i \le t <t^\delta_{i+1}$. The main result of the paper is the following theorem.
\[mainthm\] Let $g \in W^{1,1}([0,1], H^1({\Omega}))$ be such that $\|g(t)\|_\infty \le C$ for all $t \in [0,1]$ and let $g_{\varepsilon}\in W^{1,1}([0,1], H^1({\Omega}))$ be such that $\|g_{\varepsilon}(t)\|_\infty \le C$, $g_{\varepsilon}(t) \in {{{\mathcal A}}{{\mathcal F}}_{\varepsilon}({\Omega})}$ for all $t \in [0,1]$ and $$\label{bdryconv}
g_{\varepsilon}\to g \quad \mbox { strongly in }W^{1,1}([0,1], H^1({\Omega})).$$ Given the discrete evolution $\{t \to u^{\delta}_{\varepsilon,a}(t)\}$ relative to the boundary data $g_{\varepsilon}$, let $\Gamma^\delta_{{\varepsilon},a}$ and ${{\mathcal E}}^\delta_{{\varepsilon},a}$ be the associated fracture and total energy.
Then there exist $\delta_n \to 0$, $\varepsilon_n \to 0$, $a_n \to 0$, and a quasi-static evolution $\{t \to (u(t),\Gamma(t)),\,t \in [0,1]\}$ relative to the boundary data $g$, satisfying and , and such that setting $u_n:=u^{\delta_n}_{\varepsilon_n, a_n}$, $\Gamma_n:=\Gamma^{\delta_n}_{\varepsilon_n, a_n}$, ${{\mathcal E}}_n:={{\mathcal E}}^{\delta_n}_{\varepsilon_n, a_n}$, the following hold:
- if ${{\mathcal N}}$ is the set of discontinuities of ${{\mathcal H}}^1(\Gamma(\cdot))$, for all $t \in [0,1] \setminus {{\mathcal N}}$ we have $$\label{mainconvgradt}
\nabla u_n(t) \to \nabla u(t) \quad \mbox{ strongly in }L^2({\Omega};{{\mathbb R}}^2)$$ and $$\label{mainconvjumpt}
\lim_n {{\mathcal H}}^1(\Gamma_n(t))= {{\mathcal H}}^1(\Gamma(t));$$
- - for all $t \in [0,1]$ we have $$\label{mainconvent}
\lim_n {{\mathcal E}}_n(t)={{\mathcal E}}(t).$$
We conclude that we have the convergence of the total energy at each time $t \in [0,1]$, and the separate convergence of bulk and surface energy for all $t \in [0,1]$ except a countable set.
In order to prove Theorem \[mainthm\], we proceed in two steps. Firstly, we fix $a$ and let $\delta \to 0$ and ${\varepsilon}\to 0$. We obtain an evolution $\{t \to u_a(t)\,:\,t \in [0,1]\}$ such that $\nabla u^{\delta}_{{\varepsilon},a}(t) \to \nabla u_a(t)$ strongly in $L^2({\Omega};{{\mathbb R}}^2)$ for all $t$ up to a countable set and such that the following minimality property holds: for all $v \in SBV({\Omega})$ $$\label{aminimality}
\int_{\Omega}|\nabla {u_a(t)}|^2\,dx \le \int_{\Omega}|\nabla v|^2\,dx+ \mu(a)
{{\mathcal H}}^1 \left( \big( S_v \cup (\partial_D {\Omega}\cap \{v \not= g(t)\}) \big)
\setminus \Gamma_a(t) \right),$$ where $\mu:]0,\frac{1}{2}[ \to ]0,+\infty[$ is a function independent of ${\varepsilon}$ and $\delta$, such that $\mu \ge 1$, $\lim_{a \to 0} \mu(a)=1$ and $\Gamma_a(t):=\bigcup_{s \le t, s \in D} S_{{u_a(s)}} \cup (\partial_D {\Omega}\cap \{u_a(s) \not= g(s)\})$. The minimality property takes into account possible anisotropies that could be generated as $\delta$ and ${\varepsilon}\to 0$: in fact, since $a$ is fixed, we have that the angles of the triangles in ${{\mathcal T}}_{{\varepsilon},a}({\Omega})$ are between fixed values (determined by $a$), and so fractures with certain directions cannot be approximated in length. In the second step, we let $a \to 0$ and determine from $\{t \to u_a(t)\,:\,t \in [0,1]\}$ a quasi-static evolution $\{t \to u(t)\,:\,t \in [0,1]\}$ in the sense of Francfort and Larsen. Then, using a diagonal argument, we find sequences $\delta_n \to 0$, ${\varepsilon}_n \to 0$, and $a_n \to 0$ satisfying Theorem \[mainthm\].
The main difficulties arise in the first part of our analysis, namely when $\delta,{\varepsilon}\to 0$. The convergence $u^\delta_{{\varepsilon},a}(t) \to u_a(t)$ in $SBV({\Omega})$ for $t \in D \subseteq [0,1]$ countable and dense is easily obtained by means of Ambrosio’s Compactness Theorem. The minimality property derives from its discrete version using a variant of Lemma 1.2 of [@FL]: given $v \in SBV({\Omega})$, we construct $v^\delta_{{\varepsilon},a} \in {\mathcal A}_{{\varepsilon},a}({\Omega})$ such that $$\label{intrconv1}
\nabla v^\delta_{{\varepsilon},a} \to \nabla v
\quad \mbox{ strongly in }L^2({\Omega};{{\mathbb R}}^2)$$ and $$\begin{gathered}
\label{intrconv2}
\limsup_{\delta,{\varepsilon}\to 0}
{{\mathcal H}}^1 \left[ \big( S_{v^\delta_{{\varepsilon},a}} \cup
S_D^{g_{\varepsilon}^\delta(t)}(v^\delta_{{\varepsilon},a}) \big)
\setminus \Gamma^\delta_{{\varepsilon},a}(t) \right] \le \\
\le \mu(a)
{{\mathcal H}}^1 \left[ \big( S_v \cup
\big( \partial_D {\Omega}\cap \{v \not= g(t) \} \big) \big)
\setminus \Gamma_a(t) \right],\end{gathered}$$ where $g_{\varepsilon}^\delta(t):=g_{\varepsilon}(t^\delta_i)$ for $t^\delta_i \le t <t^\delta_{i+1}$. The main difference with respect to Lemma 1.2 of [@FL] is that we have to find the approximating functions $v^\delta_{{\varepsilon},a}$ in the finite element space ${\mathcal A}_{{\varepsilon},a}({\Omega})$. This can be regarded as an interpolation problem, so we try to construct triangulations ${{\bf T}}_{\varepsilon}\in {{\mathcal T}}_{{\varepsilon},a}({\Omega})$ adapted to $v$ in order to obtain and . In all the geometric operations involved, we need to avoid degeneration of the triangles of ${{\bf T}}(u^\delta_{{\varepsilon},a}(t))$ which is guaranteed from the fact that $a$ is constant: this is the principal reason to keep $a$ fixed in the first step. A second difficulty arises when $u_a(\cdot)$ is extended from $D$ to the entire interval $[0,1]$: indeed it is no longer clear whether $\nabla u^\delta_{{\varepsilon},a}(t) \to \nabla u_a(t)$ for $t \not\in D$. Since the space ${\mathcal A}_{{\varepsilon},a}({\Omega})$ is not a vector space, we cannot provide an estimate on $\|\nabla u^\delta_{{\varepsilon},a}(t)-\nabla u^\delta_{{\varepsilon},a}(s)\|$ with $s \in D$ and $s<t$: we thus cannot expect to recover the convergence at time $t$ from the convergence at time $s$. We overcome this difficulty observing that $\nabla u^\delta_{{\varepsilon},a}(t) \to \nabla \tilde{u}_a$ with $\tilde{u}_a$ satisfying a minimality property similar to and then proving $\nabla \tilde{u}_a=\nabla u_a(t)$ by a uniqueness argument for the gradients of the solutions.
The plan of the paper is the following. In Section \[prel\] we give the basic definitions and prove some auxiliary results. In Section \[devol\], we prove the existence of a discrete evolution. In Section \[convres\] we prove the convergence of the discrete evolution to a quasi-static evolution of brittle fractures in the sense of Francfort and Larsen. The proof of minimality property requires a careful analysis to which is dedicated Section \[secmin\]. In Section \[remark\] we show that the arguments of Section \[convres\] can be used to improve the convergence results for the discrete in time approximation considered in [@FL].
Preliminaries {#prel}
=============
In this section we state the notation and prove some preliminaries employed in the rest of the paper.
We will employ the following basic notation:
- ${\Omega}$ is a polygonal open subset of ${{\mathbb R}}^2$;
- $L^p({\Omega};{{\mathbb R}}^m)$ with $1 \le p < +\infty$ and $m \ge 1$ is the usual Lebesgue space of $p$-summable ${{\mathbb R}}^m$-valued functions, and $L^p({\Omega}):=L^p({\Omega};{{\mathbb R}})$;
- for all $k \ge 1$ and $1 \le p \le +\infty$, $W^{k,p}({\Omega})$ is the usual Sobolev space of functions in $L^p({\Omega})$ with distributional derivatives of order $1, \ldots, k$ in $L^p({\Omega})$; we will write $H^k({\Omega})$ for $W^{k,2}({\Omega})$;
- if $u \in W^{k,p}({\Omega})$, $\nabla u$ is its gradient;
- ${{\mathcal H}}^{1}$ is the one-dimensional Hausdorff measure;
- $\|\cdot\|_{\infty}$ denotes the sup-norm;
- if $f \in L^2({\Omega};{{\mathbb R}}^m)$, $\|f\|$ denotes the $L^2$-norm of $f$;
- for all $A \subseteq {{\mathbb R}}^2$, $|A|$ denotes the Lebesgue measure of $A$;
- if $\mu$ is a measure on ${{\mathbb R}}^2$ and $A$ is a Borel subset of ${{\mathbb R}}^2$, $\mu {\mathop{\hbox{\vrule height 7pt width .5pt depth 0pt
\vrule height .5pt width 6pt depth 0pt}}\nolimits}A$ denotes the restriction of $\mu$ to $A$, i.e. $(\mu {\mathop{\hbox{\vrule height 7pt width .5pt depth 0pt
\vrule height .5pt width 6pt depth 0pt}}\nolimits}A)(B):=\mu(B \cap A)$ for all Borel sets $B \subseteq {{\mathbb R}}^2$;
- if $\sigma \in ]0,+\infty[$, $o_\sigma$ is such that $\lim_{\sigma \to 0^+} o_\sigma=0$.
For the general theory of functions of bounded variation, we refer to [@AFP]; here we recall some basic definitions and theorems we need in the sequel. Let $A$ be an open subset of ${{\mathbb R}}^N$, and let $u: A \to {{\mathbb R}}^n$. We say that $u \in BV(A;{{\mathbb R}}^n)$ if $u \in L^1(A;{{\mathbb R}}^n)$, and its distributional derivative is a vector-valued Radon measure on $A$. We say that $u \in SBV(A;{{\mathbb R}}^n)$ if $u \in BV(A;{{\mathbb R}}^n)$ and its distributional derivative can be represented as $$Du(A)= \int_A \nabla u(x) \,dx+ \int_{A \cap S_u} (u^+(x)-u^-(x)) \otimes \nu_x
\,d{{\mathcal H}}^{N-1}(x),$$ where $\nabla u$ denotes the approximate gradient of $u$, $S_u$ denotes the set of approximate jumps of $u$, $u^+$ and $u^-$ are the traces of $u$ on $S_u$, $\nu_x$ is the normal to $S_u$ at $x$, and ${{\mathcal H}}^{N-1}$ is the $(N-1)$-dimensional Hausdorff measure. The space $SBV(A;{{\mathbb R}}^n)$ is called the space of [*special functions of bounded variation*]{}. Note that if $u \in SBV(A;{{\mathbb R}}^n)$, then the singular part of $Du$ is concentrated on $S_u$ which turns out to be countably ${{\mathcal H}}^{N-1}$-rectifiable.
The space $SBV$ is very useful when dealing with variational problems involving volume and surface energies because of the following compactness and lower semicontinuity result due to L.Ambrosio (see [@A1], [@A2], [@A3], [@AFP]).
\[SBVcompact\] Let $A$ be an open and bounded subset of ${{\mathbb R}}^N$, and let $(u_k)$ be a sequence in $SBV(A;{{\mathbb R}}^n)$. Assume that there exists $q>1$ and $c \ge 0$ such that $$\int_A |\nabla u_k|^q \,dx+ {{\mathcal H}}^{N-1}(S_{u_k})+ ||u_k||_\infty \le c$$ for every $k \in {{\mathbb N}}$. Then there exists a subsequence $(u_{k_h})$ and a function $u \in SBV(A;{{\mathbb R}}^n)$ such that $$\begin{aligned}
\label{sbvconv}
\nonumber
u_{k_h} \to u \quad {strongly \; in}\; L^1(A;{{\mathbb R}}^n); \\
\nabla u_{k_h} {\rightharpoonup}\nabla u \quad {weakly \; in}\; L^1(A;M^{N \times n}); \\
\nonumber
{{\mathcal H}}^{N-1}(S_u) \le \liminf_h {{\mathcal H}}^{N-1}(S_{u_{k_h}}).\end{aligned}$$
In the rest of the paper, we will say that $u_k \to u$ in $SBV(A;{{\mathbb R}}^n)$ if $u_k$ and $u$ satisfy . It will also be useful the following fact which can be derived from Ambrosio’s Theorem: if $u_k \to u$ in $SBV(A;{{\mathbb R}}^n)$ and if ${{\mathcal H}}^{N-1} {\mathop{\hbox{\vrule height 7pt width .5pt depth 0pt
\vrule height .5pt width 6pt depth 0pt}}\nolimits}S_{u_k} {\stackrel{\ast}{\rightharpoonup}}\mu$ weakly-star in the sense of measures, then ${{\mathcal H}}^{N-1} {\mathop{\hbox{\vrule height 7pt width .5pt depth 0pt
\vrule height .5pt width 6pt depth 0pt}}\nolimits}S_u \le \mu$ as measures. We will set $SBV(A):=SBV(A;{{\mathbb R}})$.
Let ${\Omega}$ be an open bounded subset of ${{\mathbb R}}^N$ with Lipschitz boundary, and let $\partial_D {\Omega}$ be a subset of $\partial {\Omega}$ open in the relative topology. Let $g:[0,1] \to H^1({\Omega})$ be absolutely continuous; we indicate the gradient of $g$ at time $t$ by $\nabla g(t)$, and the time derivative of $g$ at time $t$ by $\dot{g}(t)$. The main result of [@FL] is the following theorem.
\[qse\] There exists a crack $\Gamma(t) \subseteq {\overline{\Omega}}$ and a field $u(t) \in SBV({\Omega})$ such that:
- $\Gamma(t)$ increases with $t$;
- - $u(0)$ minimizes $$\int_{\Omega}|\nabla v|^2 \,dx +
{{\mathcal H}}^{N-1}( S_v \cup
\{x \in \partial_D {\Omega}: v(x) \not= g(0)(x)\})$$ among all $v \in SBV({\Omega})$ (inequalities on $\partial_D {\Omega}$ are intended for the traces of $v$ and $g$);
- - for $t>0$, $u(t)$ minimizes $$\int_{\Omega}|\nabla v|^2 \,dx
+{{\mathcal H}}^{N-1} \left( \left[ S_v \cup
\{x \in \partial_D {\Omega}: v(x) \not= g(t)(x)\} \right] \setminus \Gamma(t)
\right)$$ among all $v \in SBV({\Omega})$;
- - $S_{u(t)} \cup \{x \in \partial_D {\Omega}\,:\, u(t)(x) \not= g(t)(x) \}
\subseteq \Gamma(t)$, up to a set of ${{\mathcal H}}^{N-1}$-measure $0$.
Furthermore, the total energy $${{\mathcal E}}(t):= \int_{\Omega}|\nabla u(t)|^2 \,dx +{{\mathcal H}}^{N-1}( \Gamma(t))$$ is absolutely continuous and satisfies $${{\mathcal E}}(t)={{\mathcal E}}(0)+2 \int_0^t \int_{\Omega}\nabla u(\tau) \nabla \dot{g}(\tau) \,dx \,d\tau$$ for every $t \in [0,1]$. Finally, for any countable, dense set $I \subseteq [0,1]$, the crack $\Gamma(t)$ and the displacement $u(t)$ can be chosen so that $$\Gamma(t)= \bigcup_{\tau \in I, \tau \le t}
\left( S_{u(\tau)} \cup
\{x \in \partial_D {\Omega}\,:\, u(\tau)(x) \not= g(\tau)(x) \} \right).$$
The main tool in the proof of Theorem \[qse\] is the following result [@FL Theorem 2.1], which is useful also in our analysis.
\[jumptransfer\] Let ${\overline{\Omega}}\subseteq {\Omega}'$, with $\partial {\Omega}$ Lipschitz, and let for $r=1,\dots,i$ $(u^r_n)$ be a sequence in $SBV({\Omega}')$ such that
- $S_{u^r_n} \subseteq {\overline{\Omega}}$;
- - $|\nabla u^r_n|$ weakly converges in $L^1({\Omega}')$; and
- - $u^r_n \to u^r$ strongly in $L^1({\Omega}')$,
where $u^r \in BV({\Omega}')$ with ${{\mathcal H}}^{N-1}(S_{u^r}) <\infty$. Then for every $\phi \in
SBV({\Omega}')$ with ${{\mathcal H}}^{N-1}(S_\phi) <\infty$ and $\nabla \phi \in L^q({\Omega};{{\mathbb R}}^N)$ for some $q \in [1,+\infty[$, there exists a sequence $(\phi_n)$ in $SBV({\Omega}')$ with $\phi_n \equiv \phi$ on ${\Omega}' \setminus {\overline{\Omega}}$ such that
- $\phi_n \to \phi$ strongly in $L^1({\Omega}')$;
- - $\nabla \phi_n \to \nabla \phi$ strongly in $L^q({\Omega}')$; and
- - ${{\mathcal H}}^{N-1} \left( [S_{\phi_n} \setminus \bigcup_{r=1}^i S_{u^r_n}] \setminus
[S_{\phi} \setminus \bigcup_{r=1}^i S_{u^r}] \right) \to 0$.
In particular $$\label{jt}
\limsup_n {{\mathcal H}}^{N-1} \left( S_{\phi_n} \setminus \bigcup_{r=1}^i S_{u^r_n} \right) \le
{{\mathcal H}}^{N-1} \left( S_{\phi} \setminus \bigcup_{r=1}^i S_{u^r} \right).$$
Let $A \subseteq {{\mathbb R}}^2$ be open and bounded, and let ${{\mathcal K}}(\overline{A})$ be the set of all compact subsets of $\overline{A}$. ${{\mathcal K}}(\overline{A})$ can be endowed by the Hausdorff metric $d_H$ defined by $$d_H(K_1,K_2) := \max \left\{ \sup_{x \in K_1} {\rm dist}(x,K_2), \sup_{y \in
K_2} {\rm dist}(y,K_1)\right\},$$ with the conventions ${\rm dist}(x, \emptyset)= {\rm diam}(A)$ and $\sup
\emptyset=0$, so that $d_H(\emptyset, K)=0$ if $K=\emptyset$ and $d_H(\emptyset,K)={\rm diam}(A)$ if $K \not=\emptyset$. It turns out that ${{\mathcal K}}(\overline{A})$ endowed with the Hausdorff metric is a compact space (see e.g. [@Ro]).
Let ${\Omega}\subseteq {{\mathbb R}}^2$ be a polygonal set and let us fix two positive constants $0<c_1<c_2$. By a [*regular triangulation*]{} of ${\Omega}$ of size ${\varepsilon}$ we intend a finite family of (closed) triangles $T_i$ such that ${\overline{\Omega}}=\bigcup_i T_i$, $T_i \cap T_j$ is either empty or equal to a common edge or to a common vertex, and each $T_i$ contains a ball of diameter $c_1 {\varepsilon}$ and is contained in a ball of diameter $c_2 {\varepsilon}$.
We indicate by ${{\mathcal R}}_{\varepsilon}({\Omega})$ the family of all regular triangulations of ${\Omega}$ of size ${\varepsilon}$. It turns out that there exist $0<\vartheta_1 < \vartheta_2 <\pi$ such that for all $T$ belonging to a triangulation ${{\bf T}}\in {{\mathcal R}}_{\varepsilon}({\Omega})$, the inner angles of $T$ are between $\vartheta_1$ and $\vartheta_2$. Moreover, every edge of $T$ has length greater than $c_1 {\varepsilon}$ and lower than $c_2 {\varepsilon}$.
Let us fix a triangulation ${{\bf R}}_{\varepsilon}\in {{\mathcal R}}_{\varepsilon}({\Omega})$ for all ${\varepsilon}>0$ and let $a \in ]0,\frac{1}{2}[$. Let us consider a new triangulation ${{\bf T}}$ nested in ${{\bf R}}_{\varepsilon}$ obtained dividing each $T \in {{\bf R}}_{\varepsilon}$ into four triangles taking over every edge $[x,y]$ of $T$ a knot $z$ which satisfies $$z=tx+(1-t)y, \quad \quad t \in [a,1-a].$$ We will call these new vertices [*adaptive*]{}, the triangles obtained joining these points [*adaptive triangles*]{}, and their edges [*adaptive edges*]{} (see Fig.1).
We denote by ${{\mathcal T}}_{{\varepsilon},a}({\Omega})$ the set of all triangulations ${{\bf T}}$ constructed in this way. Note that for all ${{\bf T}}\in {{\mathcal T}}_{{\varepsilon},a}({\Omega})$ there exists $0<c_1^a<c_2^a<+\infty$ such that every $T_i \in {{\bf T}}$ contains a ball of diameter $c_1^a {\varepsilon}$ and is contained in a ball of diameter $c_2^a {\varepsilon}$. Then there exist $0<\vartheta_1^a < \vartheta_2^a <\pi$ such that for all triangles $T$ belonging to a triangulation ${{\bf T}}\in {{\mathcal T}}_{{\varepsilon},a}({\Omega})$, the inner angles of $T$ are between $\vartheta_1^a$ and $\vartheta_2^a$. Moreover, every edge of $T$ has length greater than $c_1^a {\varepsilon}$ and lower than $c_2^a {\varepsilon}$.
We will often use the following [*interpolation estimate*]{} (see [@Cia Theorem 3.1.5]). If $u \in W^{2,2}({\Omega})$ and $T \in {{\bf R}}_{\varepsilon}$, let $u_T$ denote the affine interpolation of $u$ on $T$. We have that there exists ${K}$ depending only on $c_1,c_2$ such that $$\label{interp1}
\|u_T -u\|_{W^{1,2}(T)} \le {K}{\varepsilon}\|u\|_{W^{2,2}(T)}.$$ Estimate holds also for ${{\bf T}}\in {{\mathcal T}}_{{\varepsilon},a}({\Omega})$: in this case $K$ depends on $a$.
The following lemmas will be used in Section 4.
\[simplecurve\] Let ${{\bf T}}\in {{\mathcal T}}_{{\varepsilon},a}({\Omega})$, and let $l \subseteq {\Omega}$ be a segment with extremes $p,\,q$ belonging to edges of ${{\bf T}}$. There exists a polyhedral curve $\Gamma$ with extremal points $p$ and $q$ (see Fig.2) such that $\Gamma$ is contained in the union of the edges of those $T \in {{\bf T}}$ with $T \cap l \neq \emptyset$, and such that the following properties hold:
- $\Gamma = \gamma_p \cup \gamma \cup \gamma_q$, where $\gamma$ is union of edges of ${{\bf T}}$ and $\gamma_p$, $\gamma_q$ are segments containing $p$ and $q$ respectively, and each one is contained in an edge of ${{\bf T}}$;
- - there exists a constant $c$ independent of $\epsilon$ (but depending on $a$) such that $${{\mathcal H}}^1(\Gamma)\le c \, {{\mathcal H}}^1(l).$$
Let $\{T_1, \ldots T_k\}$ be the family of triangles in ${{\bf T}}$ such that the intersection with $l$ is a segment with positive length. For every integer $1\le i \le k$, let $l_i:=T_i\cap l.$ If $l_i$ is an edge of $T_i$, we set $D_i= T_i$. Otherwise let $D_i$ be a connected component of $T_i \setminus l_i$ such that $|D_i| \leq \frac{1}{2} |T_i|.$ We claim that there exists a constant $c>0$ independent of $\epsilon$ such that $$\label{inelung}
{{\mathcal H}}^1 (\partial D_i) \leq c\, {{\mathcal H}}^1(l_i).$$ We have to analyze two possibilities, namely $D_i$ is a triangle, or $D_i$ is a trapezoid. Suppose that $D_i$ is a triangle and that $m_i$ is an edge of $D_i$. Let $\alpha$ be the angle of $D_i$ opposite to $l_i$. It is easy to prove that ${{\mathcal H}}^1(l_i) \ge {{\mathcal H}}^1(m_i) \sin\alpha$, and so $${{\mathcal H}}^1(l_i) \ge \frac{1}{3} \sin\alpha {{\mathcal H}}^1(\partial D_i).$$ Since $\vartheta_1^a\le\alpha\le\vartheta_2^a$, $\sin\alpha$ is uniformly bounded from below, and hence inequality (\[inelung\]) follows. If $D_i$ is a trapezoid, since $|D_i| \leq \frac{1}{2} |T_i|$, it follows that $T_i\setminus D_i$ is a triangle such that its edges different from $l_i$ have length greater than $\frac{1}{2} c_1^a {\varepsilon}$. Let $\alpha$ be the inner angle of $T_i\setminus D_i$ opposite to $l_i$. We have that $${{\mathcal H}}^1(l_i) \ge \frac{1}{2}\sin\alpha c_1^a {{\varepsilon}} \ge
\frac{1}{2} \sin\alpha \frac {c_1^a}{c_2^a}\frac{1}{4} {{\mathcal H}}^1(\partial D_i).$$ Since $\vartheta_1^a\le\alpha\le\vartheta_2^a$, inequality (\[inelung\]) follows.
By , we deduce that $${{\mathcal H}}^1 \left( \bigcup_{i=1}^k\partial D_i \right) \leq c\, {{\mathcal H}}^1(l);$$ moreover, since $\bigcup_{i=1}^k\partial D_i$ is arcwise connected and contains $p,q$, we conclude that there exists a curve $\Gamma \subseteq
\bigcup_{i=1}^k\partial D_i$ which satisfies the thesis.
\[guscio\] There exists a constant $c>0$ such that for every segment $l \subseteq {\Omega}$ there exists ${\varepsilon}_0$ with the following property: for every $\varepsilon \leq \varepsilon_0$, setting ${{\mathcal R}}(l):=\{T\in {{\bf R}}_{\varepsilon}:\, T \cap l \neq \emptyset\}$, we have $${{\mathcal H}}^1(\partial {{\mathcal R}}(l))\leq c {{\mathcal H}}^1(l).$$
Let $ {{\mathcal N}}_{{\varepsilon}}(l):=\{ x \in {\Omega}:\, {\rm dist}(x,l) \leq c_2 {\varepsilon}\}$. We have that $|{{\mathcal N}}_{{\varepsilon}}(l)| = {{\mathcal H}}^1(l)c_2{\varepsilon}+ \pi c_2^2{\varepsilon}^2$, and hence there exists a positive constant ${\varepsilon}_0$ such that, for every ${\varepsilon}\leq {\varepsilon}_0$, we have that $$|{{\mathcal N}}_{{\varepsilon}}(l)| \leq 2 {{\mathcal H}}^1(l) c_2 {\varepsilon}.$$ We have that ${{\mathcal R}}(l) \subseteq {{\mathcal N}}_{{\varepsilon}}(l),$ and $$\sharp {{\mathcal R}}(l) \le \frac{4}{c_1^2\pi^2} \frac{|{{\mathcal N}}_{{\varepsilon}}(l)|}{{\varepsilon}^2},$$ where $\sharp {{\mathcal R}}(l)$ denotes the number of triangles of ${{\mathcal R}}(l)$. Then, we have $${{\mathcal H}}^1(\partial {{\mathcal R}}(l)) \leq 3 c_2 {\varepsilon}\, \sharp{{\mathcal R}}(l) \leq
3 c_2 {\varepsilon}\frac{4}{c_1^2\pi^2}
\frac{|{{\mathcal N}}_{{\varepsilon}}(l)|}{{\varepsilon}^2} \leq
3 c_2^2 \frac{4}{c_1^2\pi^2} 2{{\mathcal H}}^1(l),$$ and so the proof is concluded.
Let $A \subseteq {{\mathbb R}}^2$ be open. We say that $K \subseteq A$ is polygonal (with respect to $A$), if it is the intersection of $A$ with the union of a finite number of closed segments. The following density result is proved in [@C].
\[piecedensity\] Assume that $\partial A$ is locally Lipschitz, and let $u \in SBV(A)$ such that $u \in L^2(A)$, $\nabla u \in L^2({\Omega};{{\mathbb R}}^2)$, and ${{\mathcal H}}^{1}(S_u)<+\infty$. For every ${\varepsilon}>0$, there exists a function $v \in SBV(A)$ such that
- $S_v$ is essentially closed, i.e., ${{\mathcal H}}^{1}( \overline{S_v} \setminus S_v)=0$;
- - $\overline{S_v}$ is a polyhedral set;
- - $v \in W^{k, \infty}(A \setminus \overline{S_v})$ for every $k \in {{\mathbb N}}$;
- - $||v-u||_{L^2(A)} < {\varepsilon}$;
- - $||\nabla v- \nabla u||_{L^2(A; {{\mathbb R}}^2)} < {\varepsilon}$;
- - $|{{\mathcal H}}^{1}(S_v)-{{\mathcal H}}^{1}(S_u)| <{\varepsilon}$.
Let $\partial_D {\Omega}$ be a relatively open subset of $\partial {\Omega}$ composed of edges lying in $\partial {\Omega}$. Let us consider ${\Omega}_D$ polygonal open bounded subset of ${{\mathbb R}}^2$ such that ${\Omega}_D \cap {\Omega}=\emptyset$ and $\partial {\Omega}\cap \partial {\Omega}_D= \partial_D {\Omega}$ up to a finite number of vertices. We set ${\Omega}':= {\Omega}\cup {\Omega}_D \cup \partial_D {\Omega}$. In Section 4, we will use the following result.
\[regularization\] Given $u \in SBV({\Omega}')$ with $u=0$ on ${\Omega}' \setminus {\overline{\Omega}}$ and ${{\mathcal H}}^{N-1}(S_u)<+\infty$, there exists $u_h \in SBV({\Omega}')$ such that
- $u_h=0$ in ${\Omega}' \setminus {\overline{\Omega}}$;
- - $S_{u_h}$ is polyhedral, $\overline{S_{u_h}} \subseteq {\Omega}$ and $u_h \in
W^{k,\infty}({\Omega}' \setminus \overline{S_{u_h}})$ for all $k$;
- - $u_h \to u$ strongly in $L^2({\Omega}')$ and $\nabla u_h \to \nabla u$ strongly in $L^2({\Omega}';{{\mathbb R}}^2)$;
- - for all $A$ open subset of ${\Omega}'$ with ${{\mathcal H}}^1(\partial A \cap S_u)=0$, we have $$\lim_h {{\mathcal H}}^1(A \cap S_{u_h})={{\mathcal H}}^1(A \cap S_u).$$
Using a partition of unity, we may prove the result in the case ${\Omega}:=]-1,1[ \times ]0,1[$, ${\Omega}':=]-1,1[ \times ]-1,1[$, and $\partial_D {\Omega}:=]-1,1[ \times \{0\}$. We set $w_h(x,y):=u(x,y-h)$, and let $\varphi_h$ be a cut off function with $\varphi_h=1$ on $]-1,1[ \times ]-1,\frac{h}{3}[$, $\varphi_h=0$ on $]-1,1[ \times ]\frac{h}{2},1[$, and $||\nabla \varphi_h||_\infty \le \frac{7}{h}$. Let us set $v_h:=(1-\varphi_h)w_h$. We have that $v_h=0$ in ${\Omega}' \setminus {\overline{\Omega}}$; moreover we have $$\nabla v_h=(1-\varphi_h) \nabla w_h-\nabla \varphi_h w_h.$$ Since $\nabla \varphi_h w_h \to 0$ strongly in $L^2({\Omega}';{{\mathbb R}}^2)$, we have $\nabla v_h \to \nabla u$ strongly in $L^2({\Omega}';{{\mathbb R}}^N)$. Finally, for all $A$ open subset of ${\Omega}'$ with ${{\mathcal H}}^1(\partial A \cap S_u)=0$, we have $$\lim_h {{\mathcal H}}^1(A \cap S_{v_h})={{\mathcal H}}^1(A \cap S_u).$$ In order to conclude the proof, let us apply Theorem \[piecedensity\] obtaining $\tilde{v}_h$ with polyhedral jumps in ${\Omega}$ such that $\tilde{v}_h
\in W^{k,\infty}({\Omega}' \setminus \overline{S_{\tilde{v}_h}})$, $||w_h-\tilde{v}_h||_{L^2({\Omega})}+
||\nabla w_h -\nabla \tilde{v}_h||_{L^2({\Omega};{{\mathbb R}}^2)} \le h^2$ and $|{{\mathcal H}}^{N-1}(S_{w_h}) -{{\mathcal H}}^{N-1}(S_{\tilde{v}_h})| \le h$. If we set $u_h:= \varphi_h g+(1-\varphi_h) \tilde{v}_h$, we obtain the thesis.
The discontinuous finite element approximation {#devol}
==============================================
In this section we construct a discrete approximation of quasi-static evolution of brittle fractures in linearly elastic bodies: the discretization is done both in space and time.
From now on we suppose that ${\Omega}$ is a polygonal open bounded subset of ${{\mathbb R}}^2$, and that $\partial_D {\Omega}\subseteq \partial {\Omega}$ is open in the relative topology. For all $\varepsilon>0$, we fix a triangulation ${{\bf R}}_\varepsilon \in {{\mathcal R}}_\varepsilon({\Omega})$, and suppose that $\partial_D {\Omega}$ is composed of edges of ${{\bf R}}_{\varepsilon}$ for all ${\varepsilon}$; we indicate the family of these edges by ${{\bf S}}_{\varepsilon}$.
We consider the following discontinuous finite element space. We indicate by ${{{\mathcal A}}_{\varepsilon,a}({\Omega})}$ the set of all $u$ such that there exists a triangulation ${{\bf T}}(u) \in {{{\mathcal T}}_{\varepsilon,a}({\Omega})}$ nested in ${{\bf R}}_\varepsilon$ with $u$ affine on every $T \in {{\bf T}}(u)$. For every $u \in {{{\mathcal A}}_{\varepsilon,a}({\Omega})}$, we write $\|\nabla u\|$ for the $L^2$-norm of $\nabla u$ and we indicate by $S_u$ the family of edges of ${{\bf T}}(u)$ inside ${\Omega}$ across which $u$ is discontinuous. Notice that $u \in SBV({\Omega})$ and that the notation is consistent with the usual one employed in the theory of functions with bounded variation. Let us also denote by ${{{\mathcal A}}{{\mathcal F}}_{\varepsilon}({\Omega})}$ the set of affine functions in ${\Omega}$ with respect to the triangulation ${{\bf R}}_{\varepsilon}$. Finally, given any $g \in {{{\mathcal A}}{{\mathcal F}}_{\varepsilon}({\Omega})}$, for all $u \in {{{\mathcal A}}_{\varepsilon,a}({\Omega})}$ we set $$\label{jump*}
S_D^g(u):= \{\zeta \in {{\bf S}}_{\varepsilon}\,:\, u \not=g \mbox{ on } \zeta\},$$ that is $S_D^g(u)$ denotes the edges at which the boundary condition is not satisfied. Moreover we set $$\label{jumpbis}
{S^{g}(u)}:= S_u \cup S_D^g(u)$$ Let now consider $g \in W^{1,1}([0,1];H^1({\Omega}))$ with $g(t) \in {{{\mathcal A}}{{\mathcal F}}_{\varepsilon}({\Omega})}$ for all $t \in [0,1]$. Let $\delta>0$ and let $N_\delta$ be the largest integer such that $\delta (N_\delta-1) < 1$; for $0\le i \le N_\delta-1$ we set $t_i^\delta:=i\delta$, $t^\delta_{N_\delta}:=1$ and $g_i^\delta:=g(t_i^\delta)$. The following proposition holds.
\[discrevol\] Let ${\varepsilon}>0$, $a \in ]0,\frac{1}{2}]$ and $\delta>0$ be fixed. Then for all $i=0, \ldots, N_\delta$ there exists $u^{\delta,i}_{{\varepsilon},a} \in {{{\mathcal A}}_{\varepsilon,a}({\Omega})}$ such that, setting $$\Gamma^{\delta,i}_{{\varepsilon},a}:=
\bigcup_{r=0}^i {S^{g^\delta_r}(u^{\delta,r}_{{\varepsilon},a})},$$ the following hold:
- - $\|u^{\delta,i}_{{\varepsilon},a}\|_\infty \le \|g^\delta_i\|_\infty$;
- - for all $v \in {{{\mathcal A}}_{\varepsilon,a}({\Omega})}$ we have $$\label{MSdiscr}
\|\nabla u^{\delta,0}_{{\varepsilon},a}\|^2+
{{\mathcal H}}^1 \left(
{S^{g_0^\delta}(u^{\delta,0}_{{\varepsilon},a})} \right)
\le \|\nabla v\|^2 +
{{\mathcal H}}^1\left( {S^{g_0^\delta}(v)} \right),$$ and $$\label{piecemin}
\|\nabla u^{\delta,i}_{{\varepsilon},a}\|^2 \le \|\nabla v\|^2 +
{{\mathcal H}}^1\left( {S^{g^\delta_i}(v)} \setminus
\Gamma^{\delta,i-1}_{{\varepsilon},a}\right).$$
The proof is carried out through a variational argument. Let $u^{\delta,0}_{{\varepsilon},a}$ be a minimum of the following problem $$\label{step0}
\min
\left\{ \|\nabla u\|^2+ {{\mathcal H}}^1({S^{g^\delta_0}(u)}) \right\}.$$ We set $\Gamma^{\delta,0}_{{\varepsilon},a}:= {S^{g^\delta_0}(u^{\delta,0}_{{\varepsilon},a})}$. Recursively, supposing to have constructed $u^{\delta,i-1}_{{\varepsilon},a}$ and $\Gamma^{\delta,i-1}_{{\varepsilon},a}$, let $u^{\delta,i}_{{\varepsilon},a}$ be a minimum for $$\label{stepj}
\min \left\{ \|\nabla u\|^2 +
{{\mathcal H}}^1 \left( {S^{g^\delta_i}(u)} \setminus
\Gamma^{\delta,i-1}_{{\varepsilon},a} \right) \right\}.$$ We set $\Gamma^{\delta,i}_{{\varepsilon},a}:=
{S^{g^\delta_i}(u^{\delta,i}_{{\varepsilon},a})} \cup \Gamma^{\delta,i-1}_{{\varepsilon},a}$. We claim that problems and admit a solution $u^{\delta,i}_{{\varepsilon},a}$ such that $\|u^{\delta,i}_{{\varepsilon},a}\|_{\infty} \le \|g^\delta_i\|_\infty$ for all $i=0, \ldots, N_\delta$. We prove the claim for problem , the other case being similar. Let $(u_n)$ be a minimizing sequence for problem : since $g^\delta_i$ is an admissible test function, we deduce that for $n$ large $$\|\nabla u_n\|^2+ {{\mathcal H}}^1 \left( {S^{g^\delta_i}(u_n)}
\setminus \Gamma^{\delta,i-1}_{{\varepsilon},a} \right) \le
\|\nabla g^\delta_i\|^2.$$ Moreover, we may modify $u_n$ in the following way. If $\pi$ denotes the projection in ${{\mathbb R}}$ over the interval $I:=[-\|g^\delta_i\|_\infty, \|g^\delta_i\|_\infty]$, let $\tilde{u}_n \in {{{\mathcal A}}_{\varepsilon,a}({\Omega})}$ be defined on each $T \in {{\bf T}}(u_n)$ as the affine interpolation of the values $(\pi(u_n(x_1)),\; \pi(u_n(x_2)),\; \pi(u_n(x_3))$, where $x_1, \;x_2$ and $x_3$ are the vertices of $T$. Note that by construction we have for all $n$ $$\|\tilde{u}_n\|_\infty \le \|g^\delta_i\|_\infty, \quad \quad
\|\nabla \tilde{u}_n\| \le \|\nabla u_n\|, \quad \quad
{S^{g^\delta_i}(\tilde{u}_n)} \subseteq
{S^{g^\delta_i}(u_n)},$$ so that $(\tilde{u}_n)$ is a minimizing sequence for problem . We conclude that it is not restrictive to assume $\|u_n\|_\infty \le \|g^\delta_i\|_\infty$.
Since ${{\bf T}}(u_n) \in {{{\mathcal T}}_{\varepsilon,a}({\Omega})}$, we have that the number of elements of ${{\bf T}}(u_n)$ is uniformly bounded. Up to a subsequence, we may suppose that there exists an integer $k$ such that ${{\bf T}}(u_n)$ has exactly $k$ elements $T_n^1, \ldots, T_n^k$. Using a diagonal argument we may suppose that, up to a further subsequence, there exists ${{\bf T}}=\{T^1, \ldots, T^k\} \in {{{\mathcal T}}_{\varepsilon,a}({\Omega})}$ such that $T_n^i \to T^i$ in the Hausdorff metric for all $i=1, \ldots, k$. Let us consider $T^i \in {{\bf T}}$, and let $\tilde{T}^i$ be contained in the interior of $T^i$. For $n$ large enough, $\tilde{T}^i$ is contained in the interior of $T^i_n$ and $(u_n)_{|\tilde{T}^i}$ is affine with $\int_{\tilde{T}^i} |\nabla u_n|^2 \,dx \le C$ with $C$ independent of $n$. We deduce that there exists a function $u^i$ affine on $\tilde{T}^i$ such that up to a subsequence $u_n \to u$ uniformly on $\tilde{T}^i$. Since $\tilde{T}^i$ is arbitrary, it turns out that $u^i$ is actually defined on $T^i$ and $$\int_{T^i} |\nabla u^i|^2 \,dx \le \liminf_n \int_{T_n^i} |\nabla u_n|^2 \,dx.$$ Let $u \in {{{\mathcal A}}_{\varepsilon,a}({\Omega})}$ such that $u=u^i$ on $T^i$ for every $i=1,\dots,k$: we have $$\|\nabla u\|^2 \le \liminf_n \|\nabla u_n\|^2.$$ On the other hand, it is easy to see that ${S^{g^\delta_i}(u)}$ is contained in the Hausdorff limit of ${S^{g^\delta_i}(u_n)}$, and that $${{\mathcal H}}^1 \left(
{S^{g^\delta_i}(u)} \setminus \Gamma^{\delta,i-1}_{{\varepsilon},a}
\right)
\le
\liminf_n {{\mathcal H}}^1 \left(
{S^{g^\delta_i}(u_n)}
\setminus \Gamma^{\delta,i-1}_{{\varepsilon},a}
\right).$$ We conclude that $u$ is a minimum point for the problem with $\|u\|_\infty \le \|g^\delta_i\|_\infty$. We have that point $(a)$ is proved.
Concerning point $(b)$, by construction we get ; for $i \ge 1$ we have $$\|\nabla u^{\delta,i}_{{\varepsilon},a}\|^2 +
{{\mathcal H}}^1 \left(
{S^{g^\delta_i}(u^{\delta,i}_{{\varepsilon},a})}
\setminus \Gamma^{\delta,i-1}_{{\varepsilon},a} \right) \le
\|\nabla v\|^2 +
{{\mathcal H}}^1 \left( {S^{g^\delta_i}(v)}
\setminus \Gamma^{\delta,i-1}_{{\varepsilon},a} \right)$$ for all $v \in {{{\mathcal A}}_{\varepsilon,a}({\Omega})}$, so that $$\|\nabla u^{\delta,i}_{{\varepsilon},a}\|^2 \le
\|\nabla v\|^2 +
{{\mathcal H}}^1 \left( {S^{g^\delta_i}(v)} \setminus
\Gamma^{\delta,i-1}_{{\varepsilon},a} \right),$$ and this proves point $(b)$.
\[oss\] [For technical reasons due to the asymptotic analysis of the discrete evolution $u^{\delta,i}_{{\varepsilon},a}$ when $\delta \to 0$, ${\varepsilon}\to 0$ and $a \to 0$, we define $u^{\delta,i}_{{\varepsilon},a}$ from $u^{\delta,i-1}_{{\varepsilon},a}$ through problem without requiring that the adaptive vertices determining $\Gamma^{\delta,i-1}_{{\varepsilon},a}$ remain fixed. We just penalize their possible changes if they are used to create new fracture: in fact in this case, the surface energy increases at each change of a quantity at least of order $a{\varepsilon}$. As a consequence, during the step by step minimization, it could happen that some triangles $T \in {{\mathcal T}}_{{\varepsilon},a}({\Omega})$ contain the fracture $\Gamma^{\delta,i}_{{\varepsilon},a}$ in their interior. This is in contrast with the interpretation of the triangles as elementary blocks for the elasticity problem, but being this situation penalized in the minimization process, we expect that it occurs rarely. ]{}
The following estimate is essential for the study of asymptotic behavior of the discrete evolution.
\[dener\] If $(u^{\delta,i}_{{\varepsilon},a},\Gamma^{\delta,i}_{{\varepsilon},a})$ for $i=0,\dots,N_\delta$ satisfies condition $(b)$ of Proposition \[discrevol\], setting ${{\mathcal E}}^{\delta,i}_{{\varepsilon},a}:=\|\nabla u^{\delta,i}_{{\varepsilon},a}\|^2+
{{\mathcal H}}^1 \left( \Gamma^{\delta,i}_{{\varepsilon},a}\right)$, we have for $0 \le j \le i \le N_{\delta}$ $$\label{discrenergy}
{{\mathcal E}}^{\delta,i}_{{\varepsilon},a} \le {{\mathcal E}}^{\delta,j}_{{\varepsilon},a}+
2 \sum_{r=j}^{i-1} \int_{t^\delta_{r}}^{t^\delta_{r+1}}
\int_{{\Omega}} \nabla u^{\delta,r}_{{\varepsilon},a} \nabla\dot{g}(\tau)
\,dx\,d\tau +o^\delta,$$ where $$\label{odelta}
o^\delta:= \left[
\max_{r=0,\ldots,N_\delta-1}
\int_{t^\delta_{r}}^{t^\delta_{r+1}} \|\dot{g}(\tau)\|_{H^1({\Omega})} \,d\tau
\right]
\int_0^1 \|\dot{g}(\tau)\|_{H^1({\Omega})}\,d\tau.$$
For all $0 \le j \le N_\delta-1$, by construction of $u^{\delta,j+1}_{{\varepsilon},a}$ we have that $$\begin{gathered}
\|\nabla u^{\delta,j+1}_{{\varepsilon},a}\|^2 +
{{\mathcal H}}^1 \left(
{S^{g^\delta_{j+1}}(u^{\delta,j+1}_{{\varepsilon},a})}
\setminus \Gamma^{\delta,j}_{{\varepsilon},a} \right)
\le
\|\nabla u^{\delta,j}_{{\varepsilon},a}+ \nabla (g^\delta_{j+1} - g^\delta_{j})\|^2 =\\
=
\|\nabla u^{\delta,j}_{{\varepsilon},a}\|^2 +
2 \int_{{\Omega}}
\nabla u^{\delta,j}_{{\varepsilon},a} \nabla (g^\delta_{j+1} - g^\delta_{j})
\,dx+||\nabla (g^\delta_{j+1}-g^\delta_{j})||^2.\end{gathered}$$ Notice that $$\nabla (g^\delta_{j+1}-g^\delta_{j}) =
\int_{t^\delta_{j}}^{t^\delta_{j+1}}
\nabla \dot{g}(\tau) \,d\tau,$$ so that $$\begin{gathered}
\label{ineqi}
\|\nabla u^{\delta,j+1}_{{\varepsilon},a}\|^2 +
{{\mathcal H}}^1 \left(
{S^{g^\delta_{j+1}}(u^{\delta,j+1}_{{\varepsilon},a})}
\setminus \Gamma^{\delta,j}_{{\varepsilon},a} \right) \le \\
\le \|\nabla u^{\delta,j}_{{\varepsilon},a}\|^2 +
2 \int_{t^\delta_{j}}^{t_{j+1}^\delta} \int_{{\Omega}}
\nabla u^{\delta,j}_{{\varepsilon},a} \nabla \dot{g}(\tau)\,dx \,d\tau
+e(\delta)\int_{t^\delta_{j}}^{t^\delta_{j+1}} \|\dot{g}(\tau)\|_{H^1({\Omega})}
\,d\tau,\end{gathered}$$ where $$e(\delta):= \max_{r=0,\ldots,N_\delta-1}
\int_{t^\delta_{r}}^{t^\delta_{r+1}} \|\dot{g}(\tau)\|_{H^1({\Omega})} \,d\tau.$$ From , we obtain that for all $0 \le j \le i \le N_\delta$ $$\begin{gathered}
||\nabla u^{\delta,i}_{{\varepsilon},a}||^2+ {{\mathcal H}}^1(\Gamma^{\delta,i}_{{\varepsilon},a}) \le
||\nabla u^{\delta,j}_{{\varepsilon},a}||^2 + {{\mathcal H}}^1(\Gamma^{\delta,j}_{{\varepsilon},a})+ \\
+2 \sum_{r=j}^{i-1} \int_{t^\delta_{r}}^{t^\delta_{r+1}} \int_{{\Omega}}
\nabla u^{\delta,r}_{{\varepsilon},a} \nabla \dot{g}(\tau) \,dx\,d\tau
+e(\delta) \int_{t^\delta_{j}}^{t^\delta_{i}}
||\dot{g}(\tau)||_{H^1({\Omega})} \,d\tau,\end{gathered}$$ and so the proof of point $(c)$ is complete choosing $$o^\delta:=e(\delta) \int_0^1 \|\dot{g}(\tau)\|_{H^1({\Omega})}\,d\tau.$$
The convergence result {#convres}
======================
This section is devoted to the proof of Theorem \[mainthm\]. As in Section \[devol\], let ${\Omega}$ be a polygonal open bounded subset of ${{\mathbb R}}^2$, and let $\partial_D {\Omega}\subseteq \partial {\Omega}$ be open in the relative topology. For all $\varepsilon>0$, let ${{\bf R}}_\varepsilon \in {{\mathcal R}}_\varepsilon({\Omega})$ be a regular triangulation of ${\Omega}$ such that $\partial_D {\Omega}$ is composed of edges of ${{\bf R}}_{\varepsilon}$. As in the previous section, let ${{{\mathcal A}}{{\mathcal F}}_{\varepsilon}({\Omega})}$ be the family of continuous piecewise affine functions with respect to ${{\bf R}}_{\varepsilon}$, and let ${{{\mathcal A}}_{\varepsilon,a}({\Omega})}$ be the family of functions which are affine on the triangles of some triangulation ${{\bf T}}\in {{{\mathcal T}}_{\varepsilon,a}({\Omega})}$ nested in ${{\bf R}}_{\varepsilon}$ and can jump across the edges of ${{\bf T}}$.
In the following, it will be useful to treat points at which the boundary condition is violated (see ) as internal jumps. Thus we consider ${\Omega}_D$ polygonal open bounded subset of ${{\mathbb R}}^2$ such that ${\Omega}_D \cap {\Omega}=\emptyset$ and $\partial {\Omega}\cap \partial {\Omega}_D= \partial_D {\Omega}$ up to a finite number of points; we set ${\Omega}':= {\Omega}\cup {\Omega}_D \cup \partial_D {\Omega}$. Given $u \in {{{\mathcal A}}_{\varepsilon,a}({\Omega})}$ and $g \in {{{\mathcal A}}{{\mathcal F}}_{\varepsilon}({\Omega})}$, we may extend $g$ to a function of $H^1({\Omega}')$ and $u$ to a function $\tilde{u} \in SBV({\Omega}')$ setting $\tilde{u}=g$ on ${\Omega}_D$. In this way, recalling , we have $${S^{g}(u)}=S_{\tilde{u}},$$ so that the violation of the boundary condition of $u$ can be read in the set of jumps of $\tilde{u}$. Analogously, given $u \in SBV({\Omega})$ and $g \in H^1({\Omega})$, we set $$\label{jumptris}
{S^{g}(u)}:=S_u \cup \{x \in \partial_D {\Omega}\,:\, \gamma(u)(x) \not= \gamma(g)(x)\}$$ where $\gamma$ denotes the trace operator on $\partial {\Omega}$. We may assume $g \in H^1({\Omega}')$ using an extension operator. We can then consider $\tilde{u} \in SBV({\Omega}')$ such that $\tilde{u}=u$ on ${\Omega}$, and $\tilde{u}=g$ on ${\Omega}_D$. In this way we have $${S^{g}(u)}=S_{\tilde{u}} \quad \mbox{ up to a set of ${{\mathcal H}}^1${-}measure $0$}.$$ Let us consider $g \in W^{1,1}([0,1],H^1({\Omega}))$ such that $\|g(t)\|_\infty \le C$ for all $t \in [0,1]$ and let $g_{\varepsilon}\in W^{1,1}([0,1],H^1({\Omega}))$ be such that $g_{\varepsilon}(t) \in {{{\mathcal A}}{{\mathcal F}}_{\varepsilon}({\Omega})}$ for all $t \in [0,1]$, $$\label{lib}
\|g_{\varepsilon}(t)\|_\infty \le C$$ for all $t \in [0,1]$, and for ${\varepsilon}\to 0$ $$\label{strconv}
g_{\varepsilon}\to g
\quad \mbox{ strongly in }W^{1,1}([0,1],H^1({\Omega})).$$ We indicate by $\{u^{\delta,i}_{{\varepsilon},a},\,i=0, \ldots,N_\delta\}$ the discrete evolution relative to the boundary data $g_{\varepsilon}$ given by Proposition \[discrevol\], and we denote by ${{\mathcal E}}^{\delta,i}_{{\varepsilon},a}$ its total energy as in Proposition \[dener\].
We assume that $g(\cdot)$ and $g_{\varepsilon}(\cdot)$ are defined in $H^1({\Omega}')$ (we still denote these extensions by $g(\cdot)$ and $g_h(\cdot)$), in such a way that and hold in ${\Omega}'$. Let us moreover set $g^\delta_{\varepsilon}(t):=g_{\varepsilon}(t^\delta_i)$ for all $t^\delta_i \le t <t^\delta_{i+1}$ with $i=0,\dots,N_\delta-1$ and $g^\delta_{\varepsilon}(1):=g_{\varepsilon}(1)$.
Let us make the following piecewise constant interpolation in time: $$u^\delta_{{\varepsilon},a}(t):=u^{\delta,i}_{{\varepsilon},a}
\quad \mbox{ for } t^\delta_i \le t <t^\delta_{i+1}
\quad i=0,\dots,N_\delta-1,$$ and $u^\delta_{{\varepsilon},a}(1):=u^{\delta,N_\delta}_{{\varepsilon},a}$. For all $t \in [0,1]$ we define the [*discrete fracture*]{} at time $t$ as $$\Gamma^\delta_{{\varepsilon},a}(t):= \bigcup_{s \le t}
{S^{g^\delta_{\varepsilon}(s)}(u^\delta_{{\varepsilon},a}(s))},$$ and the [*discrete total energy*]{} at time $t$ as $${{\mathcal E}}^\delta_{{\varepsilon},a}(t):=
\|\nabla u^\delta_{{\varepsilon},a}(t)\|^2+
{{\mathcal H}}^1 \left( \Gamma^\delta_{{\varepsilon},a}(t) \right).$$ We have for all $t \in [0,1]$ $$\label{boundonu}
\|u^{\delta}_{{\varepsilon},a}(t)\|_\infty \le \|g_{\varepsilon}^\delta(t)\|_\infty.$$ Moreover for all $v \in {{{\mathcal A}}_{\varepsilon,a}({\Omega})}$ we have $$\label{MSdiscr2}
\|\nabla u^{\delta}_{{\varepsilon},a}(0)\|^2
+{{\mathcal H}}^1\left( {S^{g^\delta_{\varepsilon}(0)}(u^{\delta}_{{\varepsilon},a}(0))} \right)
\le \|\nabla v\|^2 +
{{\mathcal H}}^1\left( {S^{g^\delta_{\varepsilon}(0)}(v)} \right),$$ and for all $t \in ]0,1]$ and for all $v \in {{{\mathcal A}}_{\varepsilon,a}({\Omega})}$ $$\label{piecemin2}
\|\nabla u^{\delta}_{{\varepsilon},a}(t)\|^2 \le \|\nabla v\|^2 +
{{\mathcal H}}^1 \left( {S^{g^\delta_{\varepsilon}(t)}(v)} \setminus
\Gamma^{\delta}_{{\varepsilon},a}(t) \right).$$ Finally for all $0 \le s \le t \le 1$ we have $$\label{discrenergybists}
{{\mathcal E}}^{\delta}_{{\varepsilon},a}(t) \le {{\mathcal E}}^{\delta}_{{\varepsilon},a}(s)+
2\int_{s^\delta_{i}}^{t^\delta_{i}}
\int_{{\Omega}} \nabla u^{\delta}_{{\varepsilon},a}(\tau) \nabla
\dot{g}_{\varepsilon}(\tau) \,dx\,d\tau +o^\delta_{{\varepsilon}},$$ where $t^\delta_i \le t <t^\delta_{i+1}$, $s^\delta_i \le s <s^\delta_{i+1}$ and $$\label{opiccolo}
o^\delta_{{\varepsilon}}:=
\left[
\max_{r=0,\ldots,N_\delta-1}
\int_{t^\delta_{r}}^{t^\delta_{r+1}} \|\dot{g}_{\varepsilon}(\tau)\|_{H^1({\Omega})} \,d\tau
\right] \int_0^1 \|\dot{g}_{\varepsilon}(\tau)\|_{H^1({\Omega})}.$$ For $s=0$ we obtain the following estimate from above for the discrete total energy $$\label{discrenergybis}
{{\mathcal E}}^{\delta}_{{\varepsilon},a}(t) \le {{\mathcal E}}^{\delta}_{{\varepsilon},a}(0)+
2\int_{0}^{t^\delta_{i}}
\int_{{\Omega}} \nabla u^{\delta}_{{\varepsilon},a}(\tau) \nabla
\dot{g}_{\varepsilon}(\tau) \,dx\,d\tau +o^\delta_{{\varepsilon}},$$ where $t^\delta_i \le t <t^\delta_{i+1}$.
We study the behavior of the evolution $\{t \to u^\delta_{\varepsilon,a}(t),\,t \in [0,1]\}$ varying the parameters in the following way. We let firstly ${\varepsilon}\to 0$ and $\delta \to 0$ obtaining an evolution $\{t \to u_a(t),\,t \in [0,1]\}$ relative to the boundary data $g$ with the minimality property ; then we let $a \to 0$ obtaining a quasi-static evolution of brittle fractures $\{t \to u(t)\,,t \in [0,1]\}$ relative to the boundary data $g$. Finally, by a diagonal argument we deal with $(\delta,{\varepsilon},a)$ at the same time.
In order to develop this program, we need some compactness, and so we derive a bound for the total energy ${{\mathcal E}}^\delta_{{\varepsilon},a}$. By , we have that for all $t \in [0,1]$ $$\|\nabla u^{\delta}_{\varepsilon,a}(t)\| \le \|\nabla g^\delta_{\varepsilon}(t)\| \le \tilde{C}$$ with $\tilde C$ independent of $\delta$, ${\varepsilon}$ and $t$. We deduce for all $t \in [0,1]$ $${{\mathcal E}}^{\delta}_{\varepsilon,a}(t) \le
{{\mathcal E}}^{\delta}_{\varepsilon,a}(0)+
2 \tilde{C}^2 +o^\delta_{\varepsilon}$$ Notice that ${{\mathcal E}}^\delta_{{\varepsilon},a}(0)$ is uniformly bounded as $\delta,{\varepsilon}$ vary. Moreover, by and since $\|g_{\varepsilon}(t)\|_\infty \le C$ for all $t \in [0,1]$, we have that $u^{\delta}_{\varepsilon,a}(t)$ is uniformly bounded in $L^\infty({\Omega})$ independently of $\delta,{\varepsilon}$ and $a$. Taking into account , we conclude that there exists $C'$ independent of $\delta,\varepsilon,a$ such that for all $t\in [0,1]$ $$\label{unifenergybound}
{{\mathcal E}}^{\delta}_{\varepsilon,a}(t) + \|u^{\delta}_{\varepsilon,a}(t)
\|_\infty \le C'.$$ Formula gives the desired compactness in order to perform the asymptotic analysis of the discrete evolution.
Let now consider $\delta_n \to 0$ and ${\varepsilon}_n \to 0$: by we have $$\label{Cto0}
o^{\delta_n}_{{\varepsilon}_n} \to 0,$$ where $o^{\delta_n}_{{\varepsilon}_n}$ is defined in . By Helly’s theorem on monotone functions, we may suppose that there exists an increasing function $\lambda_a$ such that (up to a subsequence) for all $t \in [0,1]$ $$\label{lambn}
\lambda_{n,a}(t):=
{{\mathcal H}}^1 \left( \bigcup_{s \le t}{S^{g^{\delta_n}_{{\varepsilon}_n}(s)}(u^{\delta_n}_{{\varepsilon}_n,a}(s))} \right)
\to \lambda_a(t).$$ Let us fix $D \subseteq [0,1]$ countable and dense with $0 \in D$.
\[eps-to-0\] For all $t \in D$ there exists ${u_a(t)}\in SBV({\Omega})$ such that up to a subsequence independent of $t$ $$u^{\delta_n}_{{\varepsilon}_n,a}(t) \to {u_a(t)}\quad \mbox{ in } SBV({\Omega}).$$ Moreover for all $t \in D$ we have $$\label{ueb2}
\|\nabla {u_a(t)}\|^2+ {{\mathcal H}}^1 \big( {S^{g(t)}({u_a(t)})} \big) +
\|{u_a(t)}\|_\infty \le C'.$$
Let us consider $t \in D$. By , we can apply Ambrosio’s Compactness Theorem \[SBVcompact\] obtaining $u \in SBV({\Omega})$ such that, up to a subsequence, $u^{\delta_n}_{{\varepsilon}_n,a}(t) \to u$ in $SBV({\Omega})$. Let us set ${u_a(t)}:=u$. Using a diagonal argument, we deduce that there exists a subsequence of $(\delta_n, {\varepsilon}_n)$ (which we still denote by $(\delta_n,{\varepsilon}_n)$) such that $u^{\delta_n}_{{\varepsilon}_n,a}(t) \to
{u_a(t)}$ in $SBV({\Omega})$ for all $t \in D$. In order to obtain inequality , we extend $u^{\delta_n}_{{\varepsilon}_n,a}(t)$ and ${u_a(t)}$ to ${\Omega}'$ setting $u^{\delta_n}_{{\varepsilon}_n,a}(t):= g^{\delta_n}_{{\varepsilon}_n}(t)$ and $u_a(t):=g(t)$ on ${\Omega}_D$; since $g^{\delta_n}_{{\varepsilon}_n}(t) \to g(t)$ on ${\Omega}_D$ strongly in $H^1({\Omega}_D)$, we have that $u^{\delta_n}_{{\varepsilon}_n,a}(t) \to u_a(t)$ in $SBV({\Omega}')$, so that we can apply Ambrosio’s Theorem, and derive from .
The following result is essential for the sequel: its proof is postponed to Section \[secmin\].
\[pminpropD\] Let $t \in D$. For all $v \in SBV({\Omega})$ we have $$\label{minpropD}
\|\nabla {u_a(t)}\|^2 \le \|\nabla v\|^2+ \mu(a) {{\mathcal H}}^1({S^{g(t)}(v)} \setminus
\bigcup_{s \le t, s \in D} {S^{g(s)}({u_a(s)})}),$$ where $\mu:]0,\frac{1}{2}[ \to ]0,+\infty[$ is such that $\lim_{a \to 0} \mu(a)=1$. Moreover, $\nabla u^{\delta_n}_{{\varepsilon}_n,a}(t) \to \nabla {u_a(t)}$ strongly in $L^2({\Omega};{{\mathbb R}}^2)$.
We now extend the evolution $\{t \to u_a(t)\,:\,t \in D\}$ to the entire interval $[0,1]$. Let us set for all $t \in [0,1]$ $$\Gamma_a(t):= \bigcup_{s \le t, s \in D} {S^{g(s)}(u_a(s))}.$$
\[aextension\] For every $t \in [0,1]$ there exists $u_a(t) \in SBV({\Omega})$ such that the following hold:
- for all $t \in [0,1]$ $$\label{ajumpt}
{S^{g(t)}(u_a(t))} \subseteq
\Gamma_a(t) \;\mbox{ up to a set of }{{\mathcal H}}^1\mbox{-measure }0,$$ and $$\label{aueb}
\|\nabla {u_a(t)}\|^2+ {{\mathcal H}}^1 \big( {S^{g(t)}({u_a(t)})} \big) +
\|{u_a(t)}\|_\infty \le C';$$
- - for all $v \in SBV({\Omega})$ $$\label{amint}
\|\nabla u_a(t)\|^2 \le \|\nabla v\|^2 +
\mu(a){{\mathcal H}}^1 \left( {S^{g(t)}(v)} \setminus \Gamma_a(t) \right);$$
- - $\nabla u_a$ is left continuous in $[0,1] \setminus D$ with respect to the strong topology of $L^2({\Omega};{{\mathbb R}}^2)$;
- - for all $t \in [0,1] \setminus {{\mathcal N}}_a$ we have that $$\nabla u^{\delta_n}_{{\varepsilon}_n,a}(t) \to \nabla u_a(t)
\quad \mbox{ strongly in }L^2({\Omega}, {{\mathbb R}}^2),$$ where ${{\mathcal N}}_a$ is the set of discontinuities of the function $\lambda_a$ defined in .
Let $t \in [0,1] \setminus D$ and let $t_n \in D $ with $t_n \nearrow t$. By , we can apply Ambrosio’s Theorem to the sequence $(u_a(t_n))$ obtaining $u \in SBV({\Omega})$ such that, up to a subsequence, $u_a(t_n) \to u$ in $SBV({\Omega})$. Let us set ${u_a(t)}:=u$. Let us extend $u_a(t_n)$ and ${u_a(t)}$ to ${\Omega}'$ setting $u_a(t_n):=g(t_n)$ and ${u_a(t)}:=g(t)$ on ${\Omega}_D$: we have $u_a(t_n) \to {u_a(t)}$ in $SBV({\Omega}')$. Since ${{\mathcal H}}^1 {\mathop{\hbox{\vrule height 7pt width .5pt depth 0pt
\vrule height .5pt width 6pt depth 0pt}}\nolimits}S_{u_a(t_n)} \le {{\mathcal H}}^1 {\mathop{\hbox{\vrule height 7pt width .5pt depth 0pt
\vrule height .5pt width 6pt depth 0pt}}\nolimits}\Gamma_a(t)$ for all $n$, as a consequence of Ambrosio’s Theorem, we deduce that ${{\mathcal H}}^1 {\mathop{\hbox{\vrule height 7pt width .5pt depth 0pt
\vrule height .5pt width 6pt depth 0pt}}\nolimits}S_{{u_a(t)}} \le {{\mathcal H}}^1 {\mathop{\hbox{\vrule height 7pt width .5pt depth 0pt
\vrule height .5pt width 6pt depth 0pt}}\nolimits}\Gamma_a(t)$. This means ${{\mathcal H}}^1 {\mathop{\hbox{\vrule height 7pt width .5pt depth 0pt
\vrule height .5pt width 6pt depth 0pt}}\nolimits}{S^{g(t)}({u_a(t)})} \le {{\mathcal H}}^1 {\mathop{\hbox{\vrule height 7pt width .5pt depth 0pt
\vrule height .5pt width 6pt depth 0pt}}\nolimits}\Gamma_a(t)$, so that holds. Moreover, for all $v \in SBV({\Omega})$, by we may write $$\begin{gathered}
\label{minbis}
\|\nabla u_a(t_n)\|^2 \le
\|\nabla v-\nabla g(t)+\nabla g(t_n)\|^2+
\mu(a) {{\mathcal H}}^1 \left( {S^{g(t)}(v)} \setminus \Gamma_a(t_n) \right) \le \\
\le \|\nabla v-\nabla g(t)+\nabla g(t_n)\|^2+
\mu(a) {{\mathcal H}}^1 \left( {S^{g(t)}(v)} \setminus \Gamma_a(t) \right) +
\mu(a) {{\mathcal H}}^1 \left( \Gamma_a(t) \setminus \Gamma_a(t_n) \right),\end{gathered}$$ so that, since by definition of $\Gamma_a(t)$ we have ${{\mathcal H}}^1(\Gamma_a(t) \setminus \Gamma_a(t_n)) \to 0$, we obtain that holds; choosing $v=u_a(t)$ and taking the limsup in , we obtain that $$\limsup_n \|\nabla u_{a}(t_n)\|^2 \le \|\nabla u_a(t)\|^2,$$ and so the convergence $\nabla u_a(t_{n}) \to \nabla u_a(t)$ is strong in $L^2({\Omega}, {{\mathbb R}}^2)$. Notice that $\nabla u_a(t)$ is uniquely determined by (\[ajumpt\]) and (\[amint\]) since the gradient of the solutions of the minimum problem $$\min \left\{ \|\nabla u\|^2 \,:\, {S^{g(t)}(u)} \subseteq \Gamma_a(t)
\mbox{ up to a set of }{{\mathcal H}}^1\mbox{-measure }0 \right\}$$ is unique by the strict convexity of the functional: we conclude that $\nabla u_a(t)$ is well defined. The same arguments prove that $\nabla u_a$ is left continuous at all $t \in [0,1] \setminus D$. Finally is a direct consequence of and of Ambrosio’s Theorem, and so points $(a)$, $(b)$, $(c)$ are proved.
Let us come to point $(d)$. Let us consider $u^{\delta_n}_{{\varepsilon}_n,a}(t)$ with $t \not\in {{\mathcal N}}_a$; we may suppose that $t \not \in D$, since otherwise the result has already been established. By Proposition \[pminpropD\] with $D':=D \cup\{t\}$ in place of $D$, we have that, up to a subsequence, $u^{\delta_n}_{{\varepsilon}_n,a}(t) \to u$ in $SBV({\Omega})$ such that $$\|\nabla u\|^2 \le \|\nabla v\|^2 +\mu(a) {{\mathcal H}}^1 \left( {S^{g(t)}(v)}
\setminus (\Gamma_a(t) \cup {S^{g(t)}(u)}) \right)$$ for all $v \in SBV({\Omega})$ and $\nabla u^{\delta_n}_{{\varepsilon}_n,a}(t) \to \nabla u$ strongly in $L^2({\Omega};{{\mathbb R}}^2)$. Let $s<t$ with $ s\in D$; by the minimality of $u^{\delta_n}_{{\varepsilon}_n,a}(s)$ and by we have $$\begin{gathered}
\|\nabla u^{\delta_n}_{{\varepsilon}_n,a}(s)\|^2 \le
\|\nabla u^{\delta_n}_{{\varepsilon}_n,a}(t)-\nabla g^{\delta_n}_{{\varepsilon}_n}(t)+
\nabla g^{\delta_n}_{{\varepsilon}_n}(s)\|^2
+\lambda_{n,a}(t)-\lambda_{n,a}(s) \le \\
\le \|\nabla u^{\delta_n}_{{\varepsilon}_n,a}(t)\|^2
+2 \sqrt{C'}\|\nabla g^{\delta_n}_{{\varepsilon}_n}(t)-\nabla g^{\delta_n}_{{\varepsilon}_n}(s)\|+ \\
+\|\nabla g^{\delta_n}_{{\varepsilon}_n}(t)-\nabla g^{\delta_n}_{{\varepsilon}_n}(s)\|^2+
\lambda_{n,a}(t)-\lambda_{n,a}(s).\end{gathered}$$ Passing to the limit for $n \to +\infty$, recalling that $g^{\delta_n}_{{\varepsilon}_n}(\tau) \to g(\tau)$ strongly in $H^1({\Omega})$ for all $\tau \in [0,1]$, we deduce $$\begin{gathered}
\|\nabla u_a(s)\|^2 \le \|\nabla u\|^2+
2\sqrt{C'}\|\nabla g(t)-\nabla g(s)\|+
\|\nabla g(t)-\nabla g(s)\|^2+
\lambda_a(t)-\lambda_a(s),\end{gathered}$$ so that, since $t$ is a point of continuity for $\lambda_a$, $\nabla u_a$ is left continuous at $t$, and $g$ is absolutely continuous, we get for $s \to t$ $$\|\nabla u_a(t)\|^2 \le \|\nabla u\|^2.$$ We conclude that $u_a(t)$ is a solution of $$\min \{\|\nabla v\|^2\,:\,
{S^{g(t)}(v)} \subseteq \Gamma_a(t) \cup {S^{g(t)}(u)}
\mbox{ up to a set of }{{\mathcal H}}^1\mbox{-measure }0\},$$ so that $\nabla u=\nabla u_a(t)$ by uniqueness of the gradient of the solution. We deduce that $\nabla u^{\delta_n}_{{\varepsilon}_n,a}(t) \to \nabla u_a(t)$ strongly in $L^2({\Omega};{{\mathbb R}}^2)$, and so the proof is complete.
We can now let $a \to 0$.
\[azero\] There exists $a_n \to 0$ such that, for all $t \in D$, $u_{a_n}(t) \to u(t)$ in $SBV({\Omega})$ for some $u(t) \in SBV({\Omega})$ such that for all $v \in SBV({\Omega})$ we have $$\label{alessminpropD}
\|\nabla u(t)\|^2 \le \|\nabla v\|^2+ {{\mathcal H}}^1({S^{g(t)}(v)} \setminus
\bigcup_{s \le t, s \in D} {S^{g(s)}(u(s))}).$$ Moreover, $\nabla u_{a_n}(t) \to \nabla u(t)$ strongly in $L^2({\Omega};{{\mathbb R}}^2)$ and $$\label{ueb3}
\|\nabla u(t) \|^2+ {{\mathcal H}}^1({S^{g(t)}(u(t))}) +
\|u(t)\|_\infty \le C'.$$
By , applying Ambrosio’s Theorem to the extensions of ${u_a(t)}$ to ${\Omega}'$ by setting ${u_a(t)}:=g(t)$ on ${\Omega}_D$, and using a diagonal argument, we find a sequence $a_n \to 0$ such that, for all $t \in D$, $u_{a_n}(t) \to u(t)$ in $SBV({\Omega})$ for some $u(t) \in SBV({\Omega})$ such that holds.
We now prove that $u(t)$ satisfies property (\[alessminpropD\]). Let $v \in SBV({\Omega})$. Let us fix $t_1 \le t_2 \le \ldots \le t_k=t$ with $t_i \in D$. We extend $v$ and $u_{a_n}(t_i)$ to ${\Omega}'$ setting $v:=g(t)$ and $u_{a_n}(t_i):=g(t_i)$ on ${\Omega}_D$ respectively. Since $u_{a_n}(t_i) \to u(t_i)$ in $SBV({\Omega}')$ for all $i=1, \ldots, k$, by Theorem \[jumptransfer\] there exists $v_n \in SBV({\Omega}')$ with $v_n=g(t)$ on ${\Omega}_D$ such that $\nabla v_n \to \nabla v$ strongly in $L^2({\Omega}';{{\mathbb R}}^2)$ and $$\label{convlength}
\limsup_n {{\mathcal H}}^1 \left( S_{v_n} \setminus \bigcup_{i=1}^k S_{u_{a_n}(t_i)} \right)
\le {{\mathcal H}}^1 \left( S_v \setminus \bigcup_{i=1}^k S_{u(t_i)} \right).$$ By we obtain $$\label{minan}
\|\nabla u_{a_n}(t)\|^2 \le \|\nabla v_n\|^2+ \mu(a_n)
{{\mathcal H}}^1 \left( S_{v_n} \setminus \bigcup_{i=1}^k S_{u_{a_n}(t_i)} \right),$$ so that passing to the limit for $n \to +\infty$ and recalling that $\mu(a) \to 1$ as $a \to 0$, we obtain $$\|\nabla u(t)\|^2 \le \|\nabla v\|^2+
{{\mathcal H}}^1 \left( S_v \setminus \bigcup_{i=1}^k S_{u(t_i)} \right).$$ Thus we get $$\|\nabla u(t)\|^2 \le \|\nabla v\|^2+
{{\mathcal H}}^1 \left( {S^{g(t)}(v)} \setminus \bigcup_{i=1}^k {S^{g(t_i)}(u(t_i))} \right).$$ Since $t_1, \ldots, t_k$ are arbitrary, we obtain . Choosing $v=u(t)$, taking the limsup in and using , we obtain $\nabla u_{a_n}(t) \to \nabla u(t)$ strongly in $L^2({\Omega};{{\mathbb R}}^2)$.
In order to deal with $\delta,{\varepsilon}$ and $a$ at the same time, we need the following lemma.
\[convD\] Let $\{u(t)\,:\,t \in D\}$ be as in Lemma \[azero\]. There exist $\delta_n \to 0$, $\varepsilon_n \to 0$, and $a_n \to 0$ such for all $t \in D$ we have $$u^{\delta_n}_{\varepsilon_n,a_n}(t) \to u(t)
\quad \mbox{ in $SBV({\Omega})$ }.$$ Moreover, for all $n$ there exists ${{\mathcal B}}_n \subseteq [0,1]$ with $|{{\mathcal B}}_n|<2^{-n}$ such that for all $t \in [0,1] \setminus {{\mathcal B}}_n$ $$\label{graduanconv}
\|\nabla u^{\delta_n}_{\varepsilon_n,a_n}(t)-\nabla u_{a_n}(t)\|
\le \frac{1}{n}.$$ Finally, we have that for all $v \in SBV({\Omega})$ $$\label{min0}
\| \nabla u(0)\|^2 +{{\mathcal H}}^1 \left( {S^{g(0)}(u(0))} \right)
\le \|\nabla v\|^2+{{\mathcal H}}^1 \left( {S^{g(0)}(v)} \right)$$ and $$\label{conv0}
{{\mathcal E}}^{\delta_n}_{{\varepsilon}_n,a_n}(0) \to \|\nabla u(0)\|^2+
{{\mathcal H}}^1 \left( {S^{g(0)}(u(0))} \right).$$
Let $(a_n)$ be the sequence determined by Lemma \[azero\]. By Lemma \[eps-to-0\], for all $n$ there exists $(\delta^n_m,\varepsilon^n_{m})$ such that for all $t \in D$ and $m \to +\infty$ we have $$u^{\delta^n_m}_{\varepsilon^n_m,a_n}(t) \to u_{a_n}(t)
\quad \mbox{ in }SBV({\Omega}),$$ and $$\nabla u^{\delta^n_m}_{\varepsilon^n_m,a_n}(t)
\to \nabla u_{a_n}(t)
\quad \mbox{ strongly in }L^2({\Omega};{{\mathbb R}}^2).$$ Moreover by Lemma \[aextension\] we have that $\nabla u^{\delta^n_m}_{\varepsilon^n_m,a_n} \to \nabla u_{a_n}$ quasi-uniformly on $[0,1]$ as $m \to +\infty$. Let ${{\mathcal B}}_n \subseteq [0,1]$ with $|{{\mathcal B}}_n| <2^{-n}$ such that $\nabla u^{\delta^n_m}_{\varepsilon^n_m,a_n} \to \nabla u_{a_n}$ uniformly on $[0,1] \setminus {{\mathcal B}}_n$ as $m \to +\infty$. We now perform the following diagonal argument. Let $D=\{t_n,\,n \ge 1\}$. Choose $m_1$ such that $$\|\nabla u^{\delta^1_{m_1}}_{\varepsilon^1_{m_1},a_1}(t_1)-
\nabla u_{a_1}(t_1)\|+
\|u^{\delta^1_{m_1}}_{\varepsilon^1_{m_1},a_1}(t_1)-u_{a_1}(t_1)\|
\le 1,$$ and $$\|\nabla u^{\delta^1_{m_1}}_{\varepsilon^1_{m_1},a_1}(t)-
\nabla u_{a_1}(t)\| \le 1
\quad \mbox{ for all $t \in [0,1] \setminus {{\mathcal B}}_1$ }.$$ Let $m_n$ be such that $$\|\nabla u^{\delta^n_{m_n}}_{\varepsilon^n_{m_n},a_n}(t_j)
-\nabla u_{a_n}(t_j)\|+
\|u^{\delta^n_{m_n}}_{\varepsilon^n_{m_n},a_n}(t_j)-u_{a_n}(t_j)\|
\le \frac{1}{n}
\;\;\mbox{ for all $j=1, \ldots, n$}$$ and $$\|\nabla u^{\delta^n_{m_n}}_{\varepsilon^n_{m_n},a_n}(t)-
\nabla u_{a_n}(t)\| \le \frac{1}{n}
\quad \mbox{ for all $t \in [0,1] \setminus {{\mathcal B}}_n$ }.$$ We may suppose that $\delta^n_{m_n} \to 0$, ${\varepsilon}^n_{m_n} \to 0$. Then $(\delta^n_{m_n},{\varepsilon}^n_{m_n},a_n)$ is the sequence which satisfies the thesis. In fact by construction and taking into account , for all $t \in D$ we have $u^{\delta^n_{m_n}}_{\varepsilon^n_{m_n},a_n}(t) \to u(t)$ in $SBV({\Omega})$; moreover the set ${{\mathcal B}}_n$ satisfies . Notice that $u^{\delta^n_{m_n}}_{{\varepsilon}^n_{m_n},a^n_{m_n}}(0)$ satisfies and so and follow by the $\Gamma$[-]{}convergence result of [@N].
Let $(\delta_n,{\varepsilon}_n,a_n)$ be the sequence determined by Lemma \[convD\]. For all $t \in [0,1]$ let us set $$\lambda_n(t):=
{{\mathcal H}}^1 \left( \Gamma^{\delta_n}_{{\varepsilon}_n,a_n}(t) \right).$$ By Helly’s theorem, we may suppose that there exist two increasing functions $\lambda$ and $\eta$ such that up to a subsequence $$\lambda_n \to \lambda \quad \mbox{ pointwise in }[0,1],$$ and $$\label{eta}
\lambda_{a_n} \to \eta \quad \mbox{ pointwise in }[0,1],$$ where $\lambda_{a_n}$ is defined as in . We now extend the evolution $\{t \to u(t)\,:\,t \in D\}$ to the entire interval $[0,1]$. Let us set for all $t \in [0,1]$ $$\Gamma(t):= \bigcup_{s \le t, s \in D} {S^{g(s)}(u(s))},$$ and let ${{\mathcal N}}$ be the set of discontinuities of ${{\mathcal H}}^1(\Gamma(\cdot))$. Notice that for all $t \in [0,1]$ $$\label{size2}
{{\mathcal H}}^1(\Gamma(t)) \le \lambda(t).$$ In fact if $t \in D$, let $t_1 \le t_2 \le \ldots \le t_k=t$ with $t_i \in D$, consider $w_n \in SBV({\Omega}';{{\mathbb R}}^k)$ defined as $$w_n(x):=(u^{\delta_n}_{\varepsilon_n,a_n}(t_1)(x), \ldots,
u^{\delta_n}_{\varepsilon_n,a_n}(t_k)(x)),$$ where we assume that $u^{\delta_n}_{\varepsilon_n,a_n}(t_i)=g^{\delta_n}_{{\varepsilon}_n}(t_i)$ on ${\Omega}_D$. We have $w_n \to w:=(u(t_1), \ldots, u(t_k))$ in $SBV({\Omega}';{{\mathbb R}}^k)$, where $u(t_i)=g(t_i)$ on ${\Omega}_D$. Note that for all $n$ we have $S_{w_n}=
\bigcup_{i=1}^k S_{u^{\delta_n}_{\varepsilon_n,a_n}(t_i)}$ so that $${{\mathcal H}}^1(S_{w_n}) \le \lambda_n(t).$$ Passing to the limit for $n \to +\infty$ and applying Ambrosio’s Theorem we get $${{\mathcal H}}^1 \left( \bigcup_{i=1}^k S_{u(t_i)} \right)
={{\mathcal H}}^1(S_w) \le \liminf_n {{\mathcal H}}^1(S_{w_n}) \le
\lambda(t);$$ we thus have $${{\mathcal H}}^1 \left( \bigcup_{i=1}^k {S^{g(t_i)}(u(t_i))} \right)
={{\mathcal H}}^1(S_w) \le \lambda(t)$$ and taking the sup over all $t_1, \ldots, t_k$, we obtain in $D$. The case $t \not \in D$ follows since ${{\mathcal H}}^1(\Gamma(\cdot))$ is left continuous by definition.
\[extension\] For every $t \in [0,1]$ there exists $u(t) \in SBV({\Omega})$ such that the following hold:
- for all $t \in [0,1]$ $$\label{jumpt}
{S^{g(t)}(u(t))} \subseteq
\Gamma(t) \; \mbox{ up to a set of }{{\mathcal H}}^1\mbox{-measure }0,$$ and for all $t \in [0,1]$ and for all $v \in SBV({\Omega})$ $$\label{mint}
\|\nabla u(t)\|^2 \le
\|\nabla v\|^2+ {{\mathcal H}}^1 \left( {S^{g(t)}(v)} \setminus \Gamma(t) \right);$$
- - $\nabla u$ is continuous in $[0,1] \setminus (D \cup {{\mathcal N}})$ with respect to the strong topology of $L^2({\Omega};{{\mathbb R}}^2)$;
- - if $\tilde{{{\mathcal N}}}$ is the set of discontinuities of the function $\eta$ defined in , for all $t \in [0,1] \setminus \tilde{{{\mathcal N}}}$ we have that $$\nabla u_{a_n}(t) \to \nabla u(t)
\quad \mbox{ strongly in }L^2({\Omega}, {{\mathbb R}}^2).$$
Finally $$\label{energybelow}
{{\mathcal E}}(t) \ge {{\mathcal E}}(0)+
2\int_0^t \int_{{\Omega}} \nabla u(\tau) \nabla \dot{g}(\tau) \,dx\,d\tau,$$ where $$\label{defenergy}
{{\mathcal E}}(t):= \|\nabla u(t)\|^2+{{\mathcal H}}^1(\Gamma(t)).$$
The definition of $u(t)$ is carried out as in Lemma \[aextension\] considering $t \in [0,1] \setminus D$, $t_n \in D$ with $t_n \nearrow t$, and the limit (up to a subsequence) of $u(t_n)$ in $SBV({\Omega})$: and hold, so that point $(a)$ is proved. It turns out that $\nabla u(t)$ is uniquely determined and that it is left continuous in $[0,1] \setminus D$. Let us consider $t \in [0,1] \setminus (D \cup {{\mathcal N}})$, and let $t_n \searrow t$. By Ambrosio’s Theorem, we have that there exists $u \in SBV({\Omega})$ with such that, up to a subsequence, $u(t_n) \to u$ in $SBV({\Omega})$. Since $t$ is a continuity point of ${{\mathcal H}}^1(\Gamma(\cdot))$, we deduce that ${S^{g(t)}(u)} \subseteq \Gamma(t)$ up to a set of ${{\mathcal H}}^1$-measure $0$. Moreover by the minimality property for $u(t_n)$ and the fact $\Gamma(t) \subseteq \Gamma(t_n)$, we have that for all $v \in SBV({\Omega})$ with $$\begin{gathered}
\|\nabla u(t_n)\|^2 \le
\|\nabla v-\nabla g(t)+\nabla g(t_n)\|^2+
{{\mathcal H}}^1 \left( {S^{g(t)}(v)} \setminus \Gamma(t_n) \right) \le \\
\le \|\nabla v-\nabla g(t)+\nabla g(t_n)\|^2+
{{\mathcal H}}^1 \left( {S^{g(t)}(v)} \setminus \Gamma(t) \right),\end{gathered}$$ and so we deduce that holds with $u$ in place of $u(t)$, and that $\nabla u(t_n) \to \nabla u$ strongly in $L^2({\Omega};{{\mathbb R}}^2)$. We obtain by uniqueness that $\nabla u=\nabla u(t)$, and so $\nabla u(\cdot)$ is continuous in $[0,1] \setminus (D \cup {{\mathcal N}})$ and this proves point $(b)$. Point $(c)$ follows in the same way of point $(d)$ of Lemma \[aextension\].
Let us come to the proof of . Given $t \in [0,1]$ and $k>0$, let $s_i^k:= \frac{i}{k}t$ for all $i=0, \ldots, k$. Let us set $u^k(s):=u(s_{i+1}^k)$ for $s_i^k< s \le s_{i+1}^k$. By (\[mint\]), comparing $u(s^k_i)$ with $u(s^k_{i+1})-g(s^k_{i+1})+g(s^k_{i})$, it is easy to see that $${{\mathcal E}}(t) \ge {{\mathcal E}}(0)+
2\int_0^t \int_{{\Omega}} \nabla u^k(\tau) \nabla \dot{g}(\tau) \,d\tau \,dx+o_k,$$ where $o_k \to 0$ as $k \to +\infty$. Since $\nabla u$ is continuous with respect to the strong topology of $L^2({\Omega};{{\mathbb R}}^2)$ in $[0,1]$ up to a countable set, passing to the limit for $k \to +\infty$ we deduce .
We are now ready to prove the main result of the paper.
Let $D$ be a countable and dense set in $[0,1]$ such that $0 \in D$, and let $(\delta_n,{\varepsilon}_n,a_n)$ and $\{t \to u(t) \in SBV({\Omega})\,:\, t \in [0,1]\}$ be the sequence and the evolution determined in Lemma \[convD\] and Lemma \[extension\]. Let us set $$u_n:=u^{\delta_n}_{\varepsilon_n, a_n}, \quad \quad
\Gamma_n:=\Gamma^{\delta_n}_{\varepsilon_n, a_n}, \quad \quad
{{\mathcal E}}_n:={{\mathcal E}}^{\delta_n}_{\varepsilon_n, a_n}.$$ Let $\overline{{{\mathcal N}}}$ be the union of the sets of discontinuities of $\eta$ and ${{\mathcal H}}^1(\Gamma(\cdot))$, where $\eta$ is defined in . Let ${{\mathcal B}}:= \bigcap_{k=1}^{+\infty} \bigcup_{h=k}^\infty {{\mathcal B}}_h$, where ${{\mathcal B}}_h$ are as in Lemma \[convD\]; since $|\bigcup_{h=k}^\infty {{\mathcal B}}_h| <2^{-k+1}$, we have $|{{\mathcal B}}|=0$. For all $t \in [0,1] \setminus ({{\mathcal B}}\cup \overline{{{\mathcal N}}})$ we claim that $$\label{convgrad2}
\nabla u_n(t) \to \nabla u(t) \quad \mbox{ strongly in }L^2({\Omega};{{\mathbb R}}^2).$$ In fact, since $t \not \in \bigcup_{h=k}^\infty {{\mathcal B}}_h$ for some $k$, by Lemma \[convD\] we have $$\lim_n \|\nabla u^{\delta_n}_{\varepsilon_n,a_n}(t)-\nabla u_{a_n}(t)\|=0;$$ for $t \not \in \overline{{{\mathcal N}}}$, by Lemma \[extension\] we have that $\nabla u_{a_n}(t) \to \nabla u(t)$ strongly in $L^2({\Omega}; {{\mathbb R}}^2)$ and so holds.
Since $g_{{\varepsilon}_n} \to g$ strongly in $W^{1,1}([0,1];H^1({\Omega}))$, we deduce that for a.e. $\tau \in [0,1]$ $$\nabla \dot{g}_{{\varepsilon}_n}(\tau) \to \nabla \dot{g}(\tau)
\quad \mbox{ strongly in }L^2({\Omega};{{\mathbb R}}^2).$$ Since ${{\mathcal E}}_n(0) \to {{\mathcal E}}(0)$ by and $o^{\delta_n}_{{\varepsilon}_n} \to 0$, by semicontinuity of the energy and by we have that for all $t \in D$ $$\label{energyconv*}
{{\mathcal E}}(t) \le \liminf_n {{\mathcal E}}_n(t) \le \limsup_n {{\mathcal E}}_n(t) \le
{{\mathcal E}}(0)+2 \int_0^t \int_{{\Omega}}
\nabla u(\tau) \nabla \dot{g}(\tau) \,dx\,d\tau.$$ In view of , we conclude that for all $t \in D$ $${{\mathcal E}}(t)=
{{\mathcal E}}(0)+2 \int_0^t \int_{{\Omega}}
\nabla u(\tau) \nabla \dot{g}(\tau) \,dx\,d\tau,$$ and since $\nabla u(\cdot)$ and ${{\mathcal H}}^1(\Gamma(\cdot))$ are left continuous at $t \not \in D$ and so ${{\mathcal E}}(\cdot)$ is, we conclude that the equality holds for all $t \in [0,1]$. As a consequence $\{t \to u(t)\,, t \in [0,1]\}$ is a quasi-static evolution of brittle fractures. Let us prove that is indeed true for all $t \in [0,1]$. In fact, if $t \notin D$, it is sufficient to prove $$\label{truelsc}
\liminf_n {{\mathcal E}}_n(t) \ge {{\mathcal E}}(t).$$ Considering $s \ge t$ with $s \in D$, by we have $${{\mathcal E}}_n(s) \le {{\mathcal E}}_n(t)+
\int_{t^{\delta_n}_{j_n}}^{s^{\delta_n}_{j_n}} \int_{{\Omega}} \nabla u_n(\tau)
\nabla \dot{g}_{{\varepsilon}_n}(\tau)\,dx\,d\tau
+o^{\delta_n}_{{\varepsilon}_n}
\quad
t^{\delta_n}_{j_n} \le t <t^{\delta_n}_{j_n+1},\;
s^{\delta_n}_{j_n} \le s <s^{\delta_n}_{j_n+1},$$ so that $$\liminf_n {{\mathcal E}}_n(t) \ge {{\mathcal E}}(s)-
\int_t^s \int_{{\Omega}} \nabla u(\tau) \nabla \dot{g}(\tau)\,dx\,d\tau.$$ Letting $s \searrow t$, since ${{\mathcal E}}(\cdot)$ is continuous, we have holds. By we deduce that ${{\mathcal E}}_n(t) \to {{\mathcal E}}(t)$ for all $t \in [0,1]$, so that point $(b)$ is proved.
We now come to point $(a)$. Since $\lambda(t) \ge {{\mathcal H}}^1(\Gamma(t))$ for all $t \in [0,1]$, by and point $(b)$, we deduce that $\lambda={{\mathcal H}}^1(\Gamma(\cdot))$ in $[0,1]$ up to a set of measure $0$. Since they are increasing functions, we conclude that $\lambda$ and ${{\mathcal H}}^1(\Gamma(\cdot))$ share the same set of continuity points $[0,1] \setminus {{\mathcal N}}$, and that $\lambda={{\mathcal H}}^1(\Gamma(\cdot))$ on $[0,1] \setminus {{\mathcal N}}$. In view of , point $(a)$ is thus established for all $t$ except $t \in ({{\mathcal B}}\cup \overline{{{\mathcal N}}}) \setminus {{\mathcal N}}$. In order to treat this case, we use the following argument. Considering the measures $\mu_n:={{\mathcal H}}^1 {\mathop{\hbox{\vrule height 7pt width .5pt depth 0pt
\vrule height .5pt width 6pt depth 0pt}}\nolimits}\Gamma_n(t)$, we have that, up to a subsequence, $\mu_n {\stackrel{\ast}{\rightharpoonup}}\mu$ weakly-star in the sense of measures, and as a consequence of Ambrosio’s Theorem we have ${{\mathcal H}}^1 {\mathop{\hbox{\vrule height 7pt width .5pt depth 0pt
\vrule height .5pt width 6pt depth 0pt}}\nolimits}\Gamma(t) \le \mu$ as measures. Since $t \notin {{\mathcal N}}$ we have $\mu_n({{\mathbb R}}^2) \to {{\mathcal H}}^1(\Gamma(t))$, and so we deduce ${{\mathcal H}}^1 {\mathop{\hbox{\vrule height 7pt width .5pt depth 0pt
\vrule height .5pt width 6pt depth 0pt}}\nolimits}\Gamma(t) =\mu$. Let us consider now $u_n(t)$; we have up to a subsequence $u_n(t) \to u$ in $SBV({\Omega})$ for some $u \in SBV({\Omega})$. Setting $u_n(t):=g^{\delta_n}_{{\varepsilon}_n}(t)$ and $u:=g(t)$ on ${\Omega}_D$, we have $u_n(t) \to u$ in $SBV({\Omega}')$, and as a consequence of Ambrosio’s Theorem, we get that ${{\mathcal H}}^1 {\mathop{\hbox{\vrule height 7pt width .5pt depth 0pt
\vrule height .5pt width 6pt depth 0pt}}\nolimits}{S^{g(t)}(u)}
\le \mu={{\mathcal H}}^1 {\mathop{\hbox{\vrule height 7pt width .5pt depth 0pt
\vrule height .5pt width 6pt depth 0pt}}\nolimits}\Gamma(t)$, that is ${S^{g(t)}(u)} \subseteq \Gamma(t)$. By Theorem \[jumptransfer\], we deduce that $u$ is a minimum for $$\min \{ \|\nabla v\|^2\,:\, {S^{g(t)}(v)} \subseteq \Gamma(t)
\mbox{ up to a set of ${{\mathcal H}}^1$-measure $0$ } \},$$ and by uniqueness of the gradient we get that $\nabla u=\nabla u(t)$, so that the proof is concluded.
Piecewise Affine Transfer of Jump and Proof of Proposition \[pminpropD\] {#secmin}
========================================================================
The proof of Proposition \[pminpropD\] is based on the following proposition, which is a variant of Theorem \[jumptransfer\] in the context of piecewise affine approximation.
\[pctransfer\] Given ${\varepsilon}_n \to 0$, let $g^r_n\in H^1({\Omega})$ be such that $g^r_n \in {{{\mathcal A}}{{\mathcal F}}_{\varepsilon_n}({\Omega})}$ and $g^r_n \to g^r$ strongly in $H^1({\Omega})$ for all $r=0,\dots,i$. If ${u^r_n}\in {{{\mathcal A}}_{\varepsilon_n,a}({\Omega})}$ is such that ${u^r_n}\to {u^r}$ in $SBV({\Omega})$ for $r=0,\dots,i$, then for all $v \in SBV({\Omega})$ with ${{\mathcal H}}^1 \left( {S^{g^i}(v)} \right) <+\infty$ and $\nabla v \in
L^2({\Omega};{{\mathbb R}}^2)$, there exists $v_n \in {{{\mathcal A}}_{\varepsilon_n,a}({\Omega})}$ such that $v_n \to v$ strongly in $L^1({\Omega})$, $\nabla v_n \to \nabla v$ strongly in $L^2({\Omega};{{\mathbb R}}^2)$ and $$\label{pctransfjump}
\limsup_n {{\mathcal H}}^1 \left( {S^{g^i_n}(v_n)} \setminus
\bigcup_{r=0}^i {S^{g^r_n}(u^r_n)}\right) \le
\mu(a) {{\mathcal H}}^1 \left( {S^{g(t)}(v)} \setminus
\bigcup_{r=0}^i {S^{g^r}(u^r)} \right),$$ where $\mu\,:\,]0;\frac{1}{2}[ \to {{\mathbb R}}$ with $\lim_{a \to 0^+} \mu(a)=1$.
In view of Proposition \[pctransfer\], we can now prove Proposition \[pminpropD\].
Notice that, in order to prove , it is sufficient to prove the existence of $\mu\,:\,]0;\frac{1}{2}[ \to {{\mathbb R}}$ with $\lim_{a \to 0^+} \mu(a)=1$ such that, given $t \in D$, for every $0=t_0 \le \ldots \le t_r \le \ldots \le t_i=t$, $t_r \in D$, for all $v \in SBV({\Omega})$ we have $$\label{mineqk*}
\|\nabla u_a(t)\|^2 \le
\|\nabla v\|^2 +
\mu(a) {{\mathcal H}}^1 \left( {S^{g(t)}(v)} \setminus
\bigcup_{r=0}^i {S^{g(t_r)}(u_a(t_r))} \right).$$ In fact, taking the sup over all possible $t_0, \ldots, t_i$, we get .
We apply Proposition \[pctransfer\] considering $g^r_n:=g^{\delta_n}_{{\varepsilon}_n}(t_r)$, $g^r:=g(t_r)$, $u^r_n:=u^{\delta_n}_{{\varepsilon}_n,a}(t_r)$, and $u^r:=u_a(t_r)$ for $r=0,\dots,i$. There exists $\mu\,:\,]0;\frac{1}{2}[ \to {{\mathbb R}}$ with $\lim_{a \to 0^+} \mu(a)=1$ such that for $v \in SBV({\Omega})$, there exists $v_n \in {{{\mathcal A}}_{\varepsilon_n,a}({\Omega})}$ with $\nabla v_n \to \nabla v$ strongly in $L^2({\Omega};{{\mathbb R}}^2)$ and $$\begin{gathered}
\limsup_n {{\mathcal H}}^1 \left( {S^{g^{\delta_n}_{{\varepsilon}_n}(t)}(v_n)} \setminus
\bigcup_{r=0}^i {S^{g^{\delta_n}_{{\varepsilon}_n}(t_r)}(u^{\delta_n}_{{\varepsilon}_n,a}(t_r))} \right) \le \\
\le
\mu(a) {{\mathcal H}}^1 \left( {S^{g(t)}(v)} \setminus
\bigcup_{r=0}^i {S^{g(t_r)}(u_a(t_r))} \right),\end{gathered}$$ Comparing $u^{\delta_n}_{{\varepsilon}_n,a}(t)$ and $v_n$ by means of , we obtain $$\begin{gathered}
\label{mminimality}
\|\nabla u^{\delta_n}_{{\varepsilon}_n,a}(t) \|^2 \le \|\nabla v_n \|^2+
{{\mathcal H}}^1 \left( {S^{g^{\delta_n}_{{\varepsilon}_n}(t)}(v_n)} \setminus \Gamma^{\delta_n}_{{\varepsilon}_n,a}(t) \right) \le \\
\le
\|\nabla v_n \|^2+
{{\mathcal H}}^1 \left( {S^{g^{\delta_n}_{{\varepsilon}_n}(t)}(v_n)} \setminus
\bigcup_{r=0}^i {S^{g^{\delta_n}_{{\varepsilon}_n}(t_r)}(u^{\delta_n}_{{\varepsilon}_n,a}(t_r))} \right),\end{gathered}$$ so that, passing to the limit for $n \to +\infty$, we obtain that holds. Moreover, we have that choosing $v=u_a(t)$, and taking the limsup in , we get that $\nabla u^{\delta_n}_{{\varepsilon}_n,a}(t)
\to \nabla u_a(t)$ strongly in $L^2({\Omega};{{\mathbb R}}^2)$.
The rest of the section is devoted to the proof of Proposition \[pctransfer\]. It will be convenient, as in Section \[convres\], to consider ${\Omega}_D$ polygonal open bounded subset of ${{\mathbb R}}^2$ such that ${\Omega}_D \cap {\Omega}=\emptyset$ and $\partial {\Omega}\cap \partial {\Omega}_D= \partial_D {\Omega}$ up to a finite number of vertices; we set ${\Omega}':= {\Omega}\cup {\Omega}_D \cup \partial_D {\Omega}$. We suppose that ${{\bf R}}_{\varepsilon}$ can be extended to a regular triangulation of ${\Omega}'$ which we still indicate by ${{\bf R}}_{\varepsilon}$.
We need several preliminary results. Let us set ${z^r_n}:=u^r_n-g^r_n$, and let us extend ${z^r_n}$ to zero on ${\Omega}_D$. Similarly, we set ${z^r}:=u^r-g^r$, and we extend ${z^r}$ to zero on ${\Omega}_D$.
Let $\sigma>0$, and let $C$ be the set of corners of $\partial_D {\Omega}$. Let us fix $G \subseteq {{\mathbb R}}$ countable and dense: we recall that for all $r=0,\dots,i$ we have up to a set of ${{\mathcal H}}^{1}$-measure zero $$S_{{z^r}}= \bigcup_{c_1,c_2 \in G} \partial^* E_{c_1}(r) \cap \partial^*E_{c_2}(r),$$ where $E_c(r):=\{x \in {\Omega}'\,:\, {z^r}(x) > c\}$ and $\partial^*$ denotes the essential boundary (see [@AFP]). Let us consider $$J_j:=\{x \in \bigcup_{r=0}^i S_{{z^r}} \setminus C\,:\,
{(z^l)^+}(x)-{(z^l)^-}(x)>\frac{1}{j} \mbox{ for some } l=0, \ldots, i\},$$ with $j$ so large that ${{\mathcal H}}^1(\bigcup_{r=0}^i S_{{z^r}} \setminus J_j) \le \sigma$. Let $U$ be a neighborhood of $\bigcup_{r=0}^i S_{{z^r}}$ such that $|U| \le \frac{\sigma}{j^2}$. Following [@FL Theorem 2.1] (see Fig.3), we can find a finite disjoint collection of closed cubes $\{Q_k\}_{k=1, \ldots,K}$ with center $x_k \in J_j$, edge of length $2r_k$ and oriented as the normal $\nu(x_k)$ to $S_{{z^{r(k)}}}$ at $x_k$, such that $\bigcup_{k=1}^K Q_k \subseteq U$ and ${{\mathcal H}}^{1}(J_j \setminus \bigcup_{k=1}^K Q_k) \le \sigma$. Moreover for all $k=1, \ldots,K$ there exists $r(k) \in \{0, \ldots, i\}$ and ${c_1(r(k))},{c_2(r(k))}>0$ such that $${{\mathcal H}}^1 \left( \left[ \bigcup_{r=0}^i S_{{z^r}} \setminus S_{{z^{r(k)}}}
\right] \cap Q_k \right) \le \sigma r_k,$$ and the following hold
- if $x_k \in {\Omega}$ then $Q_k \subseteq {\Omega}$, and if $x_k \in \partial_D {\Omega}$ then $Q_k \cap \partial_D {\Omega}=H_k$, where $H_k$ denotes the intersection of $Q_k$ with the straight line through $x_k$ orthogonal to $\nu(x_k)$;
- - ${{\mathcal H}}^{1} (S_{{z^{r(k)}}} \cap \partial Q_k)=0$;
- - $r_k \le c {{\mathcal H}}^{1} (S_{{z^{r(k)}}} \cap Q_k)$ for some $c>0$;
- - ${(z^{r(k)})^-}(x) < {c_1(r(k))}< {c_2(r(k))}< {(z^{r(k)})^+}(x)$ and ${c_2(r(k))}-{c_1(r(k))}\ge \frac{1}{2j}$;
- - ${{\mathcal H}}^{1} ([S_{{z^{r(k)}}} \setminus
\partial^* E_{{c_s(r(k))}}(r(k))] \cap Q_k)
\le \sigma r_k$ for $s=1,2$;
- - if $s=1,2$, ${{\mathcal H}}^{1} (\{y \in \partial^* E_{{c_s(r(k))}}(r(k)) \cap Q_k\,:\,
{\rm dist}(y,H_k) \ge \frac{\sigma}{2} r_k\}) <\sigma r_k$;
- - if $Q^+_k:= \{x\in Q_k \,|\, x \cdot \nu(x_k) >0\}$ and $s=1,2$ $$\label{density1}
\| 1_{E_{{c_s(r(k))}}(r(k)) \cap Q_k} -1_{Q^+_k}\|_{L^1({\Omega}')}
\le \sigma^2 r_k^2;$$
- - ${{\mathcal H}}^{1} ((S_v \setminus S_{{z^{r(k)}}}) \cap Q_k) < \sigma r_k$ and ${{\mathcal H}}^{1} (S_v \cap \partial Q_k)=0$.
Let us indicate by $R_k$ the intersection of $Q_k$ with the strip centered in $H_k$ with width $2\sigma r_k$, and let us set $V_k^\pm:=\{x_k\pm r_k e(x_k)+s \nu(x_k):\, s \in {{\mathbb R}}\} \cap R_k$, where $e(x_k)$ is such that $\{e(x_k), \nu(x_k)\}$ is an orthonormal base of ${{\mathbb R}}^2$ with the same orientation of the canonical one.
For all $B \subseteq {\Omega}'$, let us set $${{\mathcal R}}_n(B):=\{T \in {{\bf R}}_{{\varepsilon}_n} \,:\, T \cap B \not= \emptyset\}, \quad \quad
{{\mathcal T}}^k_n(B):=\{T \in {{\bf T}}({z^{r(k)}_n})\,:\, T \cap B \not= \emptyset\}.$$ In the following, we will often indicate with the same symbol a family of triangles and their support in ${{\mathbb R}}^2$, being clear from the context in which sense has to be intended. We will consider ${z^{r(k)}_n}$ defined pointwise in ${\Omega}' \setminus \overline{S}_{{z^{r(k)}_n}}$ and so the upper levels of ${z^{r(k)}_n}$ are intended as subsets of ${\Omega}' \setminus \overline{S}_{{z^{r(k)}_n}}$.
\[upperlevels\] For all $k=1, \ldots,K$ there exists $c_n^k \in [{c_1(r(k))},{c_2(r(k))}]$ such that, setting ${E_n^{k}}:=\{x \in {{{\mathcal R}}_n(Q_k)}\,:\,
{z^{r(k)}_n}(x) >c^k_n\}$, we have $$\label{coarea1}
\limsup_n \sum_{k=1}^K
{{\mathcal H}}^1 \left( \big( \partial_{{{{\mathcal R}}_n(Q_k)}} {E_n^{k}}\big) \setminus S_{{z^{r(k)}_n}} \right)=o_\sigma,$$ and $$\label{nearset1}
\limsup_n \|1_{{E_n^{k}}}-1_{Q_k^+}\|_{L^1({\Omega}')} \le
\sigma^2 r_k^2,$$ where $\partial_{{{{\mathcal R}}_n(Q_k)}}$ denotes the boundary operator in ${{{\mathcal R}}_n(Q_k)}$, and $o_\sigma \to 0$ as $\sigma \to 0$.
Note that for $n$ large we have $\bigcup_{k=1}^K {{{\mathcal R}}_n(Q_k)}\subseteq U$, so that $|\bigcup_{k=1}^K {{{\mathcal R}}_n(Q_k)}| \le \frac{\sigma}{j^2}$. By Hölder inequality and since $\|\nabla {z^r_n}\| \le C'$ for all $r=0,\dots,i$, it follows that $$\sum_{r=0}^i \int_{\{\cup_k {{{\mathcal R}}_n(Q_k)}: r(k)=r\}} |\nabla {z^r_n}| \,dx \le
\sum_{r=0}^i \|\nabla {z^r_n}\| \frac{\sqrt{\sigma}}{j} \le (i+1)C'\frac{\sqrt{\sigma}}{j}.$$ Following [@FL Theorem 2.1], we can apply coarea-formula for BV-functions (see [@AFP]) taking into account that ${z^{r(k)}_n}$ belongs to $SBV({\Omega}')$ so that the singular part of the derivative is carried only by $S_{{z^{r(k)}_n}}$: since for $n$ large the ${{{\mathcal R}}_n(Q_k)}$’s are disjoint, we obtain $$\label{coarea}
\sum_{k=1}^K \int_{{\mathbb R}}{{\mathcal H}}^1 \Big( \big( \partial E_{c,n}(r(k))
\cap {{{\mathcal R}}_n(Q_k)}\big) \setminus S_{{z^{r(k)}_n}} \Big)\, dc \le (i+1)C' \frac{\sqrt{\sigma}}{j},$$ where $E_{c,n}(r(k)):=\{x \in {\Omega}' \setminus \overline{S}_{{z^{r(k)}_n}}\,:\,
{z^{r(k)}_n}(x) >c\}$, and so $$\sum_{k=1}^K \int_{{c_1(r(k))}}^{{c_2(r(k))}}
{{\mathcal H}}^1 \Big( \big( \partial E_{c,n}(r(k))
\cap {{{\mathcal R}}_n(Q_k)}\big) \setminus S_{{z^{r(k)}_n}} \Big) \, dc
\le (i+1)C' \frac{\sqrt{\sigma}}{j}.$$ Notice that we can use the topological boundary instead of the reduced boundary of $E_{c,n}(r(k))$ in since ${z^{r(k)}_n}$ is piecewise affine, and so $\partial E_{c,n}(r(k)) \setminus
\partial^* E_{c,n}(r(k)) \not= \emptyset$ just for a finite number of $c$’s. By the Mean Value Theorem we have that there exist $c_n^k \in [{c_1(r(k))},{c_2(r(k))}]$ such that $$\sum_{k=1}^K
{{\mathcal H}}^1 \Big( \big( \partial E_{c_n^k,n}(r(k))
\cap {{{\mathcal R}}_n(Q_k)}\big) \setminus S_{{z^{r(k)}_n}} \Big) \, \le 2iC' \sqrt{\sigma},$$ and taking the limsup for $n \to +\infty$, we get . Let us come to . Since $$E_{{c_2(r(k))},n}(r(k)) \subseteq E_{c_n^k,n}(r(k)) \subseteq E_{{c_1(r(k))},n}(r(k)),$$ by we have that for $n$ large $$\|1_{E_{c_n^k,n}(r(k)) \cap Q_k} -1_{Q_k^+}\|_{L^1({\Omega}')} \le \sigma^2 r_k^2,$$ and so, since $|{{{\mathcal R}}_n(Q_k)}\setminus Q_k| \to 0$, we conclude that holds.
Fix $k \in \{1, \ldots, K\}$, and let us consider the family ${{\mathcal T}}^k_n({E_n^{k}})$. Let us modify this family in the following way. Let $T \in {{\mathcal T}}^k_n({E_n^{k}})$; we keep it if $|T \cap {E_n^{k}}|> \frac{1}{2}|T|$, and we erase it otherwise. Let ${E^{k,+}_n}$ be this new family of triangles, and let ${E^{k,-}_n}$ be its complement in ${{\mathcal T}}^k_n({{{\mathcal R}}_n(Q_k)})$.
\[en+\] For all $k=1, \ldots,K$ we have $$\label{estper}
\limsup_n \sum_{k=1}^K
{{\mathcal H}}^1 \left( \partial_{{{{\mathcal R}}_n(Q_k)}} {E^{k,+}_n}\setminus S_{{z^{r(k)}_n}} \right)
=o_\sigma,$$ and $$\label{estarea}
\limsup_n \|1_{{E^{k,+}_n}}-1_{Q_k^+}\|_1 \le 4\sigma^2 r_k^2,$$ where $o_\sigma \to 0$ as $\sigma \to 0$.
Let $T \in {{\mathcal T}}^k_n({E_n^{k}})$. Since ${z^{r(k)}_n}$ is affine on $T$, it follows that $T \cap {E_n^{k}}$ is either a triangle with at least two edges contained in the edges of $T$ or a trapezoid with three edges contained in the edges of $T$. Let $l(T)$ be the edge inside $T$ where ${z^{r(k)}_n}=c_n^k$, where $c^k_n$ is the value determining ${E_n^{k}}$ (we consider $l(T)=\emptyset$ if ${\rm int}(T) \subseteq {E_n^{k}}$). In the case $T \in {E^{k,+}_n}$ as in the case $T \in {E^{k,-}_n}$, since the angles of the triangles of ${{\bf T}}({z^{r(k)}_n})$ are uniformly bounded away from $0$ and from $\pi$, arguing as in Lemma \[simplecurve\], we deduce that keeping or erasing $T$, we increase $\partial_{{{{\mathcal R}}_n(Q_k)}} {E_n^{k}}$ of a quantity which is less than $c{{\mathcal H}}^1(l(T))$ with $c$ independent of $\varepsilon_n$. Then we have $$\sum_{k=1}^K {{\mathcal H}}^1(\partial_{{{{\mathcal R}}_n(Q_k)}} {E^{k,+}_n}\setminus \partial_{{{{\mathcal R}}_n(Q_k)}} {E_n^{k}})
\le \sum_{k=1}^K
\sum_{T \in {{\mathcal T}}^k_n({E_n^{k}})} c{{\mathcal H}}^1(l(T))
\le c \sum_{k=1}^K {{\mathcal H}}^1( \partial_{{{{\mathcal R}}_n(Q_k)}} {E_n^{k}}\setminus S_{{z^{r(k)}_n}}),$$ so that taking the limsup for $n \to +\infty$ and in view of we deduce that holds.
Let us come to . Note that $|{{\mathcal T}}^k_n(\partial Q_k^+)| \to 0$ as $n \to +\infty$. Then if $A^{k,+}_n:=\{T \in {{\bf T}}({z^{r(k)}_n})\,:\, T \subseteq {{\rm int}(Q_k^+)}\}$, for $n$ large we have $$|Q_k^+ \setminus {E^{k,+}_n}| \le |A^{k,+}_n \setminus {E^{k,+}_n}|+
|{{\mathcal T}}^k_n(\partial Q_k^+)| \le 2|Q_k^+ \setminus {E_n^{k}}|+
|{{\mathcal T}}^k_n(\partial Q_k^+)|,$$ where the last inequality follows by construction of ${E^{k,+}_n}$. Taking the limsup for $n \to +\infty$, in view of we get $$\limsup_n |Q_k^+ \setminus {E^{k,+}_n}| \le 2\sigma^2 r_k^2.$$ The inequality $\limsup_n |{E^{k,+}_n}\setminus Q_k^+| \le 2\sigma^2 r_k^2$ follows analogously.
For all $k=1, \ldots, K$ and $s \in {{\mathbb R}}$, let us set $${H_{k}}(s) := \{x + s \nu(x_k),\, x\in H_k\}.$$
\[fubini\] There exist $s^+_n \in ]\frac{\sigma}{4}r_k, \frac{\sigma}{2} r_k[$ and $s^-_n \in ]-\frac{\sigma}{2} r_k, -\frac{\sigma}{4}r_k[$ such that, setting ${H^{k,+}_n}:={H_{k}}(s^+_n)$ and ${H^{k,-}_n}:={H_{k}}(s^-_n)$ we have for $n$ large enough $${{\mathcal H}}^1({H^{k,+}_n}\setminus {E^{k,+}_n}) \le 20\sigma r_k,
\quad
{{\mathcal H}}^1({H^{k,-}_n}\cap {E^{k,+}_n}) \le 20\sigma r_k.$$
By (\[estarea\]) we can write for $n$ large $$\int_{\frac{\sigma}{4}r_k}^{\frac{\sigma}{2} r_k}
{{\mathcal H}}^1(H_k(s) \setminus {E^{k,+}_n}) \,ds \le
5 \sigma^2 r_k^2,$$ so that we get $s^+_n \in ]\frac{\sigma}{4}r_k, \frac{\sigma}{2} r_k[$ with $${{\mathcal H}}^1(H_k(s^+_n) \setminus {E^{k,+}_n}) \le 20\sigma r_k.$$ Similarly we can reason for $s^-_n$.
Let $S^{k,+}_n$ be the straight line containing ${H^{k,+}_n}$: up to replacing ${H^{k,+}_n}$ by the connected component of $S^{k,+}_n \cap {{{\mathcal R}}_n(Q_k)}$ to which it belongs, we may suppose that ${H^{k,+}_n}\setminus {E^{k,+}_n}$ is a finite union of segments $l_j^+$ with extremes $A_j$ and $B_j$ belonging to the edges of the triangles of ${{\mathcal T}}_n^k({{{\mathcal R}}_n(Q_k)})$ such that for $n$ large $${{\mathcal H}}^1({H^{k,+}_n}\setminus {E^{k,+}_n})={{\mathcal H}}^1 \left( \bigcup_{j=1}^{m} l_j^+ \right) \le 20\sigma r_k.$$ By Lemma \[simplecurve\], for all $j$ there exists a curve $L_j^+$ inside the edges of the triangles of ${{\mathcal T}}_n^k({{{\mathcal R}}_n(Q_k)})$ joining $A_j$ and $B_j$ and such that $$\label{estsimplecurve}
{{\mathcal H}}^1(L_j^+) \le c{{\mathcal H}}^1(l_j^+),$$ with $c$ independent of $\varepsilon_n$. Let us set $$\gamma^{k,+}_n:= L_1^+ \cup B_1 A_2 \cup L_2^+ \cup \cdots \cup B_{m-1}A_m \cup
L_{m}^+.$$ Similarly, let us construct $\gamma^{k,-}_n$ relative to ${H^{k,-}_n}\cap E_n^+$. Note that for $n$ large enough $\gamma^{k,+}_n \cap H_k(\sigma)=\emptyset$, $\gamma^{k,-}_n \cap H_k(-\sigma)=\emptyset$, and $\gamma^{k,+}_n \cap \gamma^{k,-}_n=\emptyset$. Let us consider the connected component ${{\mathcal C}}_k^+$ of ${{{\mathcal R}}_n(Q_k)}\setminus \gamma^{k,+}_n$ containing $H_k(\sigma)$. Similarly, let us consider the connected component ${{\mathcal C}}_k^-$ of ${{{\mathcal R}}_n(Q_k)}\setminus \gamma^{k,-}_n$ containing $H_k(-\sigma)$. For $n$ large enough, by $$\label{estsimplecurve2}
{{\mathcal H}}^1 \left( \partial_{{{{\mathcal R}}_n(Q_k)}} {{\mathcal C}}_k^+
\setminus
\bigcup_{i=1}^{m-1} B_i A_{i+1}\right)
\le c \sum_{j=1}^{m} {{\mathcal H}}^1(l_j^+) \le 20c\sigma r_k.$$ A similar estimate holds for $\partial_{{{{\mathcal R}}_n(Q_k)}} {{\mathcal C}}_k^-$.
Let ${\tilde{E}^{k,+}_n}$ be the family of triangles obtained adding to ${E^{k,+}_n}$ those $T \in {E^{k,-}_n}$ such that $T \subseteq {{\mathcal C}}_k^+$, and subtracting those $T \in {E^{k,+}_n}$ such that $T \subseteq {{\mathcal C}}_k^-$. Let ${\tilde{E}^{k,-}_n}$ be the complement of ${\tilde{E}^{k,+}_n}$ in ${{\mathcal T}}_n^k({{{\mathcal R}}_n(Q_k)})$.
We claim that there exists $C>0$ independent of $n$ such that for all $k=1, \ldots, K$ and for $n$ large $$\label{estper3}
{{\mathcal H}}^1 \left( \partial_{{{{\mathcal R}}_n(Q_k)}} {\tilde{E}^{k,+}_n}\setminus \partial_{{{{\mathcal R}}_n(Q_k)}} {E^{k,+}_n}\right) \le
C\sigma r_k.$$ In fact, let $\zeta$ be an edge of $\partial_{{{{\mathcal R}}_n(Q_k)}} {\tilde{E}^{k,+}_n}\setminus
\partial_{{{{\mathcal R}}_n(Q_k)}} {E^{k,+}_n}$, that is $\zeta$ belongs to a triangle $T$ that has been changed in the operation above described. Let us assume for instance that $T \in {E^{k,-}_n}$ and $T \subseteq {{\mathcal C}}_k^+$. If $T'$ is such that $T \cap T'=\zeta$, then $T' \in {E^{k,-}_n}$: in fact if by contradiction $T' \in {E^{k,+}_n}$, then $T' \in {\tilde{E}^{k,+}_n}$ and so we would have $\zeta \not\in \partial_{{{{\mathcal R}}_n(Q_k)}} {\tilde{E}^{k,+}_n}$ which is absurd. Similarly we get $T' \not \subseteq {{\mathcal C}}_k^+$. This means that $\zeta \subseteq \partial_{{{{\mathcal R}}_n(Q_k)}} {{\mathcal C}}_k^+$, and since the horizontal edges of $\gamma^{k,+}_n$ intersect by construction only elements of ${E^{k,+}_n}$, we deduce that $\zeta \subseteq \partial_{{{{\mathcal R}}_n(Q_k)}} {{\mathcal C}}_k^+ \setminus
\left( \cup_{i=1}^{m} A_iB_i \right)$, and by we conclude that holds.
We can summarize the previous results as follows.
\[families\] For all $k=1, \ldots, K$ there exist two families ${\tilde{E}^{k,+}_n}$ and ${\tilde{E}^{k,-}_n}$ of triangles with ${{\mathcal T}}_n^k({{{\mathcal R}}_n(Q_k)})={\tilde{E}^{k,+}_n}\cup {\tilde{E}^{k,-}_n}$, $Q_k^+ \setminus R_k \subseteq {\tilde{E}^{k,+}_n}$ and $Q_k^- \setminus R_k \subseteq {\tilde{E}^{k,-}_n}$, and such that $$\label{estper2}
\limsup_n \sum_{k=1}^K {{\mathcal H}}^1 \left( \partial_{{{{\mathcal R}}_n(Q_k)}} {\tilde{E}^{k,+}_n}\setminus S_{{z^{r(k)}_n}} \right)=
o_\sigma,$$ where $o_\sigma \to 0$ as $\sigma \to 0$. Moreover, in the case $x_k \in \partial_D {\Omega}$, we can modify ${\tilde{E}^{k,+}_n}$ or ${\tilde{E}^{k,-}_n}$ in such a way that ${\tilde{E}^{k,+}_n}\subseteq {\Omega}$ or ${\tilde{E}^{k,-}_n}\subseteq {\Omega}$.
We have that follows from and , and the fact that $\sum_{k=1}^K r_k \le c$, with $c$ independent of $\sigma$. Let us consider the case $x_k \in \partial_D {\Omega}$ with $Q_k^+ \setminus R_k \subseteq {\Omega}$ (the other case being similar). From we have that for $n$ large $\sum_{k=1}^K {{\mathcal H}}^1 \left( \partial_{{{{\mathcal R}}_n(Q_k)}} {\tilde{E}^{k,+}_n}\cap Q_k^- \right) \le o_\sigma$ because ${z^{r(k)}_n}=R_{{\varepsilon}_n} g_{h_n}(r(k))$ on $Q_k^-$ and so there are no jumps in $Q_k^-$. We can thus redefine ${\tilde{E}^{k,+}_n}$ subtracting those triangles that are in $Q_k^-$ obtaining again .
We are now in position to prove Proposition \[pctransfer\].
We work in the context of ${\Omega}'$. For all $v \in SBV({\Omega}')$ with $v=g^i$ on ${\Omega}_D$, ${{\mathcal H}}^1(S_v) <+\infty$ and $\nabla v \in L^2({\Omega}';{{\mathbb R}}^2)$, we have to construct $v_n \in SBV({\Omega}')$ such that $v_n=g^i_n$ on ${\Omega}_D$, $(v_n)_{|{\Omega}} \in {{{\mathcal A}}_{\varepsilon_n,a}({\Omega})}$, $v_n \to v$ strongly in $L^1({\Omega}')$, $\nabla v_n \to \nabla v$ strongly in $L^2({\Omega}';{{\mathbb R}}^2)$ and $$\label{mineqk}
\limsup_n {{\mathcal H}}^1 \left( S_{v_n} \setminus
\bigcup_{r=0}^i S_{u^r_n}\right) \le
\mu(a) {{\mathcal H}}^1 \left( S_{v} \setminus
\bigcup_{r=0}^i S_{u^r} \right),$$ where we suppose that $u^r_n$ and $u^r$ are extended to ${\Omega}'$ setting $u^r_n:=g^r_n$, and $u^r:=g^r$ on ${\Omega}_D$ respectively.
We set $v=g^i+w$, where $w \in SBV({\Omega}')$ with $w=0$ on ${\Omega}_D$. By density, it is sufficient to consider the case $w \in L^\infty({\Omega}')$. Up to reducing $U$, we may assume that $\|\nabla g^i\|_{L^2(U;{{\mathbb R}}^2)} < \sigma$ and $\|\nabla w\|_{L^2(U;{{\mathbb R}}^2)} <\sigma$. Let $R'_k$ be a rectangle centered in $x_k$, oriented as $R_k$, and such that $\overline{R'_k} \subset {\rm int}R_k$ and ${{\mathcal H}}^1(S_w \cap (R_k \setminus R'_k)) < \sigma r_k$. We claim that there exists ${w_\sigma}\in SBV({\Omega}')$ with ${w_\sigma}=w$ on $\bigcup_{k=1}^K R'_k$ and ${w_\sigma}=0$ in ${\Omega}_D$ such that
- - $\|w -{w_\sigma}\|+\|\nabla w-\nabla {w_\sigma}\| \le \sigma$;
- - ${{\mathcal H}}^1(S_{{w_\sigma}} \cap (Q_k \setminus R'_k)) \le o_\sigma r_k$, with $o_\sigma \to 0$ as $\sigma \to 0$;
- - ${{\mathcal H}}^1(S_{{w_\sigma}} \setminus \bigcup_{k=1}^K R_k) \le {{\mathcal H}}^1(S_w \setminus \bigcup_{k=1}^K R_k)+\sigma$;
- - $S_{{w_\sigma}} \setminus \bigcup_{k=1}^K R_k$ is union of disjoint segments with closure contained in ${\Omega}\setminus \bigcup_{k=1}^K R_k$;
- - ${w_\sigma}$ is of class $W^{2,\infty}$ on ${\Omega}\setminus \left( \bigcup_{k=1}^K R_k \cup
\overline{S_{{w_\sigma}}} \right)$.
In fact, by Proposition \[regularization\], there exists $w_m \in SBV({\Omega}')$ with $w_m=0$ in ${\Omega}' \setminus {\overline{\Omega}}$ such that $w_m \to w$ strongly in $L^2({\Omega}')$, $\nabla w_m \to \nabla w$ strongly in $L^2({\Omega}';{{\mathbb R}}^2)$, $S_{w_m}$ is polyhedral with $\overline{S_{w_m}} \subseteq {\Omega}$, $w_m$ is of class $W^{2,\infty}$ on ${\Omega}\setminus \left( \bigcup_{k=1}^K R_k \cup
\overline{S_{w_m}} \right)$, and $\lim_m {{\mathcal H}}^{1}(A \cap S_{w_m})={{\mathcal H}}^{1}(A \cap S_w)$ for all $A$ open subset of ${\Omega}'$ with ${{\mathcal H}}^{1}(\partial A \cap S_{w})=0$. It is not restrictive to assume that ${{\mathcal H}}^1(S_{w} \cap \partial R_k)=0$ and ${{\mathcal H}}^1(S_{w_m} \cap \partial R_k)=0$ for all $m$. Let $\psi_k$ be a smooth function such that $0 \le \psi_k \le 1$, $\psi_k=1$ on $R'_k$ and $\psi_k=0$ outside $R_k$. Setting $\psi:=\sum_{k=1}^K \psi_k$, let us consider $\tilde{w}_m:= \psi w+ (1-\psi) w_m$. Note that $\tilde{w}_m \to w$ strongly in $L^2({\Omega}')$, $\nabla \tilde{w}_m \to \nabla w$ strongly in $L^2({\Omega}';{{\mathbb R}}^2)$, $\tilde{w}_m=0$ in ${\Omega}_D$. Moreover, by capacity arguments, we may assume that $S_{\tilde{w}_m} \setminus \bigcup_{k=1}^K R_k$ is a finite union of disjoint segments with closure contained in ${\Omega}\setminus \bigcup_{k=1}^K R_k$. Finally, for $m \to +\infty$, we have $${{\mathcal H}}^1(S_{\tilde{w}_m} \setminus \bigcup_{k=1}^K R_k) \to
{{\mathcal H}}^1(S_{w} \setminus \bigcup_{k=1}^K R_k)),$$ $${{\mathcal H}}^1(S_{\tilde{w}_m} \cap \bigcup_{k=1}^K (Q_k \setminus R_k)) \to
{{\mathcal H}}^1(S_{w} \cap \bigcup_{k=1}^K (Q_k \setminus R_k))$$ and $
\limsup_m {{\mathcal H}}^1(S_{\tilde{w}_m} \cap (R_k \setminus R_k')) \le
2{{\mathcal H}}^1(S_w \cap (R_k \setminus R_k')) \le 2\sigma r_k.
$ Then we can take ${w_\sigma}:=\tilde{w}_m$ for $m$ large enough.
Let $S_{{w_\sigma}} \setminus \bigcup_{k=1}^K Q_k:= \bigcup_{j=1}^m l_j$, where, by capacity arguments, we can always assume that $l_j$ are disjoint segments with closure contained in ${\Omega}\setminus \bigcup_{k=1}^K Q_k$. We define a triangulation ${{\bf T}}_n \in {{\mathcal T}}_{\varepsilon_n,a}({\Omega}')$ specifying its adaptive vertices as follows. Let us consider the families ${{{\mathcal R}}_n(Q_k)}$ and ${{\mathcal R}}_n(l_j)$ for $k=1, \ldots, K$ and $j=1, \ldots, m$. Note that for $n$ large enough, ${{\mathcal R}}_n(Q_{k_1}) \cap {{\mathcal R}}_n(Q_{k_2})= \emptyset$ for $k_1 \not= k_2$, ${{\mathcal R}}_n(l_{j_1}) \cap {{\mathcal R}}_n(l_{j_2})= \emptyset$ for $j_1 \not= j_2$, and ${{\mathcal R}}_n(Q_k) \cap {{\mathcal R}}_n(l_j)= \emptyset$ for every $k,j$. We consider inside the triangles of ${{\mathcal R}}_n(Q_k)$ the adaptive vertices of ${{\bf T}}({z^{r(k)}_n})$. Passing to ${{\mathcal R}}_n(l_j)$, by density arguments it is not restrictive to assume that $l_j$ does not pass through the vertices of ${{\bf R}}_{{\varepsilon}_n}$ and that its extremes belong to the edges of ${{\bf R}}_{{\varepsilon}_n}$. Let $\zeta:=[x,y]$ be an edge of ${{\mathcal R}}_n(l_j)$ such that $l_j \cap \zeta =\{P\}$. Proceeding as in [@N], we take as adaptive vertex of $\zeta$ the projection of $P$ on $\{tx+(1-t)y:\, t \in [a,(1-a)]\}$. Connecting these adaptive vertices, we obtain an [*interpolating*]{} polyhedral curve $\tilde{l}_j$ with $$\label{estinterp}
{{\mathcal H}}^1(\tilde{l}_j) \le \mu(a) {{\mathcal H}}^1(l_j),$$ where $\mu$ is an increasing function such that $\lim_{a \to 0} \mu(a)=1$. Finally, in the remaining edges, we can consider any admissible adaptive vertex, for example the middle point.
Let us define $w_n \in SBV({\Omega}')$ in the following way. For all $Q_k$, let $w_n$ be equal to ${w_\sigma}$ on ${{\mathcal R}}_n(Q_k) \setminus R_k$, equal to the reflection of ${{w_\sigma}}_{|Q_k^+ \setminus R_k}$ with respect to $H_k(\sigma)$ on ${\tilde{E}^{k,+}_n}\cap R_k$ and equal to the reflection of ${{w_\sigma}}_{|Q_k^- \setminus R_k}$ with respect to $H_k(-\sigma)$ on ${\tilde{E}^{k,-}_n}\cap R_k$, where $\tilde{E}^{k,\pm}_n$ are defined as in Lemma \[families\]. On the other elements of ${{\bf T}}_n$, let us set $w_n={w_\sigma}$. Notice that $w_n=0$ on ${\Omega}_D$ and that inside each ${{{\mathcal R}}_n(Q_k)}$, all the discontinuities of $w_n$ are contained in $\partial_{{{{\mathcal R}}_n(Q_k)}} {\tilde{E}^{k,+}_n}\cup V_k \cup P_{{w_\sigma}}^k$, where $P_{{w_\sigma}}^k$ is the union of the polyhedral jumps of ${w_\sigma}$ in ${{\mathcal R}}_n(Q_k)$ and of their reflected version with respect to $H_k(\pm \sigma)$. By Lemma \[families\] and since $\sum_{k=1}^K {{\mathcal H}}^1(V_k \cup P_{{w_\sigma}}^k) \le o_\sigma$ with $o_\sigma \to 0$ as $\sigma \to 0$, and ${{\mathcal H}}^1 \left( \bigcup_{r=0}^i S_{{z^r}} \setminus \bigcup_{k=1}^K Q_k \right)
\le 2\sigma$, we have that $$\begin{gathered}
\limsup_n {{\mathcal H}}^1 \left( S_{w_n} \setminus \bigcup_{r=0}^i S_{{z^r_n}} \right) \le \\
\le {{\mathcal H}}^1 \left( S_{{w_\sigma}} \setminus \bigcup_{k=1}^K Q_k \right) +
\limsup_n {{\mathcal H}}^1 \left( (S_{w_n} \setminus \bigcup_{r=0}^i S_{{z^r_n}}) \cap {{{\mathcal R}}_n(Q_k)}\right) \le \\
\le {{\mathcal H}}^1 \left( S_{{w_\sigma}} \setminus \bigcup_{r=0}^i S_{{z^r}} \right)+
{{\mathcal H}}^1 \left( \bigcup_{r=0}^i S_{{z^r}} \setminus \bigcup_{k=1}^K Q_k \right)+
o_\sigma \le \\
\le {{\mathcal H}}^1 \left( S_{{w_\sigma}} \setminus \bigcup_{r=0}^i S_{{z^r}} \right)+o_\sigma,\end{gathered}$$ and since $\|\nabla {w_\sigma}\|_{L^2(U;{{\mathbb R}}^2)} \le o_\sigma$ we get for $n$ large $$\label{estimategrad}
\|\nabla w_n\|^2_{L^2(\bigcup_{k=1}^K {{{\mathcal R}}_n(Q_k)})} \le o_\sigma.$$ We now want to define an interpolation ${\tilde{w}_n}$ of $w_n$ on ${{\bf T}}_n$. Firstly, we set ${\tilde{w}_n}=0$ on all regular triangles of ${\Omega}_D$. Passing to the triangles in ${{{\mathcal R}}_n(Q_k)}$ (see fig.4), by Lemma \[guscio\], we know that for $n$ large enough, we have $${{\mathcal H}}^1(\partial {{\mathcal R}}_n(V_k)) \le c {{\mathcal H}}^1(V_k), \quad
{{\mathcal H}}^1(\partial {{\mathcal R}}_n(P_{{w_\sigma}}^k)) \le c {{\mathcal H}}^1(P_{{w_\sigma}}^k),$$ with $c$ independent of $n$. If $T \in {{\mathcal R}}_n(V_k) \cup {{\mathcal R}}_n(P_{{w_\sigma}}^k)$, we set ${\tilde{w}_n}=0$ on $T$; otherwise, we define ${\tilde{w}_n}$ on $T$ as the affine interpolation of $w_n$.
Since $\nabla {\tilde{w}_n}$ is uniformly bounded on ${{\mathcal R}}_n(H_k(\pm \sigma))$, $|{{\mathcal R}}_n(H_k(\pm \sigma))| \to 0$ and since $w_n$ is uniformly bounded in $W^{2,\infty}$ on the triangles contained in ${{\mathcal R}}_n(Q_k) \setminus {{\mathcal R}}_n(V_k \cup P_{{w_\sigma}}^k \cup H_k(\pm \sigma))$ we have by the interpolation estimate and by $$\label{estgradqint}
\limsup_n \|\nabla {\tilde{w}_n}\|^2_{L^2(\bigcup_{k=1}^K {{{\mathcal R}}_n(Q_k)})} \le o_\sigma.$$ Moreover we have $$\label{estjumpqint}
\limsup_n \sum_{k=1}^K
{{\mathcal H}}^1 \left( \Big( S_{{\tilde{w}_n}} \setminus \bigcup_{r=0}^k S_{{z^r_n}} \Big) \cap {{{\mathcal R}}_n(Q_k)}\right) \le o_\sigma.$$ Let us come to the triangles not belonging to ${{{\mathcal R}}_n(Q_k)}$ for $k=1, \ldots, K$. For all $j=1, \ldots, m$, we denote by $\hat{{{\mathcal R}}}_n(l_j)$ the family of regular triangles that have edges in common with triangles of ${{\mathcal R}}_n(l_j)$. For $n$ large we have that $\hat{{{\mathcal R}}}_n(l_{j_1}) \cap \hat{{{\mathcal R}}}_n(l_{j_2})= \emptyset$ for $j_1 \not= j_2$. On every regular triangle $T \not \in \bigcup_{k=1}^K {{{\mathcal R}}_n(Q_k)}\cup
\bigcup_{j=1}^m \hat{{{\mathcal R}}}_n(l_j)$, we define ${\tilde{w}_n}$ as the affine interpolation of ${w_\sigma}$. Since ${w_\sigma}$ is of class $W^{2,\infty}$ on $T$ and $T$ is regular, we obtain by the interpolation estimate $$\label{estgradinside}
\|{\tilde{w}_n}- {w_\sigma}\|^2_{W^{1,2}(T)} \le {K}\varepsilon_n \|{w_\sigma}\|_{W^{2,\infty}}.$$ Let us consider now those triangles that are contained in the elements of $\bigcup_{j=1}^m \hat{{{\mathcal R}}}_n(l_j)$. Following [@N], we can define ${\tilde{w}_n}$ on every $T$ in such a way that ${\tilde{w}_n}$ admits discontinuities only on $\tilde{l}_j$, and $\|\nabla {\tilde{w}_n}\|_{L^\infty(T)} \le \|\nabla {w_\sigma}\|_\infty$. Since $|\hat{{{\mathcal R}}}_n(l_j)| \to 0$ as $n \to \infty$, we deduce that $$\label{estgradjumps}
\lim_n \|\nabla {\tilde{w}_n}\|^2_{L^2(\hat{{{\mathcal R}}}_n(l_j))}=0.$$ Moreover by and since ${{\mathcal H}}^1 \left( \bigcup_{r=0}^i S_{{z^r}} \setminus \bigcup_{k=1}^K Q_k \right) \le 2\sigma$, we have $$\label{estjumpinside}
{{\mathcal H}}^1 \left( S_{{\tilde{w}_n}} \cap \bigcup_{j=1}^m \hat{{{\mathcal R}}}_n(l_j) \right) \le \mu(a)
{{\mathcal H}}^1 \left( S_{{w_\sigma}} \setminus \bigcup_{i=1}^k S_{{z^r}} \right) +o_\sigma,$$ where $o_\sigma \to 0$ as $\sigma \to 0$.
We are now ready to conclude. Let us consider ${\hat{w}_n}\in {{{\mathcal A}}_{\varepsilon_n,a}({\Omega})}$ defined as ${\hat{w}_n}:=g^i_n+{\tilde{w}_n}$. We have ${\hat{w}_n}\to g^i+{w_\sigma}$ strongly in $L^2({\Omega}')$. By , , we get $$\limsup_n \|\nabla {\hat{w}_n}\|^2 \le \|\nabla g^i+\nabla {w_\sigma}\|^2 +o_\sigma,$$ while by and we have $$\limsup_n {{\mathcal H}}^1 \left( S_{{\hat{w}_n}} \setminus \bigcup_{r=0}^i S_{{z^r_n}} \right) \le \mu(a)
{{\mathcal H}}^1 \left( S_{{w_\sigma}} \setminus \bigcup_{r=0}^i S_{{z^r}} \right)+o_\sigma.$$ Letting now $\sigma \to 0$, using a diagonal argument, we conclude that Proposition \[pctransfer\] holds.
Revisiting the approximation by Francfort and Larsen {#remark}
====================================================
In this section we show how the arguments of Section \[convres\] may be used to deal with the discrete in time approximation of quasi-static growth of brittle fractures proposed by Francfort and Larsen in [@FL]. More precisely, we prove that there is strong convergence of the gradient of the displacement (in particular convergence of the bulk energy) and convergence of the surface energy at all times of continuity of the length of the crack; moreover there is convergence of the total energy at any time.
We briefly recall the notation employed in [@FL]. Let $I_\infty$ be countable and dense in $[0,1]$, and let $I_n:=\{0=t^n_0 \le \ldots \le t^n_n=1\}$ such that $(I_n)$ is an increasing sequence of sets whose union is $I_\infty$. Let ${\Omega}\subseteq
{{\mathbb R}}^N$ be a Lipschitz bounded domain, and let $\partial {\Omega}=\partial {\Omega}^c_f \cup \partial {\Omega}_f$, where $\partial {\Omega}^c_f$ is open in the relative topology. Let ${\Omega}' \subseteq {{\mathbb R}}^N$ be open and such that $\overline{{\Omega}} \subseteq {\Omega}'$, and let $g \in W^{1,1}([0,1];H^1({\Omega}'))$. At any time $t^n_k$, Francfort and Larsen consider $u^n_k$ minimizer of $$\int_{{\Omega}} |\nabla v|^2 \,dx+ {{\mathcal H}}^{N-1} \left( S_v \setminus \left[
\bigcup_{0 \le j \le k-1} S_{u^n_j} \cup \partial {\Omega}_f \right] \right)$$ in $\{v \in SBV({\Omega}'):v=g(t^n_k)\mbox{ in }{\Omega}'\setminus \overline{{\Omega}}\}$. Setting $u^n(t):=u^n_k$ for $t \in [t^n_k,t^n_{k+1}[$, and $\Gamma^n(t):=
\bigcup_{s \le t, s \in I_n} S_{u^n(s)} \cup \partial {\Omega}_f$, they prove that $$\label{estabovec}
{{\mathcal E}}^n(t) \le {{\mathcal E}}^n(0)+2\int_0^{t^n_k} \int_{{\Omega}} \nabla u^n(\tau) \nabla \dot{g}(\tau) \,dx\,d\tau
+o_n, \quad\quad t\in [t^n_k,t^n_{k+1}[,$$ where ${{\mathcal E}}^n(t):=\int_{{\Omega}} |\nabla u^n(t)|^2 \,dx+ {{\mathcal H}}^{N-1}\left( \Gamma^n(t) \right)$ and $o_n \to 0$ as $n \to +\infty$. Using Theorem \[jumptransfer\], they obtain a subsequence of $(u^n(\cdot))$, still denoted by the same symbol, such that $u^n(t) \to u(t)$ in $SBV({\Omega}')$ and $\nabla u^n(t) \to \nabla u(t)$ strongly in $L^2({\Omega}';{{\mathbb R}}^N)$ for all $t \in I_\infty$, with $u(t)$ a minimizer of $$\int_{{\Omega}} |\nabla v|^2 \,dx+ {{\mathcal H}}^{N-1} \left( S_v \setminus \Gamma(t) \right),$$ where $\Gamma(t):= \bigcup_{s \in I_{\infty}, s \le t} S_{u(s)} \cup \partial {\Omega}_f$. The evolution $\{t \to u(t),\,t \in I_\infty\}$ is extended to the whole $[0,1]$ using the approximation from the left in time.
We can now use the arguments of Section \[convres\]. Following Lemma \[extension\], it turns out that for all $t \in [0,1]$ $$\label{estbelowc}
{{\mathcal E}}(t) \ge {{\mathcal E}}(0)+2\int_0^t \int_{\Omega}\nabla u(\tau) \nabla \dot{g}(\tau) \,dx\,d\tau.$$ Moreover, by the Transfer of Jump and the uniqueness argument of Lemma \[aextension\], we have that $\nabla u^n(t) \to \nabla u(t)$ strongly in $L^2({\Omega}';{{\mathbb R}}^N)$ for all $t \not\in {{\mathcal N}}$, where ${{\mathcal N}}$ is the (at most countable) set of discontinuities of the pointwise limit $\lambda$ of ${{\mathcal H}}^{N-1}(\Gamma(\cdot))$ (which exists up to a further subsequence by Helly’s Theorem). Then we pass to the limit in obtaining $${{\mathcal E}}(t) \le {{\mathcal E}}(0)+2\int_0^t \int_{{\Omega}} \nabla u(\tau) \nabla \dot{g}(\tau) \,dx\,d\tau;$$ moreover, following the proof of Theorem \[mainthm\], we have that for all $t \in [0,1]$ $${{\mathcal E}}(t) \le \liminf_n {{\mathcal E}}_n(t) \le \limsup_n {{\mathcal E}}_n(t)={{\mathcal E}}(0)
+2\int_0^t \int_{{\Omega}} \nabla u(\tau) \nabla \dot{g}(\tau) \,dx\,d\tau,$$ and taking into account we get the convergence of the total energy at any time. Since $\nabla u^n(t) \to \nabla u(t)$ strongly in $L^2({\Omega}';{{\mathbb R}}^N)$ for every $t \in I_\infty$, we deduce that $\lambda={{\mathcal H}}^{N-1}(\Gamma(\cdot))$ on $I_\infty$, so that the convergence of the surface energy holds in $I_\infty$. The extension to the continuity times for ${{\mathcal H}}^{N-1}(\Gamma(\cdot))$ follows like in the final part of the proof of Theorem \[mainthm\].
ACKNOWLEDGMENTS
The authors wish to thank Gianni Dal Maso for having proposed them the problem, and Gianni Dal Maso and Gilles A. Francfort for many helpful and interesting discussions.
[99]{}
: A compactness theorem for a new class of functions of bounded variations. [*Boll. Un. Mat. Ital.*]{}[**3-B**]{} (1989), 857-881.
: Existence theory for a new class of variational problems. [*Arch. Ration. Mech. Anal.*]{}[**111**]{} (1990), 291-322.
: A new proof of the SBV compactness theorem. [*Calc. Var.*]{}[**3**]{} (1995), 127-137.
: [*Functions of bounded variations and Free Discontinuity Problems*]{}. Clarendon Press, Oxford, 2000.
: Numerical experiments in revisited brittle fracture. [*J. Mech. Phys. Solids*]{},[**48-4**]{} (2000), 797-826.
: [*Opérateurs maximaux monotones et semi-groupes de contractions dans les espaces de Hilbert*]{}. North-Holland, Amsterdam, 1973.
: [*The Finite Element Method for Elliptic Problems*]{}, North Holland, Amsterdam (1978).
: Strong approximation of $GSBV$ functions by piecewise smooth functions. [*Ann. Univ. Ferrara - Sez VII - Sc. Mat.*]{}[**43**]{} (1997), 27-49.
: A model for the quasi-static growth of brittle fractures: existence and approximation results. [*Arch. Ration. Mech. Anal.*]{}[**162**]{} (2002), 101-135.
: Existence and convergence for quasi-static evolution in brittle fracture. [*To appear.*]{}
: Revisiting brittle fractures as an energy minimization problem. [*J. Mech. Phys. Solids*]{}[**46**]{} (1998), 1319-1342.
: Numerical methods for free-discontinuity problems based on approximations by $\Gamma$[-]{}convergence. PhD Thesis, SISSA/ISAS, (2001).
:[*Hausdorff Measures*]{}. Cambridge University Press, Cambridge, 1970.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We report the results of our theoretical study and analysis of earlier experimental data for the g-factor tensor components of the ground $^2\Pi_{1/2}$ state of free PbF radical. The values obtained both within the relativistic coupled-cluster method combined with the generalized relativistic effective core potential approach and with our fit of the experimental data from \[R.J. Mawhorter, B.S. Murphy, A.L. Baum, T.J. Sears, T. Yang, P.M. Rupasinghe, C.P. McRaven, N.E. Shafer-Ray, L.D. Alphei, J.-U. Grabow, Phys. Rev. A 84, 022508 (2011); A. Baum, B.S. thesis, Pomona College, 2011\]. The obtained results agree very well with each other but contradict the previous fit performed in the cited works. Our final prediction for g-factors is $G_{\parallel}= 0.081(5)$, $G_{\perp}=-0.27(1)$.'
author:
- 'L.V. Skripnikov'
- 'A.N. Petrov'
- 'A.V. Titov'
- 'R.J. Mawhorter'
- 'A.L. Baum'
- 'T.J. Sears'
- 'J.-U. Grabow'
title: 'Verification of g-factors for lead monofluoride ground state, PbF'
---
Introduction
============
Lead monofluodide, PbF, molecule is one of prospective systems to search for the electron electric dipole moment ([[*e*]{}EDM]{}). It was studied and discussed during three decades in many papers including [@Kozlov:87; @Dmitriev:92; @Shafer-Ray:06; @Shafer-Ray:08E; @Baklanov:10; @Petrov:13]. It was recently shown in Ref. [@Alphei:11] that some “enhanced” (coincidental) near-degeneracy for the levels of opposite parity in the ground rotational state $J=1/2$ for $^{207}$PbF of the ground electronic state $^2\Pi_{1/2}$ [@Shafer-Ray:08E] takes place that is caused by the near cancellation between the shifts in the energies of these levels due to omega-type doubling and the magnetic hyperfine interaction. This can lead to suppression of systematic errors in an experiment.
In Ref. [@Skripnikov:14c] we have calculated the parameters (more generally, the characteristics of atoms in compounds [@Lomachuk:13; @Titov:14a; @Skripnikov:15b]\]) required to interpret the experimental energy shift in terms of the [[*e*]{}EDM]{} and other effects of simultaneous violation of space parity (P) and/or time-reversal invariance (T) including the P-odd anapole moment [@Alphei:11] and the T,P-odd pseudoscalar-scalar electron-nucleus neutral current interaction for the ground $^2\Pi_{1/2}$ state. For instance, the effective electric field in PbF was found to be greater than or equal to those in the other transition element compounds considered (1.7 times larger than in HfF$^+$ [@Petrov:07a; @Fleig:13], 1.4 larger than in PtH$^+$ [@Skripnikov:09], and 1.1 larger than in WC [@Lee:13a] and TaN [@Skripnikov:15c])).
In the present paper our aim is to study the PbF g-factor for the $^2\Pi_{1/2}$ term which is required for preparation of experiments on the molecule [@Shafer-Ray:06; @Yang:13; @Petrov:14]. Up to now the g-factors have been measured in Ref. [@Mawhorter:11; @Baum:11] only. Previous theoretical estimations and calculations of g-factors have been performed in Refs. [@Kozlov:87; @Dmitriev:92; @Baklanov:10].
Molecular Hamiltonian
=====================
We represent the molecular Hamiltonian for $^{208}$PbF as [@Petrov:13]: $${\rm \bf H}_{\rm mol} = {\rm \bf H}_{\rm rot} + {\rm \bf H}_{\rm hfs} + {\rm \bf H}_{1} + {\rm \bf H}_{ext} .$$ Here ${\rm \bf H}_{\rm rot}$ is the rotational Hamiltonian and ${\rm \bf H}_{\rm hfs}$ is the hyperfine interaction between electrons and nuclei. ${\rm \bf H}_{1}$ includes the nuclear spin – rotational interaction and also effectively takes into account the rotational and hyperfine interactions between $^2\Pi_{1/2}$ and other electronic states. ${\rm \bf H}_{ext}$ describes the interaction of the molecule with an external magnetic field $\mathbf{B}$. Parameters for ${\rm \bf H}_{\rm rot}$, ${\rm \bf H}_{\rm hfs}$, and ${\rm \bf H}_{1}$ are taken from Ref. [@Petrov:13]. For ${\rm \bf H}_{ext}$ we have: $$\label{Bext}
{\rm \bf H}_{ext} = \mu_{B}~\mathbf{B}\cdot\widehat{G}\cdot \mathbf{S}^{\prime} -g_1\mu_{N}~\mathbf{B}\cdot \mathbf{I_1}$$ Here $\mathbf{S}^{\prime}$ is effective spin defined by the following equations: $\mathbf{S}^{\prime}_{\hat{n}}|\Omega> = \Omega|\Omega>$, $\mathbf{S}^{\prime}_{\pm}|\Omega=\mp 1/2> = |\Omega=\pm 1/2>$, $\mathbf{S}^{\prime}_{\pm}|\Omega=\pm 1/2> = 0$ [@Kozlov:87; @Dmitriev:92], $\mathbf{I_1}$ is the angular-momentum operator of the fluorine nuclei, $\mu_{B}$ and $\mu_{N}$ are Bohr and nuclear magnetons respectively, and $g_1=5.25773$ is the $^{19}$F nuclear $g-$factor.
In the molecular frame coordinate system the tensor contractions $$\begin{gathered}
\label{contraction}
\mathbf{B} \cdot\widehat{G}\cdot\mathbf{S}^{\prime}=G_{||}\mathbf{B}_{0}\mathbf{S}_{0}^{\prime}-G_{\perp}(\mathbf{B}_{1}\mathbf{S}_{-1}^{\prime}
+\mathbf{B}_{-1}\mathbf{S}_{1}^{\prime}) \end{gathered}$$ are determined by the body-fixed $g-$factors $$\begin{aligned}
\label{Gpar}
G_{\parallel} &=&\frac{1}{\Omega} \langle \Psi_{^2\Pi_{1/2}} |\hat{L}^e_{\hat{n}} - g_{S} \hat{S}^e_{\hat{n}} |\Psi_{^2\Pi_{1/2}} \rangle, \end{aligned}$$ $$\begin{aligned}
\label{Gperp}
G_{\perp} &=&\langle \Psi_{^2\Pi_{1/2}} |\hat{L}^e_{+} - g_{S} \hat{S}^e_{+} |\Psi_{^2\Pi_{-1/2}} \rangle, \end{aligned}$$ where ${\vec{L}}^e$ and ${\vec{S}}^e$ are the electronic orbital and electronic spin momentum operators, respectively; $g_{S} = -2.0023$ is a free$-$electron $g$-factor; $\hat{n}$ is the unit vector along the molecular axis directed from Pb to F.
In this paper the parameters $G_{\parallel}$ and $G_{\perp}$ are obtained (i) by using Eqs. (\[Gpar\],\[Gperp\]) from calculation of the [*electronic*]{} wavefunction $\Psi_{^2\Pi_{1/2}}$ and (ii) by fitting the experimentally observed transitions reported in Ref. [@Baum:11].
Methods
=======
The matrix elements (\[Gpar\],\[Gperp\]) were calculated using the computational scheme similar to that used by us in Ref. [@Skripnikov:14c]. The basis set for Pb was taken from Ref. [@Skripnikov:14c]. For F the aug-ccpVQZ basis set [@Kendall:92] with two removed g-type basis functions was employed. The Pb$-$F internuclear distance was set to 3.9 a.u., which is close to the experimental datum 3.8881(4) a.u. [@Lumley:77], which was later confirmed by Ref. [@Ziebarth:98]. Inner core $1s-4f$ electrons of lead were excluded from the correlation calculation using the “valence” semi-local version of the generalized relativistic effective core potential (GRECP) approach [@Mosyagin:10a; @Titov:99]. Note that the approach allows one to account for the Breit interaction very effectively [@Petrov:04b; @Mosyagin:06amin; @Mosyagin:10a]. All the other 31 electrons were included into the calculation. Electron correlation effects were considered within the relativistic two-component coupled-cluster approach with accounting for single and double cluster amplitudes, CCSD, as well as single, double and perturbative triple cluster amplitudes, CCSD(T). Note that the matrix element (\[Gperp\]) is off-diagonal. Therefore, it was calculated within the linear-response two-component coupled-cluster method with single and double cluster amplitudes [@Kallay:5]. The coupled-cluster calculations were performed using the [dirac12]{} [@DIRAC12] and [mrcc]{} [@MRCC2013] codes. Matrix elements of the operators corresponding to (\[Gpar\],\[Gperp\]) over the molecular spinors were calculated with the code developed in Refs. [@Skripnikov:11a; @Skripnikov:13b; @Skripnikov:13c; @Skripnikov:14b; @Skripnikov:15b; @Petrov:14].
To obtain the experimental values for $G_{\parallel}$ and $G_{\perp}$ we have performed two fits using the data from Ref. [@Baum:11]. In “fit 1” the Zeeman shifts of $J = 1/2$ to $J = 3/2$ transitions for the ground vibrational level of $^2\Pi_{1/2}$ electronic state are obtained by numerical diagonalization of the molecular Hamiltonian (${\rm \bf H}_{\rm mol}$) on the basis set of the electronic-rotational wavefunctions. The scheme of the calculation is similar to that employed in Refs. [@Petrov:11; @Petrov:13; @Lee:13a]. Only the $G_{\parallel}$ and $G_{\perp}$ parameters were optimized. The other parameters of ${\rm \bf H}_{\rm mol}$ were taken from Ref. [@Petrov:13]. In “fit 2” we have reproduced the scheme described in Ref. [@Mawhorter:11].
Results and discussion
======================
The results of our calculations of g-factors for the PbF ground state together with the results of previous studies are given in Table \[TResults\]. One can see that the value of $G_{\parallel}$ is stable with respect to improvement of the electron correlation treatment in the present study (from CCSD to CCSD(T) level).
[ l c c]{} Method & $G_{\parallel}$ & $G_{\perp}$\
SCF $^a$, [@Kozlov:87] & 0.034 $< G_{\parallel} <$ 0.114 & -0.438 $< G_{\perp} <$ -0.269\
SCF $^a$, [@Dmitriev:92] & 0.114 & -0.438\
13e-SODCI$^b$, [@Baklanov:10]& 0.082 & -0.319\
\
31e-CCSD, this work & 0.081 & -0.274\
31e-CCSD(T), & 0.081 & —\
this work &&\
\
Experiment, [@Mawhorter:11] & 0.12 & -0.38\
Experiment + fit 1, & 0.081 & -0.269\
this work &&\
Experiment + fit 2, & 0.085 & -0.271\
this work &&\
\
$^a$ SCF, self consistent field.\
$^b$ 13-electron SODCI, spin-orbit direct configuration interaction, [@Baklanov:10]. Outer-core electrons $5s^25p^65d^{10}$ of Pb are excluded from the correlation treatment.\
$G_{\parallel}$ and $G_{\perp}$ values obtained by fit 1 and fit 2 (see Methods section) are also given in Table \[TResults\]. The deviations of our fits from the observed Zeeman shifts are given in Table \[spectr208\]. For the last seven transitions the shifts are reproduced with deviations which are much larger than the declared experimental accuracy. One is inclined to suspect that the accuracy is overestimated for these transitions. We note however, that the experimental $(\Delta U/B)_{\rm obs}$ values for Zeeman components that only differ (model independent) in sign (e.g. F$_L$, MF$_L$ $\to$ F$_U$, MF$_U$ = 1, 0 $\to$ 2, 1 vs. 1, 0 $\to$ 2, -1; 1, 1 $\to$ 2, 2 vs. 1, -1 $\to$ 2, -2; 1, 1 $\to$ 2, 0 vs. 1, -1 $\to$ 2, 0; 0, 0 $\to$ 1, 1 vs. 0, 0 $\to$ 1, -1) agree within their error bars, which indicates correct accuracy estimations. It is also the case that the deviations for those pairs are systematic and not statistical. It seems that the F$_L$ $\to$ F$_U$ = 1 $\to$ 2 pattern is predicted to be somewhat too narrow while the F$_L$ $\to$ F$_U$ = 0 $\to$ 1 pattern is somewhat too wide.
We also note that the $G_{\parallel}$ = 0.085, $G_{\perp}$ = -0.271 parameters obtained in fit 2 differ substantially from the $G_{\parallel}$ = 0.12, $G_{\perp}$ = -0.38 values obtained by the same method and reported in the Ref. [@Mawhorter:11]. Our results here show good agreement between $G_{\parallel}$ and $G_{\perp}$ obtained in fit 1, fit 2, and the [*ab initio*]{} calculation. While both Ref. [@Mawhorter:11] values are higher by a common factor of $\sim$1.4–1.5, the origin of the discrepancies is not clear at present and will require further investigation.
Our final values for the g-factors are $G_{\parallel}$ = 0.081(5) and $G_{\perp}$ = -0.27(1). It should be noted that these smaller g-factor values and their improved accuracy together favor the experimental search for the electron electric dipole moment and other parity-violating and related effects [@Borschevsky:13; @Flambaum:2013] in PbF due to the additional suppression of systematic errors.
Unsplit line (MHz) F$_L$ F$_U$ MF$_L$ MF$_U$ $(\Delta U/B)_{\rm obs}$ [@Baum:11] $\delta_1$ $\delta_2$
-------------------- ------- ------- -------- -------- ------------------------------------- ------------ ------------
18414.588 1 2 -1 -1 0.0665(13) -40 -30
0 0 -0.00050(93) 107 50
1 1 -0.0635(13) 9 0
18462.193 0 1 0 0 0.00032(90) -89 -32
0 -1 -0.1369(30) -4 5
0 1 0.1363(29) -2 11
18497.136 1 1 -1 -1 0.00766(86) -150 -70
1 1 -0.00729(92) 119 33
1 0 -0.1428(17) 0 -15
0 -1 -0.1328(21) -33 -46
0 1 0.1345(13) 29 29
-1 0 0.1427(9) 2 16
22574.934 1 2 -1 -1 -0.03864(27) 28 100
0 0 -0.00005(90) 2 5
1 1 0.03851(9) -14 -87
1 0 0.1023(36) 64 62
0 -1 0.07296(25) -264 -207
-1 -2 0.03411(60) -211 -85
1 2 -0.03406(49) 206 80
0 1 -0.07267(52) 229 178
-1 0 -0.10323(79) -546 -529
22691.931 0 1 0 -1 0.11114(48) 371 433
0 1 -0.11133(41) -346 -414
: Observed Zeeman shifts $(\Delta U/B)_{\rm obs}$ (MHz/Gauss) of the $J = 1/2$ to $J = 3/2$ transitions for $^{208}$Pb$^{19}$F [@Baum:11]. The number in parenthesis gives two standard deviation error of the final digits of precision. The subscripts $U$ and $L$ refer to the upper and lower energy level of the transition, respectively. F is the total angular momentum of PbF, MF is its projection on laboratory axis. The deviation of $n$-th fit is given by $\delta_n= (\Delta U/B)_{\rm fit} - (\Delta U/B)_{\rm obs}$ in units of the last digit of precision.
\[spectr208\]
Acknowledgement {#acknowledgement .unnumbered}
===============
The molecular calculations were partly performed at the Supercomputer “Lomonosov”. This work is supported by the SPbU Fundamental Science Research grant from Federal budget No. 0.38.652.2013 and RFBR Grant No. 13-02-01406. L.S. is also grateful to the grant of President of Russian Federation No.MK-5877.2014.2 and Dmitry Zimin “Dynasty” Foundation. J.-U.G. acknowledges funding from the Deutsche Forschungsgemeinschaft (DFG) and the Land Niedersachsen, and R.J.M. and A.L.B. are grateful for research support provided by the Pomona College Sontag Fellowship Program.
[37]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, , , , , ****, ().
, , , , , , , ****, ().
, ****, ().
, , , ****, (), .
, , , , ****, ().
, , , , ****, ().
, , , , , , , , , , , ****, ().
, , , , ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, , , , ****, ().
, ****, ().
, , , , ****, ().
, , , , , , , ****, ().
, , , , (), .
, , , , ****, (), <http://link.aps.org/doi/10.1103/PhysRevA.87.014101>.
, , , , , , , , , , ****, ().
, , , , , , , , , , ****, ().
,
, , , ****, ().
, ****, ().
, , , , ****, (), ISSN .
, , , ****, ().
, ****, ().
, , , , ****, ().
, , , , ****, ().
, ****, ().
.
.
, , , , , ****, ().
(), .
, , , ****, ().
, ****, ().
, ****, ().
, , , , , ****, ().
, , , , ****, ().
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In this paper, we study a class of quasilinear Schrödinger equation of the form $$\begin{aligned}
-\varepsilon^2\Delta u+V(x)u-\varepsilon^2(\Delta(|u|^{2\alpha}))|u|^{2\alpha-2}u
&=&\lambda|u|^{q-2}u+|u|^{2^*(2\alpha)-2}u,\quad\mbox{in }{\mathbb{R}}^N,\end{aligned}$$ where $\varepsilon>0,~\lambda>0,~q\geq2,~\alpha>1/2$ are constants, $N\geq3$. By using change of variable and variational approach, the existence of positive solution which has a local maximum point and decays exponentially is obtained. $\\$[**Mathematics Subject Classification (2000)**]{} [35J10 35J20 35J25]{}
author:
- Zhouxin Li
- Yimin Zhang
title: 'Solutions for a class of quasilinear Schrödinger equations with critical exponents term [^1] '
---
Introduction {#sec:1}
============
Let $l$ and $h$ be real functions of pure power forms, it is interesting to consider the existence of solutions to the following quasilinear Schrödinger equation $$\label{sch1}
i\partial_tz=-\varepsilon^2\Delta z+W(x)z-l(|z|^2)z-k\varepsilon^2\Delta h(|z|^2)h'(|z|^2)z,\quad x\in{\mathbb{R}}^N,N\geq3,$$ where $W(x)$ is a given potential, $k$ is a real constant, $\varepsilon>0$ is a real parameter. It has many applications in physics according to different types of $h$. For example, in [@Ku-81], it was used in plasma physics with $h(s)=s$, and was used in [@Ri-94] to models the self-channeling of high-power ultrashort laser in matter with $h(s)=(1+s)^{1/2}$. Readers can refer to [@Liu-Wang-03; @Liu-Wang2-03] for references of more applications of it.
In this paper, we assume that $h(s)=s^{\alpha}$, $l(s)=\lambda
s^{(q-2)/2}+s^{2^*(\alpha)-1}$, where $\lambda>0,~q\geq2,~\alpha>1/2,~2^*=2N/(N-2)$ are constants. If we consider standing waves solutions of the form $z(x,t)=\exp(-iEt/\varepsilon)u(x)$, then $z(x,t)$ satisfies equation (\[sch1\]) if and only if the function $u(x)$ solves the equation $$\begin{aligned}
\label{sch2}
-\varepsilon^2\Delta u&+&V(x)u-k\alpha\varepsilon^2(\Delta(|u|^{2\alpha}))|u|^{2\alpha-2}u \nonumber\\
&=&\lambda|u|^{q-2}u+|u|^{2^*(2\alpha)-2}u,\quad u>0,x\in{\mathbb{R}}^N,\end{aligned}$$ where $V(x)=W(x)-E$ is the new potential function.
In case $k=0$, equation (\[sch2\]) is a semilinear elliptic equation which has been extensively studied in the past two decades. In recent years, the quasilinear case $k\neq0$ arose a lot of interest to mathematical researchers. From [@Liu-Wang2-03], we know that the constant $2^*(2\alpha)>2^*$ is thought to behave like the critical exponent to equation (\[sch2\]). When $\alpha=1$ (i.e. $h(s)=s$) and $\varepsilon=1$, this kind of problems with different types of nonlinearities $l(s)$ at sub-“critical growth”, i.e. at sub-$2^*(2\alpha)$ growth, have been widely studied, see [@Ch-Gu-05; @Co-Je-04; @Do-Se-09; @Liu-Wang2-03] and so on. In [@Liu-Wang2-03], by using a changing of variable, they transform the equation to a semilinear one, then the existence of solutions was obtained via variational methods under different types of potentials $V(x)$. This method is significant and was widely used in the studies of this kind of problems. For general $\alpha$, there are few results for this case, as far as we know, just [@aw; @aw2; @Liu-Wang-03; @Mo-06-2]. In [@Liu-Wang-03], for $\alpha>\frac{1}{2}$, $l(s)=\lambda s^{\frac{p-1}{2}}$ and $2<p+1<2^*(2\alpha)$, the existence of solution of equation (\[sch2\]) without critical term was obtained by using the method of Lagrange multiplier. In [@Mo-06-2], the existence of at least one or sometimes two standing wave solutions for $\alpha>\frac{1}{2}$ and $l(s)=\mu f(x) s^{\frac{p-1}{2}}$ was obtained through fibreing method. Employing the change variable method just as [@Liu-Wang2-03], the authors of [@aw] obtained the existence of at least one positive solution for $\alpha>\frac{1}{2}$ and general $l(s)$ by using variational approached. Moreover, in [@aw2], for $V(x)=\lambda$, $\alpha>\frac{1}{2}$ and $l(s)=s^{\frac{p-1}{2}}$, they obtained the unique existence of positive radial solution under some suitable conditions.
Problem at “critical growth”, i.e. at $2^*(2\alpha)$ growth rate, was studied by Moameni [@Mo-06] with a general $\varepsilon>0$ and $\alpha=1$. It was assumed in [@Mo-06] that $V(x)=0$ on an annule, which enable the avoidance of proving the compact embedding near the origin. In [@llw], the existence of radially symmetric solution was obtained for $\varepsilon>0$ small enough. Other kinds of such problems at “critical growth” were studied in [@do-Mi-So-07; @llw; @Zh-Wa-Sh] and the references therein. But all these are studied for $\alpha=1$, for general $\alpha$, there is no results according to what we know.
In this paper, our aim is to study the existence of positive solutions of (\[sch2\]) with general $\alpha>1/2$ and at $2^*(2\alpha)$ growth. Problem of (\[sch2\]) at $2^*(2\alpha)$ growth has two difficulties. Firstly, the embedding $H^1({\mathbb{R}}^N)\hookrightarrow
L^{2^*(2\alpha)}({\mathbb{R}}^N)$ is not compact, so it is hard to prove the Palais-Smale ((PS) in short) condition. Secondly, even if we can obtain the compactness result of (PS) sequence, it is only holds at some level of positive upper bound, it is difficult for us to prove that the functional has such minimax level.
We assume that $V(x)$ is locally Hölder continuous and
(V)
: $\exists~ V_\infty>V_0>0$ such that $\min_{x\in{\mathbb{R}}^N}V(x)=V_0$ and $\lim_{|x|\to\infty}V(x)=V_\infty$.
Note that assumption (V) allows zero be the minimum point of $V(x)$. This is different to the assumptions on $V(x)$ in [@Mo-06].
Under assumption (V), we define a space $$X:=\bigg\{u\in H^{1}({\mathbb{R}}^N):\int_{{\mathbb{R}}^N}V(x)u^2<\infty\bigg\}$$ with the norm $
\|u\|_X^2=\int_{{\mathbb{R}}^N}|\nabla u|^2+\int_{{\mathbb{R}}^N}V(x)u^2.
$
For simplicity of notation, we let $m=2\alpha$, $\bar q=q/m$ and $k\alpha=1$ in this paper. Set $$g(t)=\lambda|t|^{q-2}t+|t|^{2^*m-2}t\quad\mbox{and}\quad G(t)=\int_0^tg(s)ds.$$ We formulate problem (\[sch2\]) in variational structure in the space $X$ as follows: $$I(u)=\frac{\varepsilon^2}{2}\int_{{\mathbb{R}}^N}(1+m|u|^{2(m-1)})|\nabla u|^2+\frac{1}{2}\int_{{\mathbb{R}}^N}V(x)u^2-\int_{{\mathbb{R}}^N}G(u)$$ Note that $I$ is lower simicontinuous on $X$, we define that $u\in X$ is a [*weak solution*]{} for (\[sch2\]) if $u\in X\cap L^{\infty}({\mathbb{R}}^N)$ and it is a critical point of $I$.
Firstly, for an arbitrary $\varepsilon>0$, we have:
\[main1\] Assume that $q\in(2m,2^*m)$ and that condition (V) holds.
$\\$ Case 1: $1< m<2$. Assume that one of the following conditions holds:
(i)
: $\bar q>4+\frac{2}{m}$ if $N=3$;
(ii)
: $\bar q>2+\frac{2}{m}$ if $N=4$;
(iii)
: $\bar q>\frac{4}{3}+\frac{2}{m}$ if $N=5$;
(iv)
: $\bar q>2$ if $N\geq6$.
Case 2: $m\geq2$. Assume that one of the following conditions holds:
(v)
: $\bar q>5$ if $N=3$;
(vi)
: $\bar q>3$ if $N=4$;
(vii)
: $\bar q>\frac{7}{3}$ if $N=5$;
(viii)
: $\bar q>2$ if $N\geq6$.
Then for $\varepsilon>0$ small enough, problem (\[sch2\]) has a positive weak solution $u_\varepsilon\in X\cap L^{\infty}({\mathbb{R}}^N)$ with $$\lim_{\varepsilon\to0}\|u_\varepsilon\|_{X}=0,\quad\mbox{and}\quad
u_\varepsilon(x)\leq C\exp(-\frac{\beta}{\varepsilon}|x-x_\varepsilon|).$$ where $C>0,\beta>0$ are constants, $x_\varepsilon\in{\mathbb{R}}^N$ is a local maximum point of $u_\varepsilon$.
Next, we consider the case $\varepsilon=1$. We have the following result:
\[main2\] Assume that all conditions in Theorem \[main1\] hold and that $\varepsilon=1$, then problem (\[sch2\]) has a positive weak solution $u_1\in X\cap L^{\infty}({\mathbb{R}}^N)$.
This paper is organized as follows. In section 2, we first use a change of variable to reformulate the problem, then we modify the functional in order to regain the (PS) condition. In section 3, we prove that the functional satisfies the (PS) condition, this is a crucial job of this paper. Finally, in section 4, we prove the main theorems, which involves the construction of a mountain pass level at a certain high.
Preliminaries {#sec:2}
==============
Since $I$ is lower semicontinuous on $X$, we follow the idea in [@Co-Je-04; @Liu-Wang2-03] and make the change of variables $v=f^{-1}(u)$, where $f$ is defined by $$\begin{aligned}
\left\{
\begin{array}{ll}
f(0)=0, & \hbox{} \\
f'(v)=(1+m|f(v)|^{2(m-1)})^{-1/2}, & \hbox{on $[0,+\infty)$;} \\
f(v)=-f(-v), & \hbox{on $(-\infty,0]$.}
\end{array}
\right.\end{aligned}$$ The above function $f(t)$ and its derivative satisfy the following properties (see [@aw; @aw2; @Liu-Wang2-03]):
\[lem.f\] For $m>1$, we have
\(1) $f$ is uniquely defined, $C^2$ and invertible;
\(2) $|f'(t)|\leq1$ for all $t\in{\mathbb{R}}$;
\(3) $|f(t)|\leq |t|$ for all $t\in{\mathbb{R}}$;
\(4) $f(t)/t\to1$ as $t\to0$;
\(5) $|f(t)|\leq m^{1/2m}|t|^{1/m}$ for all $t\in{\mathbb{R}}$;
\(6) $\frac{1}{m}f(t)\leq tf'(t)\leq f(t)$ for all $t>0$;
\(7) $f(t)/\sqrt[m]{t}\to m^{1/2m}$ as $t\to+\infty$;
According to [@Do-Se-09] (see Corollary 2.1 and Proposition 2.2 in it, note that the embedding in Corollary 2.1 of [@Do-Se-09] is also compact.), we have:
\[lem.imbed\] The map: $v\mapsto f(v)$ from $X$ into $L^r({\mathbb{R}}^N)$ is continuous for $1\leq r\leq 2^*m$, and is locally compact for $1\leq r< 2^*m$.
Using this change of variable, we rewrite the functional $I(u)$ to: $$J(v)=I(f(v))=\frac{\varepsilon^2}{2}\int_{{\mathbb{R}}^N}|\nabla v|^2
+\frac{1}{2}\int_{{\mathbb{R}}^N}V(x)f^2(v)-\int_{{\mathbb{R}}^N}G(f(v)).$$ The critical point of $J$ is the weak solution of equation $$\begin{aligned}
\label{Sch3}
-\varepsilon^2\Delta v+V(x)f(v)f'(v)=g(f(v))f'(v),\quad x\in{\mathbb{R}}^N.\end{aligned}$$
Now we define a suitable modification of the functional $J$ in order to regain the Palais-Smale condition. In this time, we make use of the method in [@Pi-Fe-96].
Let $l$ be a positive constant such that $$\label{def.l}
l=\sup\{s>0: \frac{g(t)}{t}\leq\frac{V_0}{k}\mbox{ for every }0\leq t\leq s\}$$ for some $k>\theta/(\theta-2)$ with $\theta\in(2m,q]$. We define the functions: $$\gamma(s)=\left\{
\begin{array}{ll}
g(s), & \hbox{$s>0$;} \\
0, & \hbox{$s\leq0$.}
\end{array}
\right.\quad
\bar\gamma(s)=\left\{
\begin{array}{ll}
\gamma(s), & \hbox{$0\leq s\leq l$;} \\
\frac{V_0}{k}s, & \hbox{$s>l$.}
\end{array}
\right.$$ and $$p(x,s)=\chi_R(x)\gamma(s)+(1-\chi_R(x))\bar\gamma(s),$$ $$P(x,s)=\int_0^sp(x,t)dt,$$ where $\chi_R$ denotes the characteristic function of the set $B_R$ (the ball centered at 0 and with radius $R$ in ${\mathbb{R}}^N$), $R>0$ is sufficiently large and such that $$\min_{B_R}V(x)<\min_{\partial B_R}V(x).$$ By definition, the function $p(x,s)$ is measurable in $x$, of class $C$ in $s$ and satisfies:
(p$_1$)
: $0<\theta P(x,s)\leq p(x,s)s$ for every $x\in B_R$ and $s\in{\mathbb{R}}^+$.
(p$_2$)
: $0\leq2P(x,s)\leq p(x,s)s\leq\frac{1}{k}V(x)s^2$ for every $x\in B_R^c:={\mathbb{R}}^N\setminus B_R$ and $s\in{\mathbb{R}}^+$.
Now we study the existence of solutions for the deformed equation: $$\begin{aligned}
\label{Sch4}
-\varepsilon^2\Delta v+V(x)f(v)f'(v)=p(x,f(v))f'(v),\quad x\in{\mathbb{R}}^N.\end{aligned}$$ The corresponding functional of (\[Sch4\]) is given by $$\bar J(v)=\frac{\varepsilon^2}{2}\int_{{\mathbb{R}}^N}|\nabla v|^2
+\frac{1}{2}\int_{{\mathbb{R}}^N}V(x)f^2(v)-\int_{{\mathbb{R}}^N}P(x,f(v)).$$ For $v\in X$, since $$\begin{aligned}
\label{lem.X.1}
|\nabla (|f(v)|^m)|^2=\frac{m^2|f(v)|^{2(m-1)}}{1+m|f(v)|^{2(m-1)}}|\nabla v|^2\leq m|\nabla v|^2,\end{aligned}$$ we infer that $|f(v)|^m\in X$. By Sobolev inequality, we have $$\begin{aligned}
\label{lem.X.2}
\|f(v)\|_{2^*m} = \||f(v)|^{m}\|_{2^*}^{1/m}\leq C\|\nabla(|f(v)|^{m})\|_2^{1/m}\leq C\|v\|_{X}^{1/m}.\end{aligned}$$ It results that $f(v)\in L^{2^*m}({{\mathbb{R}}^N})$. Using interpolation inequality, we obtain that $f(v)\in L^{q}({{\mathbb{R}}^N})$. Thus $\bar J$ is well defined on $X$. Let $(v_n)\subset X, v\in X$ with $v_n\to v$ in $X$. Then from Lemma \[lem.imbed\], we infer that $V(x)f^2(v_n)\to V(x)f^2(v)$ in $L^1({{\mathbb{R}}^N})$ and that $f(v_n)\to f(v)$ in $L^{q}({{\mathbb{R}}^N})$. Thus $\bar J$ is continuous on $X$. $\bar J$ is Gateaux-differentiable in $X$ and the G-derivative is $$\begin{aligned}
\langle \bar J'(v),\varphi\rangle&=&\varepsilon^2\int_{{\mathbb{R}}^N}\nabla v\nabla \varphi+
\int_{{\mathbb{R}}^N}V(x)f(v)f'(v)\varphi \\
&&-\int_{{\mathbb{R}}^N}p(x,f(v))f'(v)\varphi,
\quad\forall\varphi\in X.\end{aligned}$$ Then if $v\in X\cap L^\infty({\mathbb{R}}^N)$ is a critical point of $\bar J$, and $v(x)\leq a:=f^{-1}(l),~\forall x\in B_R^c$, we have $u=f(v)\in X\cap L^\infty({\mathbb{R}}^N)$ (note that we have $|u|\leq|v|$ and $|\nabla u|\leq|\nabla v|$ by the properties of $f$) is a solution of (\[sch2\]).
Compactness of (PS) sequence {#sec:3}
=============================
In this section, we show that the functional $\bar J$ satisfies (PS) condition, this is a crucial job, its proof is composed of four steps. Let $S$ denotes the best Sobolev constant, we have
\[lem.PS\] Assume that condition (V) holds and $q\in(2m,2^*m)$. Then $\bar J$ satisfies (PS) condition at level $c_\varepsilon<\frac{1}{Nm}\varepsilon^N S^{N/2}$.
[**Proof**]{}Let $(v_n)\in E$ be a (PS) sequence of $\bar J$ at level $c_\varepsilon$, that is, $(v_n)$ satisfies: $$\begin{aligned}
\label{ps-j-1}
\bar J(v_n)&=&\frac{\varepsilon^2}{2}\int_{{\mathbb{R}}^N}|\nabla v_n|^2+\frac{1}{2}\int_{{\mathbb{R}}^N}V(x)f^2(v_n) \nonumber\\
&&-\int_{{\mathbb{R}}^N}P(x,f(v_n))=c_\varepsilon+o(1),\end{aligned}$$ and $$\begin{aligned}
\label{ps-j-2}
\langle\bar J'(v_n),\varphi\rangle&=&\varepsilon^2\int_{{\mathbb{R}}^N}\nabla v_n\nabla \varphi+
\int_{{\mathbb{R}}^N}V(x)f(v_n)f'(v_n)\varphi \nonumber\\
&&-\int_{{\mathbb{R}}^N}p(x,f(v_n))f'(v_n)\varphi
=o(1)\|\varphi\|_X,\quad \forall \varphi\in X.\end{aligned}$$
We divide the proof into four steps.
Step 1: the sequence $
\int_{{\mathbb{R}}^N}(|\nabla v_n|^2+V(x)f^2(v_n))
$ is bounded. Multiplying (\[ps-j-1\]) by $\theta$ ($\theta$ is given in section 2) and using ($p_1$)-($p_2$), we get $$\begin{aligned}
&&\frac{\theta\varepsilon^2}{2}\int_{{\mathbb{R}}^N}|\nabla v_n|^2+\frac{\theta}{2}\int_{{\mathbb{R}}^N}V(x)f^2(v_n)\\
&&\quad\leq \int_{B_R}p(x,f(v_n))f(v_n)+\frac{\theta}{2k}\int_{B_R^c}V(x)f^2(v_n)+\theta c_\varepsilon+o(1).\end{aligned}$$ On the other hand, taking $\varphi=f(v_n)/f'(v_n)$ in (\[ps-j-2\]), we get $$\begin{aligned}
&& \int_{{\mathbb{R}}^N}\varepsilon^2\bigg(1+\frac{m(m-1)|f(v_n)|^{2(m-1)}}{1+m|f(v_n)|^{2(m-1)}}\bigg)|\nabla v_n|^2+\int_{{\mathbb{R}}^N}V(x)f^2(v_n) \\
&&\quad = \int_{{\mathbb{R}}^N}p(x,f(v_n))f(v_n)+o(\|v_n\|_X)
\geq \int_{B_R}p(x,f(v_n))f(v_n)+o(1)\|v_n\|_X.\end{aligned}$$ Conbining the above two inequalities, we get $$\begin{aligned}
\label{lem2-6-0}
&&(\frac{\theta}{2}-m)\varepsilon^2\int_{{\mathbb{R}}^N}|\nabla v_n|^2+(\frac{\theta}{2}-\frac{\theta}{2k}-1)\int_{{\mathbb{R}}^N}V(x)f^2(v_n) \nonumber\\
&&\quad\leq \frac{\theta\varepsilon^2}{2} \int_{{\mathbb{R}}^N}|\nabla v_n|^2-
\int_{{\mathbb{R}}^N}\varepsilon^2\bigg(1+\frac{m(m-1)|f(v_n)|^{2(m-1)}}{1+m|f(v_n)|^{2(m-1)}}\bigg)|\nabla v_n|^2 \nonumber\\
&&\qquad +\frac{\theta}{2}\int_{{\mathbb{R}}^N}V(x)f^2(v_n)-\frac{\theta}{2k}\int_{B_R^c}V(x)f^2(v_n)-\int_{{\mathbb{R}}^N}V(x)f^2(v_n) \nonumber\\
&&\quad\leq \theta c_\varepsilon+o(1)+o(1)\|v_n\|_X.\end{aligned}$$ Since $\theta>2m$ and $k>\frac{\theta}{\theta-2}$, we get the conclusion from (\[lem2-6-0\]).
Step 2: for every $\delta>0$, there exists $R_1\geq R>0$ such that $$\begin{aligned}
\label{lem2-6-1}
\limsup_{n\to\infty}\int_{B_{2R_1}^c}(|\nabla v_n|^2+V(x)f^2(v_n))<\delta.\end{aligned}$$ We consider a cut-off function $\psi_{R_1}=0$ on $B_{R_1}$, $\psi_{R_1}=1$ on $B_{2R_1}^c$, $|\nabla \psi_{R_1}|\leq C/R_1$ on ${\mathbb{R}}^N$ for some constant $C>0$. On one hand, taking $\varphi=f(v_n)/f'(v_n)$, we compute $\langle\bar J'(v_n),\varphi\psi_{R_1}\rangle$ and get $$\begin{aligned}
\label{lem2-6-2}
o(1)\|v_n\|_X &=& \int_{{\mathbb{R}}^N}\varepsilon^2\bigg(1+\frac{m(m-1)|f(v_n)|^{2(m-1)}}{1+m|f(v_n)|^{2(m-1)}}\bigg)|\nabla v_n|^2\psi_{R_1} \nonumber\\
&&+\int_{{\mathbb{R}}^N}\varepsilon^2\varphi\nabla v_n\nabla \psi_{R_1}
+\int_{{\mathbb{R}}^N}V(x)f^2(v_n)\psi_{R_1} \nonumber\\
&&-\int_{{\mathbb{R}}^N}p(x,f(v_n))f(v_n)\psi_{R_1} \nonumber\\
&\geq& \int_{{\mathbb{R}}^N}\varepsilon^2|\nabla v_n|^2\psi_{R_1}
+\int_{{\mathbb{R}}^N}\varepsilon^2\varphi\nabla v_n\nabla \psi_{R_1} \nonumber\\
&&+(1-\frac{1}{k})\int_{{\mathbb{R}}^N}V(x)f^2(v_n)\psi_{R_1}.\end{aligned}$$ On the other hand, by Hölder inequality, $$\begin{aligned}
\label{lem2-6-3}
\bigg|\int_{{\mathbb{R}}^N}\varphi\nabla v_n\nabla \psi_{R_1}\bigg|\leq\frac{C}{R_1}\|\nabla v_n\|_{L^2({\mathbb{R}}^N)}\|\varphi\|_{L^2({\mathbb{R}}^N)}.\end{aligned}$$ Note that $\|\nabla v_n\|_{L^2({\mathbb{R}}^N)}$ is bounded, and $$\begin{aligned}
\label{lem2-6-4}
\|\varphi\|_{L^2({\mathbb{R}}^N)}^2
&=&\int_{{\mathbb{R}}^N}f^2(v_n)(1+m|f(v_n)|^{2(m-1)}) \nonumber\\
&=&\int_{{\mathbb{R}}^N}f^2(v_n)+m\int_{{\mathbb{R}}^N}|f(v_n)|^{2m},\end{aligned}$$ by (\[lem.X.2\]), $\|\varphi\|_{L^2({\mathbb{R}}^N)}$ is bounded also. Therefore, it follows from (\[lem2-6-2\])-(\[lem2-6-4\]) that $$\limsup_{n\to\infty}\int_{B_{2R_1}^c}(|\nabla v_n|^2+V(x)f^2(v_n))\leq\frac{C}{R_1}$$ for $R_1$ sufficiently large, this yields (\[lem2-6-1\]).
Step 3, there exists $v\in X$ such that $$\begin{aligned}
\label{lem2-6-5}
\lim_{n\to\infty}\int_{{\mathbb{R}}^N}p(x,f(v_n))f(v_n)=\int_{{\mathbb{R}}^N}p(x,f(v))f(v).\end{aligned}$$ Firstly, by step 1, there exists $v\in X$ such that up to a subsequence, $v_n\to v$ weakly in $X$ and $v_n\to v$ a.e. in ${\mathbb{R}}^N$. Since we may replace $v_n$ by $|v_n|$, we assume $v_n\geq0$ and $v\geq0$. By (\[lem2-6-1\]), for any $\delta>0$, there exists $R_1>0$ sufficiently large such that $$\limsup_{n\to\infty}\int_{B_{2R_1}^c}(|\nabla v_n|^2+V(x)f^2(v_n))\leq k\delta.$$ Therefore, by ($p_2$) we have $$\label{lem2-6-6}
\limsup_{n\to\infty}\int_{B_{2R_1}^c}p(x,f(v_n))f(v_n)
\leq \limsup_{n\to\infty}\int_{B_{2R_1}^c}\frac{V(x)}{k}f^2(v_n)\leq\delta,$$ and by Fatou Lemma, $$\label{lem2-6-7}
\int_{B_{2R_1}^c}p(x,f(v))f(v)\leq\delta.$$ Secondly, we prove that $$\label{lem2-6-8}
\int_{B_{2R_1}}p(x,f(v_n))f(v_n)\to\int_{B_{2R_1}}p(x,f(v))f(v).$$ Then from this, (\[lem2-6-6\])-(\[lem2-6-7\]), and the arbitrariness of $\delta$, we get (\[lem2-6-5\]). In fact, since $(v_n)$ is bounded in $X$, we have $(f(v_n))$ is bounded also. Thus there exists a $w\in X$ such that $f(v_n)\rightharpoonup w$ in $X$, $f(v_n)\to w$ in $L^r(B_{R_1})$ for $1\leq r<2^*$ and $f(v_n)\to w$ a.e. in $B_{R_1}$. According to (\[lem.X.1\]), $(|f(v_n)|^m)$ is also bounded in $X$. By a normal argument, we have $|f(v_n)|^m\rightharpoonup |w|^m$ in $X$, $|f(v_n)|^m\to |w|^m$ in $L^r(B_{R_1})$ for $1\leq r<2^*$ and $|f(v_n)|^m\to |w|^m$ a.e. in $B_{R_1}$. Applying Lions’ concentration compactness principle [@PLL-85] to $(|f(v_n)|^m)$ on $\bar B_{R_1}$, we obtain that there exist two nonnegative measures $\mu,~\nu$, a countable index set $K$, positive constants $\{\mu_k\},~ \{\nu_k\},~ k\in K$ and a collection of points $\{x_k\},~ k\in K$ in $\bar{B}_{R_1}$ such that for all $k\in K$,
(i)
: $\nu=|w|^{2^*m}+\sum\limits_{k\in K}\nu_k\delta_{x_k}$;
(ii)
: $\mu=|\nabla (|w|^m)|^{2}+\sum\limits_{k\in K}\mu_k\delta_{x_k}$;
(iii)
: $\mu_k\geq S\nu_k^{2/2^*}$,
where $\delta_{x_k}$ is the Dirac measure at $x_k$, $S$ is the best Sobolev constant. We claim that $\nu_k=0$ for all $k\in K$. In fact, let $x_k$ be a singular point of measures $\mu$ and $\nu$, as in [@Li-Zou-98], we define a function $\phi\in
C_0^{\infty}({\mathbb{R}}^N)$ by $$\phi(x)=\left\{
\begin{array}{ll}
1, & \hbox{$B_{\rho}(x_k)$;} \\
0, & \hbox{${\mathbb{R}}^N\backslash B_{2\rho}(x_k)$;} \\
\phi\geq0,|\nabla\phi|\leq\frac{1}{\rho}, & \hbox{$B_{2\rho}(x_k)\backslash B_{\rho}(x_k)$.}
\end{array}
\right.$$ where $B_{\rho}(x_k)$ is a ball centered at $x_k$ and with radius $\rho>0$. We take $\varphi=\phi f(v_n)/f'(v_n)$ as test functions in $\langle \bar J'(v_n),\varphi\rangle$ and get $$\begin{aligned}
\label{prop4-3-1}
&&\int_{{\mathbb{R}}^N} \varepsilon^2\bigg(1+\frac{m(m-1)|f(v_n)|^{2(m-1)}}{1+m|f(v_n)|^{2(m-1)}}\bigg)|\nabla v_n|^2\cdot\phi\nonumber\\
&&\quad +\int_{{\mathbb{R}}^N}\varepsilon^2\nabla v_n\nabla\phi\cdot f(v_n)/f'(v_n)
+\int_{{\mathbb{R}}^N}V(x)f^2(v_n)\phi\nonumber\\
&&\qquad -\int_{{\mathbb{R}}^N}p(x,f(v_n))f(v_n)\phi=o(1)\|v_n\phi\|_X.\end{aligned}$$ Then Lions’ concentration compactness principle implies that $$\begin{aligned}
\label{prop4-3-2}
\int_{B_{R_1}}|\nabla |f(v_n)|^m|^2\phi\to\int_{B_{R_1}}\phi
d\mu,\quad\int_{B_{R_1}}|f(v_n)|^{2^*m}\phi\to\int_{B_{R_1}}\phi d\nu.\end{aligned}$$ Since $x_k$ is singular point of $\nu$, by the continuity of $f$, we have $$f(v_n(x))|_{(B_{2\rho}\setminus\{x_k\})}\to\infty$$ as $\rho\to0$. Thus $$1+\frac{m(m-1)|f(v_n)|^{2(m-1)}}{1+m|f(v_n)|^{2(m-1)}}=m-o(\rho).$$ on $B_{2\rho}$ for $\rho$ sufficiently small. Then by (\[lem.X.1\]) we get from (\[prop4-3-1\]) that $$\begin{aligned}
\label{prop4-3-3}
&&\int_{B_{R_1}}\varepsilon^2\phi d\mu-\int_{B_{R_1}}\phi d\nu\nonumber\\
&&\quad=\lim_{n\to\infty}\bigg[\int_{B_{R_1}}\varepsilon^2|\nabla |f(v_n)|^m|^2\phi-\int_{B_{R_1}}|f(v_n)|^{2^*m}\phi\bigg]\nonumber\\
&&\quad\leq\lim_{n\to\infty}\bigg[\int_{B_{R_1}} m\varepsilon^2|\nabla v_n|^2\phi-\int_{B_{R_1}}|f(v_n)|^{2^*m}\phi\bigg]\nonumber\\
&&\quad\leq\lim_{n\to\infty}\bigg[\int_{B_{R_1}}
\varepsilon^2\bigg(1+\frac{m(m-1)|f(v_n)|^{2(m-1)}}{1+m|f(v_n)|^{2(m-1)}}\bigg)|\nabla v_n|^2\phi\nonumber\\
&&\qquad+o(\rho)\int_{B_{R_1}}\varepsilon^2|\nabla v_n|^2\phi-\int_{B_{R_1}}|f(v_n)|^{2^*m}\phi\bigg]\nonumber\\
&&\quad\leq\lim_{n\to\infty}\bigg[-\int_{B_{R_1}} \varepsilon^2\nabla v_n\nabla \phi\cdot f(v_n)/f'(v_n)
+\lambda\int_{B_{R_1}} |f(v_n)|^q\phi\nonumber\\
&&\qquad+o(\rho)\int_{B_{R_1}}\varepsilon^2|\nabla v_n|^2\phi+o(1)\|v_n\phi\|_X\bigg].\end{aligned}$$ We prove that the last inequality in (\[prop4-3-3\]) tends to zero as $\rho\to0$. By Hölder inequality, we have $$\begin{aligned}
\label{prop4-3-4}
&&\lim_{n\to\infty}\bigg|\int_{B_{R_1}} \nabla v_n\nabla\phi\cdot f(v_n)/f'(v_n)\bigg|\nonumber\\
&&\quad\leq\limsup_{n\to\infty}\bigg(\int_{B_{R_1}}|\nabla v_n|^2\bigg)^{1/2}\cdot
\bigg(\int_{B_{R_1}} |[f(v_n)/f'(v_n)]\cdot\nabla\phi|^2\bigg)^{1/2}.\end{aligned}$$ Since $|f(v_n)/f'(v_n)|^2=f^2(v_n)+m|f(v_n)|^{2m}$, using Hölder inequality we have $$\begin{aligned}
&&\lim_{n\to\infty}\int_{B_{R_1}} |[f(v_n)/f'(v_n)]\cdot\nabla\phi|^2 \\
&&\quad \leq C\rho\big(\|w\|^2_{L^{2^*}(B_{2\rho}(x_j))}+\|w\|^2_{L^{2^*m}(B_{2\rho}(x_j))}\big)
\to0\end{aligned}$$ as $\rho\to0$. Thus we obtain that the right hand side of (\[prop4-3-4\]) tends to 0. On the other hand, since $q\in(2m,2^*m)$, by Lemma \[lem.imbed\], we can prove that $g(x,h(v_n))h(v_n)\phi\to g(x,w)w\phi$ in $L^1({B_{R_1}})$ and $\int_{B_{R_1}} g(x,w)w\phi\to0$ as $\rho\to0$. All these facts imply that the last inequality in (\[prop4-3-3\]) tends to zero as $\rho\to0$. Thus $\nu_k\geq\varepsilon^2\mu_k$. This means that either $\nu_k=0$ or $\nu_k\geq \varepsilon^N S^{N/2}$ by virtue of Lions’ concentration compactness principle. We claim that the latter is impossible. Indeed, if $\nu_k\geq \varepsilon^N S^{N/2}$ holds for some $k\in K$, then $$\begin{aligned}
c_\varepsilon&=&\lim_{n\to\infty}\bigg{\{}\bar J(v_n)-\frac{1}{2m}\langle \bar J'(v_n),f(v_n)/f'(v_n)\rangle\bigg{\}}\nonumber\\
&\geq&\lim_{n\to\infty}\bigg{\{}(\frac{1}{2m}-\frac{1}{2^*m})\int_{{\mathbb{R}}^N}|f(v_n)|^{2^*m}\bigg{\}}
\geq(\frac{1}{2m}-\frac{1}{2^*m})\int_{{\mathbb{R}}^N} d\nu \nonumber\\
&\geq&(\frac{1}{2m}-\frac{1}{2^*m})\int_{{\mathbb{R}}^N}|w|^{2^*m}+(\frac{1}{2m}-\frac{1}{2^*m})S^{N/2}\varepsilon^N
\geq\frac{1}{Nm}\varepsilon^N S^{N/2},\nonumber\end{aligned}$$ which is a contradiction. Thus $\nu_k=0$ for all $k\in K$, and it implies that $\|f(v_n)\|_{L^{2^*m}(B_{R_1})}$ $\to\|w\|_{L^{2^*m}(B_{R_1})}$. By the uniform convexity of $L^{2^*m}(B_{R_1})$, we have $f(v_n)\to w$ strongly in $L^{2^*m}(B_{R_1})$. Finally, since $p(x,f(v_n))f(v_n)$ is sub-$(2^*m)$ growth on $B_{2{R_1}}\setminus B_{R_1}$, we conclude that (\[lem2-6-8\]) holds. This proves (\[lem2-6-5\]).
Step 4: $(v_n)$ is compact in $X$. Since we have (\[lem2-6-5\]), the proof of the compactness is trivial. This completes the proof of the lemma. $\quad\square$
Proof of main results {#sec:4}
======================
Before we prove Theorem \[main1\], we will show firstly some properties about the change variable $f$.
\[lem.f1\] Let $f_1(v)=|f(v)|^m/v,~v\neq0$ and $f_1(0)=0$, then $f_1$ is continuous, odd, nondecreasing and $$\begin{aligned}
\label{lem.f.limit}
\lim_{v\to0}f_1(v)=0, \quad\mbox{and } \lim_{|v|\to+\infty}|f_1(v)|=\sqrt{m}.\end{aligned}$$
[**Proof**]{}In fact, by (6) of Lemma \[lem.f\], $$f'_1(v)=v^{-2}(m|f(v)|^{m-2}f(v)f'(v)v-|f(v)|^m)\geq0,$$ so $f_1$ is nondecreasing. By (4) of Lemma \[lem.f\], $f_1(v)\to0$ as $v\to0$. Finally, according to Hospital Principle, $$\lim_{v\to+\infty}f_1(v)=\lim_{v\to+\infty}\frac{|f(v)|^m}{v}=\lim_{v\to+\infty}m|f(v)|^{m-2}f(v)f'(v)=\sqrt{m}.$$ This shows that (\[lem.f.limit\]) holds. $\quad\Box$
\[lem.f2\] There exists $d_0>0$ such that $$\lim_{v\to+\infty}(\sqrt{m}v-f^m(v))\geq d_0.$$
[**Proof**]{}Assume that $v>0$. Since by (6) of Lemma \[lem.f\], $f(v)\leq mf'(v)v$, we have $$\begin{aligned}
\label{lem.f2.1}
\sqrt{m}v-f^m(v)&\geq&\sqrt{m}v-mf^{m-1}(v)f'(v)v\nonumber\\
&=&\frac{\sqrt{1+mf^{2(m-1)}(v)}-\sqrt{m}f^{m-1}(v)}{\sqrt{1+mf^{2(m-1)}(v)}}\sqrt{m}v \nonumber\\
&=&\frac{\sqrt{m}v}{\big(\sqrt{1+mf^{2(m-1)}(v)}+\sqrt{m}f^{m-1}(v)\big)\sqrt{1+mf^{2(m-1)}(v)}} \nonumber\\
&\geq& \frac{\sqrt{m}v}{2(1+mf^{2(m-1)}(v))}\geq\frac{f^m(v)}{4mf^{2(m-1)}(v)} \nonumber\\
&=&\frac{1}{4mf^{m-2}(v)}:=d(m,v).\end{aligned}$$ In the last inequality, we have used the fact that $\sqrt{m}v\geq f^m(v)$ and that $mf^{2(m-1)}(v)>1$ for $v>0$ sufficiently large.
If $1< m< 2$, then $d(m,v)\to+\infty$ as $v\to+\infty$. If $m=2$, then $d(m,v)=1/8$. If $m>2$, we claim that $\sqrt{m}v- f^m(v)\to0$ as $v\to+\infty$ is impossible. In fact, assume on the contrary, then using Hospital Principle, we get $$\begin{aligned}
0&\leq& \lim_{v\to+\infty}\frac{\sqrt{m}v- f^m(v)}{f^{2-m}(v)}
=\lim_{v\to+\infty}\frac{\sqrt{m}-mf^{m-1}(v)f'(v)}{(2-m)f^{1-m}(v)f'(v)}\\
&=& \lim_{v\to+\infty}\frac{m}{(2-m)f^{1-m}(v)\big(\sqrt{m}\sqrt{1+mf^{2(m-1)}(v)}+mf^{m-1}(v)\big)}\\
&=&\frac{1}{2(2-m)}<0.\end{aligned}$$ This is a contradiction. Thus for all $m>1$, there exists $d_0>0$ such that there holds $$\lim_{v\to+\infty}(\sqrt{m}v-f^m(v))\geq d_0.$$ This completes the proof. $\quad\Box$
\[lem.f3\] We have
\(i) If $1< m <2$, then $$\lim_{v\to+\infty}\frac{\sqrt{m}v-f^m(v)}{f^{2-m}(v)}=\frac{1}{2(2-m)}.$$
\(ii) If $m\geq2$, then $$\lim_{v\to+\infty}\frac{\sqrt{m}v-f^m(v)}{\log {f(v)}}\leq\left\{
\begin{array}{ll}
\frac{1}{2}, & \hbox{$m=2$;} \\
0, & \hbox{$m>2$.}
\end{array}
\right.$$
[**Proof**]{}Firstly, we prove part (i). According to (\[lem.f2.1\]) in Lemma \[lem.f2\], we have $\sqrt{m}v-f^m(v)\to+\infty$ as $v\to+\infty$. Thus by Hospital Principle, we get $$\lim_{v\to+\infty}\frac{\sqrt{m}v-f^m(v)}{f^{2-m}(v)}
=\lim_{v\to+\infty}\frac{\sqrt{m}-mf^{m-1}(v)f'(v)}{(2-m)f^{1-m}(v)f'(v)}
=\frac{1}{2(2-m)}.$$
Next, we prove part (ii). If there exists a constant $C>0$ such that $\sqrt{m}v-f^m(v)\leq C$, then the conclusion holds. Otherwise, assume that $\sqrt{m}v-f^m(v)\to+\infty$ as $v\to+\infty$. Then again by Hospital Principle, we have $$\begin{aligned}
\lim_{v\to+\infty}\frac{\sqrt{m}v-f^m(v)}{\log{f(v)}}
=\lim_{v\to+\infty}\frac{\sqrt{m}-mf^{m-1}(v)f'(v)}{f'(v)/f(v)}
=\left\{
\begin{array}{ll}
\frac{1}{2}, & \hbox{$m=2$;} \\
0, & \hbox{$m>2$.}
\end{array}
\right.\end{aligned}$$ This completes the proof. $\quad\Box$
To prove Theorem \[main1\], it is crucial to prove that $\bar J$ has the mountain pass level $c_\varepsilon<\frac{1}{Nm}\varepsilon^N S^{N/2}$. Let us consider the following family of functions in [@Br-Ni-83] $$v^*_\omega(x)=\frac{[n(n-2)\omega^2]^{(n-2)/4}}{[\omega^2+|x|^2]^{(n-2)/2}},$$ which solves the equation $-\Delta u=u^{2^*-1}$ in ${\mathbb{R}}^N$ and satisfies $\|\nabla v^*_\omega\|^2_{L^2}=\|v^*_\omega\|^{2^*}_{L^{2^*}}=S^{N/2}$. Let $\omega$ be such that $2\omega<R$ and let $\eta_\omega(x)\in[0,1]$ be a positive smooth cut-off function with $\eta_\omega(x)=1$ in $B_\omega$, $\eta_\omega(x)=0$ in $B_R\setminus B_{2\omega}$. Let $v_\omega=\eta_\omega v^*_\omega$. For all $\omega>0$, there exists $t^\omega>0$ such that $\bar J(t^\omega v_\omega)<0$ for all $t>t^\omega$. Define the class of paths $$\Gamma=\{\gamma\in C([0,1],X):~\gamma(0)=0,\gamma(1)=t^\omega v_\omega\},$$ and the minimax level $$c_\varepsilon=\inf_{\gamma\in\Gamma}\max_{t\in[0,1]}\bar J(\gamma(t)).$$ Let $t_\omega$ be such that $$\bar J(t_\omega v_\omega)=\max_{t\geq0}\bar J(tv_\omega).$$ Note that the sequence $(v_\omega)$ is uniformly bounded in $X$, then if $\bar J(t_\omega v_\omega)\to0$ as $t_\omega\to0$, we are done; on the other hand, if $t_\omega\to+\infty$, then $\bar J(t_\omega v_\omega)\to-\infty$, which is impossible, so it remains to consider the case where the sequence $(t_\omega)$ is upper and lower bounded by two positive constants. According to [@Br-Ni-83], we have, as $\omega\to0$, $$\|\nabla v_\omega\|^2_{L^{2}}=S^{N/2}+O(\omega^{N-2}),\quad
\|v_\omega\|^{2^*}_{L^{2^*}}=S^{N/2}+O(\omega^{N}).$$ Let $a\in(0,\frac{\varepsilon^{(N-2)/2}}{2\sqrt{m}})$, $b\in(\frac{2\varepsilon^{(N-2)/2}}{\sqrt{m}},+\infty)$ be such that $t_\omega\in[a,b],~\forall \omega\in(0,\omega_0)$, where $\omega_0>0$ small enough. By computing $\frac{d}{dt}\bar J(tv_\omega)=0$, we obtain $t_\omega=\frac{\varepsilon^{(N-2)/2}}{\sqrt{m}}+o(1)$. Let $$H(v)=-\frac{1}{2}V(x)f^2(v)+\frac{\lambda}{q}|f(v)|^q-\frac{1}{2^*m}|\sqrt{m}v|^{2^*}+\frac{1}{2^*m}|f(v)|^{2^*m},$$ then by (\[lem.f.limit\]) and (4) of Lemma \[lem.f\], for $m>1$, we have $$\lim_{|v|\to+\infty}H(v)/|v|^{2^*}=0,\quad\mbox{and}\quad \lim_{v\to0}H(v)/v^2=-\frac{1}{2}V(x).$$ Thus $H(v)$ is sub-$(2^*)$ growth.
The following proposition is important to the computation of a mountain pass level $c_\varepsilon<\frac{1}{Nm}\varepsilon^nS^{N/2}$.
\[prop.H\] Under the assumptions of Theorem \[main1\], there exists a function $\tau=\tau(\omega)$ such that $\lim_{\omega\to0}\tau(\omega)=+\infty$ and for $\omega$ small enough, $$\begin{aligned}
\int_{{\mathbb{R}}^N}H(t_\omega v_\omega)\geq\tau(\omega)\cdot\omega^{N-2}.\end{aligned}$$
[**Proof**]{}We divide the proof into three steps.
Step 1: we prove that $$\begin{aligned}
\label{prop.H.1}
\frac{1}{\omega^{N-2}}\int_{B_\omega}H(t_\omega v_\omega)\geq\tau_1(\omega)\end{aligned}$$ with $\lim_{\omega\to0}\tau_1(\omega)=+\infty$.
By the definition of $v_\omega$, for $x\in B_\omega$, there exist constants $c_2\geq c_1>0$ such that for $\omega$ small enough, we have $$\label{prop.H.2}
c_1\omega^{-(N-2)/2}\leq v_\omega(x)\leq c_2\omega^{-(N-2)/2}.$$ and $$\label{prop.H.3}
c_1\omega^{-(N-2)/2}\leq f^m(v_\omega(x))\leq c_2\omega^{-(N-2)/2}.$$ On one hand, by (7) of Lemma \[lem.f\], (\[prop.H.3\]) and the continuity of $V(x)$ in $\bar B_\omega$, there exists $C_1>0$ such that $$\label{prop.H.4}
\int_{B_\omega}V(x)f^2(t_\omega v_\omega)
\leq C_1\omega^{N-\frac{2}{m}\frac{N-2}{2}}
= C_1\omega^{(\frac{2^*}{2}-\frac{1}{m})(N-2)}.$$ Similarly, there exists $C_2>0$ such that $$\label{prop.H.5}
\int_{B_\omega}f^q(t_\omega v_\omega)
\geq C_2\omega^{N-\frac{q}{m}\frac{N-2}{2}}
=C_2\omega^{(\frac{2^*}{2}-\frac{\bar q}{2})(N-2)},$$ where $\bar q=q/m$. On the other hand, using Hölder inequality, we have $$\begin{aligned}
\label{prop.H.6}
&&\frac{1}{2^*m}\int_{B_\omega}\bigg[(\sqrt{m}t_\omega v_\omega)^{2^*}
-(f^{m}(t_\omega v_\omega))^{2^*}\bigg] \nonumber\\
&&\quad\leq\frac{1}{m}\int_{B_\omega}(\sqrt{m}t_\omega v_\omega)^{2^*-1}
[\sqrt{m}t_\omega v_\omega-f^m(t_\omega v_\omega)] \nonumber\\
&&\quad \leq\frac{1}{m}\bigg(\int_{B_\omega}(\sqrt{m}t_\omega v_\omega)^{2^*}\bigg)^{(2^*-1)/2^*}
\bigg(\int_{B_\omega}[\sqrt{m}t_\omega v_\omega-f^m(t_\omega v_\omega)]^{2^*}\bigg)^{1/2^*}.\end{aligned}$$
[**Case 1**]{}: $1< m<2$. From (\[prop.H.6\]) and (i) of Lemma \[lem.f3\], we obtain that there exists $C_3>0$ such that $$\begin{aligned}
\label{prop.H.7}
&&\frac{1}{2^*m}\int_{B_\omega}\bigg[(\sqrt{m}t_\omega v_\omega)^{2^*}
-(f^{m}(t_\omega v_\omega))^{2^*}\bigg] \nonumber\\
&&\quad \leq C_3 \omega^{[N-(\frac{2}{m}-1)\frac{N-2}{2}2^*]\frac{1}{2^*}}
=C_3\omega^{(1-\frac{1}{m})(N-2)}.\end{aligned}$$ Combining (\[prop.H.4\]), (\[prop.H.5\]) and (\[prop.H.7\]), we have $$\begin{aligned}
&&\frac{1}{\omega^{N-2}}\int_{B_\omega}H(t_\omega v_\omega) \nonumber\\
&&\quad\geq
-C_1\omega^{(\frac{2^*}{2}-\frac{1}{m}-1)(N-2)}
+C_2\omega^{(\frac{2^*}{2}-\frac{\bar q}{2}-1)(N-2)}
-C_3\omega^{-\frac{1}{m}(N-2)}:=\tau_1(\omega).\end{aligned}$$ It is obvious that $\frac{2^*}{2}-\frac{1}{m}-1>-\frac{1}{m}$. No matter which one of conditions (i)-(iv) in Theorem \[main1\] holds, we all have $\frac{2^*}{2}-\frac{\bar q}{2}-1<-\frac{1}{m}$. It results that $\tau_1(\omega)\to+\infty$ as $\omega\to0$.
[**Case 2**]{}: $m\geq2$. Note that for any $\delta\in(0,m)$, $\lim_{v\to+\infty}\log{f(v)}/f^{\delta}(v)=0,$ we have $\log{f(v)}\leq f^{\delta}(v)$ for $v>0$ large enough. Thus for $\omega>0$ small enough, from (\[prop.H.6\]) and (ii) of Lemma \[lem.f3\], we get $$\begin{aligned}
\label{prop.H.8}
&&\frac{1}{2^*m}\int_{B_\omega}\bigg[(\sqrt{m}t_\omega v_\omega)^{2^*}
-(f^{m}(t_\omega v_\omega))^{2^*}\bigg] \nonumber\\
&&\quad \leq C_3' \omega^{[N-\frac{\delta}{m}\frac{N-2}{2}2^*]\frac{1}{2^*}}
=C_3'\omega^{\frac{1}{2}(1-\frac{\delta}{m})(N-2)}.\end{aligned}$$ Combining (\[prop.H.4\]), (\[prop.H.5\]) and (\[prop.H.8\]), we have $$\begin{aligned}
&&\frac{1}{\omega^{N-2}}\int_{B_\omega}H(t_\omega v_\omega) \nonumber\\
&&\quad\geq
-C_1\omega^{(\frac{2^*}{2}-\frac{1}{m}-1)(N-2)}
+C_2\omega^{(\frac{2^*}{2}-\frac{\bar q}{2}-1)(N-2)}
-C_3'\omega^{-\frac{1}{2}(1+\frac{\delta}{m})(N-2)}:=\tau_1(\omega).\end{aligned}$$ Since $m\geq2$, we have $\frac{2^*}{2}-\frac{1}{m}-1>-\frac{1}{2}(1+\frac{\delta}{m})$. No matter which one of conditions (v)-(viii) in Theorem \[main1\] holds, there exists a $\delta=\delta(N,\bar q)>0$ (depends on $N$ and $\bar q$) small enough such that $\frac{2^*}{2}-\frac{\bar q}{2}-1<-\frac{1}{2}(1+\frac{\delta}{m})$. It results that $\tau_1(\omega)\to+\infty$ as $\omega\to0$.
Case 1 and case 2 show that (\[prop.H.1\]) holds.
Step 2: we prove that there exists $C_4>0$ such that $$\begin{aligned}
\label{prop.H.9}
\frac{1}{\omega^{N-2}}\int_{B_{2\omega}\setminus B_\omega}H(t_\omega v_\omega)\geq -C_4\omega^{(\frac{2^*}{2}-\frac{1}{m}-1)(N-2)}:=\tau_2(\omega).\end{aligned}$$
Note that for $x\in B_{2\omega\setminus B_\omega}$, we have $$\begin{aligned}
\label{prop.H.10}
v_\omega(x)\leq v^*_\omega(x) \leq c_2\omega^{-(N-2)/2}.\end{aligned}$$ Since $\eta_\omega$ is a positive smooth cut-off function, without lost of generality, we may assume that $\eta_\omega$ is such that $$\int_{B_{2\omega}\setminus B_\omega}|v_\omega|^{2^*}
\leq \int_{B_{2\omega}\setminus B_\omega}V(x)f^2(v_\omega).$$ Thus by (\[lem.f.limit\]) and (\[prop.H.10\]), we have $$\begin{aligned}
&&\frac{1}{\omega^{N-2}}\int_{B_{2\omega}\setminus B_\omega}H(t_\omega v_\omega)\\
&&\quad \geq-\frac{1}{2\omega^{N-2}}\int_{B_{2\omega}\setminus B_\omega}V(x)f^2(t_\omega v_\omega)
-\frac{1}{2^*m\omega^{N-2}}\int_{B_{2\omega}\setminus B_\omega}|\sqrt{m}t_\omega v_\omega|^{2^*} \\
&&\quad \geq-\frac{C_5}{\omega^{N-2}}\int_{B_{2\omega}\setminus B_\omega}V(x)f^2(t_\omega v_\omega)\\
&&\quad \geq-C_4\omega^{N-\frac{2}{m}\frac{N-2}{2}-(N-2)}
= -C_4\omega^{(\frac{2^*}{2}-\frac{1}{m}-1)(N-2)},\end{aligned}$$ where $C_4>0$, $C_5>0$ are constants. This shows that (\[prop.H.9\]) holds.
Step 3: to conclude, let $\tau(\omega)=\tau_1(\omega)+\tau_2(\omega)$, we have $\tau(\omega)\to+\infty$ as $\omega\to0$. This implies the conclusion of the proposition. $\quad\Box$
$\\$ By (4) and (7) of Lemma \[lem.f\], it is easy to verify that $\bar J$ has the Mountain Pass Geometry. Lemma \[lem.PS\] shows that $\bar J$ satisfies (PS) condition. We prove that $\bar J$ has the mountain pass level $c_\varepsilon<\frac{1}{Nm}\varepsilon^NS^{N/2}$. Let $$F(t)=\frac{\varepsilon^2}{2}\|\nabla(tv_\omega)\|_{L^2}^2-\frac{1}{2^*m}\|\sqrt{m}tv_\omega\|_{L^{2^*}}^{2^*}.$$ Then we have $$\begin{aligned}
\label{main1.1}
F(t)\leq F(t_0)=\frac{1}{Nm}\varepsilon^NS^{N/2}+O(\omega^{N-2}),\quad\forall t\geq0,\end{aligned}$$ where $t_0=\frac{\varepsilon^{(N-2)/2}}{\sqrt{m}}$. By (\[main1.1\]) and (\[prop.H\]), we have $$\begin{aligned}
\bar J(t_\omega v_\omega)
&=&F(t_\omega v_\omega)-\int_{{\mathbb{R}}^N}H(t_\omega v_\omega) \\
&\leq& \frac{1}{Nm}\varepsilon^NS^{N/2}+O(\omega^{N-2})-\tau(\omega)\omega^{N-2} \\
&<&\frac{1}{Nm}\varepsilon^NS^{N/2}.\end{aligned}$$ This shows that $\bar J(v)$ has a nontrivial critical point $v_\varepsilon\in X$, which is a weak solution of (\[Sch4\]).
We prove that $v_\varepsilon$ is also a weak solution of (\[Sch3\]). Firstly, we can argue as the proof of Proposition 2.1 in [@Pi-Fe-96] to obtain that $$\lim_{\varepsilon\to0}\max_{x\in\partial B_R}{v_\varepsilon(x)}=0.$$ Thus there exists $\varepsilon_0>0$ such that for all $\varepsilon\in(0,\varepsilon_0)$, we have $v_\varepsilon(x)\leq a:=f^{-1}(l),~\forall ~|x|= R$, where $l$ is given in (\[def.l\]). Secondly, we prove that $$\label{main1.2}
v_\varepsilon(x)\leq a,
\quad\forall~\varepsilon\in(0,\varepsilon_0)\mbox{ and }\forall ~x\in{\mathbb{R}}^N\setminus B_R.$$ Taking $$\varphi=\left\{
\begin{array}{ll}
(v_\varepsilon-a)^+, & \hbox{$x\in{\mathbb{R}}^N\setminus B_R$;} \\
0, & \hbox{$x\in B_R$.}
\end{array}
\right.$$ as a test function in $\langle\bar J'(v_\varepsilon),\varphi\rangle=0$, we get $$\begin{aligned}
\label{main1.3}
&&\varepsilon^2\int_{{\mathbb{R}}^N\setminus B_R}|\nabla (v_\varepsilon-a)^+|^2 \nonumber\\
&&\quad +\varepsilon^2\int_{{\mathbb{R}}^N\setminus B_R}\bigg(V(x)-\frac{p(x,f(v_\varepsilon))}{f(v_\varepsilon)}\bigg)
f(v_\varepsilon)f'(v_\varepsilon)(v_\varepsilon-a)^+=0.\end{aligned}$$ By ($p_2$), we have $$V(x)-\frac{p(x,f(v_\varepsilon))}{f(v_\varepsilon)}>0,\quad\forall x\in{\mathbb{R}}^N\setminus B_R.$$ Therefore, all terms in (\[main1.3\]) must be equal to zero. This implies $v_\varepsilon\leq a$ in ${\mathbb{R}}^N\setminus B_R$. This proves (\[main1.2\]). Thus $v_\varepsilon$ is a solution of Problem (\[Sch3\]).
To complete the proof, we deduce as the proof for Theorem 4.1 in [@Li-11] to obtain that $v_\varepsilon|_{B_R}\in L^{\infty}(B_R)$. Thus $u_\varepsilon=f(v_\varepsilon)\in X\cap L^{\infty}({\mathbb{R}}^N)$ is a nontrivial weak solution of (\[sch2\]). Finally, by Proposition \[prop.u\] in the following, we have $
\lim_{\varepsilon\to0}\|u_\varepsilon\|_X=0
$ and $
u_\varepsilon(x)\leq Ce^{-\frac{\beta}{\varepsilon}|x-x_\varepsilon|}.
$ This completes the proof. $\quad\Box$
We prove the norm estimate and the exponential decay.
\[prop.u\] Let $v_\varepsilon\in X\cap L^{\infty}({\mathbb{R}}^N)$ be a solution of (\[Sch3\]) and let $u_\varepsilon=f(v_\varepsilon)$, then we have $$\lim_{\varepsilon\to0}\|u_\varepsilon\|_X=0,\quad\mbox{and}\quad
u_\varepsilon(x)\leq Ce^{-\frac{\beta}{\varepsilon}|x-x_\varepsilon|},$$ where $C>0$, $\beta>0$ are constants.
[**Proof**]{}Firstly, let $x_0\in B_R$ be such that $V(x_0)=V_0$. Define $J_{0}: X\to {\mathbb{R}}$ by $$J_{0}(v)=\frac{1}{2}\int_{{\mathbb{R}}^N}|\nabla v|^2+\frac{1}{2}\int_{{\mathbb{R}}^N}V_0f^2(v)
-\int_{{\mathbb{R}}^N}G(f(v)).$$ Let $$c_0=\inf_{\gamma\in\Gamma_0}\sup_{t\in[0,1]}J_{0}(\gamma(t)),$$ $$\Gamma_0=\{v\in C([0,1],X):\gamma(0)=0, J_{0}(\gamma(1))<0\}.$$ Similar to the proof for estimate (2.4) in [@Pi-Fe-96] (or Lemma 3.1 in [@Sq-03]), we can show that $
c_\varepsilon\leq \varepsilon^Nc_0+o(\varepsilon^N)
$ by using the change of coordinates $y=(x-x_0)/\varepsilon$. Arguing as for (\[lem2-6-0\]), and by virtue of this energy estimate, we obtain $$\|v_\varepsilon\|_X\leq\frac{\theta c_\varepsilon}{\min\{(\frac{\theta}{2}-m)\varepsilon^2,(\frac{\theta}{2}-\frac{\theta}{2k}-1)\}}
\leq\frac{2\theta c_0}{\theta-2m}\varepsilon^{N-2}+o(\varepsilon^{N-2})$$ for $\varepsilon>0$ sufficient small. Let $u_\varepsilon=f(v_\varepsilon)$, then $u_\varepsilon\neq0$. Note that $|\nabla u_\varepsilon|\leq |v_\varepsilon|$ and $|u_\varepsilon|\leq|v_\varepsilon|$, we get $
\lim_{\varepsilon\to0}\|u_\varepsilon\|_X=0.
$
Secondly, similar to the proof for Theorem 4.1 in [@Li-11], we conclude that $v_\varepsilon\in L^{\infty}({\mathbb{R}}^N)$ and by [@La-Ur-68], we have $v_\varepsilon\in C^{1,\alpha}(B_R)$. Now let $x_\varepsilon$ denote the maximum point of $v_\varepsilon$ in $B_R$ and let $$\sigma:=\sup\{s>0:g(t)< V_0t \mbox{ for every } t\in[0,s]\}.$$ Then $v_\varepsilon(x_\varepsilon)\geq f^{-1}(\sigma)$ for $\varepsilon>0$ small. In fact, assume that $v_\varepsilon(x_\varepsilon)<f^{-1}(\sigma)$ for some $\varepsilon>0$ sufficiently small. According to the definition of $l$ (see (\[def.l\])) and $\sigma$, we have $v_\varepsilon(x)\leq f^{-1}(l)<f^{-1}(\sigma)$ (note that $k>1$ in (\[def.l\])), $\forall x\in {\mathbb{R}}^N\setminus B_R$. Thus $$V(x)-\frac{g(f(v_\varepsilon))}{f(v_\varepsilon)}>0,\quad\forall x\in{\mathbb{R}}^N.$$ Since $v_\varepsilon=f^{-1}(u_\varepsilon)$ is a critical point of $J_\varepsilon$, we choose $\varphi=f(v_\varepsilon)/f'(v_\varepsilon)$ as a test function in $\langle J'_\varepsilon(v_\varepsilon),\varphi\rangle=0$ and get $$\begin{aligned}
0&=& \varepsilon^2\int_{{\mathbb{R}}^N}\bigg(1+\frac{m(m-1)|f(v_n)|^{2(m-1)}}{1+m|f(v_n)|^{2(m-1)}}\bigg)|\nabla v_\varepsilon|^2 \\
&&+\int_{{\mathbb{R}}^N}V(x)f^2(v_\varepsilon)-\int_{{\mathbb{R}}^N}g(f(v_\varepsilon))f(v_\varepsilon) \\
&=& \varepsilon^2\int_{{\mathbb{R}}^N}\bigg(1+\frac{m(m-1)|f(v_n)|^{2(m-1)}}{1+m|f(v_n)|^{2(m-1)}}\bigg)|\nabla v_\varepsilon|^2 \\
&&+\int_{{\mathbb{R}}^N}\bigg(V(x)-\frac{g(f(v_\varepsilon))}{f(v_\varepsilon)}\bigg)f^2(v_\varepsilon).\end{aligned}$$ It turns out that all terms in the above equality must be equal to zero, which means that $v_\varepsilon\equiv0$, a contradiction.
Now let $w_\varepsilon(x)=v_{\varepsilon}(x_\varepsilon+\varepsilon x)$, then $w_\varepsilon$ solves the equation $$\begin{aligned}
-\Delta w_\varepsilon+V(x_\varepsilon+\varepsilon x)f(w_\varepsilon)f'(w_\varepsilon)=g(f(w_\varepsilon))f'(w_\varepsilon), \quad x\in{\mathbb{R}}^N.\end{aligned}$$ Note that $\lim_{t\to0^+}\frac{f(t)f'(t)}{t}=1$ by the properties of $f$ and that $w_\varepsilon(x)\to0$ as $|x|\to+\infty$, we have, there exists $R_0>0$ such that for all $|x|\geq R_0$, $$\label{thm-3}
f(w_\varepsilon(x))f'(w_\varepsilon(x))\geq\frac{3}{4}w_\varepsilon(x)$$ and $$\label{thm-4}
g(f(w_\varepsilon(x)))f'(w_\varepsilon(x))\leq\frac{V_0}{2}w_\varepsilon(x).$$ Let $\varphi(x)=Me^{-\beta|x|}$ with $\beta^2<\frac{V_0}{4}$ and $Me^{-\beta R_0}\geq w_\varepsilon(x)$ for all $|x|=R_0$. It is easy to verify that for $x\neq0$, $$\label{thm-5}
\Delta \varphi\leq\beta^2\varphi.$$ Now define $\psi_\varepsilon=\varphi-w_\varepsilon$. Using (\[thm-3\])-(\[thm-5\]), we have $$\begin{aligned}
\left\{
\begin{array}{ll}
-\Delta \psi_\varepsilon+\frac{V_0}{4}\psi_\varepsilon\geq0, & \hbox{in $|x|\geq R_0$;} \\
\psi_\varepsilon\geq0, & \hbox{in $|x|=R_0$;} \\
\lim_{|x|\to\infty}\psi_\varepsilon=0. & \hbox{}
\end{array}
\right.\end{aligned}$$ By the maximum principle, we have $\psi_\varepsilon\geq0$ for all $|x|\geq R_0$. Thus, we obtain that for all $|x|\geq R_0$, $$w_\varepsilon(x)\leq\varphi(x)\leq Me^{-\beta|x|}.$$ Using the change of variable, we have that for all $|x|\geq R_0$, $$v_\varepsilon(x)=w_\varepsilon(\varepsilon^{-1}(x-x_\varepsilon))\leq Me^{-\frac{\beta}{\varepsilon}|x-x_\varepsilon|}.$$ Then by the regularity of $v_\varepsilon$ on $B_R$ and note that $f(t)\leq t$ for all $t\geq0$, we have $$u_\varepsilon(x)\leq Ce^{-\frac{\beta}{\varepsilon}|x-x_\varepsilon|}$$ for some $C>0$. This completes the proof. $\quad\square$
$\\$ We consider the following equation $$\begin{aligned}
\label{sch5}
-\Delta u&+&V(x)u-k\alpha(\Delta(|u|^{2\alpha}))|u|^{2\alpha-2}u \nonumber\\
&=&\lambda|u|^{q-2}u+|u|^{2^*(2\alpha)-2}u,\quad u>0,x\in{\mathbb{R}}^N.\end{aligned}$$ Let $y=\varepsilon x$ with $\varepsilon\in(0,\varepsilon_0)$, $\varepsilon_0$ is given by Theorem \[main1\], then we can transform (\[sch5\]) into $$\begin{aligned}
\label{sch6}
-\varepsilon^2\Delta u&+&\bar V(y)u-k\alpha\varepsilon^2(\Delta(|u|^{2\alpha}))|u|^{2\alpha-2}u \nonumber\\
&=&\lambda|u|^{q-2}u+|u|^{2^*(2\alpha)-2}u,\quad u>0,y\in{\mathbb{R}}^N.\end{aligned}$$ Here $\bar V(y)=V(\frac{y}{\varepsilon})$ still has the properties given in assumption (V). Thus according to Theorem \[main1\], (\[sch6\]) has a positive weak solution $u_\varepsilon(y)$ in $X\cap L^\infty({\mathbb{R}}^N)$, this implies that (\[sch5\]) has a positive weak solution $u_1(x)=u_\varepsilon(\varepsilon x)$. $\quad\Box$
Adachi, S., Watanabe, T.: G-invariant positive solutions for a quasilinear Schröinger equation. Adv. Diff. Eqns. **16**, 289-324 (2011)
Adachi, S., Watanabe,T.: Uniqueness of the ground state solutions of quasilinear Schröinger equations. Nonlinear Anal. **75**, 819-833 (2012)
Brézis, H., Nirenberg, L.: Positive solutions of nonlinear elliptic equations involving critical Sobolev exponents. Comm. Pure Appl. Math. **36**, 437-477 (1983)
Chen, J., Guo,B.: Multiple nodal bound states for a quasilinear Schrödinger equation. J. Math. Phys. **46**, 123502, 11pp (2005)
Colin, M., Jeanjean, L.: Solutions for a quasilinear Schrödinger equation: a dual approach. Nonlinear Anal. **56**, 213-226 (2004)
do Ó, J. M., Miyagaki, O., Soares, S.:Soliton solutions for quasilinear Schrödinger equations: the critical exponential case. Nonlinear Anal. **67**, 3357-3372 (2007)
Do Ó, J. M., Severo, U.: Quasilinear Schrö dinger equations involving concave and convex nonlinearities. Comm. Pure Appl. Anal. **8**, 621-644 (2009)
Kurihura, S.: Large-amplitude quasi-solitions in superfluid films. J. Phys. Soc. Jpn **50**, 3262-3267 (1981)
Ladyzhenskaya, O. A., Uraltseva, N. N.: Linear and quasilinear elliptic equations. Academic Press, New York (1968)
Li, S., Zou, W.: Remarks on a class of elliptic problems with critical exponents. Nonlinear Anal. **12**(6), 769-774 (1998)
Li, Z.: Existence of nontrivial solutions for quasilinear elliptic equations at critical growth. Appl. Math. Comput. **218**, 76-87 (2011)
Lions, P. L.: The concentration-compactness principle in the calculus of varations, The locally compact case, part I and part II. Rev Mat Iberoamericana **1**, 145-201, 223-283 (1985)
Liu, J., Wang, Z.-Q.: Soliton solutions for quasilinear Schrödinger equations, I. Proc. Amer. Math. Soc. **131**, 441-448 (2003)
Liu, J., Wang, Y., Wang, Z.-Q.: Soliton solutions for quasilinear Schrödinger equations, II. J. Differential Equations **187**, 473-493 (2003)
Liu, X., Liu, J., Wang, Z.-Q.: Ground states for quasilinear Schröinger equations with critical growth. Calc. Var., DOI 10.1007/s00526-012-0497-0
Moameni, A.: On the existence of standing wave solutions to quasilinear Schrödinger equations. Nonlinearity **19**, 937-957 (2006)
Moameni, A.: Existence of soliton solutions for a quasilinear Schrödinger equation involving critical exponent in ${\mathbb{R}}^N$. J. Differential Equations **229**, 570-587 (2006)
del Pino, M., Felmer, P. L.: Local mountain passes for semilinear elliptic problems in unbounded domains. Calc. Var. Partial Differ. Equ. **4**, 121-137 (1996)
Ritchie, B.: Relativistic self-focusing and channel formation in laser-plasma interactions. Phys. Rev. E **50**, 687-689 (1994)
Squassina, M.: Spike solutions for a class of singularly perturbed quasilinear elliptic equation. Nonlinear Anal. **54**, 1307-1336 (2003)
Zhang, Y., Wang, Y., Shen, Y.: Solutions for quasilinear Schrödinger equations with critical Sobolev-Hardy exponents. Comm. Pure Appl. Anal. **10**, 1037-1054 (2011)
[^1]: Supported by NSFC under grant numbers 11201488 and 11101418.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Nickel-doped graphene has been synthesized from propylene on a Ni(111) surface and studied using scanning tunneling microscopy (STM) and density functional theory (DFT). It is established that nickel centers are formed during graphene synthesis on the Ni(111) surface by both chemical vapor deposition (CVD) and temperature-programmed growth (TPG); apparently, they are always present in graphene synthesized on Ni(111). The centers are observed in STM images as single defects or defect chains and identified by DFT calculations as Ni atoms in carbon bivacancies. These nickel atoms are positively charged and may be of interest for single-atom catalysis. The incorporated Ni atoms should remain in graphene after the detachment from the substrate since they bound more strongly with carbon atoms in graphene than with substrate nickel atoms.'
author:
- 'S. L. Kovalenko'
- 'T. V. Pavlova'
- 'B. V. Andryushechkin'
- 'G. M. Zhidomirov'
- 'K. N. Eltsov'
bibliography:
- 'Ni\_defect\_Gr\_arXiv.bib'
title: 'Ni-doped Epitaxial Graphene Monolayer on the Ni(111) Surface'
---
\[sec:intro\]Introduction
=========================
Defects in graphene are known to affect significantly its properties. In particular, their presence may cause the shift of Fermi level, opening locally the energy gap. Moreover, defects can influence the carrier mobility, conductivity, magnetic and mechanical properties of graphene; and play the role of active centers in chemical reactions [@2011Banhart]. Doping graphene with nitrogen and boron also affects the Fermi level position (these processes can be considered as p-doping [@2013Cattelan; @2013Gebhardt] and n-doping [@2011Usachov; @2012Velez-Fort], respectively), leads to the energy gap opening ($\approx$0.2 eV for nitrogen [@2011Usachov]), and changes the carrier mobility in graphene [@2010Guo; @2013Wang].
Graphene doped with individual metal atoms has recently been studied for use as a single atom catalyst (SAC), in particular, in the hydrogen evolution reaction (HER) [@2015Qiu; @2018Zhang] and oxygen reduction reactions (ORR) [@2018Zhang; @2019Mao]. Moreover, for single-atom catalysis, not only free graphene is interesting, but also graphene on a substrate (i.e. Ni(111) [@2019Mao]), since the presence of a substrate allows one to change the activation barrier of reactions.
The graphene/Ni(111) system attracts a lot of interest (see review [@2014Dahal]) due to the small (1.2%) lattices mismatch that potentially makes possible the fabrication of the large area graphene crystals of high quality. Many researchers [@2012Jacobson; @2014Bianchini; @2015Smerieri; @2013Patera; @2019Carnevali; @2015Garcia-Lekue; @2017Parreiras; @2017Kovalenko] observed characteristic objects in scanning tunneling microscopy (STM) images of graphene synthesized on Ni(111), which looked much brighter than graphene structural defects.
The observation of such features (both individual objects and chains) was reported for graphene synthesized on Ni(111) by chemical vapor deposition (CVD) from toluene [@2012Jacobson] or ethylene [@2014Bianchini], although, their structure was not considered. Similar objects were observed in graphene synthesized by CVD from ethylene and interpreted as possible impurities (molecules from the residual atmosphere of ultrahigh vacuum chamber) deposited on graphene structural defects [@2015Smerieri]. In Ref. [@2013Patera], bright objects with a concentration of 1–2$\%$ were observed in graphene formed on the Ni(111) surface from ethylene by CVD at the temperature range of 400–500$^{\circ}$C. The interpretation assumed a possible replacement of carbon atoms by nickel ones, however, the structure details were not discussed. Recently, it has been shown that the incorporation of nickel atoms into the graphene lattice does occur during CVD synthesis of graphene on Ni(111), moreover, these atoms are visualized in STM as bright points [@2019Carnevali].
Note that similar bright objects were also observed on graphene islands synthesized by temperature programmed growth (TPG) on the Ni(111) surface from ethylene and propylene [@2015Garcia-Lekue; @2017Parreiras]. In our previous study [@2017Kovalenko], we considered TPG synthesis of graphene on Ni(111) and assigned the bright features in STM images with nickel defects, i.e. with nickel atoms occupying mono- or bivacancies. However, we were not able to establish the exact structure of the objects by comparing the theoretical and experimental STM images.
In the present work, graphene was synthesized from propylene on the Ni(111) surface by both CVD and TPG methods. In both cases, identical objects similar to the defects described in [@2012Jacobson; @2014Bianchini; @2013Patera; @2015Smerieri; @2019Carnevali; @2015Garcia-Lekue; @2017Parreiras; @2017Kovalenko] were found in STM images. We performed a DFT simulation of the transition from one structure (nickel in monovacancies) to another structure (nickel in bivacancies) and found that the main nickel defect in graphene is the nickel atom in the carbon bivacancy. Since nickel defects are present in graphene synthesized by both CVD and TPG, we believe that they are always formed during the graphene synthesis on the Ni(111) surface. It should also be noted that metal atoms occupying carbon bivacancies in graphene were already considered theoretically in [@2013WangLu; @2015Qiu; @2016Xu; @2018Zhang; @2019Mao]. Nickel defects have a positive charge and are promising from the point of view of single-atom catalysis [@2019Mao].
\[sec:method\]Methods
=====================
\[ssec:exp\]Experimental Techniques
-----------------------------------
TPG of graphene was carried out in two stages: (i) propylene was adsorbed from 2 mm capillary (dose of 500 Langmuir) on a purified Ni(111) surface at room temperature; (ii) sample was annealed at $500^\circ$C in UHV without propylene flux until a monolayer graphene coverage was formed. CVD of graphene was performed by supplying propylene at a pressure of $3\times 10^{-5}$Torr and temperature of $500^\circ$C on the clean Ni(111) surface. The clean Ni(111) surface was prepared by standard cycles, including etching by 1 keV Ar+ ions and subsequent annealing at $600^\circ$C for 15 min. All technological processes and surface analysis were performed in an ultrahigh-vacuum setup with a base pressure better than $1\times 10^{-10}$Torr. STM measurements were carried out at room temperature using scanning tunneling microscope GPI300[@SigmaScan].
\[ssec:com\]Theoretical methods
-------------------------------
All DFT calculations were performed in the spin-polarized version, implemented in the VASP software package [@1993Kresse; @1996Kresse]. Generalized gradient approximation (GGA) and exchange-correlation PBE functional [@1996Perdew] were applied (Van der Waals interactions were taken into account using the Grimme method [@2006Grimme]). The Ni(111) surface was modeled as a set of repeated 6$\times$6 cells, each containing four layers of nickel atoms, separated by a vacuum gap of 17[Å]{}. The atomic arrangements in the lower two layers were fixed during calculations, whereas the atoms from other layers could relax. The graphene structure was optimized on the upper nickel layer so that carbon atoms were in the on-top positions (above nickel atoms) and in the fcc positions (on-top/fcc configuration). The activation barrier was calculated using the nudged elastic band (NEB) method [@1998NEB], implemented in the VASP package. The reciprocal-space partition into a 6$\times$6$\times$1 grid was carried out using the G-centered k-point grid. NEB calculations were carried out applying a 3$\times$3$\times$1 k-point grid. STM images were simulated using the Hive STM program [@2008Vanpoucke] in the Tersoff–Hamann approximation [@1985Tersoff].
Results and Discussion
======================
Figure \[fig1\] shows an STM image of the Ni(111) surface almost completely covered with graphene. There is a triangular nickel island at the center of the STM image; it is surrounded with graphene located on the preceding nickel terrace. The rays emerging from the graphene island are one-dimensional residues of nickel islands, which pass eventually into separate points (nickel atoms). This Ni(111) surface structure is characteristic for TPG graphene, when heating is terminated before a continuous graphene layer is formed.
![STM image ($936\times936$[Å]{}$^2$, $I_{t}=0.2$ nA, $U_{s}=-1128$ mV) of the Ni(111) surface, almost completely covered with graphene. There is a two-dimensional nickel island at the center, with emerging rays in the form of one-dimensional nickel islands, transformed eventually into individual nickel atoms.[]{data-label="fig1"}](fig11.eps){width="\linewidth"}
To compare nickel and structural defects, Fig. \[fig2\]a presents an STM image of TPG graphene, which contains defects of both types. Nickel defects are brighter than the structural ones (see Fig. 2a). Note that these structural defects (Fig. \[fig2\]a) are very rare, because graphene grows epitaxially during TPG, and structural defects are scarce in it. Similar nickel defects are observed in CVD graphene (Fig. \[fig2\]b). The minimum distance between the defects is about 5[Å]{} (which makes two periods of Ni(111) lattice) for both graphene synthesis techniques.
![(a) STM image ($88\times88$[Å]{}$^2$, $I_{t}=1.3$nA, $U_{s}=-7$ mV) of the Gr/Ni(111) surface obtained by TPG. There are nickel defects on the left and a structural extended defect (indicated by arrow) on the right. (b) STM image ($88\times88$[Å]{}, $I_{t}=2.1$nA, $U_{s}=-2$mV) of Gr/Ni(111) surface obtained by CVD. Note that the nickel defects formed by (a) TPG and (b) CVD are similar.[]{data-label="fig2"}](fig2.eps){width="\linewidth"}
It was shown previously [@2017Kovalenko] that the bright points in STM images are nickel defects: nickel atoms occupying a monovacancy (Ni-MV) or a bivacancy (Ni-BV). However, we could not reveal the exact defect structure by comparing the STM images, because Ni atoms in mono- and bivacancies yield almost identical theoretical STM images. Now, we have investigated the defect transformation from Ni-MV to Ni-BV using DFT and established the defect type.
Figure \[fig3\] shows the minimum-energy path for a Ni atom from the monovacancy to the bivacancy. Among two possible models of a Ni-MV defect (carbon atom vacancy in the on-top or fcc position), we chose the model, in which the Ni atom replaces the C atom in the on-top position. This model is energetically more favorable (by 0.69 eV) than that where the Ni atom occupies a vacancy in the fcc position, because the Ni adatom in the on-top position forms a bond with the substrate Ni atom located below. The defect expansion in graphene from mono- to bivacancy occurs due to the detachment of the carbon atom neighboring the vacancy and its dissolution under the first substrate layer (in the octa position). The occurrence of reaction barrier is due to the energy cost for breaking the bond between specified carbon atom and two neighboring C atoms in graphene. This energy cost is partially compensated for by the following processes: (i) formation of a bond between the distant carbon atom and six substrate atoms, (ii) formation of an additional bond between the Ni adatom and carbon atom in graphene (the Ni adatom in the Ni-BV model is bound with four C atoms rather than with three, as in the Ni-MV model), and (iii) formation of a bond between the Ni adatom and three substrate atoms. As a result of the formation of new bonds, the Ni adatom is stabilized in a bivacancy. The energy of the final state (Ni-BV) is lower, even despite the fact that it is more favorable for carbon atom to be in graphene than to be dissolved under the first Ni(111) layer. The activation energy ($E_a$) of the Ni-MV $\rightarrow$ Ni-BV transition is 1.65 eV (see Fig. \[fig3\]).
![Energy diagram of the transition of a Ni adatom from a monovacancy to a bivacancy with carbon atom detachment from graphene and its dissolution under the first Ni(111) layer. Substrate Ni atoms, Ni atom in bivacancy, and C atoms are colored gray, blue, and red, respectively.[]{data-label="fig3"}](fig3.eps){width="\linewidth"}
The average time ($\tau$) of the Ni-MV $\rightarrow$ Ni-BV transition at a temperature $T$ can be estimated using the following formula from the transition-state theory [@2012Cui]: $$\tau = \frac{h}{k_B T} \exp\Biggl\{\frac{E_a}{k_B T}\Biggr\}, \label{eq:1}$$ where $h$ is the Planck’s constant, $k_B$ is the Boltzmann constant, and $E_a$ is the transition activation energy. When graphene synthesis is performed at a temperature of $500^\circ$C, the average time of the Ni-MV $\rightarrow$ Ni-BV transition is about 0.01 s. Therefore, if nickel atoms incorporate into graphene with occupation of monovacancies during the synthesis, the sample-heating time is sufficient for them to pass to a more stable state (in bivacancies).
Thus, the main type of nickel defects in Gr/Ni(111) is Ni-BV. Figure \[fig4\] shows the model of a Ni-BV defect (Fig. \[fig4\]a) and theoretical STM image (Fig. \[fig4\]b), which is in agreement with the experimental image (Fig. \[fig4\]c). The distance between the nearest Ni-BV defects cannot be smaller than two lattice periods of the Ni(111) face; otherwise, the defect structure is violated. This statement is in agreement with the experimental STM images, in which the minimum distance between defects equals to two nickel lattice periods (see Fig. \[fig2\]).
![(a) Structure of a defect with Ni atom in graphene bivacancy on Ni(111) (substrate Ni atoms, Ni atom in bivacancy, and C atoms are colored gray, blue, and red, respectively). (b) Theoretical STM image of the defect ($U_s = -50$mV). (c) Experimental STM image ($36\times15$[Å]{}$^2$, $I_{t}=1.4$nA, $U_{s}=-64$mV) of graphene with incorporated nickel atoms.[]{data-label="fig4"}](fig4.eps){width="\linewidth"}
According to the Bader charge analysis, a Ni adatom in a graphene bivacancy on the Ni(111) surface is charged positively with lack of 0.7 electron in comparison with neutral Ni atom. Some part of the Ni adatom charge is transferred to the nearest carbon atoms, thus forming a negatively charged region around the Ni adatom. The atoms of the upper Ni(111) layer are also charged positively, and their average charge is 0.1e. The surface Ni atoms transfer their charge to carbon atoms in graphene; therefore, the charge distribution of the Ni-BV defect in graphene formed on a substrate differs from that in freestanding graphene [@2019Mao].
We also considered models of two structures that can be formed when detaching graphene from the Ni(111) surface: (i) a Ni adatom remains on Ni(111) surface (Fig. \[fig6\]a) or (ii) a Ni-BV defect remains in freestanding graphene (Fig. \[fig6\]b). A comparison of the total energies of these two models shows that, if graphene is detached from the substrate, Ni atoms remain in graphene; this model is energetically more favorable (by 2.50 eV). Thus, the Ni adatom is bound with graphene stronger than with the Ni(111) surface atoms. Therefore, the Ni defects formed during graphene synthesis on Ni(111) can be retained when graphene is detached from the substrate, as was confirmed experimentally [@2015Qiu].
![Structural models of graphene at a distance of 10.6[Å]{} above the Ni(111) surface: (a) Ni adatom is on the Ni(111) surface in the fcc position and (b) Ni adatom occupies a bivacancy in graphene.[]{data-label="fig6"}](fig6.eps){width="0.85\linewidth"}
Conclusions
===========
The nickel defects formed in graphene synthesized on Ni(111) by CVD and TPG methods are identical. When analyzing the Ni(111) surface during TPG, we were able to obtain an STM image of intermediate structures and establish that nickel atoms remain in graphene (looking like bright points in the image). According to DFT, the Ni adatom in graphene is more stable in a carbon bivacancy rather than in a monovacancy. Our calculations show that the configuration with a Ni atom in monovacancy transforms into the configuration with a Ni atom in bivacancy due to the dissolution of one carbon atom into the nickel bulk over a period of about 0.01 s at a graphene synthesis temperature. We have found also that the Ni adatom is charged positively, and there is a negatively charged region around it. Such positively charged centers in graphene have the potential for single-atom catalysis. The Ni adatom in a bivacancy is more strongly bound with graphene rather than with the substrate; therefore, when graphene is detached, it remains in graphene bivacancy. Ni-BV centers on the Ni(111) surface may also be of interest. The presence of these centers in graphene on a substrate opens ways to study single-atom catalysts by STM and atomic force microscopy, observing in situ transformations of individual molecules.
Acknowledgements
================
This study was supported by the Russian Foundation for Basic Research, project no. 16-29-06426. We also thank the Joint Supercomputer Center of RAS for providing the computing power.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Parton evolution with the rapidity essentially is a branching diffusion process. We describe the fluctuations of the density of partons which affect the properties of QCD scattering amplitudes at moderately high energies. We arrive at different functional forms of the latter in the case of dipole-nucleus and dipole-dipole scattering.'
author:
- |
[*Stéphane Munier*]{}\
Centre de physique théorique, École polytechnique, CNRS, Palaiseau, France
title: On Parton Number Fluctuations
---
Quantum chromodynamics at high density
======================================
Quantum chromodynamics in the high-energy/high-density regime is a very rich field from a theoretical viewpoint since it involves genuinely nonlinear physics, and nontrivial fluctuations. The latter are deemed [*a goldmine for modern physics*]{} [@scapparone]. From a phenomenological viewpoint, there is a wealth of data from different experiments which await interpretation. (For a review, we refer the reader to the recent textbook by Kovchegov and Levin [@kovchegov2012quantum]).
Electron-proton (or better, nucleus) scattering is maybe the best experiment to probe QCD in this regime, as was done at HERA, an $e^\pm p$ facility. The electron interacts with the proton through a quark-antiquark pair, which appears as a quantum fluctuation of a (virtual) photon of the Weizsäcker-Williams field of the electron. The probability amplitudes for these fluctuations follow from a simple QED calculation. The $q\bar q$ pair is a color dipole, and hence electron-hadron scattering may be related to dipole-hadron scattering. If one looks at events in which the $q\bar q$ pair has a small-enough size (as compared to the typical size of a hadron), as is possible by selecting longitudinally-polarized highly-virtual photons, then perturbative QCD may be used as a starting point to compute some properties of the dipole-hadron scattering amplitudes.
As for the interaction of protons and/or nuclei as is currently performed at the LHC, the observables need to be carefully chosen if one wants to be able to predict cross sections from first principles – at least in the present state of the art of the theory. Indeed, one needs a hard momentum scale to justify the use of perturbation theory, and the latter must be found in the final state in the form of e.g. the transverse momentum of a jet. It turns out that an observable such as $p_\perp$-broadening in proton-nucleus collisions, namely the transverse momentum distribution of single jets, may also be related to the dipole-nucleus amplitude.
We will first review the formulation of the rapidity evolution of the dipole-nucleus scattering amplitude in QCD in the high-energy limit. The latter is given by the Balitsky-Kovchegov (BK) equation. We will relate the known shape of its solution to gluon-number fluctuations in the quantum evolution of the dipole. We will then be able to predict the form of geometric scaling for dipole-dipole scattering, which turns out to be different from the solution to the BK equation.
Dipole-nucleus scattering
=========================
Let us start with the scattering of a dipole off a nucleus at relatively low energy. The forward elastic amplitude $T$ is a function of the dipole size $r_0$, which is given by the McLerran-Venugopalan model: $$T(r_0)=1-e^{-\frac{r_0^2 Q_A^2}{4}}.
\label{eq:MV}$$ This formula resums multiple exchanges of pairs of gluons between the bare dipole and the nucleus (see Fig. \[Fig:dipolenucleus\]a). $Q_A$ is the saturation momentum of the nucleus. Equation (\[eq:MV\]) essentially means that a dipole of size larger than $\sim 1/Q_A$ is absorbed ($T\sim 1$), while the nucleus is transparent to dipoles of size smaller than $\sim 1/Q_A$. For our purpose, we may approximate $T(r_0)$ by the step function $\Theta(\ln r_0^2 Q_A^2/4)$.
[m[0.2]{} | m[0.2]{} m[1cm]{} m[0.2]{} ]{} ![Particular graphs contributing to the dipole-nucleus scattering amplitude at low (a) and high (b) energy in the restframe of the nucleus. \[Fig:dipolenucleus\] ](munier_stephane_fig1a.eps "fig:"){width="20.00000%"} & ![Particular graphs contributing to the dipole-nucleus scattering amplitude at low (a) and high (b) energy in the restframe of the nucleus. \[Fig:dipolenucleus\] ](munier_stephane_fig1b1.eps "fig:"){width="20.00000%"} &$\underset{\begin{minipage}[c]{1.2cm}{\footnotesize large $N_c$}\end{minipage}}{\huge \simeq}$ &![Particular graphs contributing to the dipole-nucleus scattering amplitude at low (a) and high (b) energy in the restframe of the nucleus. \[Fig:dipolenucleus\] ](munier_stephane_fig1b2.eps "fig:"){width="20.00000%"}\
\
& &
Going to higher energies $\sqrt{s}$ by increasing the rapidity of the dipole, the scattering process gets dominated by high-occupancy quantum fluctuations of the initial dipole (see Fig. \[Fig:dipolenucleus\]b). The rapidity ($y\equiv \ln s $) dependence of the amplitude is given by the Balitsky-Kovchegov (BK) equation $$\partial_y T(r_0,y)=\bar\alpha\int \frac{d^2 r_1}{2\pi}\frac{r_0^2}{r_1^2(r_0-r_1)^2}
\left[T(r_1,y)+T(r_0-r_1,y)-T(r_0,y)
-T(r_1,y)T(r_0-r_1,y)\right]
\label{eq:BK}$$ ($\bar\alpha\equiv\frac{\alpha_s N_c}{\pi}$), whose large-$y$ solutions are traveling waves, namely fronts which translate (almost) unchanged in shape towards negative values of the $\ln r_0^2$ variable as the rapidity increases. The linear part of this equation (the first three terms in the r.h.s.) form the BFKL equation, whose kernel possesses as eigenfunctions the power functions $|r_0|^{2\gamma}$, the corresponding eigenvalues being $\bar\alpha\chi(\gamma)$, where $\chi(\gamma)\equiv 2\psi(1)-\psi(\gamma)-\psi(1-\gamma)$. Introducing the particular eigenvalue $\chi(\gamma_0)$, where $\gamma_0$ is such that $\chi^\prime(\gamma_0)=\chi(\gamma_0)/\gamma_0$, the shape of $T$ as a function of the dipole size $r_0$ in the region $T\ll 1$ and the $y$-dependence of the saturation scale read $$T(r_0,y)\underset{r_0\ll 1/Q_s(y)}{\sim}
\ln \frac{1}{r_0^2 Q_s^2(y)} e^{\gamma_0 \ln ({r_0^2 Q_s^2(y)})}\ \ \text{and}
\ \
Q_s^2(y)/Q_A^2\simeq e^{\bar\alpha\chi^\prime(\gamma_0)y}.
\label{eq:TdA}$$
The BK equation (\[eq:BK\]) can be established in the framework of the dipole model (see e.g. [@kovchegov2012quantum]), where gluons are replaced by zero-size $q\bar q$ pairs. In this model, the Fock state of the incoming dipole which is “seen” by the nucleus at the time of the interaction is built from successive independent splittings of dipoles. At a given rapidity $y$, the latter Fock state can be thought of as a collection of $n$ dipoles, generated by a splitting process which belongs to a class of processes generically called [*branching diffusion*]{}.
The main point we wanted to make at this conference and in Ref. [@Mueller:2014fba] was that $T$ has an elegant and useful probabilistic interpretation in the dipole picture: [*It represents the probability that the largest dipole present in the Fock state of the incoming $q\bar q$ pair at the time of the interaction has a size which is larger than the inverse nuclear saturation momentum, $1/Q_A$.*]{} Indeed, according to the McLerran-Venugopalan model, a given dipole interacts with the nucleus only if its size is larger than $1/Q_A$, hence it is necessary and sufficient that at least one of the dipoles in the Fock state be larger than $1/Q_A$ for the scattering to take place. Thus solving the BK equation amounts to understanding the statistics of the extremal particles in a branching random walk (BRW). Our first task is to recover the shape of the amplitude (\[eq:TdA\]), previously obtained through a analysis of the BK equation, from the latter statistics.
We observe that the extremal particle in a BRW has fluctuations which can originate only from two places: From the first stages of the rapidity evolution, when the overall number of dipoles is small and thus subject to large statistical fluctuations (we shall call this type of stochasticity “front fluctuations”), and from the tip of the distribution, where by definition, particle numbers keep small. Elsewhere, the evolution is essentially deterministic since it acts on a large number of objects.
[m[0.45]{} m[0.55]{}]{} ![\[Fig:d1\] (a) Sketch of the dipole distribution (as seen at a given impact parameter) in a particular realization of the evolution to the rapidity $y$ of an initial dipole of size $r_0$. (b) Model for the evolution of a given realization, with the “front” and “tip” fluctuations described in the text. ](munier_stephane_fig2a.eps "fig:"){width="40.00000%"} & ![\[Fig:d1\] (a) Sketch of the dipole distribution (as seen at a given impact parameter) in a particular realization of the evolution to the rapidity $y$ of an initial dipole of size $r_0$. (b) Model for the evolution of a given realization, with the “front” and “tip” fluctuations described in the text. ](munier_stephane_fig2b.eps "fig:"){width="50.00000%"}\
&
The effect of the front fluctuations is to shift the particle distribution by $\Delta$. We conjecture[^1] that the distribution of $\Delta$ is $\propto e^{-\gamma_0\Delta}$. The effect of the tip fluctuations is instead to send randomly particles ahead of the front by $\delta$. We conjecture the same exponential law $\propto e^{-\gamma_0\delta}$.
We introduce our notations in Fig. \[Fig:d1\]. According to the previous discussion, in a particular event, the scattering occurs if $x_0+\bar X(y)+\Delta+\delta\geq 0$. Hence the amplitude $T$ simply is the average of this condition over $\Delta$ and $\delta$: $$T\propto \int_{0}^{+\infty}
d\delta\,e^{-\gamma_0\delta}
\int_{0}^{+\infty}
d\Delta\,e^{-\gamma_0\Delta}
\Theta(x_0+\bar X(y)+\Delta+\delta)\propto(-x_0-\bar X(y))e^{\gamma_0(x_0+\bar X(y))}.$$ Switching back to the QCD variables, we recover the expression of $T$ given in Eq. (\[eq:TdA\]). We conclude that [*the shape of the dipole-nucleus scattering amplitude as a function of the dipole size is directly related to the event-by-event fluctuations of the size of the largest dipole*]{}, which in turn stem from the fluctuations of the numbers of gluons produced in the QCD evolution.
Dipole-dipole scattering
========================
While the dipole-nucleus amplitude probes the statistics of the largest dipole in the quantum evolution, the physics of dipole-dipole scattering is a bit different: Indeed, since the elementary amplitude (for dipoles of respective sizes $r_0$ and $R_0$) at zero rapidity is essentially $T(r_0,R_0)\sim \alpha_s^2\delta(\ln r_0^2/R_0^2)$, it is the very shape of the dipole number distribution that is actually probed (Fig. \[Fig:dd\]). So in order to compute the shape of the amplitude, we need on one hand the probability distribution of the front fluctuations used before, and on the other hand the shape of the dipole number density from the deterministic evolution. We also need to implement [*saturation*]{} in the evolution (see Fig. \[Fig:dd\]c) to comply with the unitarity constraint $T\leq 1$. All in all, we obtain $$T(r_0,y)\underset{r_0\ll 1/Q_s(y)}{\sim}
\ln^2 \frac{1}{r_0^2 Q_s^2(y)} e^{\gamma_0 \ln ({r_0^2 Q_s^2(y)})}
\ \ \text{where}
\ \
Q_s^2(y)R_0^2\simeq e^{\bar\alpha\chi^\prime(\gamma_0)y}.
\label{eq:Tdd}$$ Interestingly enough, it differs from the dipole-nucleus case; compare Eq. (\[eq:TdA\]) to Eq. (\[eq:Tdd\]). This is the main prediction of the way of looking at QCD evolution we have promoted at this conference and in Ref. [@Mueller:2014fba].
[m[0.22]{} m[0.35]{} m[0.35]{}]{} ![\[Fig:dd\] (a) Graph contributing to dipole-dipole scattering at high energies. (b) Sketch of the evolution of the dipole number density, model including fluctuations. (c) The same, but with saturation. ](munier_stephane_fig3a.eps "fig:"){width="20.00000%"} & ![\[Fig:dd\] (a) Graph contributing to dipole-dipole scattering at high energies. (b) Sketch of the evolution of the dipole number density, model including fluctuations. (c) The same, but with saturation. ](munier_stephane_fig3b.eps "fig:"){width="32.00000%"} & ![\[Fig:dd\] (a) Graph contributing to dipole-dipole scattering at high energies. (b) Sketch of the evolution of the dipole number density, model including fluctuations. (c) The same, but with saturation. ](munier_stephane_fig3c.eps "fig:"){width="32.00000%"}\
& &
We refer the reader to [@Mueller:2014fba] for the details, references, and more results, in particular on the finite-$y$ corrections to the saturation scale in both the dipole-dipole and dipole-nucleus cases.
[99]{} E. Scapparone, plenary talk at this conference.
Y.V. Kovchegov and E. Levin. . Cambridge Monographs on Particle Physics, Nuclear Physics and Cosmology. Cambridge University Press, 2012.
A. H. Mueller and S. Munier, Phys. Lett. B [**737**]{} (2014) 303. A. H. Mueller and S. Munier, “Phenomenological picture of fluctuations in branching random walks,” arXiv:1404.5500 \[cond-mat.dis-nn\] to appear in Phys. Rev. E (2014).
[^1]: Arguments in favor of this conjecture were presented in Ref. [@Mueller:2014gpa].
| {
"pile_set_name": "ArXiv"
} |
UDC 512.542
**On Thompson’s conjecture for alternating and symmetric groups**
I.B. Gorshkov [^1]
**1. Introduction**
In this paper, all groups are finite. Let $G$ be a finite group. If $g\in G$ that $g^G$ denotes the conjugacy class in $G$ containing $g$ and $|g^G|$ its size, $C_G(g)$ denotes the centralizer of $g$ in $G$. We use $N(G)$ for the set $\{n| \exists g\in G$ such that $|g^G|=n\}$. Thompson in 1987 posted the following conjecture with respect to $N(G)$.
**Thompson’s Conjecture (see [@Kour], Question 12.38)**. [*If $L$ is a finite simple non Abelian group, $G$ is a finite group with trivial center, and $N(G)=N(L)$, then $G\simeq L$.*]{}
Thompson’s conjecture was proved valid for many finite simple groups of Lie type. Let $Alt_n$ be the alternating group of degree $n$, $Sym_n$ be the symmetric group of degree $n$. Alavi and Daneshkhah proved that the groups $Alt_n$ with $n=p$, $n=p+1$, $n=p+2$ and $p$ prime are characterized by $N(G)$ (see [@AD]). Recently, the groups $Alt_{10}, Alt_{16}$, and $Alt_{22}$ were proved valid for this conjecture (see [@VasT], [@Gor], [@Xu]).
Our main result is the following.
If $G$ is a finite group such that $N(G)=N(Alt_n), n\geq5$ or $N(G)=N(Sym_n)$, $n\geq23$, then $G$ is non-solvable.
**2. Notation and preliminary results**
Given a finite group $G$, denote by $\pi(G)$ the set of prime divisors of its order.
\[pi\] Let $G$ and $H$ are finite groups, center of $G$ is trivial and $N(G) = N(H)$. Then $\pi(G)\subseteq \pi(H)$.
*Proof.* In the same way as [@VasT Lemma 3].
\[factor\] Let $K$ be a normal subgroup of a finite group $G$, and $\overline{G}=G/K$. If $\overline{x}$ is the image of an element $x$ of $G$ to $\overline{G}$, them $|\overline{x}^{\overline{G}}|$ divides $|x^G|$. Moreover, if $(|x|,|K|)=1$, then $C_{\overline{G}}(\overline{x})=C_G(x)K/K$.
\[centr\] Let $x,y\in G, (|x|,|y|)=1$, $xy=yx$. Then $C_G(xy)=C_G(x)\cap
C_G(y)$.
*Proof.* The proof is trivial.
\[Fitting\] Let $A$ act via automorphisms on an Abelian group $G$, and suppose that $(|G|,|A|)=1$. Then $G=C_G(A)\times[G,A]$.
\[Frobenius\] Let $g$ act via automorphisms on an Abelian group $G$, and suppose that $(|G|,|g|)=1$. Then $|g|$ divides $|[G,g]|-1$.
\[vas\] If $G$ is finite group and there exists a prime $p\in \pi(G)$ such that $p^2$ does not divide $|x^G|$ for all $x$ in $G$. Then a Sylow $p$-subgroup of $G$ is Abelian.
*Proof.* In the same way as [@VasT Lemma 4].
Let $\Theta\subseteq \mathbb{N}, |\Theta|<\infty$, $\Gamma(\Theta)$ is directed graph, the set vertices is equal to $\Theta$ and $\overrightarrow{ab}$ if $a$ divides $b$, $h(\Theta)$ is length of a maximal path in the graph $\Gamma(\Theta)$. Let $\Gamma(G)=\Gamma(N(G))$, $V_i\in\{Alt_i, Sym_i\}$, $G$ be a finite solvable group such that $N(G)=N(V_n)$ where $n\geq5$ if $V_n\simeq Alt_n$ and $n\geq 23$ if $V_n\simeq Sym_n$, $\Omega=\{t| n/2<t\leq n, t -$ prime $\}$, $p$ be maximal number of $\Omega$. From Lemma \[pi\] we obtain that $\pi(G)\supseteq\pi(V_n)$ in particular $\Omega \subseteq \pi(G)$.
\[omega\] If $n>1361$, then $|\Omega|>\log_2(n!/(p)!)$,
*Proof.* Let $\phi (x)$ be the number of primes less or equal $x$, From [@Buh Section 35, §1] it follows that $0,921\cdot (x/\ln(x))<
\phi(x) < 1,106\cdot (x/\ln(x))$ for $10<x$. Hence $|\Omega|=\phi(n)-\phi(n/2)\geq \frac{0,921\cdot n}{\ln(n)}- \frac{1,106\cdot
n/2}{\ln(n/2)}$. It is proved in [@Baker] that $n-p<n^{0,525}$. Using this assertions we get that lemma is valid if $n\geq1000000$. If $1361\leq n<1000000$ then using [@GAP] we obtain our assertion. The lemma is proved.
\[ConClass\] Let $t\in \Omega$, $\alpha\in N(G)$ and $t\not \in \pi(\alpha)$. Then $\alpha=|V_n|/(t|C|)$ or $\alpha=|V_n|/(|V_{t+i}||B|)$ where $C=C_{V_{n-t}}(g)$, for any $g\in V_{n-t}$, $t+i\leq n$, $B=C_{V_{n-t-i}}(h), h\in V_{n-t-i}$ and $h$ moves $n-t-i$ points.
*Proof.* The proof is trivial.
Let $\Phi_t=\{\alpha\in N(G)| \alpha=|V_n|/(t|C|)$, $C=C_{V_{n-t}}(g)$, for any $g\in V_{n-t}\}$, $\Psi_t=\{\alpha\in N(L)| \alpha=|V_n|/(|V_{t+i}||B|$, $i\geq0, t+i<n-1$, $B=C_{V_{n-t-i}(g)}, g\in V_{n-t-i}$ and $g$ moves $n-t-i$ points $\}$. Let us remark that the definitions of sets $\Phi$ and $\Psi$ do not imply that $t$ is prime.
\[hz\]
1. [If $n-t=2$, then $h(\Psi_t)\leq 1$ ]{}
2. [If $n-t=3$, then $h(\Psi_t)\leq 2$ ]{}
3. [If $n-t=4$, then $h(\Psi_t)\leq 3$ ]{}
4. [If $n-t=5$, then $h(\Psi_t)\leq 5$ ]{}
5. [If $n-t=6$, then $h(\Psi_t)\leq 6$ ]{}
6. [If $n-t=7$, then $h(\Psi_t)\leq 8$ ]{}
7. [If $n-t=8$, then $h(\Psi_t)\leq 11$ ]{}
8. [If $n-t=9$, then $h(\Psi_t)\leq 14$ ]{}
9. [If $n-t=10$, then $h(\Psi_t)\leq 18$ ]{}
10. [If $n-t=11$, then $h(\Psi_t)\leq 21$ ]{}
11. [If $n-t=12$, then $h(\Psi_t)\leq 26$ ]{}
12. [If $n-t=13$, then $h(\Psi_t)\leq 30$ ]{}
13. [If $n-t=18$, then $h(\Psi_t)\leq 69$ ]{}
*Proof.* By definition we have $h(\Psi_t)\leq \sum_{i\in\{1,..,n-t\}}h(R_i)$ where $R_i=\{|g^{V_i}|, g\in V_i,$ and $g$ moved $i$ point $\}$. Let us use [@GAP] for calculation $R_i$. The lemma is proved.
\[abelian\] If $G$ is solvable group, then Hall $\Omega$-subgroup is an Abelian.
*Proof.* It follows from Lemma \[vas\] that Sylow $t$-subgroup of $G$ is Abelian for any $t\in \Omega$. Assume that Hall $\Omega$-subgroup of $G$ is not Abelian. It follows that there exists $\{t_i, t_j\}\subset\Omega$ such that a Hall $\{t_i,t_j\}$-subgroup $H$ of $G$ is non Abelian. Let $R<G$ be a maximal normal subgroup such that $\overline{H}=HR/R$ is non Abelian. Hence in group $\overline{G}=G/R$ there exist normal $t$-subgroup $T$, $t\in\{t_1, t_2\}$, such that there exist $g\in \overline{G}, |g|\in\{t_1,t_2\}\setminus \{t\}$ and $[T,g]\neq 1$. By Lemma \[Fitting\], $T=C_T(g)\times[T,g]$. Using Lemma \[Frobenius\] we get that $|g|$ divides $[T,g]-1$. But $|g|$ not divides $t-1$, therefore $|[T,g]|>t$. From Lemma \[factor\] it follows that $|g^{\overline{G}}|_t>t$ and there exists $n\in N(G)$ such that $t^2$ divides $n$; a contradiction. The lemma is proved.
\[abelian2\] Let $G$ by solvable. If there exists $r$ prime such that $p+1<2r\leq n$ then a Hall $\{r\}\cup\Omega$-subgroup is isomorphic direct product of $T\times R$ where $T$ is a Hall $\Omega$-subgroup and $R$ is a Sylow $r$-subgroup. In particular, if $g\in T$ then $|g^G|$ is not divisible by $r$.
*Proof.* The proof is analogous to that of Lemma \[abelian\].
\[solvable\] Let $T$ be a Hall $\Omega$-subgroup of $G$, $\Theta=\{|g^G|$ for all $g\in T\}$. If $|\Omega|>h(\Theta)$, then $G$ are non-solvable.
*Proof.* Assume that $G$ is solvable and $|\Omega|>h(\Theta)$. From Lemma \[abelian\] it follows that a Hall $\Omega$-subgroup of $G$ is Abelian. Let $g_1\in T, |g_1|=t_1, t_1\in \Omega$ is minimal of $\Omega$. It follows that there exist $r_1\in C_G(g_1)$ such that intersection $C_G(r_1)\cap T$ is Hall $\Omega$-subgroup of $C_G(r_1)$, and $t_2$ divides $|r_1^G|$ where $t_2$ is minimal of $\Omega\setminus\{t_1\}$. Hence there exist $g_2\in T, |g_2|=t_2$ such that $C_G(g_1)\neq C_G(g_2)$. Since $T$ is Abelian from Lemma \[centr\] it follows that $|(g_1g_2)^G|>|g_1^G|$ and $|g_1^G|$ divides $|(g_1g_2)^G|$. Repeating this procedure $|\Omega|$ times we obtain a set $\Sigma=\{g_1, g_1g_2, g_1g_2g_3,..., g_1g_2...g_{|\Omega|}\}$ and $|g_1^G| | |(g_1g_2)^G| |...||g_1g_2...g_{|\Omega|}|$. Thus $h(\Theta)\geq |\Omega|$; a contradiction. The lemma is proved.
**3. Proof of the main theorem**
If $n>1361$ then $G$ is non-solvable.
*Proof.* Assume that $G$ is solvable group. Let $T$ be a Hall $\Omega$-subgroup of $G$. Using Lemma \[ConClass\] we get that $|g^G|\in\Psi_p$ for any $g\in T$. Let us show that $|\Omega|>h(\Psi_p)$. Let $h_1,...,h_k\in \Psi_p$ such that $h_1|h_2|...|h_k$. Therefore $2h_1\leq h_2, 2h_2\leq h_3...2h_{k-1}\leq h_k$. If $l\in\Psi_p$ then $l|n!/(p)!$. Hence $h(\Psi_p)\leq log_2(n!/(p)!)$. Using Lemma \[omega\] we get $|\Omega|>log_2(n!/(p)!\geq h(\Psi_p)$. From Lemma \[solvable\] if follow that $G$ is non-solvable; a contradiction. The lemma is proved.
If $22<n\leq1361$ then $G$ is non-solvable.
*Proof.* Assume that $G$ is solvable group. Let $T$ be a Hall $\Omega$-subgroup of $G$. Using Lemma \[ConClass\] we get that $|g^G|\in\Psi_p$ for any $g\in T$. From Lemma \[abelian2\] it follows that $|g^G|$ is not divisible by $r$ where $g\in T$ and $p+1<2r\leq n$, $r$ is prime. Hence $|g^G|=|V_n|/(V_{2r+i}|C|)\in\Psi_{2r}$ where $0\leq i \leq n-2r+2$, and $C=C_{V_{n-2r-i}(g)}, g\in V_{n-t-i}, 0\leq i\leq n-2r+2$. Let $\Theta=\{|g^G|$ for all $g\in T\}$. Combining Lemma \[hz\], [@GAP], and Lemma \[solvable\], we get a contradiction. The lemma is proved.
If $5\leq n\leq22$ then $G$ is non-solvable.
*Proof.* The assertion follows from [@AD], [@VasT] and [@Gor].
[1]{} V. D. Mazurov, E. I. Khukhro, Eds., The Kourovka Notebook: Unsolved Problems in Group Theory, Russian Academy of Sciences Siberian Division, Institute of Mathematics, Novosibirsk, Russia, 18th edition, 2014.
S. H. Alavi and A.Daneshkhah, A new characterization of alternating and symmetric groups, Journal of Applied Mathematics and Computing, 17(1), pp. 245–258, 2005.
A. V. Vasil’ev, On Thompson’s conjecture, Sibirskie Elektronnye Matematicheskie Izvestiya, vol. 6, pp. 457–464, 2009.
I. B. Gorshkov, Thompson’s conjecture for simple groups with a connected prime graph, Algebra and Logic, 51(2), pp. 111–127, 2012. M. Xu, Thompson’s conjecture for alternating group of degree 22, Frontiers of Mathematics in China, 8(5), pp. 1227– 1236, 2013.
I. M. Isaacs, Finite group theory, Graduate Studies in Mathematics, 92. American Mathematical Society, Providence, RI, 2008. xii+350 pp.
A. A. Buhshtab, Number theory // Moskow. 1966
R. C. Baker, The difference between consecutive primes, II. Proceeding of the London Mathematical Society 83(3), pp. 532-562, 2001.
The GAP Group, GAP — Groups, Algorithms, and Programming, Version 4.4, 2004; (http://www.gap-system.org).
I.B. Gorshkov\
Sobolev Institute of Mathematics,\
4 Acad. Koptyug avenue, 630090, Novosibirsk, Russia\
Universidade de Sao Paulo, Brazil\
E-mail address: [email protected]
[^1]: The work is supported by Russian Science Foundation (project 14-21-00065)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider a new setting of online clustering of contextual cascading bandits, an online learning problem where the underlying cluster structure over users is unknown and needs to be learned from a random prefix feedback. More precisely, a learning agent recommends an ordered list of items to a user, who checks the list and stops at the first satisfactory item, if any. We propose an algorithm of *CLUB-cascade* for this setting and prove a $T$-step regret bound of order $\tilde{O}(\sqrt{T})$. Previous work [@li2016contextual] corresponds to the degenerate case of only one cluster, and our general regret bound in this special case also significantly improves theirs. We conduct experiments on both synthetic and real data, and demonstrate the effectiveness of our algorithm and the advantage of incorporating online clustering method.'
author:
- |
Shuai Li\
The Chinese University of Hong Kong\
[email protected]
bibliography:
- 'CCB.bib'
title: Online Clustering of Contextual Cascading Bandits
---
6. Conclusions
==============
In this paper, we bring up a new problem of online clustering of contextual cascading bandits, where the algorithm has to explore the unknown cluster structure on users under a prefix feedback of the recommended item list. We propose a CLUB-cascade algorithm based on the principle of optimism in face of uncertainty and prove a cumulative regret bound, whose degenerate case improves the existing results. The experiments conducted on both synthetic and real datasets demonstrate the advantage of incorporating online clustering.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'W. Bednarek & J. Pabich'
date: 'Received ; accepted '
subtitle: 'Application to Her X-1 and Sco X-1'
title: 'Optical depths for gamma-rays in the radiation field of a star heated by external X-ray source in LMXBs'
---
Introduction
============
Recently TeV $\gamma$-rays have been observed from a few massive X-ray binaries (e.g. Aharonian et al. 2005a, Albert et al. 2006). Up to now such $\gamma$-ray emission has not been reported from the low mass X-ray binaries (see e.g. observations of SS 433, Aharonian et al. 2005b), which are very luminous X-ray sources ($L_{\rm X}\sim 10^{36-39}$ erg s$^{-1}$, see the catalog of LMXB Liu et al. 2007). However, the class of LMXBs as a whole has not been monitored intensively by the modern Cherenkov telescopes up to now. In this paper we try to argue that very compact LMXBs are also promissing targets for TeV $\gamma$-ray astronomy.
LMXBs likely contain a neutron star which accretes matter from a companion star. Sometimes X-ray emission shows also pulsations with the rotational period of the neutron star (Nagase 1989, Karino 2007). Some LMXBs show also optical modulation with the period of the binary system which can be interpreted as a result of external heating of the stellar surface by the X-ray source (e.g. Her X-1, Sco X-1).
The mechanism for particle acceleration and $\gamma$-ray production in the massive TeV $\gamma$-ray binaries is not at present clear. Two general scenarios are considered. In the first one, energetic pulsar creates relativistic wind which collides with the strong wind of the massive star creating a shock wave (e.g. Maraschi & Treves 1981, Vestrand & Eichler 1982, Harding & Gaisser 1990, Tavani & Arons 1997, Kirk et al. 1999). Electrons can be accelerated up to the TeV energies in the shock acceleration scenario or in the not well known mechanism occurring in the pulsar wind. In the second scenario, a compact object accretes matter from the massive star. Electrons can be accelerated in the shock waves created in the stream of matter expelled from the inner part of the accretion disk (a jet) (e.g. Levinson & Blandford 1996, Georganopoulos et al. 2002, Romero et al. 2002) or at the border in which accreting matter is stopped by the rotating neutron star magnetosphere (Bednarek 2009a,b). The second scenario can also occur in the case of low mass X-ray binaries. Therefore, it is possible that acceleration of particles to TeV energies should also occur within the low mass X-ray binaries. Note that TeV $\gamma$-ray emission is observed from variety of cosmic sources (jets of active galaxies, vicinity of pulsars, supernova remnants, clusters of stars) which significantly differ in structure and physical parameters. So then, even if parameters of objects responsible for acceleration of particles within HMXBs and LMXBs may differ significantly, this does not exclude acceleration of particles to TeV energies in LMXBs. However the efficiency of particle acceleration may differ. The main purpose of searching for $\gamma$-ray emission from LMXBs with the present and future Cherenkov telescopes (e.g. CTA) will be to put light on this basic question. The aim of this paper is to show what are the most favourable conditions for detection of TeV $\gamma$-ray signal from the class of very compact LMXBs.
Efficient $\gamma$-ray production within massive binaries likely occurs due to the comptonization by relativistic electrons of the soft radiation produced by the massive star. Such strong radiation is not present inside the low mass binaries where a normal star has characteristic surface temperature below $\sim 10^4$ K. Therefore, in general, efficient TeV $\gamma$-ray production is not expected there. However, in very compact LMXBs, a low mass star can be extensively heated by the X-ray source produced due to accretion of matter onto companion compact object. As a result, some parts of the stellar surface can reach temperature comparable to those observed in the case of high mass stars (i.e. a few $10^4$ K). This soft radiation provides enough target for relativistic electrons, accelerated inside such compact low mass binaries, for production of TeV $\gamma$-rays. In this paper we calculate the optical depths for $\gamma$-ray photons in the radiation field of irradiated low mass star in order to check whether $\gamma$-ray production in such systems can occur efficient and whether produced $\gamma$-rays can escape from the vicinity of the star. Note, that the cross sections for $\gamma$-ray production in the inverse Compton scattering (ICS) and $\gamma$-ray absorption in collision with soft photons are comparable. Therefore, directions within the binary system in which significant absorption of $\gamma$-rays can occur should also correspond to directions of efficient $\gamma$-ray production in ICS process.
9.5truecm
Soft radiation from irradiated star
===================================
The X-ray emission produced in the vicinity of a compact source in some close low mass X-ray binaries can reach the values of $\sim 10^{36-39}$ erg s$^{-1}$. A part of this X-ray emission can illuminate the surface of a companion star. As a result, the surface temperature rises significantly. We calculate the temperature profile onto the surface of such irradiated star by a point X-ray source located at the distance, $H$, from the center of the star. We assume that all X-ray emission absorbed by the stellar surface is irradiated as a black body emission with characteristic temperature $T(z)$, where $z$ is the distance measured from the direction defined by the center of the star and the X-ray source and a specific point onto the stellar surface (see Fig. 1). Simple considerations turn out to the temperature profile on the stellar surface of the type, $$\begin{aligned}
T(z) = \left(T_\star^4 + {{L_{\rm X}\cos\beta/4\pi \sigma_{\rm SB}R_\odot^2}\over{z^2 +
\left(H - \sqrt{R_\star^2 - z^2}\right)^2}}\right)^{1/4}\nonumber\\
\approx
% \left(1.3\times 10^{15} + {{1.4\times 10^{41}(L_{38}\cos\beta/R_\odot^2)}\over{z^2 +
%(H -\sqrt{R_\star^2 - z^2})^2}}\right)^{1/4} {\rm K},
7.3\times 10^4 \left({{L_{38}\cos\beta}\over{z^2 + \left(H -\sqrt{R_\star^2 - z^2}\right)^2}}\right)^{1/4} {\rm K},
\label{eq1}\end{aligned}$$ where distances are defined in Fig. 1, $L_{\rm X} = 10^{38}L_{38}$ erg s$^{-1}$ is the luminosity of the X-ray source, $\sigma_{\rm SB}$ is the Stefan-Boltzmann constant, $\beta = \pi - \mu - \nu$, $\cos\mu = z/R_\star$, $\tan\nu = \left(H - \sqrt{R_\star^2 - z^2}\right)/z$, and $R_\star$ is the stellar radius. The distances $H$, $R$, $R_\star$, and $z$ are expressed in units of the Solar radius $R_\odot = 7\times 10^{10}$ cm. The value of $z$ can change in the range from 0 to $z_{\rm max} = R_\star\sqrt{H^2 - R_\star^2}/H$.
14.3truecm
The distribution of temperature on the surface of such irradiated star seen by the external observer located at different places outside the star is shown in Fig. 2. Two effects determine this soft radiation field at a specific place: (a) the distance, $R$, from the stellar surface determines the solid angle intercepted by the star; (b) the angle, $\alpha$, at which the hot region is seen (measured in respect to the direction defined by the X-ray source and the star). Note, that for reasonable parameters of the low mass X-ray binary system (e.g. $L_{\rm X} = 10^{38}$ erg s$^{-1}$, and $H = 2 R_\star$), the surface temperature can increase due to irradiation of the star up to a few $10^4$ K. This is comparable to the surface temperature of the massive stars within the high mass X-ray binaries. Therefore, we expect that the optical depths for TeV $\gamma$-rays in the soft radiation of such irradiated low mass stars can also reach substantial values, exceeding unity in some cases.
Since the $\gamma$-ray photon injected at a specific place sees the hot part of the star at different angles, the soft radiation field seen by the $\gamma$-ray photon changes significantly during its propagation in the vicinity of the star (due to the change of the viewing angle and the distance from the stellar surface). In Fig. 2 we show the example radiation field from the stellar surface as seen by the $\gamma$-ray photon at different propagation distances $L$ assuming that the primary $\gamma$-ray is injected at the X-ray source. Note that, for the propagation angles $\alpha_\gamma > 90^{\rm o}$ (see Fig. 1), the solid angle subtracted by the star at first increases since the $\gamma$-ray approaches the stellar surface. For small angles $\alpha_\gamma$, the solid angle subtracted by the star continuously drops with the $\gamma$-ray propagation path and the hot region becomes less visible.
The optical depths for gamma-rays
=================================
The optical depths for $\gamma$-ray photons injected at an arbitrary distance from the surface of the star in the case of its fixed surface temperature were calculated for the first time in the general case by Bednarek (1997,2000). In respect to previous calculations (e.g. Moskalenko et al. 1993), Bednarek took also into account dimensions of the star, which allows their application to very compact binaries such as e.g. Cyg X-3. Such calculations are necessary in the case of the IC $e^\pm$ pair cascades developing within the massive binaries since in principle secondary $e^\pm$ pairs and $\gamma$-rays can appear everywhere within the binary. They can also fall onto the surface of the companion star. The optical depths for $\gamma$-rays calculated in the radiation field of the star with specific parameters can be easily re-scaled for the cases of massive stars with other surface temperatures and radii (Bednarek 2009a). For example, $\gamma$-ray photons with energies, $E_\gamma^{\rm o}$, propagating at specific distance $D$ from the star, and in direction (defined by the angle $\alpha$), approaching close to the star with specific parameters ($T_{\rm o}$ and $R_{\rm o}$) are related to the optical depths around arbitrary stars with $T_\star$ and $R_\star$ in the following way, $$\begin{aligned}
\tau({{E_\gamma^{\rm o}}\over{S_{\rm T}}},T_\star,R_\star,D,\alpha)
= S_{\rm T}^3S_{\rm R}\tau(E_\gamma^{\rm o}, T_{\rm o}, R_{\rm o},D,\alpha)
\label{eq1}\end{aligned}$$ where $S_{\rm T} = T_\star/T_{\rm o}$, $S_{\rm R} = R_\star/R_{\rm o}$, and the distance $D$ is measured in stellar radii. These $\gamma$-ray optical depths have been frequently discussed in the context of the massive stars inside the high mass X-ray binaries recently detected in the TeV $\gamma$-rays (see e.g. Böttcher & Dermer 2005, Bednarek 2006a, Dubus 2006).
12.2truecm
Here we are interested in a more complicated scenario which can be appropriate for the low mass X-ray binary systems. In LMXBs the companion stars produce relatively weak soft radiation field due to its low surface temperature resulting from nuclear burning. However, in the presence of a nearby strong X-ray source, the surface temperature of a star can significantly increase due to the irradiation process (see section 2). Therefore, relativistic electrons injected not far from the surface of the low mass star in a compact X-ray binary can also suffer strong energy losses on the ICS process. Primary $\gamma$-rays and secondary cascade $\gamma$-rays can be efficiently absorbed in this soft radiation field. In order to check whether such processes may become important, we calculate the optical depths for $\gamma$-rays in such more complicated scenario following the standard prescription, $$\tau=\int_\ell dl \int d\epsilon d\Omega n(l, \epsilon, \Omega) \sigma_{\gamma\gamma}(\epsilon,\theta)(1-\cos\theta), \label{eq:tau}
\label{eq2}$$ where $n(l, \epsilon, \Omega)$ is the differential density of soft photons with energy $\epsilon$ which arrive from the low mass star inside the solid angle $\Omega$ to instantaneous location of the $\gamma$-ray photon at the propagation distance $l$, $\sigma_{\gamma\gamma}$ is the $e^\pm$ pair production cross section, and $\theta$ is the angle between the momentum vectors of the gamma-ray and soft photon. $\ell$ denotes the path along propagation direction of the gamma-ray photon in the soft radiation field.
6.2truecm
We investigate the optical depths for $\gamma$-rays as a function of their energies and other free parameters describing the geometry of the picture, such as the injection distance $R$ and the angle $\alpha$ (see Fig. 1). The low mass stars with different radii and surface temperatures are considered. For the purpose of the example calculations, the X-ray luminosity of the compact object (a neutron star) is fixed on $L_{\rm X} = 10^{38}$ erg s$^{-1}$. The optical depths as a function of $\gamma$-ray photon energy for selected injection angles and distances from the companion star are shown in Fig. 3. The calculations show that the optical depths are significant for some range of investigated energies of $\gamma$-ray photons. They are clearly above unity provided that $\gamma$-rays are injected within $R\sim 2R_\star$ from the center of the star at a part of the hemisphere containing the star. Note that the optical depths increase also for stars with lower radii in the case of similar compactness of the binary system (expressed as the ratio of the injection distance and the radius of the star). This is due to the fact that irradiating X-ray source is closer to the stellar surface in such cases. Note that the amount of power supplied to the stellar surface from the X-ray source increases as a square of its distance but the optical depths depend only linearly with this distance. Due to the closer X-ray source, the surface is heated to higher temperature.
In Fig. 4, we show how the optical depths depend on the angle $\alpha$ (which is the angle between the direction defined by the location of the X-ray source and the center of the star and the injection place of $\gamma$-rays and the center of the star, see Fig. 1). Such situation may correspond to the case of injection of $\gamma$-rays from different places of the jet launched from the compact object. Note, that for larger angles $\alpha$ the hottest region on the stellar surface is also seen under larger angles. Therefore, the effective radiation field from the stellar surface seen from the injection place of the $\gamma$-ray drops significantly. As a consequence the optical depths for $\gamma$-ray photons are lower (compare e.g. Figs. 4a and 4c).
In summary, we conclude that the optical depths for $\gamma$-rays can also exceed unity in the case of very compact low mass X-ray binary systems. Therefore, we predict that these LMXBs can become sources of GeV-TeV $\gamma$-rays as in the case of high mass X-ray binaries, provided that the acceleration mechanism of electrons to TeV energies still works.
8.5truecm
Gamma-ray optical depth light curves
====================================
19.truecm
We investigate also how the $\gamma$-ray optical depths can change in the case of a low mass compact binary system in which the X-ray source is on the circular orbit around the solar type star. In this case the angle $\beta$ between the direction towards the observer and the direction defined by the X-ray source and the star can change significantly depending on the inclination angle $\alpha$ of the binary system. $\beta$ is determined in this case by the phase $\Phi$ of the X-ray source (for geometrical situation see Fig. 5).
The example $\gamma$-ray optical depth light curves are calculated for the range of the inclination angles of the binary system and specific energies of $\gamma$-ray photons injected in the direction towards the observer. The results for the X-ray source on a circular orbit with the radius $R_{\rm b} = 2R_\star$ and the injection place of $\gamma$-rays also at the X-ray source are shown in Fig. 6. Note the characteristic optical depth light curves in the case of the non-eclipsing systems with the inclination angles different than zero. In such case, the optical depths reach the maximum when the X-ray source (and also $\gamma$-ray source) is behind the star (in respect to the observer). The optical depths at the peak of the maximum are larger for larger inclination angles of the binary system. For the considered parameters of the binary system, the optical depths are above unity at least for some range of phases close to the phase $\Phi = 0.5$. In the case of the eclipsing binaries the observer should see two peaks in the optical depth light curve due to the eclipse of the X-ray and $\gamma$-ray source by the star. Note, that at the phases at which the $\gamma$-ray optical depth light curves reach the maximum, the hot, irradiated part of the stellar surface is not well visible. Therefore, it is natural to expect unticorrelation between the maximum of $\gamma$-ray light curve and the maximum in the optical light curve from such low mass binary systems. This prediction might be useful in order to increase the probability of the detection of the GeV-TeV $\gamma$-ray signal from the compact X-ray luminous LMXBs.
Since the assumed orbit of the X-ray source (and $\gamma$-ray source) is circular, the light curve looks symmetric. This light curve can be significantly modified when the distance of the X-ray source changes with the phase of the binary. In order to have impression about the influence of the ellipticity of the orbit, we calculate the optical depth light curves for other distances of the X-ray and $\gamma$-ray sources from the star. In Fig. 7, we investigate the optical depths for the orbits of the X-ray source in the range $R_{\rm b} = 2-3R_\star$ and distances of the $\gamma$-rays source in the range $R = 1.5-3R_\star$. Note significant change of the optical depths with these two basic parameters $R_{\rm b}$ and $R$. However, for the whole range of these parameters, the optical depths are always above unity for some range of phases. Therefore, we expect that intensive absorption (and also production) of $\gamma$-ray photons might occur in the case of compact low mass X-ray binary systems. However in the realistic case, the spectrum of $\gamma$-ray photons should be formed in the complicated cascade process due to a highly anisotropic soft radiation created by such irradiated star. This cascade scenario will be clearly more complicated than considered up to now for the $\gamma$-ray production inside the massive binary systems (see e.g. early results by Bednarek 2000).
For comparison, we also show the $\gamma$-ray light curves for the case of stellar mass companion which is in the giant phase (other parameters of the binary as in Fig. 7b). Due to larger distance of the X-ray source from the surface, the optical depths are clearly lower than in the case of the binary system with the solar mass star on the Main Sequence. However, still for small range of phases and relatively large inclination angles of the binary system, the optical depths overcome unity. Therefore, in some geometrical situations also these more geometrically extended binaries (i.e. with longer orbital periods) can provide conditions for efficient production of GeV-TeV $\gamma$-rays.
Calculated here $\gamma$-ray optical depth light curves allow us to conclude that the largest fluxes of GeV $\gamma$-rays should be expected in the phases where the optical depths are clearly above unity. In such cases the cascading effects degrade a part of the primary $\gamma$-ray spectrum in the TeV energy range due to efficient cascading. On the other hand, the largest fluxes of TeV $\gamma$-rays are expected in phases where the optical depths are close to unity. In such case, primary TeV $\gamma$-rays are efficiently produced and not severely absorbed in the IC $e^\pm$ pair cascades. Based on the analysis of Figs. 6 and 7, we expect the largest fluxes of the GeV $\gamma$-rays when the X-ray source (a compact object) is behind the normal star (except the eclipsing binaries). The largest fluxes of the TeV $\gamma$-rays are expected for the intermediate phases, i.e. just before and after the largest fluxes of GeV $\gamma$-rays.
The example cases of LMXBs
==========================
As an example, we show the $\gamma$-ray optical depth light curves for two LMXB systems which show optical modulation with the binary system period. This modulation is interpreted as due to the irradiation of the companion star by the X-ray source (e.g. Milgrom 1976). They have been reported in the past as a possible TeV-PeV $\gamma$-ray sources.
19.truecm
Her X-1
-------
The famous LMXB, Her X-1, is characterised by the X-ray luminosity $L_{\rm x} = 6\times 10^{37}$ erg s$^{-1}$ (White et al. 1983). The companion star has the radius $R_\star = 3.86R_\odot$, the mass $M_\star = 2M_\odot$ with a neutron star on an orbit with the radius $R_{\rm b} = 8.61R_\odot$. The inclination of the binary system has been estimated on $\alpha = 80^{+8}_{-5}$ degrees and its orbital period is 1.7 days (Nagase 1989).
Her X-1 has been claimed in the past as a pulsed TeV $\gamma$-ray source by the Whipple Collaboration (e.g. Gorham et al. 1986, Lamb et al. 1988). However, later analysis have not confirmed these reports (Reynolds et al. 1991). The modern Cherenkov telescope collaborations have not reported any observations of this source recently.
We have calculated the $\gamma$-ray optical depth light curves for the parameters of Her X-1. It is assumed that the source of primary $\gamma$-rays (produced by accelerated electrons) is close to the X-ray source. The results are shown in Fig. 8 for selected energies of $\gamma$-ray photons. It is clear that the optical depths are greater than unity for some relatively small range of phases in the case of $\gamma$-rays with energies in the range $E_\gamma = 0.1-1$ TeV. The optical depths can reach the maximum values $\tau\sim 3$. Therefore, the possible cascading effects have to be taken into account when calculating detailed $\gamma$-ray spectrum escaping towards the observer. However, the primary $\gamma$-ray spectrum should not be very severally degraded at the TeV $\gamma$-ray energies as expected in the case of massive binary systems such as LS 5039 or LSI 61 303. We predict that the largest fluxes of GeV-TeV $\gamma$-rays should appear in the range of phases just before and after the eclipse of the X-ray source by the companion star, i.e. in the ranges: $\Phi\sim 0.35-0.45$ and $0.55-0.65$. Note that zero phase corresponds to the situation where the X-ray compact source is in front of the companion star. Moreover, this $\gamma$-ray emission should be unticorrelated with the optical emission coming from a part of the stellar surface irradiated by the X-ray source. In fact, evidences of the optical modulation with the period of the binary system Her X-1 has been clearly detected in the past.
For the injection place of primary $\gamma$-rays which significantly differ from the location of the X-ray source (e.g. farther from the X-ray source along the jet or at the shock structure), the phase patterns of the $\gamma$-ray optical depths should of course change accordingly.
Sco X-1
-------
The second binary system, Sco X-1, has significantly different parameters from Her X-1. It contains a low mass star, $M_\star = 0.5 M_\odot$, and a neutron star (Liu et al. 2007). Sco X-1 belongs to the class of Z sources with three states of X-ray emission characterised by the luminosities $L_{\rm x} = (4-12)\times 10^{37}$ erg s$^{-1}$ (Hasinger & van der Klis 1989). The accretion disk around the neutron star in this binary system launches a jet. Therefore, the object belongs also to the class of microquasars. The inclination angle of the binary system has been estimated in the range: $\alpha = 44^{\rm o}\pm 6^{\rm o}$ (Fomalont et al. 2001). The period of the binary system is 18,9 hrs (Gottlieb et al. 1975). In our calculations we apply the radius of the companion star $R_\star = 1R_\odot$ (evolved subgiant, Steeghys & Casares 2002), the radius of the binary system $R_{\rm b} = 2.14r_\odot = 1.5\times 10^{11}$ cm, the inclination of the binary system $\alpha = 44^{\rm o}$, and the power of the X-ray source $10^{38}$ erg s$^{-1}$.
Sco X-1 has been also claimed as a TeV-PeV $\gamma$-rays source in the 80-ties but with the low significance (e.g. Brazier et al. 1990, Tonwar et al. 1991). Also these results have not been confirmed by latter observations.
As for Her X-1, the $\gamma$-ray optical light curves have been calculated as a function of energy of $\gamma$-ray photons. Note, that Sco X-1 is more compact binary system than Her X-1. Therefore, the range of phases, for which the $\gamma$-ray optical depths are above unity, is significantly broader. Also, the absolute values of the optical depths are larger. So then, we expect much stronger modification of the escaping $\gamma$-ray spectra by IC $e^\pm$ pair cascading effects. Due to these reasons, the TeV $\gamma$-ray emission can be severally attenuated for the phases close to $\sim 0.5$ and the TeV $\gamma$-ray light curve may show two minima, first due to inefficient production in the region close to $\sim 0.0$ phase and the second due to strong absorption in the region close to the phase $\sim 0.5$. In contrast, the GeV $\gamma$-ray light should show one broad maximum centered on the phase $0.5$. These conclusions are derived assuming the circular orbit of the compact object around the companion star. The more detailed analysis is at present not possible due to the lack of information on the details of the orbit of the compact object in Sco X-1 (i.e. ellipticity, phase of periastron).
It is not clear at present whether the companion star in Sco X-1 is really efficiently irradiated by the X-ray source. Early calculations by Milgrom (1976) suggest that irradiation is important and consistent with observational results at that time. In these calculations relatively small inclination angles has been considered (up to $i\sim 6^o$). The analysis of more recent observations turns to rather intermediate inclination angles (Fomalont et al. 2001,Steeghys & Casares 2002) which is applied in this paper. The present lack of strong modulation of UV emission might be understood assuming that significant part of the hot region on the stellar surface is shadowed by the large accretion disk in this system. Note, that the outer radius of this accretion disk has been estimated on $\sim 6\times 10^{10}$ cm (Vrtilek et al. 1991) which is comparable to the separation of stars in this binary.
Discussion and Conclusion
=========================
10.truecm
We showed that the optical depths for TeV $\gamma$-rays, injected inside the LMXBs close to the surface of the companion stars irradiated by the X-ray sources, are larger than unity for specific locations of the observer. It is easy to show that due to similar cross section for the $\gamma-\gamma\rightarrow e^\pm$ absorption process and the cross section for the ICS process of soft photons by relativistic electrons, also the optical depths for electrons on these same stellar radiation should have similar values. Therefore, it is expected that the production and cascading of $\gamma$-rays can be efficient in the LMXBs in which the companion star is effectively heated by the close X-ray source. However, the optical depths in the case of irradiated low mass stars are generally lower than the optical depths expected in the HMXBs detected in the TeV $\gamma$-rays (such as LS 5039 or LSI 303 +61, e.g. Bednarek 2006a). So then, the cascading effects in the LMXBs should not occur so efficiently as it was shown for those sources (Bednarek 2006b, 2007, Sierpowska-Bartosik & Torres 2007, 2009). Due to these substantial optical depths, we expect that the GeV-TeV $\gamma$-ray emission can be also efficiently produced within compact LMXBs in which the compact X-ray source strongly irradiates companion star provided that electrons are accelerated to TeV energies also inside LMXBs. The mechanism of particle acceleration inside LMXBs can either differ from that one operating within the HMXBs or can be quite similar. In fact, the TeV $\gamma$-rays are observed from different types of sources (e.g. jets of AGNs, vicinity of pulsars, or shocks of supernova remnants). Therefore, different conditions around compact objects in HMXBs and LMXBs can also turn to acceleration of particles to TeV energies. On the other hand, TeV $\gamma$-ray emission inside LMXBs may be produced in similar scenarios as proposed for HMXBs (e.g. accreting neutron star scenario proposed recently by Bednarek 2009b). However, as we have shown above the TeV $\gamma$-ray emission from LMXBs should be limited to specific narrow range of phases which are expected to be different from the range of phases observed in the HMXBs due to differences in the soft radiation field (see considered example cases of Her X-1 and Sco X-1).
Based on the results of calculations of the optical depths and a simple application of the law of gravity, we can envisage which LMXBs are expected to be potential GeV-TeV $\gamma$-ray sources. We conclude that in the case of the solar type star in the evolutionary phase on the Main Sequence ($R_\star = R_\odot$ and $M_\star = M_\odot$) and the orbit with the radius $R_{\rm b} < 3R_\star$, the $\gamma$-rays have a chance to be efficiently absorbed which means also their efficient production in IC process by electrons. This will occur provided that the period of the binary system is shorter than $P_{\rm b}\sim 17 R_\star^{3/2}M_\star^{-1/2}$ hrs, where $R_\star$ and $M_\star$ are in the solar units. Note that many LMXBs have the orbital periods within this limit (Liu et al. 2007).
Based on the analysis of the optical depth light curves we conclude that the emission patterns of escaping $\gamma$-rays from the low mass binary system strongly depend on its inclination angle. We predict that in the simplest considered model of a point like source of primary $\gamma$-rays on a circular orbit around irradiated low mass star, the observer located at:
- [*Small inclination angles*]{}: detects significant GeV-TeV $\gamma$-ray emission only from very compact binaries in which the compact object is at the distance of less than 2 stellar radii from the companion star (compare the triple-dot dashed lines with the thin dotted lines in Figs. 6 and 7). In this case the modulation of the UV emission from an irradiated companion star with the period of the binary system may be significantly reduced since the hottest parts of the stellar surface are not well visible.
- [*Intermediate inclination angles*]{}: detects a single pulse of $\gamma$-ray emission with the maximum corresponding to the phases when the X-ray source is behind the companion star. In the case the observer should detect significant modulation of the UV emission from the stellar surface modulated with the period of the binary system.
- [*Large inclination angles and eclipsing binaries*]{}: detects two narrow pulses of $\gamma$-rays just before and after the eclipse of the X-ray source by the companion star. In this case the modulation of the UV emission should be clearly observed but in some cases can be also strongly suppressed due to the obscuration of the hottest parts on the stellar surface by the large accretion disk.
The $\gamma$-ray emission features from LMXBs can differ from those observed in HMXBs. The differences are due to the fact that in general the optical depths for $\gamma$-rays in the radiation field of irradiated stars in LMXBs are lower in respect to these calculated in the case of HMXBs. We want to mention that the specific binary system may in fact differ significantly from the simplified picture discussed in this paper. The main geometrical complications can be introduced by the inhomogeneous irradiation of the star by X-ray source (anisotropic X-ray source) and the elliptic orbit of the X-ray source around the companion star. Moreover, our analysis base on the assumption that the possible cascade initiated by primary electrons and $\gamma$-rays develop mono-directionally through the binary system. In fact, the structure of the magnetic field of the star can influence the paths of secondary cascade $e^\pm$ pairs. As a result, the escaping secondary $\gamma$-rays can form complicated structures on the sky (see Sierpowska & Bednarek 2005).
Someone can wonder whether absorption of $\gamma$-rays should not occur efficiently in the radiation field of the X-ray source on the surface of the accreting neutron star. We show below that this is not the case provided that $\gamma$-rays are injected at some reasonable distance from the X-ray source. Assuming the characteristic dimension of the X-ray source equal to $\sim 10^5$ cm (the order of the polar cap on the neutron star surface) and the luminosity of the X-ray source $10^{38}$ erg s$^{-1}$, we estimate the characteristic temperature of the X-ray source on $T_{\rm X}\approx 6\times 10^7 L_{38}^{1/4}R_5^{-1/2}$ K, where $R_{\rm x} = 10^5R_5$ cm is the characteristic dimension of the X-ray source. Note, that for simplicity we assumed the black body type emission from the X-ray source. We estimate the distance, $D$, from such X-ray source at which the optical depths in the radiation of such X-ray source become lower than unity from $\tau = D\sigma_{\gamma\gamma\rightarrow e^\pm}n_{\rm x}$, where $\sigma_{\gamma-\gamma\rightarrow e^\pm}$ is the cross section for $e^\pm$ pair creation in $\gamma-\gamma$ collision, $n_{\rm x}\approx 20T^3(R_x/D)^2$ cm$^{-3}$ is density of X-ray photons at the distance $D$. Applying that the cross section for $\gamma-\gamma$ collision is inversely proportional to energy of $\gamma$-ray photon, we estimate that the optical depth becomes lower than unity for distances larger than $D\sim 10^8$ cm and energies of $\gamma$-rays above $\sim 10$ GeV. Note moreover that most of the $\gamma$-rays move in the outward direction in respect to the location of the X-ray source. Thus, the geometrical effects can additionally decrease the optical depths in the X-rays. This estimate on the possible location of the $\gamma$-ray source is clearly consistent with the dimensions of the binary system.
Based on similar estimations it is possible to show that the radiation field from the inner part of the accretion disk around the neutron star may not prevent escape of $\gamma$-ray photons even if they are created within or close to this region. For example, in the case of Her X-1 the inner disk temperature is estimated in the range $\sim (1.8-2.5\times 10^4)$ K (Sazonov 2009) and the disk inner radius is at the distance of the order of a few $\sim 10^8$ cm from the neutron star (Ghosh & Lamb 1979). We estimate the mean free path for $\gamma$-rays propagating in such radiation on $\sim 5\times 10^9$ cm. This is clearly larger than dimension of the disk inner radius. Therefore, $\gamma$-rays can escape from the inner part of the accretion disk without significant absorption. Note, that the region of particle acceleration may be located at some distance from the hot neutron star surface and the inner accretion disk but still inside the binary systems. For example, there are models which propose acceleration of particles within the jets which are observed in the case of some LMXBs (e.g. Sco X-1).
We can envisage realistic situation in which the absorption of TeV $\gamma$-rays in the radiation of the X-ray source can be safely neglected. For example, in the case of Her X-1, the X-ray source can be produced on the neutron star surface due to accretion of matter. However, the acceleration of electrons (and production of the primary $\gamma$-rays) can occur close to or in a very turbulent, magnetized region at the inner accretion disk radius where the pressure of the accreting matter is balanced by strongly magnetized rotating neutron star magnetosphere (the surface magnetic field of Her X-1 is estimated on $2.9\times 10^{12}$ G, e.g. Karino 2007). Then, the inner radius of the accretion disk in Her X-1 is expected at the distance of a few $10^8$ cm from the neutron star. In the case of Sco X-1 production of $\gamma$-rays can occur inside a part of the jet at some distance from the neutron star and the inner disk but still inside the binary system.
This work is supported by the Polish MNiSzW grant N N203 390834.
[99]{}
Aharonian, F. et al. 2005a Science 309, 746 Aharonian, F. et al. 2005b A&A 439, 635 Albert, J. et al. 2006 Science 312, 1771 Bednarek, W. 2000 A&A 363, 646 Bednarek, W. 2006a MNRAS 368, 579 Bednarek, W. 2006b MNRAS 371, 1737 Bednarek, W. 2007 A&A 464, 259 Bednarek, W. 2009a A&A 495, 919 Bednarek, W. 2009b MNRAS 397, 1420 Böttcher, M., Dermer, C.D.2005 ApJ 634, 81 Brazier, K.T.S. et al. 1990 A&A 232, 383 Dubus, G. 2006 A&A 451, 9 Fomalont, E.B., Geldzahler, B.J., Bradshaw, C.F. 2001 ApJ 512, L121 Georganopoulos, M., Aharonian, F.A., Kirk, J.G. 2002 A&A 388, L25 Gorham, P.W. et al. 1986 ApJ 308, L11 Ghosh, P., Lamb, F. 1979 ApJ 234, 296 Gottlieb, E.W., Wright, E.L., Lillier, W. 1975 ApJ 195, L33 Harding, A.K., Gaisser, T.K. 1990 ApJ 358, 561 Hasinger, G., van der Klis, M. 1989 A&A 225, 79 Karino, S. 2007 PASJ 59, 961 Kirk, J.G., Ball, L., Skjaeraasen, O. 1999 APh 10, 31 Lamb, R.C. et al. 1988 ApJ 328, L13 Levinson, A., Blandford, R. 1996 ApJ 456, L29 Liu, Q.Z. et al. 2007 A&A 469, 807 Maraschi, L., Treves, A. 1981 MNRAS 194, 1 Milgrom, M. 1976 ApJ 208, 191 Moskalenko, I.V., Karakuła, S., Tkaczyk, W. 1993, MNRAS 260, 681 Nagase, F. 1989 PASJ 41, 1 Reynolds, P.T. et al. 1991 ApJ 382, 640 Romero, G.E., Kaufman Bernado M.M., Mirabel, F. 2002 A&A 393, 61 Sierpowska, A., Bednarek, W. 2005 MNRAS 356, 711 Sierpowska-Bartosik, A., Torres, D. 2007 ApJ 671, 145 Sierpowska-Bartosik, A., Torres, D. 2009 ApJ 693, 1462 Tavani, M., Arons, J. 1997 ApJ 477, 439 Tonwar, S.C., Gopalakrishnan, N.V., Gupta, S.K., Rajeev, M.R., Sreekantan, B.V., Srivatsan, R. 1991 PRL 67, 2248 Sazonov, A.N. 2009, arXiv:0912.0706 Steeghys, D., Casares, J. 2002 ApJ 568, 273 Vestrand, W.T., Eichler, D. 1982 ApJ 261, 251 Vrtilek, S.D., Penninx, W., Raymond, J.C., Verbunt, F., Hertz, P., Wood, K., Lewin, W.H., Mitsuda, K. 1991 ApJ 376, 278 White, N.E., Swank, J.H., Holt, S.S. 1983 ApJ 270, 711
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate lamellar three-phase patterns that form during the directional solidification of ternary eutectic alloys in thin samples. A distinctive feature of this system is that many different geometric arrangements of the three phases are possible, contrary to the widely studied two-phase patterns in binary eutectics. Here, we first analyze the case of stable lamellar coupled growth of a symmetric model [*ternary*]{} eutectic alloy, using a Jackson-Hunt type calculation in thin film morphology, for arbitrary configurations, and derive expressions for the front undercooling as a function of velocity and spacing. Next, we carry out phase-field simulations to test our analytic predictions and to study the instabilities of the simplest periodic lamellar arrays. For large spacings, we observe different oscillatory modes that are similar to those found previously for binary eutectics and that can be classified using the symmetry elements of the steady-state pattern. For small spacings, we observe a new instability that leads to a change in the sequence of the phases. Its onset can be well predicted by our analytic calculations. Finally, some preliminary phase-field simulations of three-dimensional growth structures are also presented.'
author:
- Abhik Choudhury
- Mathis Plapp
- Britta Nestler
bibliography:
- 'main\_aps.bib'
title: 'Theoretical and numerical study of lamellar eutectic three-phase growth in ternary alloys'
---
Introduction
============
Eutectic alloys are of major industrial importance because of their low melting points and their interesting mechanical properties. They are also interesting for physicists because of their ability to form a large variety of complex patterns, which makes eutectic solidification an excellent model system for the study of numerous nonlinear phenomena.
In a binary eutectic alloy, two distinct solid phases co-exist with the liquid at the eutectic point characterized by the eutectic temperature $T_{\rm E}$ and the eutectic concentration $C_{\rm E}$. If the global sample concentration is close to the eutectic concentration, solidification generally results in composite patterns: alternating lamellae of the two solids, or rods of one solid immersed in a matrix of the other, grow simultaneously from the liquid. The fundamental understanding of this pattern-formation process was established by Jackson and Hunt (JH) [@JH]. They calculated approximate solutions for spatially periodic lamellae and rods that grow at constant velocity $v$, and established that the average front undercooling, that is, the difference between the average front temperature and the eutectic temperature, follows the relation $$\Delta T = K_1 v \lambda + \frac{K_2}{\lambda},
\label{JHlaw}$$ where $\lambda$ is the width of one lamella pair (or the distance between two rod centers), $v$ is the velocity of the solidification front, and $K_1$ and $K_2$ are constants whose value depends on the volume fractions of the two solid phases and various materials parameters [@JH]. The two contributions in Eq. (\[JHlaw\]) arise from the redistribution of solute by diffusion through the liquid and the curvature of the solid-liquid interfaces, respectively.
The front undercooling is minimal for a characteristic spacing $$\lambda_{JH}=\sqrt{\frac{K_2}{K_1v}}.
\label{lambdamin}$$ The spacings found in experiments in massive samples are usually distributed in a narrow range around $\lambda_{JH}$ [@Trivedi91]. However, other spacings can be reached in directional solidification experiments by imposing a solidification velocity that varies with time. In this way, the stability of steady-state growth can be probed [@Akamatsu]. In agreement with theoretical expections [@Manneville], steady-state growth is stable over a range of spacings that is limited by the occurrence of dynamic instabilities. For low spacings, a large-scale lamella (or rod) elimination instability is observed [@Faivre]. For high spacings, the type of instability that can be observed depends on the sample geometry. For thin samples, various oscillatory instabilities and a tilt instability can occur, depending on the alloy phase diagram and the sample concentration. Beyond the onset of these instabilities, stable tilted patterns as well as oscillatory limit cycles can be observed in both experiments and simulations [@Akamatsu; @Sarkissian]. For massive samples, a zig-zag instability occurs for lamellar eutectics [@Akamatsu04; @Parisi08], whereas rods exhibit a shape instability [@Parisi10].
In summary, pattern formation in binary eutectics is fairly well understood. However, most materials of practical importance have more than two components. Therefore, eutectic solidification in multicomponent alloys has received increasing attention in recent years. A particularly interesting situation arises in alloy systems that exhibit a ternary eutectic point, at which four phases (three solids and the liquid) coexist. At such a quadruple point, three binary “eutectic valleys”, that is, monovariant lines of three-phase coexistence, meet. The existence of three solid phases implies that there is a far greater variety of possible structures, even in thin samples. Indeed, for two solids $\alpha$ and $\beta$, an array $\alpha\beta\alpha\beta\ldots$ is the only possibility for a composite pattern in a thin sample; the only remaining degree of freedom is the spacing. With an additional $\gamma$ solid, an infinite number of distinct periodic cycles with different sequences of phases are possible. The simplest cycles are $\alpha\beta\gamma\alpha\beta\gamma\ldots$ and $\alpha\beta\alpha\gamma\alpha\beta\alpha\gamma\ldots$ and permutations. Clearly, cycles of arbitrary length, and even non-periodic configurations are possible. An interesting question is then which configurations, if any, will be favored.
In preliminary works, the occurrence of lamellar structures has been reported in experiments in massive samples [@Kerr; @Durand; @Cooksey; @Holder; @Rinaldi; @Ruggiero; @McCartney]. The spatio-temporal evolution in ternary eutectic systems was observed in thin samples (quasi-2D experiments) in both metallic [@In-Bi-Sn] and organic systems [@AMPDDCNPGSCN]. In both cases, the simultaneous growth of three distinct solid phases from the liquid with a $\left(\alpha \beta \alpha \gamma\right)$, (named ABAC in Ref. [@AMPDDCNPGSCN]) stacking was observed. Measurements in both cases revealed that $\lambda^{2}v$ was approximately constant, in agreement with the JH scaling of Eq. (\[lambdamin\]).
On the theoretical side, models that extend the JH analysis from binary to ternary eutectics for three different growth morphologies (rods and hexagon, lamellar, and semi-regular brick structures) were proposed by Himemiya [*et al.*]{} [@TPEG]. The relation between front undercooling and spacing is still of the form given by Eq. (\[JHlaw\]), with constants $K_1$ and $K_2$ that depend on the morphology. The differences between the minimal undercoolings for different morphologies were found to be small. No direct comparison to experiments was given.
Finally, ternary eutectic growth has also been investigated by phase-field methods in Refs. [@Apel; @Hecht], who have studied different stacking sequences formed by $\alpha$ = Ag$_{2}$Al, $\beta$ = ($\alpha$ Al) and $\gamma$ = Al$_{2}$Cu in the ternary system Al-Cu-Ag, while transients in the ternary eutectic solidification of a transparent In-Bi-Sn alloy were studied both by phase field modeling and experiments [@In-Bi-Sn].
The purpose of the present paper is to carry out a more systematic investigation of lamellar ternary eutectic growth. The main questions we wish to address are (i) can an extension of the JH theory adequately describe the properties of ternary lamellar arrays and reveal the differences between cycles of different stacking sequences, and (ii) what are the instabilities that can occur in such patterns. To answer these questions, we develop a generalization of the JH theory to ternary eutectics which is capable of describing the front undercoolings of periodic lamellar arrays with arbitrary stacking sequence. Its predictions are systematically compared to phase-field simulations. We use a generic thermodynamically consistent phase-field model [@Garcke; @Stinner]. While this model is known to exhibit several thin-interface effects which limit its accuracy [@Karma96; @Almgren99; @Kim98; @Karma+01; @Eschebaria+04], we show here that we can obtain a very satisfying agreement between theory and simulations if the solid-liquid interfacial free energy is evaluated numerically. In particular, the minimum-undercooling spacings are accurately reproduced for all stacking sequences that we have simulated.
The model is then used to systematically investigate the instabilities of lamellar arrays, in particular for large spacings. We find that, as for binary eutectics, the symmetry elements of the steady-state array determine the possible instability modes. Whereas the calculation of a complete stability diagram is not feasible due to the large number of independent parameters, we find and characterize several new instability modes. Besides these oscillatory modes that are direct analogs of the ones observed in binary eutectics, we also find a new type of instability which occurs at small spacings: cycles in which the same phase appears more than once can undergo an instability during which one of these lamellae is eliminated; the system therefore transits to a different (simpler) cycle. Furthermore, we also find that the occurrence of this type of instability can be well predicted by our generalized JH theory.
The remainder of the paper is organized as follows. In Sec. 2, we develop the generalized JH theory for ternary eutectics and calculate the undercooling-spacing relationships for several simple cycles. In Sec. 3, the phase-field model is outlined and its parameters are related to the ones of the theory. Sec. 4 presents the simulation results concerning both steady-state growth and its instabilities. In Sec. 5, we briefly discuss questions related to pattern selection and present some preliminary simulations in three dimensions. Sec. 6 concludes the paper.
\[sec:theory\]Theory
====================
We consider a ternary alloy system consisting of components $A,B$ and $C$, which can form three solid phases $\alpha,
\beta$, and $\gamma$ upon solidification from the liquid $l$. The concentrations of the components (in molar fractions) are denoted by $c_A, c_B$ and $c_C$ and fulfill the constraint $$c_A + c_B + c_C = 1.
\label{Eqn-Constraint}$$ This obviously implies that there are only two independent concentration fields.
As is customary, isothermal sections of the ternary phase diagram can be conveniently displayed in the Gibbs simplex. We are interested in alloy systems that exhibit a ternary eutectic point: four-phase coexistence between three solids and the liquid. The isothermal cross-section at the ternary eutectic temperature is displayed in Figure \[Figure1\], here for the particular example of a completely symmetric phase diagram.
![(Color Online) Projection of the ternary phase diagram for a model symmetric ternary eutectic system on the Gibbs simplex. The triangle at the center is the tie-triangle at the eutectic temperature where four phases $\alpha,\beta,\gamma$, and $l$ are in equilibrium. The diagram also contains the information on three-phase equilibria. The liquidus lines corresponding to each of these equilibria (“eutectic valleys”) are shown by dotted lines which meet at the center of the simplex, which is also the concentration of the liquid at which all the three solid phases and the liquid are at equilibrium.[]{data-label="Figure1"}](figure1.eps){width="8cm"}
The concentration of the liquid is located in the center of the simplex ($c_A=c_B=c_C=1/3$), and the three solid phases are located at the corners of the eutectic tie triangle. For higher temperatures, no four-phase coexistence is possible, but each pair of solid phases can coexist with the liquid (three-phase coexistence). Each of these three-phase equilibria is a eutectic, and the loci of the liquid concentrations at three-phase coexistence as a function of temperature form three “eutectic valleys” that meet at the ternary eutectic point. On each of the sides of the simplex (with the temperature as additional axis), a binary eutectic phase diagram is found.
The key point for the following analysis is the temperature of solid-liquid interfaces, which depends on the liquid concentration, the interface curvature, and the interface velocity. The dependence on the concentration is described by the liquidus surface, which is a two-dimensional surface over the Gibbs simplex. This surface can hence be characterized by two independent liquidus slopes at each point. For each phase $\nu$ ($\nu=\alpha,\beta,\gamma$), we choose the two liquidus slopes with respect to the minority components. Thus, for the $\alpha$ phase, the interface temperature is given by the generalized Gibbs-Thomson relation, $$T_{\rm int}^\alpha-T_{\rm E} = m_B^\alpha(c_B-c_B^E) + m_C^\alpha(c_C-c_C^E)
-\Gamma_\alpha \kappa - \frac{v_n}{\mu_{\rm int}^\alpha},
\label{GibbsThomson}$$ where $c_B$ and $c_C$ are the concentrations in the liquid adjacent to the interface, $c_B^E$ and $c_C^E$ their values at the ternary eutectic point, and $m^\alpha_B = \left. \frac{d T_\alpha}{d c_B}
\right|_{c_C={\rm const}}$ and $m^\alpha_C = \left. \frac{d T_\alpha}{d c_C}
\right|_{c_B={\rm const}}$ the liquidus slopes taken at the ternary eutectic point. Furthermore, $\Gamma_\alpha=\tilde\gamma_{\alpha l} T_{\rm E}/L_\alpha$ is the Gibbs-Thomson coefficient, with $\tilde\gamma_{\alpha l}$ the solid-liquid surface tension and $L_\alpha$ the latent heat of fusion per unit volume, and $\mu_{\rm int}^\alpha$ is the mobility of the $\alpha$-liquid interface. For the typical (slow) growth velocities that can be attained in directional solidification experiments, the last term, which represents the kinetic undercooling of the interface, is very small. It will therefore be neglected in the following. The expression for the other solid phases are obtained by cyclic permutation of the indices.
In the spirit of the original Jackson-Hunt analysis, for the calculation of the diffusion field in the liquid, the concentration differences between solid and liquid phases are assumed to be constant and equal to their values at the ternary eutectic point. Since we are interested in ternary coupled growth, which will take place at temperatures close to $T_E$, this should be a good approximation. Thus, we define $$\Delta c_{j}^{\nu} = c_{j}^{l} - c_{j}^{\nu}\qquad \mbox{with} \qquad j=A,B,C \qquad
\mbox{and} \qquad \nu = \alpha, \beta, \gamma.$$ In this approximation, the Stefan condition at a $\nu$-$l$ interface, which expresses mass conservation upon solidification, reads $$\partial_n c_j = -\frac{v_n}{D} \Delta c_{j}^{\nu},
\label{Eqn-Stefan}$$ where $\partial_n c_j$ denotes the partial derivative of $c_j$ in the direction normal to the interface, $v_n$ is the normal velocity of the interface (positive for a growing solid), and $D$ is the chemical diffusion coefficient, for simplicity assumed to be equal for all the components.
We consider a general periodic lamellar array with $M$ repeating units consisting of phases $(\nu_0, \nu_1, \nu_2, \ldots, \nu_{M-1})$ where each $\nu_{i}$ represents the name of one solid phase $\left(\alpha, \beta , \gamma\right)$ in the sequence, with a repeat distance (lamellar spacing) $\lambda$. The width of the $j$-th single solid phase region is $\left(x_{j}-x_{j-1}\right)\lambda$, with $x_{0}=0$ and $x_{M}=1$, and the sum of all the widths corresponding to any given phase is its volume fraction $\eta_{\nu}$. The eutectic front is assumed to grow in the $z$ direction with a constant velocity $v$.
=15.5cm
Concentration fields
--------------------
First, we consider the diffusion fields of the components $A,B,C$ ahead of a growing eutectic front. For the calculation of the concentration fields, the front is supposed to be planar, as in the sketches of Figure \[Figure2\]. We make the following Fourier series expansion for $c_A$ and $c_B$ $$c_X = \sum_{n=-\infty}^{\infty} X_n e^{ik_n x-q_nz} + c_X^\infty, \quad
X=A,B.
\label{Eqn-Series_cX}$$ The third concentration $c_C$ follows from the constraint of Eq. (\[Eqn-Constraint\]). In Eq. (\[Eqn-Series\_cX\]), $k_n=2\pi n/\lambda$ are wave numbers and $q_n$ can be determined from the solutions of the stationary diffusion equation $$v \partial_z c_X + D \nabla^2 c_X = 0,$$ which yields $$q_n = \frac{v}{2D} + \sqrt{k_n^2+\left(\frac{v}{2D}\right)^2}.$$ For all the modes $n\neq 0$, we thus have $q_n \simeq |k_n|$ for small Peclet number ${\rm Pe} = \lambda/\ell \ll 1$ with $\ell=2D/v_n$, which will always be the case for slow growth. The mode $n=0$ describes the concentration boundary layer which is present at off-eutectic concentrations, and which has a characteristic length scale of $\ell$.
To determine the coefficients $X_n$ in the above Fourier series, we assume the eutectic front to be at the $z=0$ position. Using the Stefan condition in Eq. (\[Eqn-Stefan\]) and taking the derivative of $c_X$ with respect to the $z$-coordinate $$\partial_z c_X|_{z=0} = \sum_{n=-\infty}^\infty - q_n X_n e^{ik_n x},$$ integration across one lamella period $\lambda$ of arbitrary partitioning of phases gives $$\begin{aligned}
q_n X_n \delta_{nm} \lambda &=& \frac{2}{\ell}
\sum_{j=0}^{M-1} \int_{x_j \lambda}^{x_{j+1}\lambda}
e^{-ik_m x} \Delta c_{X}^{\nu_j}\: dx,
\label{ftransform}\end{aligned}$$ so that the coefficients $X_n, n\in {{\if mm {\rm I}\mkern -3mu{\rm N}\else \leavevmode
\hbox{I}\kern -.17em \hbox{N} \fi}}$ in the series ansatz, Eq. (\[Eqn-Series\_cX\]) follow $$\begin{aligned}
\qquad X_n &=& \frac{4}{\ell q_n \lambda k_n} \sum_{j=0}^{M-1}
\Delta c_{X}^{\nu_j} e^{-ik_n \lambda(x_{j+1}+x_j)/2} \sin(k_n \lambda (x_{j+1}+ x_j)/2).\end{aligned}$$ Applying symmetry arguments for the sinus and cosinus functions, we can formulate real combinations of these coefficients if we additionally take the negative summation indices into account. We obtain $$X_n + X_{-n} = \frac{8}{\ell q_n \lambda k_n} \sum_{j=0}^{M-1}
\Delta c_{X}^{\nu_j} \cos(k_n \lambda(x_{j+1}+x_j)/2) \sin(k_n \lambda (x_{j+1}+ x_j)/2),$$ $$i(X_n - X_{-n}) = \frac{8}{\ell q_n \lambda k_n} \sum_{j=0}^{M-1}
\Delta c_{X}^{\nu_j} \sin(k_n \lambda(x_{j+1}+x_j)/2) \sin(k_n \lambda (x_{j+1}-x_j)/2).$$ Herewith, Eq. (\[Eqn-Series\_cX\]) reads $$\begin{aligned}
c_X = c_X^\infty + X_0 + \sum_{j=0}^{M-1} \sum_{n=1}^\infty
\frac{8}{\ell q_n \lambda k_n}
\cos(k_n \lambda(x_{j+1}+x_j)/2) \sin(k_n \lambda (x_{j+1}+ x_j)/2)\cos(k_n x) \nonumber \\
+ \sum_{j=0}^{M-1} \sum_{n=1}^\infty
\frac{8}{\ell q_n \lambda k_n}
\sin(k_n \lambda(x_{j+1}+x_j)/2) \sin(k_n \lambda (x_{j+1}-x_j)/2)\sin(k_n x).\end{aligned}$$ The general expression for the mean concentration $\langle c_X \rangle_m$ ahead of the $m$-th phase of the phase sequence can be calculated to yield $$\begin{aligned}
\langle c_X \rangle _m &=& \frac{1}{(x_{m+1} -x_m)\lambda} \int_{x_m \lambda}^{x_{m+1}\lambda}
c_X dx \nonumber \\
&=& c_X^\infty + X_0 + \frac{1}{x_{m+1} - x_m} \sum_{n=1}^{\infty} \sum_{j=0}^{M-1}
\Big\{ \frac{16}{\lambda^2 k_n^2 \ell q_n} \Delta c_{X}^{\nu_j} \sin[\pi n(x_{m+1}-x_m)] \times
\nonumber \\
&& \hspace*{1.5cm} \times \: \sin[\pi n(x_{j+1} - x_j)] \cos[\pi n(x_{m+1}+x_m - x_{j+1}-x_j)]
\Big\}.
\label{Eqn-Mean_cX_general}\end{aligned}$$ For a repetitive appearance of a phase $\nu$ in the phase sequence, the mean concentration of component $X$ ahead of this phase follows by taking the weighted average of all the lamellae of phase $\nu$, $$\langle c_X \rangle_{\nu} = \frac{\sum_{m=0}^{M-1} \langle c_X \rangle_m (x_{m+1} - x_m) \delta_{{\nu_m} \nu}}
{\sum_{m=0}^{M-1} (x_{m+1}-x_m) \delta_{{\nu_m} \nu}} \qquad \mbox{with} \qquad
\delta_{{\nu_m} \nu} = \left\{
\begin{array}{cc}
0 & \mbox{for} \: \nu = \nu_m \\
1 & \mbox{for} \: \nu \ne \nu_m.
\end{array}
\right.$$
Average front temperature
-------------------------
The average front temperature is now found by taking the average of the Gibbs-Thomson equation along the front, separately for each phase ($\alpha, \beta$ and $\gamma$): $$\begin{aligned}
\Delta T_\nu &=& T_E - T_\nu = -m_B^\nu (\langle c_B \rangle_\nu - c_B^E)
- m_C^\nu (\langle c_C \rangle_\nu - c_C^E) + \Gamma_\nu \langle \kappa \rangle_\nu, \label{Eqn-undercooling}\end{aligned}$$ for $\nu = \alpha, \beta, \gamma$. Here, $\langle \kappa \rangle_\nu$ is the average curvature of the solid-liquid interface which can be evaluated by exact geometric relations to be $$\begin{aligned}
\langle \kappa \rangle_\nu &=& \frac{\sum_{m=0}^{M-1}\langle \kappa \rangle_m (x_{m+1}-x_m) \delta_{{\nu_m}\nu}}{\sum_{m=0}^{M-1}
(x_{m+1}-x_m)\delta_{{\nu_m}\nu}}\end{aligned}$$ and $$\langle \kappa \rangle _m = \frac{\sin \theta_{\nu_{m}\nu_{m+1}} + \sin \theta_{\nu_m \nu_{m-1}}}
{(x_{m+1} - x_m) \lambda}.$$ Here, $\theta_{\nu_m \nu_{m-1}}$ are the contact angles that are obtained by applying Young’s law at the trijunction points. More precisely, $\theta_{\nu_{m}\nu_{m+1}}$ is the angle, at the triple point (identified by the intersection of the two solid-liquid interfaces and the solid-solid one), between the tangent to the $\nu_{m}-l$ interface and the horizontal (the $x$ direction). For a triple point with the phases $\nu_{m},\nu_{m+1}$ and liquid, the two contact angles $\theta_{\nu_{m} \nu_{m+1}}, \theta_{\nu_{m+1}\nu_{m}}$ satisfy the following relations, obtained from Young’s law, $$\frac{\tilde{\gamma}_{\nu_{m+1}l}}{\cos(\theta_{\nu_{m} \nu_{m+1}})}=\frac{\tilde{\gamma}_{\nu_{m}l}}{\cos(\theta_{\nu_{m+1} \nu_{m}})}=\frac{\tilde{\gamma}_{\nu_{m} \nu_{m+1}}}{\sin(\theta_{\nu_{m}\nu_{m+1}} + \theta_{\nu_{m+1}\nu_{m}})}.$$ Note that, in general, $\theta_{\nu_{m} \nu_{m+1}} \neq \theta_{\nu_{m+1}\nu_{m}}$.
A short digression is in order to motivate the closure of our system of equations. Although we have not given the explicit expressions, the coefficients $A_0$ and $B_0$ can be simply calculated by using Eq. (\[ftransform\]) with $n=0$. However, to carry out this calculation, the width of each lamella has to be given. If these widths are chosen consistent with the lever rule, that is, the cumulated lamellar width of phase $\nu$ corresponds to the nominal volume fraction of phase $\nu$ for the given sample concentration $c_A^\infty$, $c_B^\infty$, and $c_C^\infty$, the use of Eq. (\[ftransform\]) yields $X_0=c_X^E-c_X^\infty$ ($X=A,B,C$). However, this result is incorrect: the concentrations of the solids are not equal to the equilibrium concentrations at the eutectic temperature because solidification takes place at a temperature below $T_{\bf E}$. Therefore, the true volume fractions depend on the solidification conditions. Their determination would require a self-consistent calculation which is exceedingly difficult. Therefore, we will take the same path as Jackson and Hunt in their original paper [@JH]: we will assume that the volume fractions of the three phases are fixed by the lever rule at the eutectic temperature, but we will treat the amplitudes of the two boundary layers, $A_0$ and $B_0$, as unknowns. As in Ref. [@JH], one can expect that the difference to the true solution is of order $\rm Pe$ and therefore small for slow solidification.
With this assumption, the equations developed above can now be used in two ways. For [*isothermal solidification*]{}, the temperatures of all interfaces must be equal to the externally set temperature, and the three equations $\Delta T_\nu = \Delta T$ for $\nu = \alpha,\beta,\gamma$, can be used to determine the three unknowns $A_0, B_0$ and the velocity $v$ of the solid-liquid front. All of these quantities will be a function of the lamellar spacing $\lambda$. In [*directional solidification*]{}, the growth velocity in steady state is fixed and equal to the speed with which the sample is pulled from a hot to a cold region. The third unknown is now the total front undercooling. In the classic Jackson-Hunt theory for binary eutectics, the system of equations is closed by the hypothesis that the average undercoolings of the two phases are equal. This is only an approximation which is quite accurate for eutectics with comparable volume fractions of the two solids, but becomes increasingly inaccurate when the volume fractions are asymmetric [@Sarkissian]. We will use the same approximation for the ternary case here, and set $\Delta T_\alpha=\Delta T_\beta = \Delta T_\gamma=\Delta T$. This then leads to expressions for $\Delta T$ as a function of the growth speed $v$ and the lamellar spacing $\lambda$.
Examples
--------
### Binary systems
As a benchmark for both our calculations and simulations, we consider binary eutectic systems with components $A$ and $B$ and with three phases: $\alpha, \beta$, and liquid.
Setting $x_0=0, x_1 = \eta_\alpha$, $x_2 = 1$, and applying Eq. (\[Eqn-Mean\_cX\_general\]) gives $$\begin{aligned}
\langle c_X \rangle_\alpha &=& c_X^\infty + X_0 + \frac{1}{\eta_\alpha} \sum_{n=1}^{\infty}
\Big\{ \frac{16}{\lambda^2 k_n^2 \ell q_n} \left(\Delta c_{X}^{\alpha} - \Delta c_{X}^{\beta} \right)
\sin^2(\pi n\eta_\alpha) \Big\} \\
&\cong & c_X^\infty + X_0 + \frac{2 \lambda}{\eta_\alpha \ell} {\cal P}(\eta_\alpha) \Delta c_X
\qquad \mbox{and} \\
\langle c_X \rangle_\beta &=& c_X^\infty + X_0 - \frac{2 \lambda}{(1-\eta_\alpha) \ell} {\cal P}(1-\eta_\alpha)
\Delta c_X
\label{mean_cA_binary}\end{aligned}$$ with $k_n = 2 \pi n/\lambda, q_n \approx k_n, \lambda / \ell \ll 1,
\Delta c_X = \Delta c_{X}^{\alpha} - \Delta c_{X}^{\beta}$, and the dimensionless function $$\begin{aligned}
{\cal P}(\eta) &=& \sum_{n=1}^\infty \frac{1}{(\pi n)^3}\sin^2(\pi n \eta)
\label{Eqn_Pdef}\end{aligned}$$ which has the properties ${\cal P}(\eta) = {\cal P} (1-\eta)={\cal P}(\eta -1)$.
Furthermore, Eq. (\[Eqn-undercooling\]) together with $\ell = 2D/v$ leads to $$\begin{aligned}
\Delta T_\alpha = -m_B^\alpha B_0 - \frac{\lambda v}{\eta_\alpha D}{{\cal P}}(\eta_\alpha)m_B^\alpha\Delta c_B + \Gamma_\alpha \langle\kappa\rangle_\alpha,\\
\Delta T_\beta = -m_A^\beta A_0 - \frac{\lambda v}{\eta_\beta D}{{\cal P}}(\eta_\beta)m_A^\beta\Delta c_A + \Gamma_\beta \langle\kappa\rangle_\beta,\end{aligned}$$ where $\langle\kappa\rangle_\alpha = 2 \sin \theta_{\alpha \beta}/(\eta_\alpha \lambda)$ and $\langle\kappa\rangle_\beta = 2 \sin \theta_{\beta \alpha}/(\eta_\beta \lambda)$. In addition, for a binary alloy $B_0=-A_0$. The unknown $A_{0}$ and the global front undercooling are determined using the assumption of equal interface undercoolings, $\Delta T_\alpha=\Delta T_\beta$. The result is identical to the one of the Jackson-Hunt analysis.
### Ternary Systems
Next, we study ternary systems with three components ($A, B, C$) and four phases ($\alpha, \beta, \gamma$ and liquid). We start with the configuration $\left(\alpha\beta\gamma\alpha\beta\gamma \ldots\right)$, sketched in Figure \[Figure4\].
We set $x_0=0, x_1 = \eta_\alpha, x_2=\eta_\alpha + \eta_\beta = 1 - \eta_\gamma$ and $x_3 = 1$ and apply Eq. (\[Eqn-Mean\_cX\_general\]). This yields $$\begin{aligned}
\label{abc_caalpha}
\langle c_X \rangle_\alpha &=& c_X^\infty + X_0 +
\frac{2 \lambda}{\eta_\alpha \ell} \Big( {\cal P}(\eta_\alpha) \Delta c_{X}^{\alpha}
+ {\cal Q} (\eta_\alpha,\eta_\beta) \Delta c_{X}^{\beta} +
{\cal Q}(\eta_\alpha, \eta_\gamma) \Delta c_{X}^{\gamma} \Big)\\
\langle c_X \rangle_\beta &=& c_X^\infty + X_0 +
\frac{2 \lambda}{\eta_\beta \ell} \Big( {\cal Q}(\eta_\beta, \eta_\alpha)
\Delta c_{X}^{\alpha}
+ {\cal P} (\eta_\beta) \Delta c_{X}^{\beta} +
{\cal Q}(\eta_\beta, \eta_\gamma) \Delta c_{X}^{\gamma} \Big) \\
\label{abc_cabeta}
\langle c_X \rangle_\gamma &=& c_X^\infty + X_0 +
\frac{2 \lambda}{\eta_\gamma \ell} \Big( {\cal Q}(\eta_\gamma, \eta_\alpha)
\Delta c_{X}^{\alpha}
+ {\cal Q} (\eta_\gamma, \eta_\beta) \Delta c_{X}^{\beta} +
{\cal P}(\eta_\gamma) \Delta c_{X}^{\gamma} \Big).
\label{abc_cagamma}\end{aligned}$$ Here, we have used $X=A, B, C$ and ${{\cal P}}$ is the function defined in Eq. (\[Eqn\_Pdef\]), and $$\begin{aligned}
{{\cal Q}}(\eta_{\nu_i}, \eta_{\nu_j}) &=& \sum_{n=1}^\infty \frac{1}{(\pi n)^3}
\sin(\pi n \eta_{\nu_i}) \sin(\pi n \eta_{\nu_j}) \cos[\pi n (\eta_{\nu_i} + \eta_{\nu_j})]\end{aligned}$$ ${{\cal P}}(\eta_{\nu_i})$ and ${{\cal Q}}(\eta_{\nu_i},\eta_{\nu_j})$ fulfill the properties ${{\cal P}}(\eta_{\nu_i}) = -{{\cal Q}}(\eta_{\nu_i}, -\eta_{\nu_i})$ and ${{\cal Q}}(\eta_{\nu_i},\eta_{\nu_j}) = {{\cal Q}}(\eta_{\nu_j},\eta_{\nu_i})$.\
For simplicity, we now consider a completely symmetric ternary eutectic configuration: a completely symmetric ternary phase diagram (that is, any two phases can be exchanged without changing the phase diagram) and equal phase fractions $\eta_\alpha = \eta_\beta = \eta_\gamma=\frac{1}{3}$, which implies $c_X^\infty = c_X^E$. As a consequence, $X_0 = 0$, and Eq. (\[abc\_caalpha\]) simplifies to $$\begin{aligned}
\langle c_A \rangle_\alpha - c_A^E &=& \frac{2 \lambda}{\eta_\alpha \ell} {{\cal P}}(\eta_\alpha)
(\Delta c_{A}^{\alpha} - \Delta c_{A}^{\beta}) \\
\langle c_B \rangle_\alpha - c_B^E &=& \frac{ \lambda {\cal P}(\eta_\alpha) }{\eta_\alpha \ell} \Big(\Delta c_{B}^{\alpha}
- \Delta c_{B}^{\beta}\Big) \\
\langle c_C \rangle_\alpha - c_C^E &=& \frac{\lambda {\cal P}(\eta_\alpha) }{\eta_\alpha \ell} \Big(\Delta c_{C}^{\alpha}
- \Delta c_{C}^{\gamma} \Big),\end{aligned}$$
for the three components. Since, in this case, all phases have the same undercooling by symmetry, the front undercooling is simply given by $$\Delta T = - \frac{2 \lambda v}{\eta_\alpha D} {{\cal P}}(\eta_\alpha)
m_B^\alpha \Delta c_B + \Gamma_\alpha \langle\kappa\rangle_\alpha$$ where $\langle\kappa\rangle_\alpha = \tfrac{2}{\eta_\alpha \lambda}(\sin \theta_{\alpha \beta} + \sin \theta_{\alpha \gamma})$. The terms $\Delta c_{B}^{\alpha} - \Delta c_{B}^{\beta}$ and $\Delta c_{C}^{\alpha} - \Delta c_{C}^{\gamma}$ are identical. For convenience, we write the preceding equation using the term we already use for the binaries namely $\Delta c_{B}=\Delta c_{B}^{\alpha} - \Delta c_{B}^{\beta}$.
Next, we discuss again a ternary eutectic alloy with three components and four phases, but now for the phase cycle $\left(\alpha\beta\alpha\gamma\alpha\beta\ldots \right)$.
Furthermore, we suppose that the two lamellae of the $\alpha$ phase have equal width $\lambda\eta_\alpha/2$. The average concentrations $\langle c_{X}\rangle_{m}$ are deduced from the general expression in Eq.\[Eqn-Mean\_cX\_general\] and read $$\begin{aligned}
\langle c_X \rangle_\alpha &=& c_X^\infty + X_0 +
\frac{2 \lambda}{\tfrac{\eta_\alpha}{2} \ell} \Big(
{{\cal S}}(\eta_\alpha, \eta_\beta) \Delta c_{X}^{\alpha}
+ {\cal Q} (\tfrac{\eta_\alpha}{2},\eta_\beta) \Delta c_{X}^{\beta} +
{\cal Q}(\tfrac{\eta_\alpha}{2}, \eta_\gamma) \Delta c_{X}^{\gamma} \Big) \\
\label{abac_caalpha}
\langle c_X \rangle_\beta &=& c_X^\infty + X_0 +
\frac{2 \lambda}{\eta_\beta \ell} \Big( 2 {\cal Q}(\eta_\beta, \tfrac{\eta_\alpha}{2})
\Delta c_{X}^{\alpha}
+ {\cal P} (\eta_\beta) \Delta c_{X}^{\beta} + {{\cal R}}(\eta_\beta, \eta_\gamma)
\Delta c_{X}^{\gamma} \Big) \\
\label{abac_cabeta}
\langle c_X \rangle_\gamma &=& c_X^\infty + X_0 +
\frac{2 \lambda}{\eta_\gamma \ell} \Big( 2 {\cal Q}(\eta_\gamma, \tfrac{\eta_\alpha}{2})
\Delta c_{X}^{\alpha} + {{\cal R}}(\eta_\gamma, \eta_\beta) \Delta c_{X}^{\beta} +
{\cal P}(\eta_\gamma) \Delta c_{X}^{\gamma} \Big),
\label{abac_cagamma}\end{aligned}$$ where $X=A, B, C$. Furthermore, we have introduced the short notations $$\begin{aligned}
{{\cal R}}(\eta_{\nu_i},\eta_{\nu_j}) &=& \sum_{n=1}^\infty \frac{1}{(\pi n)^3} \sin(\pi n \eta_{\nu_i})
\sin(\pi n \eta_{\nu_j}) \cos(\pi n)\\
{{\cal S}}(\eta_{\nu_i},\eta_{\nu_j}) &=& \sum_{n=1}^\infty \frac{1}{(\pi n)^3}
\sin^2(\pi n \eta_{\nu_i}/2)
\{1 + \cos(\pi n)\cos[\pi n (\eta_{\nu_j} - \eta_{\nu_i})]\}.\end{aligned}$$ From the general formulation of the Gibbs-Thomson equation in Eq. (\[Eqn-undercooling\]), we determine the undercoolings, $$\begin{aligned}
\label{abag-1lamella}
\Delta T_\alpha &=& -m_{B}^{\alpha}\left(B_{0}+\frac{4 \lambda}{\eta_\alpha \ell} \Big(
{\cal S}(\eta_\alpha, \eta_\beta) \Delta c_{B}^{\alpha}
+ {\cal Q} (\tfrac{\eta_\alpha}{2},\eta_\beta) \Delta c_{B}^{\beta} +
{\cal Q}(\tfrac{\eta_\alpha}{2}, \eta_\gamma) \Delta c_{B}^{\gamma} \Big)\right) \nonumber \\
&+& -m_{C}^{\alpha}\left(C_{0}+\frac{4 \lambda}{\eta_\alpha \ell} \Big(
{\cal S}(\eta_\alpha, \eta_\beta) \Delta c_{C}^{\alpha}
+ {\cal Q} (\tfrac{\eta_\alpha}{2},\eta_\beta) \Delta c_{C}^{\beta} +
{\cal Q}(\tfrac{\eta_\alpha}{2}, \eta_\gamma) \Delta c_{C}^{\gamma} \Big)\right) \nonumber \\
&+& \Gamma_\alpha \dfrac{2\left(\sin \theta_{\alpha \beta} + \sin \theta_{\alpha \gamma}\right)}{\eta_{\alpha}\lambda}\\
\Delta T_\beta &=& -m_{A}^{\beta}\left(A_{0}+\frac{2 \lambda}{\eta_\beta \ell} \Big( 2 {\cal Q}(\eta_\beta, \tfrac{\eta_\alpha}{2})
\Delta c_{A}^{\alpha}
+ {\cal P} (\eta_\beta) \Delta c_{A}^{\beta} + {\cal R}(\eta_\beta, \eta_\gamma)
\Delta c_{A}^{\gamma} \Big)\right) \nonumber \\
&+& -m_{C}^{\beta}\left(C_{0}+\frac{2 \lambda}{\eta_\beta \ell} \Big( 2 {\cal Q}(\eta_\beta, \tfrac{\eta_\alpha}{2})
\Delta c_{C}^{\alpha}
+ {\cal P} (\eta_\beta) \Delta c_{C}^{\beta} + {\cal R}(\eta_\beta, \eta_\gamma)
\Delta c_{C}^{\gamma} \Big)\right) \nonumber \\
&+& \Gamma_\beta \dfrac{2\sin \theta_{\beta \alpha}}{\eta_{\beta}\lambda}\\
\Delta T_\gamma &=& -m_{A}^{\gamma}\left(A_{0}+\frac{2 \lambda}{\eta_\gamma \ell} \Big( 2 {\cal Q}(\eta_\gamma, \tfrac{\eta_\alpha}{2})
\Delta c_{A}^{\alpha} + {\cal R}(\eta_\gamma, \eta_\beta) \Delta c_{A}^{\beta} +
{\cal P}(\eta_\gamma) \Delta c_{A}^{\gamma} \Big)\right) \nonumber \\
&+& -m_{B}^{\gamma}\left(B_{0}+\frac{2 \lambda}{\eta_\gamma \ell} \Big( 2 {\cal Q}(\eta_\gamma, \tfrac{\eta_\alpha}{2})
\Delta c_{B}^{\alpha} + {\cal R}(\eta_\gamma, \eta_\beta) \Delta c_{B}^{\beta} +
{\cal P}(\eta_\gamma) \Delta c_{B}^{\gamma} \Big)\right) \nonumber \\
&+& \Gamma_\gamma \dfrac{2\sin \theta_{\gamma \alpha}}{\eta_{\gamma}\lambda}.\end{aligned}$$ For a symmetric phase diagram (all slopes equal, $m_{X}^{\nu_{i}}=m$) one can show using the assumption of equal undercooling of all phases that an expression for the global interface undercooling can be derived as $\Delta T = 1/3(\Delta T_{\alpha} + \Delta T_{\beta} + \Delta T_{\gamma})$ by elimination of the constants $A_{0}, B_{0}$ and $C_{0}$ using the relation $(A_{0} + B_{0} + C_{0})=0$.
Discussion {#theoretical_discussion}
----------
A point which merits closer attention is the question which of all the possible steady-state configurations exhibits the lowest undercooling. Whereas the general idea that a eutectic system will always select the state of lowest undercooling is wrong (see Sec. \[sec\_selection\] below), an information about this point constitutes nevertheless a useful starting point. Whereas the general solution to this problem is non-trivial, in the following we present some partial insights.
Let us, for the sake of discussion, first compute the average total curvature undercooling $\Delta T_\kappa$ of an arbitrary arrangement. Consider a configuration of period M having $M_a$ lamella of the $\alpha$ phase, $M_b$ lamella of the $\beta$ phase, and $M_c$ lamella of the $\gamma$ phase, where the integers $M_a$, $M_b$, and $M_c$ add up to M. In a system where all the solid-liquid and solid-solid surface tensions are identical, the total average curvature undercooling $\Delta T_{\kappa}^\nu$ of each phase $\nu$ is, $$\begin{aligned}
\Delta T_{\kappa}^\alpha &=& \Gamma_\alpha \dfrac{2\sin \theta}{\lambda} \dfrac{M_a}{\eta_\alpha}\\
\Delta T_{\kappa}^\beta &=& \Gamma_\beta \dfrac{2\sin \theta}{\lambda} \dfrac{M_b}{\eta_\beta}\\
\Delta T_{\kappa}^\gamma &=& \Gamma_\gamma \dfrac{2\sin \theta}{\lambda} \dfrac{M_c}{\eta_\gamma}.\end{aligned}$$ It is remarkable that the average curvature undercooling is independent of the individual widths of each lamella, but depends only on the total volume fraction and the number of lamellae of the specific phase. Furthermore, it is quite clear from the above examples that the final expression for the global average interface undercooling can always be written in the same form as Eq. (\[JHlaw\]). The second term of this expression (that is, the one proportional to $1/\lambda$) can be computed for the case where all Gibbs-Thomson coefficients and liquidus slopes are equal, and reads $$\begin{aligned}
\dfrac{K_2}{\lambda} &=& \dfrac{\Delta T_{\kappa}^\alpha + \Delta T_{\kappa}^\beta + \Delta T_{\kappa}^\gamma}{3} \nonumber \\
&=& \Gamma \dfrac{2\sin \theta}{3\lambda}\left ( \dfrac{M_a}{\eta_\alpha} + \dfrac{M_b}{\eta_\beta} + \dfrac{M_c}{\eta_\gamma}\right).
\label{average_curvature}\end{aligned}$$ For the special case of a completely symmetric phase diagram and a sample at the eutectic composition, Eqn.(\[average\_curvature\]) yields $$\begin{aligned}
\dfrac{K_2}{\lambda} &=& \Gamma \dfrac{2\sin \theta}{\lambda}\left(M_a+M_b+M_c\right),\end{aligned}$$ where we have used the fact that $\eta_\alpha = \eta_\beta = \eta_\gamma =1/3$. Using, $M_a+M_b+M_c=M$, $\dfrac{K_2}{\lambda} = \Gamma \dfrac{2\sin \theta}{\left(\lambda\right/M)}$. Thus, we see that the magnitude of this term per unit lamella in an arrangement is the same for all the possible arrangements, irrespective of the individual widths of the lamella and the relative positions of the lamellae in a configuration. Moreover, we see that for a general off-eutectic composition, choosing the number of lamellae in the ratio $\eta_\alpha:\eta_\beta:\eta_\gamma$ renders the average curvature undercoolings of all the three phases equal. This condition is, however, relevant only for the special case of identical solid-solid and solid-liquid surface tensions and equal liquidus slopes of the phases. For the case when the solid-liquid and solid-solid surface tensions are unequal, the curvature undercooling is no longer independent of the arrangement of the lamella in the configuration. Hence, the problem of determining the minimum undercooling configuration is complex and no general expression regarding the number, position and widths of lamellae can be derived.
Another point is worth mentioning. Under the assumption that the volume fractions of the solid phases are fixed by the lever rule, the width of the three lamellae in the $\alpha\beta\gamma$ cycle is uniquely fixed by the alloy concentration. However, for the $\alpha\beta\alpha\gamma$ cycle, and more generally for any cycle with $M>3$, this is not the case any more because there have to be at least two lamellae of the same phase in the cycle. Whereas the cumulated width of these lamellae is fixed by the global concentration, the width of each individual lamella is not. For example, in the $\alpha\beta\alpha\gamma$ cycle at the eutectic concentration $c_A^\infty=c_B^\infty=c_C^\infty=1/3$, all the configurations $(\xi,1/3,1/3-\xi,1/3)$ for $0<\xi<1/3$ are admissible, where the notation $(\cdot,\cdot,\ldots)$ is a shorthand for the list of the lamella widths $x_{n+1}-x_n$. The number $\xi$ is an internal degree of freedom that can be freely chosen by the system. With our method, the global front undercooling can be calculated for any value of $\xi$. For the $\alpha\beta\alpha\gamma$ cycle, we found that the configuration with equal widths of the $\alpha$ phases ($\xi=1/6$) was the one with the minimum average front undercooling. This gives a strong indication that this value is stable, and that perturbations of $\xi$ around this value should decay with time. Hence, the analytic expressions given above for the $\alpha\beta\alpha\gamma$ cycle, which are for $\xi=1/6$, should be the relevant ones.
Phase-field Model
=================
Model
-----
A thermodynamically consistent phase-field model is used for the present study [@Garcke; @Stinner]. The equations are derived from an entropy functional of the form $$\begin{aligned}
{\cal
S}\left(e,{\mbox{\boldmath{$c$}}},{\mbox{\boldmath{$\phi$}}}\right)&=&\int_{\Omega}\left(s\left(e,{\mbox{\boldmath{$c$}}},
{\mbox{\boldmath{$\phi$}}}\right)-\left(\epsilon a\left({\mbox{\boldmath{$\phi$}}},\nabla {\mbox{\boldmath{$\phi$}}}\right) +
\dfrac{1}{\epsilon}w\left({\mbox{\boldmath{$\phi$}}}\right)\right)\right)d\Omega,
\label{functional}\end{aligned}$$ where $e$ is the internal energy density, ${\mbox{\boldmath{$c$}}}$ = $(c_{i})_{i=1}^{K}$ is a vector of concentration variables, $K$ being the number of components, and ${\mbox{\boldmath{$\phi$}}}=\left(\phi_{\alpha}\right)_{\alpha=1}^{N}$ is a vector of phase-field variables, $N$ being the number of phases present in the system. ${\mbox{\boldmath{$\phi$}}}$ and ${\mbox{\boldmath{$c$}}}$ fulfill the constraints $$\sum_{i=1}^{K}c_{i}=1 \qquad \mbox{and} \qquad
\sum_{\alpha=1}^{N}\phi_{\alpha}=1,
\label{constraint}$$ so that these vectors always lie in $K-1$- and $N-1$-dimensional planes, respectively. Moreover, $\epsilon$ is the small length scale parameter related to the interface width, $s\left(e,{\mbox{\boldmath{$c$}}},{\mbox{\boldmath{$\phi$}}}\right)$ is the bulk entropy density, $a\left({\mbox{\boldmath{$\phi$}}},\nabla{\mbox{\boldmath{$\phi$}}}\right)$ is the gradient entropy density and $w\left({\mbox{\boldmath{$\phi$}}}\right)$ describes the surface entropy potential of the system for pure capillary-force-driven problems.
We use a multi-obstacle potential for $w\left({\mbox{\boldmath{$\phi$}}}\right)$ of the form $$\begin{aligned}
w\left({\mbox{\boldmath{$\phi$}}}\right)=\left\{
\begin{array}{ll}
\dfrac{16}{\pi^{2}}\displaystyle \sum_{\substack{
\alpha, \beta = 1 \\
(\alpha < \beta)}}^{N, N}
\sigma_{\alpha \beta}\phi_{\alpha}\phi_{\beta}+\displaystyle \sum_{\substack{
\alpha, \beta, \gamma= 1 \\
(\alpha < \beta < \gamma)}}^{N, N, N}
\sigma_{\alpha \beta \gamma}\phi_{\alpha}\phi_{\beta}\phi_{\gamma}, \quad
& \textrm{if} \hspace{0.1cm} {\mbox{\boldmath{$\phi$}}}\in \sum\\
\infty, & \textrm{elsewhere}
\\
\end{array}
\right.
\label{FunctionW}\end{aligned}$$ where $\sum=\{ {\mbox{\boldmath{$\phi$}}}\, | \sum_{\alpha=1}^{N}\phi_{\alpha} = 1$ and $\phi_{\alpha} \geq 0 \}$, $\sigma_{\alpha \beta}$ is the surface entropy density and $\sigma_{\alpha \beta
\gamma}$ is a term added to reduce the presence of unwanted third or higher order phase at a binary interface (see below for details).
The gradient entropy density $a\left({\mbox{\boldmath{$\phi$}}},\nabla {\mbox{\boldmath{$\phi$}}}\right)$ can be written as $$\begin{aligned}
a\left({\mbox{\boldmath{$\phi$}}},\nabla {\mbox{\boldmath{$\phi$}}}\right)=
\begin{array}{ll}
\displaystyle \sum_{\substack{
\alpha, \beta = 1 \\
(\alpha < \beta)}}^{N,N}
\sigma_{\alpha \beta}\left[a_{c}\left(q_{\alpha \beta}\right)\right]^{2}\lvert
q_{\alpha \beta}\lvert^{2},
\end{array}\end{aligned}$$ where $q_{\alpha \beta}=\left(\phi_{\alpha}\nabla
\phi_{\beta}-\phi_{\beta}\nabla \phi_{\alpha}\right)$ is a vector normal to the $\alpha \beta$ interface. The function $a_{c}\left(q_{\alpha \beta}\right)$ describes the form of the anisotropy of the evolving phase boundary. For the present study, we assume isotropic interfaces, and hence $a_{c}\left(q_{\alpha \beta}\right) = 1$. Evolution equations for ${\mbox{\boldmath{$c$}}}$ and ${\mbox{\boldmath{$\phi$}}}$ are derived from the entropy functional through conservation laws and phenomenological maximization of entropy, respectively [@Garcke; @Stinner]. A linearized temperature field with positive gradient $G$ in the growth direction ($z$ axis) is imposed and moved forward with a velocity $v$, $$\begin{aligned}
T = T_{0} + G(z-vt)
\label{Tfield}\end{aligned}$$ where $T_{0}$ is the temperature at $z=0$ at time $t=0$. The evolution equations for the phase-field variables read $$\begin{aligned}
\omega \epsilon \partial_{t}\phi_{\alpha}=\epsilon \left(\nabla \cdot a,_{\nabla
\phi_{\alpha}}\left({\mbox{\boldmath{$\phi$}}},\nabla
{\mbox{\boldmath{$\phi$}}}\right)-a,_{\phi_{\alpha}}\left({\mbox{\boldmath{$\phi$}}},\nabla
{\mbox{\boldmath{$\phi$}}}\right)\right)-\dfrac{1}{\epsilon}w,_{\phi_{\alpha}}
\left({\mbox{\boldmath{$\phi$}}}\right)-\dfrac{f,_{\phi_{\alpha}}({\mbox{\boldmath{$c$}}},{\mbox{\boldmath{$\phi$}}};T)}{T} - \Lambda,
\label{Evolution}\end{aligned}$$ where $\Lambda$ is the Lagrange multiplier which maintains the constraint of Eq. (\[constraint\]) for ${\mbox{\boldmath{$\phi$}}}$, and the constant $\omega$ is the relaxation time of the phase fields. Furthermore, $a,_{\nabla \phi_{\alpha}}$, $a,_{\phi_{\alpha}}$, $w,_{\phi_{\alpha}}$ and $f,_{\phi_{\alpha}}$ indicate the derivatives of the respective entropy densities with respect to $\nabla \phi_\alpha$ and $\phi_\alpha$. The function $f({\mbox{\boldmath{$c$}}},{\mbox{\boldmath{$\phi$}}};T)$ in Eq. (\[Evolution\]) describes the free energy density, and is related to the entropy density $s({\mbox{\boldmath{$c$}}},{\mbox{\boldmath{$\phi$}}};T)$, through the relation $f({\mbox{\boldmath{$c$}}},\phi;T) = e({\mbox{\boldmath{$c$}}}, \phi;T) - Ts({\mbox{\boldmath{$c$}}},{\mbox{\boldmath{$\phi$}}};T)$, where $e({\mbox{\boldmath{$c$}}}, \phi;T)$ is the internal energy density. The free energy density is given by the summation over all bulk free energy contributions $f_\alpha({\mbox{\boldmath{$c$}}};T)$ of the individual phases in the system. We use an ideal solution model, $$\begin{aligned}
f({\mbox{\boldmath{$c$}}},{\mbox{\boldmath{$\phi$}}};T) = \sum_{i=1}^{K}\left(T c_{i}\ln c_{i} +
\sum_{\alpha=1}^{N}c_{i}L_{i}^{\alpha}\dfrac{\left(T-T_{i}^{\alpha}\right)}{T_{i
}^{\alpha}}
h_{\alpha}\left({\mbox{\boldmath{$\phi$}}}\right)\right),
\label{freeen1}\end{aligned}$$ where $$\begin{aligned}
f_{\alpha}({\mbox{\boldmath{$c$}}};T)=\sum_{i=1}^{K} \left( T c_{i}\ln c_{i}+ c_{i}L_{i}^{\alpha}\dfrac{\left(T-T_{i}^{\alpha}
\right)}{T_{i}^{\alpha}} \right)
\label{freeen2}\end{aligned}$$ is the free energy density of the $\alpha$ solid phase, and $$\begin{aligned}
f_{l}({\mbox{\boldmath{$c$}}};T)=T\sum_{i=1}^{K}\left(c_{i}\ln\left(c_{i}\right)\right)
\label{freeen3}\end{aligned}$$ is the one of the liquid. The parameters $L_{i}^{\alpha}$ and $T_{i}^{\alpha}$ denote the latent heats and the melting temperatures of the $i^{th}$ component in the $\alpha$ phase, respectively. We choose the liquid as the reference state, and hence $L_{i}^{l}=0$.
The function $h_\alpha({\mbox{\boldmath{$\phi$}}})$ is a weight function which we choose to be of the form $h_{\alpha}\left({\mbox{\boldmath{$\phi$}}}\right)=\phi_{\alpha}^{2}\left(3-2\phi_{\alpha}\right)$. Thus, $f=f_\alpha$ for $\phi_\alpha=1$. Other interpolation functions involving other components of the ${\mbox{\boldmath{$\phi$}}}$ vector could also be used, but here we restrict ourselves to this simple choice.
The evolution equations for the concentration fields are derived from Eq. (\[functional\]), $$\begin{aligned}
\partial_{t}c_{i}=-\nabla \cdot \left(M_{i0}({\mbox{\boldmath{$c$}}},
{\mbox{\boldmath{$\phi$}}})\nabla\dfrac{1}{T}+\sum_{j=1}^{K}M_{ij}\left({\mbox{\boldmath{$c$}}}, {\mbox{\boldmath{$\phi$}}}\right)\nabla\left(\dfrac{1}{T}
\dfrac{\partial f({\mbox{\boldmath{$c$}}},{\mbox{\boldmath{$\phi$}}};T)}{\partial c_{j}}\right)\right).\end{aligned}$$ By a convenient choice of the mobilities $M_{ij}\left({\mbox{\boldmath{$c$}}}, {\mbox{\boldmath{$\phi$}}}\right)$, self- and interdiffusion in multicomponent systems (including off-diagonal terms of the diffusion matrix) can be modelled. Here, however, we limit ourselves to a diagonal diffusion matrix with all individual diffusivities being equal, which can be achieved by choosing $$\begin{aligned}
M_{ij}({\mbox{\boldmath{$c$}}}, {\mbox{\boldmath{$\phi$}}})=D_i({\mbox{\boldmath{$\phi$}}}) c_{i}\left(\delta_{ij}-c_{j}\right) \\
M_{i0} ({\mbox{\boldmath{$c$}}}, {\mbox{\boldmath{$\phi$}}})= M_{0i}({\mbox{\boldmath{$c$}}}, {\mbox{\boldmath{$\phi$}}}) =
-\sum_{\alpha=1}^{N}\sum_{j=1}^{K}M_{ji}h_{\alpha}\left({\mbox{\boldmath{$\phi$}}}\right)L_{i}^{
\alpha}.\end{aligned}$$ The terms $M_{i0} ({\mbox{\boldmath{$c$}}}, {\mbox{\boldmath{$\phi$}}})= M_{0i}({\mbox{\boldmath{$c$}}}, {\mbox{\boldmath{$\phi$}}})$ are the mobilities for the concentration current of the component $i$ due to a temperature gradient. The diffusion coefficient is taken as a linear interpolation between the phases, $D_i({\mbox{\boldmath{$\phi$}}}) = \sum_{\alpha=1}^{N} D_{i}^{\alpha} \phi_\alpha$, where $D_{i}^{\alpha}$ is the non-dimensionalized diffusion coefficient of the $i^{th}$ component in the $\alpha$ phase, using the liquid diffusivity $D^l$ as the reference, where the diffusivities of all the components in the liquid phase are assumed to be equal. In the simulations we assume zero diffusivity in the solid, and take the effective diffusivity to be $D_i({\mbox{\boldmath{$\phi$}}}) = D^l \phi_l$. The quantity $d^{*}= \sigma / \left(R/v_{m}\right)$ is used as the reference length scale in the simulations, where the molar volume $v_m$ is assumed to be independent of the concentration. Here, $\sigma$ is one of the surface entropy density parameters introduced in Eq. (\[FunctionW\]), and the surface entropies of all the phases are assumed to be equal. The reference time scale is chosen to be $t^* = {d^{*}}^{2}/D^l$. The temperature scale is the eutectic temperature corresponding to the three phase stability regions at the three edges of the concentration simplex and is denoted by $T^*$ while the energy scale is given by $RT^*/v_{m}$.
Relation to sharp-interface theory
----------------------------------
In order to compare our phase-field simulations to the theory outlined in Sec. 2, we need to relate the parameters of the phase-field model to the quantities needed as input for the theory. For some, this is straightforward. For example, all the parameters of the phase diagram (liquidus slopes, coexistence temperatures etc.) can be deduced from the free energy densities of Eqs. (\[freeen1\])–(\[freeen3\]) in the standard way. For others, the correspondence is less immediate. In the following, we will discuss in some detail two quantities that are crucial for the theory: the surface free energies and the latent heats, both needed to calculate the Gibbs-Thomson coefficients in Eq. (\[GibbsThomson\]).
The surface free energy $\tilde\sigma_{\alpha\beta}$ is defined as the interface excess of the thermodynamic potential density that is equal in two coexisting phases. For alloys, this is not the free energy, but the grand potential. Indeed, the equilibrium between two phases is given by $K$ conditions for $K$ components: $K-1$ chemical potentials (because of the constraint of Eq. (\[Eqn-Constraint\]), only $K-1$ chemical potentials are independent) as well as $f-\sum_{i=1}^{K-1} \mu_ic_i$, which is the grand potential, have to be equal in both phases. This is the mathematical expression of the common tangent construction for binary alloys and the common tangent plane construction for ternary alloys.
The grand potential excess has several contributions. Since $f=e-Ts$, we need to consider the entropy excess. Both the gradient term in the phase fields and the potential $w({\mbox{\boldmath{$\phi$}}})$ present in the entropy functional give a contribution inside the interface. If, along an $\alpha\beta$ interface, all the other phase fields remain exactly equal to zero, then this contribution can be calculated analytically. However, this is generally not the case: in the interface, the phase fields $\phi_\nu$, $\nu\ne\alpha,\beta$ can be different from zero, which corresponds to an “adsorption” of the other phases. Since the grand potential excess has to be calculated along the equilibrium profile of the fields, the presence of extra phases modifies the value of $\tilde\sigma_{\alpha\beta}$. The three-phase terms proportional to $\sigma_{\alpha\beta\gamma}$ have been included in the potential function to reduce (or even eliminate) the additional phases. However, the total removal of these phases requires to choose high values of $\sigma_{\alpha \beta \gamma}$. Such high values (>10 times the binary constant $\sigma_{\alpha\beta}$) cause the interface to become steeper near the regions of triple points and lines in 2D and 3D, respectively, which is a natural consequence of the fact that the higher order term affects only the points inside the phase-field simplex where three phases are present. The thinning of the interfaces leads to undesirable lattice pinning, which could only be circumvented by a finer discretization. This, however, would lead to a large increase of the computation times. Therefore, if computations are to remain feasible, we have to accept the presence of additional phases in the interfaces.
Furthermore, there is also a contribution due to the chemical part of the free energy functional. This contribution, identified for the first time in Ref. [@Kim], arises from the fact that the concentrations inside the interface (which are fixed by the condition of constant chemical potentials) do not, in general, follow the common tangent plane, as illustrated schematically in Figure \[Figure6\].
![(Color Online) Illustration of the existence of an excess interface energy contribution from the chemical free energy. Upper panel: the concentration inside the interfacial region does not necessarily follow the common tangent line. Here, the two convex curves are the free energy densities of the individual phases in contact, the straight line is the common tangent, and the thick non-monotonous line is the concentration along a cut through the interface. Lower panel: the grand chemical potential in the interface differs from the one obtained by a weighted sum of the bulk phase free energies, where the weighting coefficients are the interpolating functions of the order parameters.[]{data-label="Figure6"}](figure6.eps){width="8cm"}
Therefore, there is a contribution to the surface free energy which is given by the following expressions. For binary eutectic systems ($N = 3$ phases, ${\mbox{\boldmath{$\phi$}}}=\left(\phi_{\alpha},\phi_{\beta},\phi_{l}\right)$; $K=2$ components ${\mbox{\boldmath{$c$}}}=\left(c_{A},c_{B}\right)$), the vector ${\mbox{\boldmath{$c$}}}$ is one-dimensional and we define the concentration ($c_{A}$) to be the independent field ${\mbox{\boldmath{$c$}}}=\left(c_{A},1-c_{A}\right)$. Then, we have $$\begin{aligned}
{\Delta f}_{\textrm{chem}}\left({\mbox{\boldmath{$\phi$}}},{\mbox{\boldmath{$c$}}};T\right) = f\left({\mbox{\boldmath{$\phi$}}}, {\mbox{\boldmath{$c$}}};
T\right) - f_{l} - \mu_{A}(T)\left(c_{A}-c^{l}_{A}\right),\end{aligned}$$ where $\mu_{A}\left(T\right)=\dfrac{\partial
f\left({\mbox{\boldmath{$\phi$}}},{\mbox{\boldmath{$c$}}};T\right)}{c_{A}}$ is the chemical potential of component A. For ternary eutectic systems ($N = 4$ phases, ${\mbox{\boldmath{$\phi$}}}=
\left(\phi_{\alpha},\phi_{\beta},\phi_{\gamma},\phi_{l}\right)$; $K=3$ components, ${\mbox{\boldmath{$c$}}}=\left(c_{A},c_{B},c_{C}\right)$), the vector ${\mbox{\boldmath{$c$}}}$ is two-dimensional and with the concentrations of $A, B$ as the independent concentration fields, we get ${\mbox{\boldmath{$c$}}}=\left(c_{A},c_{B},1-c_{A}-c_{B}\right)$ and the chemical free energy excess becomes
$$\begin{gathered}
{\Delta f}_{\textrm{chem}}\left({\mbox{\boldmath{$\phi$}}},{\mbox{\boldmath{$c$}}};T\right) = f\left({\mbox{\boldmath{$\phi$}}}, {\mbox{\boldmath{$c$}}};
T\right) - f_{l} - \left(\mu_{A}(T)\right)\left(c_{A}-c^{l}_{A}\right)- \left(\mu_{B}\left(T\right)\right)\left(c_{B}-c^{l}_{B}\right).\end{gathered}$$
The entire surface excess can thus be written as the following $$\begin{aligned}
\widetilde{\sigma}_{\alpha l} &=& \int_{x} \Big (T \epsilon a\left({\mbox{\boldmath{$\phi$}}},\nabla
{\mbox{\boldmath{$\phi$}}}\right) + \frac{T}{\epsilon}{w\left({\mbox{\boldmath{$\phi$}}}\right)} + {\Delta f}_{\textrm{chem}}\left({\mbox{\boldmath{$\phi$}}},{\mbox{\boldmath{$c$}}};T\right)\Big) dx\end{aligned}$$ where $x$ is the coordinate normal to the interface, and the integral is taken along the equilibrium profile ${\mbox{\boldmath{$\phi$}}}(x)$, ${\mbox{\boldmath{$c$}}}(x)$. This integral cannot be calculated analytically. Therefore, we determine the surface free energy numerically. To this end, we perform one-dimensional simulations to determine the equilibrium profiles of concentration and phase fields, and insert the solution into the above formula to calculate $\tilde\sigma$. For these simulations, the known bulk values of the concentration fields are used as boundary conditions. To accurately calculate the surface excesses, it is important to include the contribution of the adsorbed phases. For this, the above calculations are performed by letting a small amount of these phases equilibrate at the interface of the major phases. Since the adsorbed phases equilibrate with very different concentrations compared to that of the bulk phases, the domain is chosen large enough such that the chemical potential change of the bulk phases during equilibration is kept negligibly low.
Another important quantity which is required as an input in the theoretical expressions is the latent heat of fusion $L^{\alpha}$ of the $\alpha$ phase. We follow the thermodynamic definition for the latent heat of transformation $L^{\alpha}$,
$$\begin{aligned}
L^{\alpha} &=& T_{E} \left(s^{l} - s^{\alpha}\right), \\
\textrm{with} \hspace{0.25cm} s &=& -\left(\dfrac{\partial f\left({\mbox{\boldmath{$\phi$}}},{\mbox{\boldmath{$c$}}};T\right)}{\partial T}\right) \\
\textrm{and in particular} \hspace{0.25cm} s^{l} &=& \sum_{i=1}^{K}c^{l}_{i}ln\left(c^{l}_{i}\right)\\
\textrm{and} \hspace{0.25cm} s^{\alpha} &=&
\sum_{i=1}^{K}c^{\alpha}_{i}\dfrac{L_{i}^{\alpha}}{T_{i}^{\alpha}} +
c^{\alpha}_{i}ln\left(c^{\alpha}_{i}\right),\end{aligned}$$
where the concentrations of the phases are taken from the phase diagram at the eutectic temperature.
Finally, let us give a few comments on the interface mobility $\mu_{\rm int}$ that appears in Eq. (\[GibbsThomson\]). In early works [[@Caginalp]]{}, it was shown that an expression for this mobility in terms of the phase-field parameters can be easily derived in the sharp-interface limit in which the interface thickness tends to zero. Later on, Karma and Rappel [@Karma96] proposed the thin-interface limit, in which the interface width remains finite, but much smaller than the mesoscopic diffusion length of the problem. This limit relaxes some of the stringent requirements of the sharp-interface method for the achievement of quantitative simulations. Additionally, this method introduces a correction term to the original expression for the interface mobility, which makes it possible to carry out simulations in the vanishing interface kinetics (infinite interface mobility) regime.
Clearly, such modifications of the interface kinetics are also present in our model, where they arise both from the presence of adsorbed phases in the interface and from the structure of the concentration profile through the interface. Furthermore, it is well known that solute trapping also occurs in phase-field models of the type used here [@Ahmad98]. Since the interface profile can only be evaluated numerically, and since several phase-field and concentration variables need to be taken into account, it is not possible to evaluate quantitatively the contribution of these effects to the interface mobility. However, this lack of knowledge does not decisively impair the present study since we are mainly interested in undercooling versus spacing curves at a fixed interface velocity. At constant velocity, the absolute value of the interface undercooling contains an unknown contribution from the interface kinetics, but the relative comparison between steady states of different spacings remains meaningful. In addition, even though our simulation parameters correspond to higher growth velocities than typical experiments, it will be seen below that the value of the kinetic undercooling in our simulations is small. This indicates once more that our comparisons remain consistent.
Simulation results
==================
In this section, we compare data extracted from phase-field simulations with the theory developed in Sec.\[sec:theory\], for the case of coupled growth of the solid phases in directional solidification. The simulation setup is sketched in Figure \[Figure7\].
![Simulation setup for the phase-field simulations of binary and ternary eutectic systems. We impose a temperature gradient $G$ along the $z$ direction and move it with a fixed velocity. The average interface position follows the isotherms at steady state in case of stable lamellar coupled growth.[]{data-label="Figure7"}](figure7.eps){width="8cm"}
Periodic boundary conditions are used in the transverse direction, while no-flux boundary conditions are used in the growth direction. The box width in the transverse direction directly controls the spacing $\lambda$. The box length in the growth direction is chosen several times larger than the diffusion length. The diffusivity in the solid is assumed to be zero. A non-dimensional temperature gradient, G is imposed in the growth direction and moved with a velocity $v$, such that the temperature field is given by Eq. (\[Tfield\]).
The outline of this section is as follows: first, we will briefly sketch how we extract the front undercooling from the simulation data. Then, this procedure will be validated by comparisons of the results to analytically known solutions as well as to data for binary alloys, for which well-established benchmark results exist. We start the presentation of our results on ternary eutectics by a detailed discussion of the two simplest possible cycles, $\alpha\beta\gamma$ and $\alpha\beta\alpha\gamma$. We compare the data for undercooling as a function of spacing to our analytical predictions and determine the relevant instabilities that limit the range of stable spacings. Finally, we also discuss the behavior of more complicated cycles, for sequences up to length $M=6$.
Data extraction {#Data-extraction}
---------------
At steady state, the interface velocity matches the velocity of the isotherms. The undercooling of the solid-liquid interface is extracted at this stage by the following procedure. First, a vertical line of grid points is scanned until the interface is located. Then, the precise position of the interface is determined as the position of the level line $\phi_\alpha=\phi_\beta$ for an $\alpha\beta$-interface (and in an analogous way for all the other interfaces). This is done by calculating the intersection of the phase-field profiles of the corresponding phases, which are extrapolated to subgrid accuracy by polynomial fits. In the presence of adsorbed phases at the interface, the two major phases along the scan line are used for determining the interface point. The major phases are determined from the maximum values that a particular order parameter assumes along the scan line. The temperature at a calculated interface point is then given by Eq. (\[Tfield\]).
In order to test both our data extraction methods and our calculations of the surface tensions, we have performed the following consistency check. For an alloy with a symmetric phase diagram at the eutectic concentration, a lamellar front has an equilibrium position when a small temperature gradient ($G=0.001$) is applied to the system at zero growth speed. Since the concentration in the liquid is uniform for a motionless front, according to the Gibbs-Thomson relation the interface shapes should just be arcs of circles. This was indeed the case in our simulations, and the fit of the interface shapes with circles has allowed us to obtain the interface curvature and the contact angles with very good precision. The extraction of the data is illustrated in Figure \[Figure8\].
We fit the radius and the coordinates of the circle centers. Then, the angle at the trijunction point $\theta$ is deduced from geometrical relations, with $d = a + b$ and $a = \dfrac{R_a^2 - R_b^2 + d^{2}}{2d}$, $$\begin{aligned}
b = d-a\\
\theta =
\cos^{-1}\left(\dfrac{a}{R_a}\right)+\cos^{-1}\left(\dfrac{b}{R_b}
\right).\end{aligned}$$ The meaning of the lengths $a$ and $b$ is given in Figure \[Figure8\].
Validation: Binary Systems
--------------------------
For comparison with the $\Delta T - \lambda$ relationship known from Jackson-Hunt(JH) theory, we create two binary eutectic systems by choosing suitable parameters $L_{i}^{\alpha}$ and $T_{i}^{\alpha}$ in the free energy density $f\left({\mbox{\boldmath{$\phi$}}},{\mbox{\boldmath{$c$}}};T\right)$. A symmetric binary eutectic system, shown in Figure \[Figure9a\], is created by
To create an asymmetric binary eutectic system, shown in Figure \[Figure9b\], we choose
The numbers $L_{i}^{\alpha}$, $T_{i}^{\alpha}$ are chosen such that the widths of each of the (lens-shaped) two-phase coexistence regions remain reasonably broad, and that the approximation of using the values of concentration difference between the solidus and liquidus $\left(\Delta c_{\nu}^{l}\right)$ at the eutectic temperature for the theoretical expressions holds for a good range of undercoolings. This implies that the value of the $L_{i}^{\alpha}$ should not be too small. Conversely, a too high value is also not desirable since for large values of $L_{i}^{\alpha}$ the chemical contribution to the surface free energy becomes large, which leads to very steep and narrow interface profiles.
\(a) \[Jacktab0505\] (b)
We perform simulations at two different velocities V = 0.01 and V = 0.02, with a mesh size $\Delta x = 1.0$ and the parameter set $ \epsilon = 4.0, D^{l}_{A} = D^{l}_{B} = D^{l}_{C} = D^{l}_{D} = 1.0,
\sigma_{\alpha \beta} = \sigma_{\alpha l} = \sigma_{\beta l} = 1.0,
\sigma_{\alpha \beta \gamma} = 10.0$. To give an idea of the order of magnitude of the corresponding dimensional quantities, we remark that if we assume the melting temperatures to be around 1700K and the other values to correspond to the Ni-Cu system used in the study of Warren [*et al.*]{} [@Warren], the length scale $d^*$ for the case of the binary eutectic system turns out to be around 0.2 nm and the time scale 0.04 ns.
The corresponding parameters for the sharp-interface theory are given in Table \[JacksonBin\]. The comparisons between our numerical results and the analytic theory are shown in Figs. \[Figure10a\] and \[Figure10b\].
Consistent differences can be observed in the undercooling values between our data and the predictions from JH theory for both systems. The difference in undercoolings is smaller at lower velocities, which hints at the presence of interface kinetics. We find indeed that when we change the relaxation constant in the phase-field evolution equation by about 50 %, the difference between the predicted and measured undercoolings is removed for the case of the considered symmetric binary phase diagram. This clearly shows that the interface kinetics is not negligible. It seems difficult, however, to obtain a precise numerical value for its magnitude in the framework of the present model.
The spacing at minimum undercooling, however, is reproduced to a good degree of accuracy (error of 5 %), while the minimum undercooling has a maximum error of 10 %. It should also be noted that the JH theory only is an approximation for the true front undercooling. Results obtained both with boundary integral [@Sarkissian] and quantitative phase-field methods [@Plapp] have shown that, whereas the prediction for the minimum undercooling spacing is excellent, errors of 10 % for the value of the undercooling itself are typical. If the JH curve is drawn without taking into account the additional chemical contributions to the surface tension, a completely different result is obtained, with minimum undercooling spacings that are largely different from the simulated ones. We can therefore conclude that we have captured the principal corrections.
In addition, we have performed equilibrium measurements of the angles at the trijunction point and of the radius of curvature of the lamellae as described in the preceding sub-section (\[Data-extraction\]) for the symmetric eutectic system. The contact angles differ from the ones predicted by Young’s equilibrium conditions only by a value of 0.2 degrees. The theoretical (from the Gibbs-Thomson equation) and measured undercoolings differ in the third decimal, with an error of 0.1 %.
Ternary Systems: Parameter set
------------------------------
\
We use a symmetric ternary phase diagram. The following matrices list the parameters $L_{i}^{\alpha}$,$T_{i}^{\alpha}$ in the free energy $f\left({\mbox{\boldmath{$\phi$}}},{\mbox{\boldmath{$c$}}};T\right)$ that were used to create a symmetric ternary eutectic system, shown in Figure \[Figure1\]. We perform simulations with the parameter set $\epsilon = 8.0, \Delta x = 1.0, D^{l}_{A} =
D^{l}_{B} = D^{l}_{C} = 1.0 , \sigma_{\alpha \gamma} = \sigma_{\beta \gamma} =
\sigma_{\gamma \beta} = \sigma_{\alpha l} = \sigma_{\beta l} = \sigma_{\gamma l}
= 1.0, \sigma_{\alpha \beta l} = \sigma_{\alpha \beta \gamma} =
\sigma_{\alpha \gamma l} = \sigma_{\beta \gamma l} =10.0$ and compare with the theoretical expressions using the input parameters listed in Table \[Jacktabtern\].
$\widetilde{\sigma}_{\alpha l} = \widetilde{\sigma}_{\beta l} = \widetilde{\sigma}_{\gamma l}$ 1.194035
-------------------------------------------------------------------------------------------------------------- ----------
$\widetilde{\sigma}_{\alpha \beta} = \widetilde{\sigma}_{\alpha \gamma} = \widetilde{\sigma}_{\beta \gamma}$ 1.430923
$\theta_{\alpha \beta} = \theta_{\beta \alpha} = \theta_{\gamma \alpha}$ = $\theta_{\alpha \gamma}$ 36.81
$L^{\alpha} = L^{\beta} = L^{\gamma}$ 1.33
$m_{B}^\alpha=m_{C}^{\alpha}$ -0.91
$m_{A}^{\beta} = m_{C}^{\beta}$ -0.91
$m_{A}^{\gamma} = m_{B}^{\gamma}$ -0.91
: Input parameters for the theoretical relations for the ternary eutectic system.
\[Jacktabtern\]
Simple cycles: steady states and oscillatory instability
--------------------------------------------------------
We first perform simulations to isolate the regime of stable lamellar growth for the configuration $\alpha \beta \gamma$. For this regime, we measure the average interface undercooling and compare it to our theoretical predictions. The results are shown in Figure \[Figure11\].
![Comparison between theoretical analysis and phase-field simulations at two different velocities for the arrangement $\left(\alpha \beta
\gamma\right)$ of ternary eutectic solids at $V=0.005$ and $V=0.01$. The demarcation shows the regions of stable lamellar growth and the critical spacing beyond which we observe amplified oscillatory behavior. There is a small region named “Damped Oscillations”, which is a region where oscillations occur but die down slowly with time.[]{data-label="Figure11"}](figure11.eps){width="8cm"}
The agreement in the undercoolings is much better than for the binary eutectic systems, with a smaller dependence of undercoolings on the velocities. Consequently, both the spacing at minimum undercooling (error 4 % for V=0.005 and 6 % for V=0.01) and the minimum undercooling (error of 1-2 %), match very well with the theoretical relationships, as shown in Figure \[Figure11\]. The equilibrium angles at the triple point also agree with the ones predicted from Young’s law to within an error of 0.3 degrees, while the radius of curvature matches that from the Gibbs-Thomson relationship with negligible error (<0.5 %).
It should be noted that the steady lamellae remain straight, contrary to the results of Ref. [@Hecht], where a spontaneous tilt of the lamellae with respect to the direction of the temperature gradient was reported. This difference is due to the different phase diagrams: we are using a completely symmetric phase diagram and equal surface tensions for all solid-liquid interfaces, whereas [@Hecht] uses the thermophysical data of a real alloy.
Next, we are interested in the stability range of three-phase coupled growth. From general arguments, we expect a long-wavelength lamella elimination instability (Eckhaus-type instability) to occur for low spacings, as in binary eutectics [@Akamatsu04]. Here, we will focus on oscillatory instabilities that occur for large spacings. It is useful to first recall a few facts known about binary eutectics, where all the instability modes have been classified [@Sarkissian; @Akamatsu]. Lamellar arrays in binary eutectics are characterized (in the absence of crystalline anisotropy) by the presence of two mirror symmetry planes that run in the center of each type of lamellae, as sketched in Figure \[Figure12\](a). Instabilities can break certain of these symmetries while other symmetry elements remain intact [@Coullet]. In binary eutectics, the oscillatory 1-$\lambda$-O mode is characterized by an in-phase oscillation of the thickness of all $\alpha$ (and $\beta$) lamellae; both mirror symmetry planes remain in the oscillatory pattern. In contrast, in the 2-$\lambda$-O mode, one type of lamellae start to oscillate laterally, whereas the mirror plane in the other type of lamellae survives; this leads to a spatial period doubling. Finally, in the tilted pattern both mirror planes are lost.
![(Color Online) In a periodic arrangement of lamellae, we can identify certain lines of symmetry, as shown in (a) for a binary eutectic. Similarly, for the case of the two simplest configurations, (b) $\alpha \beta \gamma$ and (c) $\alpha \beta \alpha \gamma$ in a symmetric ternary eutectic system, such planes of symmetry exist. While in the case of a binary eutectic, the lines are mirror symmetry axes (shown by dash-dotted lines), in the special case of a symmetric ternary phase diagram, one can also identify quasi-mirror lines (dashed lines) where we retrieve the original configuration after a spatial reflection and an exchange of two phases. Only quasi-mirror lines exist in the $\alpha \beta \gamma$ arrangement, which are shown in (b), while both true- and quasi-mirror planes exist in the $\alpha \beta \alpha \gamma$ arrangement as shown in (c).[]{data-label="Figure12"}](figure12.eps){width="8cm"}
It is therefore important to survey the possible symmetry elements in the ternary case. At first glance, there seems to be no symmetry plane in the pattern. However, for our specific choice of phase diagram, new symmetry elements not present in a generic phase diagram exist: mirror symmetry planes combined with the exchange of two phases. Consider for example the $\beta$ phase in the center of Figure \[Figure12\](b): if the system is reflected at its center, and then the $\alpha$ and $\gamma$ phases are exchanged, we recover the original pattern. At the eutectic concentration, there are three such symmetry planes running in the center of each lamella, and three additional ones running along the three solid-solid interfaces. Off the eutectic point, two of these planes survive if any two of the three phases have equal volume fractions.
Guided by these considerations, we can conjecture that there are two obvious possible instability modes, sketched in Figure \[Figure13\].
In the first, called mode 1 in the following, two symmetry planes survive: the width of one lamella oscillates, whereas the two other phases form a “composite lamella” that oscillates in opposition of phase; the interface in the center of this composite lamella does not oscillate at all and constitutes one of the symmetry planes. In the second (mode 2), the lateral position of one of the lamellae oscillates with time, whereas the other two phases oscillate in opposition of phase to form a “composite lamella” that oscillates laterally but keeps an almost constant width. There is no symmetry plane left in this mode.
The stability range of the coupled growth regime of the lamellar arrangement is indicated in Figure \[Figure11\]. Steady lamellar growth is stable from below the minimum undercooling spacing up to a point where an oscillatory instability occurs. In the region marked “damped oscillations”, oscillatory motion of the interfaces was noticed, but died out with time. Above a threshold in spacing, oscillations are amplified. We monitored the modes that emerged, and found indeed good examples for the two theoretically expected patterns, shown in Figure \[Figure14\].
Mode 1 is favored for off-eutectic concentrations in which one of the lamellae is wider than the two others, such as ${\mbox{\boldmath{$c$}}}=(0.32,0.32,0.36)$. Indeed, in that case the (unstable) steady-state pattern exhibits the same symmetry planes as the oscillatory pattern. This mode can also appear when one lamella is [*smaller*]{} than the two others, see Figure \[Figure14c\]. We detect mode 2 at the eutectic concentration, see Figure \[Figure15b\]. However, a “mixed mode” can also occur, in which no symmetry plane survives, but the three trijunctions oscillate laterally with phase differences that depend on the concentration and possibly on the spacing, see Figs. \[Figure14b\] and \[Figure15c\].
\
Let us now turn to the $\alpha \beta \alpha \gamma$ cycle. We perform simulations for two different velocities $V = 0.01$ and $V = 0.005$. The comparison of the measurements with the theoretical analysis for steady-state growth is shown in Figure \[Figure16\]. For the purpose of analysis, predictions from the theory for both arrangements ($\alpha \beta \gamma$ and $\alpha \beta \alpha \gamma$) are also shown. Here again, the minimum undercooling spacings match those of the theory to a good degree of accuracy (error 5%, V=0.005). However, the undercooling is lower than the one predicted by JH-theory, with a discrepancy of 4% for the case of $V=0.005$, Figure \[Figure16a\]. For $V=0.01$, Figure \[Figure16b\], simulations were not possible for a sufficient range of $\lambda$ to determine the minimum undercooling, because the width of the narrowest lamellae became comparable to the interface width $\epsilon \simeq 8.0$ before the minimum was reached. However, the general trend of the data follows the predictions of the theory for both velocities. This was also the case for simulations carried out at an off-eutectic concentration ${\mbox{\boldmath{$c$}}}= (0.32, 0.34, 0.34)$ at a velocity of $V=0.005$, for the same configuration $\alpha\beta\alpha\gamma$.
Concerning the oscillatory instabilities at large spacings, it is useful to consider again the symmetry elements. For this cycle, there are two real symmetry planes in the steady-state pattern that run through the centers of the $\beta$ and $\gamma$ lamellae. Note that these symmetries would exist even for unsymmetric phase diagrams and unequal surface tensions. Therefore, by analogy with binary eutectics, one may expect oscillatory modes that simply generalize the 1-$\lambda$-O and 2-$\lambda$-O modes of binary eutectics, see Figure \[Figure17a\]. Indeed, for our simulations at the eutectic concentration, we retrieve the 1-$\lambda$-O type oscillation, figure \[Figure18a\] as in our hypothesis (figure \[Figure17a\]).
This oscillatory instability can be quantitatively monitored by following the lateral positions of the solid-solid interfaces with time. More specifically, we extract the width of the $\beta$ phase as a function of the growth distance $z$. This is then fitted with a damped sinusoidal wave of the type $A_{0}+A\exp(-Bz)\cos((2\pi z/L)+D)$. The damping coefficient $B$ is obtained from a curve fit and plotted as a function of the spacing $\lambda$. The onset of the instability is characterized by the change in sign of the damping coefficient.
For the off-eutectic concentration we get two modes (figure \[Figure18\]). While (figure \[Figure18b\]) corresponds well to our hypothesis to the 2-$\lambda$-O type oscillation (figure \[Figure17b\]), we also observe another mode as shown in figure \[Figure18c\], which combines elements of the two modes: both the width and the lateral position of the $\alpha$ lamellae oscillate.
Lamella elimination instability
-------------------------------
For the $\alpha\beta\alpha\gamma$ cycle, there is also a new instability, which occurs for low spacings. We find that all spacings below the minimum undercooling spacing, as well as some spacings above it, are unstable with respect to lamella elimination: the system evolves to the $\alpha \beta \gamma$ arrangement by eliminating one of the $\alpha$ lamellae, both at eutectic and off-eutectic concentrations. The points plotted to the left of the minimum in Figure \[Figure16b\] are actually unstable steady states that can be reached only when the simulation is started with strictly symmetric initial conditions and the correct volume fractions of the solid phases.
This instability can actually be well understood using our theoretical expressions. As already mentioned before, when we consider the cycle $\alpha \beta \alpha \gamma$ at the eutectic concentration with a lamella width configuration $(\xi , 1/3 , 1/3 -\xi, 1/3 )$, the global average front undercooling attains a minimum for the symmetric pattern $\xi=1/6$. However, the global front undercooling is not the most relevant information for assessing the front stability. More interesting is the undercooling of an [*individual*]{} lamella, because this can give information about its evolution. More precisely, consider the undercooling of one of the $\alpha$ lamellae as a function of $\xi$. If the undercooling increases when the lamella gets thinner, then the lamella will fall further behind the front and will eventually be eliminated. In contrast, if the undercooling decreases when the lamella gets thinner, then the lamella will grow ahead of the main front and get larger. A similar argument has been used by Jackson and Hunt for their explanation of the long-wavelength elimination instability [@JH]. It should be pointed out that the new instability found here is not a long-wavelength instability, since it can occur even when only one unit cell of the cycle is simulated.
Following the above arguments, we have calculated the growth temperature of the first $\alpha$ lamella as a function of $\xi$ using the general expressions in Eq. \[abag-1lamella\]. In Figure \[Figure19\], we plot the variation of $\partial \Delta T/\partial \xi$ at $\xi = 1/6$, as a function of $\lambda$. The point at which $\partial \Delta T/\partial \xi$ becomes positive then indicates the transition to a stable $\alpha\beta\alpha\gamma$ cycle. This criterion is in good agreement with our simulation results. This argument can also be generalized to more complicated cycles (see below).
![Plot of $\partial \Delta T/\partial \xi$, taken at $\xi=1/6$ versus $\lambda$ for the $\alpha_{1} \beta \alpha_{2} \gamma$ cycle, where $\Delta T$ is the undercooling of the $\alpha_1$ lamella and $\xi$ its width (relative to $\lambda$), calculated by our analytical expressions in the volume fraction configuration. $(\xi , 1/3 , 1/3 -\xi, 1/3 )$ at V=0.01. The cycle is predicted to be unstable to lamella elimination if $\partial \Delta T/\partial \xi<0$. The $\lambda$ at which $\partial \Delta T/\partial \xi$ changes sign is the critical point beyond which the $\alpha \beta \alpha \gamma$ arrangement is stable with respect to a change to the sequence $\alpha \beta \gamma$ through a lamella elimination.[]{data-label="Figure19"}](figure19.eps){width="8cm"}
Longer cycles
-------------
Let us now discuss a few more complicated cycles. The simple cycles we have simulated until now were such that during stable coupled growth the widths of all the lamellae corresponding to a particular phase were the same. This changes starting from period $M=5$, where the configuration $\alpha \beta \alpha \beta \gamma$ is the only possibility (up to permutations). If we consider this cycle at the eutectic concentration and note the configuration of lamella widths as $(\xi,1/3-\xi,1/3-\xi,\xi,1/3)$ and compute the average front undercooling by our theoretical expressions, we find that the minimum occurs for $\xi$ close to $0.12$. In addition, for this configuration, the undercooling of any asymmetric configuration (permutation of widths of lamellae) is higher than the one considered above. If we rewrite symbolically this configuration as $\alpha_1,\beta_2,\alpha_2,\beta_1,\gamma$, it is easy to see that this configuration has two symmetry axes of the same kind as discussed in the preceding subsection: mirror reflection and exchange of the phases $\alpha$ and $\beta$. One of them runs along the interface between $\beta_2$ and $\alpha_2$, and the other one in the center of the $\gamma$ lamella.
Not surprisingly, our simulation results confirm the importance of this symmetry. The volume fractions in steady-state growth are close to those that give the minimum average front undercooling, see Figs. \[Figure20b\] and \[Figure20c\].
Additionally, we observe oscillations in the width of the largest $\gamma$ phase and oscillations in the widths and the lateral position of the smaller lamellae of the $\alpha$ and $\beta$ phases, while the interface between the larger $\alpha$ and $\beta$ phase remains straight, such that the combination of all the $\alpha$ and $\beta$ lamellae oscillates in width as one “composite lamella”. Thus, the symmetry elements of the underlying steady state are preserved in the oscillatory state.
For smaller spacings, this configuration is unstable, and the sequence changes to the $\beta\alpha\gamma$ arrangement as shown in Figure \[Figure20a\] by two successive lamella eliminations. It is noteworthy that we did not find any unstable sequence which switches to the $\alpha\beta\alpha\gamma$ arrangement, which again can be understood from the presence of the symmetry. Indeed, a symmetrical evolution would result in a change to a configuration $\alpha_1\beta_1\gamma$ or $\beta_2\alpha_2\gamma$, but precludes the change to a configuration of period length $M=4$.
Going on to cycles with period $M=6$, the first arrangement we consider is $\alpha_1 \beta_1 \alpha_2 \beta_2 \alpha_1 \gamma$, where we name the lamellae for eventual discussion and ease in description according to the symmetries. Indeed, this arrangement has two exact mirror symmetry planes in the center of the $\alpha_2$ and the $\gamma$ phases. We find that, if we calculate the average interface undercooling curves by varying the widths of individual lamella with the constraint of constant volume fraction, by choosing different $\xi$, in the width configuration $(\xi,1/6,1/3-2\xi,1/6,\xi,1/3)$, the average undercooling at the growth interface is minimal for the configuration $(1/9,1/6,1/9,1/6,1/9,1/3)$. This arrangement has the highest undercooling curve among the arrangements we have considered, shown in Figure \[Figure21a\].
It also has a very narrow range of stability, and we could isolate only one spacing which exhibits stable growth for $\lambda = 240$, Figure \[Figure22d\]. Unstable arrangements near the minimum undercooling spacing evolve to the $\alpha \beta \gamma$ arrangement, Figure \[Figure22a\], while for other unstable configurations we obtain the arrangements in Figure \[Figure22b\] and Figure \[Figure22c\] as the stable growth forms corresponding to $\lambda=150$ and $\lambda=180$ respectively.
Apart from the (trivial) period-doubled arrangement $\alpha \beta \gamma \alpha \beta \gamma$, another possibility for $M=6$ is $\alpha \beta \gamma \alpha \gamma \beta$ with a volume fraction configuration $(1/6,1/6,1/6,1/6,1/6,1/6)$. Simulations of this arrangement show that there exists a reasonably large range of stable lamellar growth, and hence we could make a comparison between simulations and the theory. We find similar agreement between our measurements and theory as we did previously for the arrangements $\alpha\beta\gamma$ and $\alpha\beta\alpha\gamma$. The plot in Figure \[Figure21\] shows the theoretical predictions of all the arrangements we have considered until now.
Discussion {#discussion}
----------
It should by now have become clear that there exists a large number of distinct steady-state solution branches, each of which can exhibit specific instabilities. In addition, the stability thresholds potentially depend on a large number of parameters: the phase diagram data (liquidus slopes, coexistence concentration), the surface tensions (assumed identical here), and the sample concentration. Therefore, the calculation of a complete stability diagram that would generalize the one for binary eutectics of Ref. [@Sarkissian] represents a formidable task that is outside the scope of the present paper. Nevertheless, we can deduce from our simulations a few guidelines that can be useful for future investigation.
Lamellar steady-state solutions can be grouped into three classes, which respectively have (I) equal number of lamellae of all three phases (such as $\alpha\beta\gamma$ and $\alpha\beta\gamma\alpha\gamma\beta$), (II) equal number of lamellae for two phases (such as $\alpha\beta\alpha\gamma$), and (III) different numbers of lamellae for each phase.
For equal global volume fractions of each phase (as in most of our simulations), class III will have the narrowest stability ranges because of the simultaneous presence of very large and very thin lamella in the same arrangement, which make these patterns prone to both oscillatory and lamella elimination instabilities.
Any cycle in which a phase appears more than once can transit to another, simpler one by eliminating one lamella of this phase. This lamella instability always appears for low spacings below a critical value of the spacing that depends on the cycle. The possibility of a transition, however, depends also on the symmetries of the pattern. For instance, the arrangement $\alpha \beta \alpha \beta \alpha \gamma$, if unstable, can transform into the $\alpha \beta \gamma$, $\alpha \beta \alpha \gamma$ or the $\alpha \beta \alpha \beta \gamma$ arrangements, while for an arrangement $\alpha \beta \alpha \beta \gamma$, it is impossible to evolve into the $\alpha \beta \alpha \gamma$ arrangement if the symmetry of the pattern is preserved by the dynamics.
For large spacings, oscillatory instabilities occur and can lead to the emergence of saturated oscillatory patterns of various structures. The symmetries of the steady states seem to determine the structure of these oscillations, but no thorough survey of all possible nonlinear states was carried out.
Some remarks on pattern selection {#sec_selection}
=================================
Up to now, we have investigated various regular periodic patterns and their instabilities. The question which, if any, of these different arrangements, is favored for given growth conditions, is still open. From the results presented above, we can already conclude that this question cannot be answered solely on the basis of the undercooling-vs-spacing curves. Indeed, we have shown that by appropriately choosing the initial conditions, any stable configuration can be reached, regardless of its undercooling. This is also consistent with experiments and simulations on binary eutectics [@Akamatsu; @Parisi10]. To get some additional insights on what happens in extended systems, we conducted some simulations of isothermal solidification where the initial condition was a random lamellar arrangement. More precisely, we initialize a large system with lamellae of width $\lambda=25$ and choose a random sequence of phases such that two neighboring lamellae are of different phases as shown in Figure \[Figure23a\]. The global probabilities of all the phases are $1/3$, which corresponds to the eutectic concentration, and the temperature is set to $T=0.785$.
Under isothermal growth conditions, one would expect that, at a given undercooling, the arrangement with highest local velocity would be the one that is chosen. However, in order for the front to adopt this pattern, a rearrangement of the phase sequence is necessary. In our simulations, we find that lamella elimination was possible (and indeed readily occurred). In contrast, there is no mechanism for the creation of new lamellae in our model, since we did not include fluctuations that could lead to nucleation, and the model has no spinodal decomposition that could lead to the spontaneous formation of new lamellae, as in Ref. [@Plapp+02]. As a result, some of the lamellae became very large in our simulations, which led to a non-planar growth front, as shown in Figure \[Figure23b\]. No clearcut periodic pattern emerged, such that our results remain inconclusive.
We believe that lamella creation is an important mechanism required for pattern adjustment. In 2D, nucleation is the only possibility for the creation of new lamellae. In contrast, in 3D, new lamellae can also form by branching mechanisms without nucleation events, since there are far more geometrical possibilities for two-phase arrangements [@Moulinet; @Walker]. Therefore, we also conducted a few preliminary simulations in 3D.
The cross sections of the simulated systems are $150 \times 150$ grid points for results in Figs. \[Figure24\](a) and (c), and $90 \times 90$ grid points for the system Figure \[Figure24\](b). The longest run took about 7 weeks on 80 processors, for the simulation of the pattern in Figure \[Figure24\](a). This long simulation time is due to the fact that the pattern actually takes a long time to settle down to a steady state; the total solidification distance was of the order of 800 grid points. The other simulations required less time to reach reasonably steady states. The patterns shown in Figs. \[Figure24\] (a) and (c) start from random initial conditions (very thin rods of randomly assigned phases), the former with the symmetric phase diagram used previously, the latter with a slightly asymmetric phase diagram constructed with the changed parameters listed below,
The picture of Figure \[Figure24\](b) corresponds to a pattern resulting of a simulation which is started with two isolated rods of $\alpha$ and $\beta$ in a matrix of $\gamma$, with an off-eutectic concentration of ${\mbox{\boldmath{$c$}}}=(0.3,0.4,0.3)$.
As shown in Figure \[Figure24\], many different steady-state patterns are possible in 3D. Not surprisingly, the type of pattern seen in the simulations depends on the concentration and on the phase diagram. Patterns very similar to Figure \[Figure24\](c) have recently been observed in experiments in the Al-Ag-Cu ternary system [@Ratke]. It should be stressed that our pictures have been created by repeating the simulation cell four times in each direction in order to get a clearer view of the pattern. This means that in a larger system, the patterns might be less regular. Furthermore, we certainly have not exhausted all possible patterns. A more thorough investigation of the 3D patterns and their range of stability is left as a subject for future work.
Conclusion and outlook
======================
In this paper, we have generalized a Jackson-Hunt analysis for arbitrary periodic lamellar three-phase arrays in thin samples, and used 2D phase-field simulations to test our predictions for the minimum undercooling spacings of the various arrangements. For the model used here the value of the interface kinetic coefficient cannot be determined, which leads to some incertitude on the values of the undercooling, but this does not influence our principal findings. When the correct values of the surface free energy (that take into account additional contributions coming from the chemical part of the free energy density) are used for the comparisons with the theory, we find good agreement for the minimum undercooling spacings for all cycles investigated. Moreover, we find that, as in binary eutectics, all cycles exhibit oscillatory instabilities for spacings larger than some critical spacing. The type of oscillatory modes that are possible are determined by the set of symmetry elements of the underlying steady state.
We have repeatedly made use of symmetry arguments for a classification of the oscillatory modes. In certain cases, the symmetry is exact and general, which implies that the corresponding modes should exist for arbitrary phase diagrams and thus be observable in experiments. For instance, the mirror symmetry lines in the middle of the $\alpha$ lamellae in the $\alpha\beta\alpha\gamma$ arrangement exist even for non-symmetric phase diagrams and unequal surface tensions, and hence the corresponding oscillatory patterns and their symmetries should be universal. In other cases, we have used a symmetry element which is specific to the phase diagram used in our simulations: a mirror reflection, followed by an exchange of two phases. For a real alloy, this symmetry obviously can never be exactly realized because of asymmetries in the surface tensions, mobilities, and liquidus slopes, and therefore some of the oscillatory modes found here might not be observable in experiments. However, their occurrence cannot be completely ruled out without a detailed survey, and we expect certain characteristics to be quite robust. For instance, we have repeatedly observed that two neighboring lamellae of different phases can be interpreted as a “composite lamella” that exhibits a behavior close to the one of a single lamella in a binary eutectic pattern. Such behavior could appear even in the absence of special symmetries, and thus be generic.
Furthermore, a new type of instability (absent in binary eutectics) was found, where a cycle transforms into a simpler one by eliminating one lamella. We interpret this instability, which occurs for small spacings, through a modified version of our theoretical analysis. It is linked to the existence of an extra degree of freedom in the pattern if a given phase appears more than once in the cycle. We have not determined the full stability diagram that would be the equivalent of the one given in Ref. [@Sarkissian] for binary eutectics, because of the large number of independent parameters involved in the ternary problem.
We have made a few attempts to address the question of pattern selection, with inconclusive results both in 2D and 3D. In 2D, the process of pattern adjustment was hindered by the absence of a mechanism for lamella creation, and in 3D the system sizes that could be attained were too small. Based on the findings for binary eutectics, however, we believe that there is no pattern selection in the strong sense: for given processing conditions, the patterns to be found may well depend on the initial conditions and/or on the history of the system. This implies that the arrangement with the minimum undercooling may not necessarily be the one that emerges spontaneously in large-scale simulations or in experiments.
The most interesting direction of research for the future is certainly a more complete survey of pattern formation in 3D and a comparison to experimental data. To this end, either the numerical efficiency of our existing code has to be improved, or a more efficient model that generalizes the model of Ref. [@Plapp] to ternary alloys has to be developed.
Acknowledgements
================
This work was financially supported by three sources: Centre National d’Etudes Spatiales (France), CCMSE (Center for Materials Science and Engineering) funded by the state of Baden-Wüttemberg, Germany and the European Fond for Regional Development (EFRE), and the DFG (German Research Foundation -project number NE822/14-1).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show that on integral normal separated schemes whose function field is separably closed, for each pair of points the intersection of the resulting local schemes is local. This extends a result of Artin from rings to schemes. The argument relies on the existence of certain modifications in inverse limits. As an application, we show that Čech cohomology coincides with sheaf cohomology for the Nisnevich topology. Along the way, we generalize the characterization of contractible curves on surfaces by negative-definiteness of the intersection matrix to higher dimensions, using bigness of invertible sheaves on non-reduced schemes.'
address: 'Mathematisches Institut, Heinrich-Heine-Universität, 40204 Düsseldorf, Germany'
author:
- Stefan Schröer
title: Total separable closure and contractions
---
Introduction {#introduction .unnumbered}
============
[\[Introduction\]]{}
This paper deals with the Zariski topology for a class of schemes that are in general highly non-noetherian, yet arise from noetherian schemes in a canonical way: We say that an integral scheme $X$ is *totally separably closed* if it is normal and the function field $F={\mathscr{O}}_{X,\eta}=\kappa(\eta)$ is separably closed. As abbreviation one also says that $X$ is a *TSC scheme*. Each integral scheme $X_0$ has a total separable closure $X={\operatorname{TSC}}(X_0)$, defined as the integral closure with respect to a chosen separable closure $F=F_0^{{\operatorname{sep}}}$ of the function field $F_0$. Such schemes $X$ are *everywhere strictly local*. In other words, all local rings ${\mathscr{O}}_{X,x}$, $x\in X$ are strictly local rings, that is, henselian with separably closed residue field. One may regard them as analogues of *Prüfer schemes*, where all local rings are valuation rings.
TSC schemes have some relevance with respect to the étale topology. Indeed, M. Artin [@Artin; @1971] used them to prove that Čech cohomology equals sheaf cohomology for the étale topology over affine schemes $X={\operatorname{Spec}}(R)$. This result immediately extends to schemes with the *AF property*, which means that any finite subset admits an affine open neighborhood. Note, however, that by [@Benoist; @2013], Corollary 2, the AF property is equivalent to quasiprojectivity for normal schemes that are separated and of finite type over a ground field. Actually, Artin used algebraic closure rather that separable closure, but this makes no difference for the underlying topological spaces. See Huneke’s overview [@Huneke; @2011] for the role of absolute integral closure in commutative algebra.
One crucial step in Artin’s arguments is to show that affine integral TSC schemes $X$ have the following surprising property, which is of purely topological nature: For any pair of points $u,v\in X$ the intersection of local schemes $${\operatorname{Spec}}({\mathscr{O}}_{X,u})\cap{\operatorname{Spec}}({\mathscr{O}}_{X,v})\subset X$$ remains a local scheme. If we endow the underlying set $X$ with the order relation $x\leq y \Leftrightarrow x\in\overline{\{y\}}$, the above property means that the supremum $\sup(u,v)$ exists for all pairs of points $u,v\in X$. This strange property almost never holds on noetherian schemes $X_0$, and intuitively means that in inverse limits $X={\varprojlim}X_\lambda$, common generizations of $u_\lambda,v_\lambda\in X_\lambda$ are totally “ripped apart”. In some sense, this is a topological incarnation of the result of Schmidt that a field with two different henselian valuations is separably closed ([@Schmidt; @1933], Satz 3. See [@Engler; @and; @Prestel; @2005], Theorem 4.4.1 for a modern account). The main goal of this paper is to establish Artin’s result in full generality:
[**(See Theorem \[local\])**]{} For any separated integral TSC scheme $X$, the intersections ${\operatorname{Spec}}({\mathscr{O}}_{X,u})\cap{\operatorname{Spec}}({\mathscr{O}}_{X,v})\subset X$ are local for all points $u,v\in X$.
In [@Schroeer; @2017], I already obtained this for total separable closures of schemes $X_0$ that are separated and of finite type over a ground field $k$. The arguments rely on modifications and contractions in inverse limits $X={\varprojlim}X_\lambda$, and do not apply in mixed characteristics. Here we modify our approach, and reduce the problem to proper schemes over excellent Dedekind domains. We then use different modifications $X'$ and contractions ${\tilde{X}}$ in inverse limits so that Artin’s result applies to the TSC scheme ${\tilde{X}}$, which is constructed to have the AF property. This is enough to conclude for the original TSC scheme $X$.
To carry this out, we have to analyze the existence of suitable *modifications* and *contractions*. On algebraic surfaces $X$, a curve $E=E_1+\ldots+E_r$ is contractible to points if and only if the intersection matrix $\Phi=(E_i\cdot E_j)$ is negative-definite. This observation goes back to Mumford, Artin and Deligne, in various forms of generality. Note that in general the contractions $r:X{\rightarrow}Y$ yield algebraic spaces rather than schemes. The following generalization to higher dimensions seems to be of independent interest:
[**(See Theorem \[big\])**]{} Let $X$ be a normal scheme that is proper over an excellent Dedekind domain $R$, and $E=E_1+\ldots+E_r$ be a Weil divisor contained in a closed fiber for the structure morphism $X{\rightarrow}{\operatorname{Spec}}(R)$. If $E$ is is contractible to points, then for each effective Cartier divisor $D=\sum m_iE_i$, the invertible sheaf ${\mathscr{O}}_D(-D)$ on $D$ is big.
Here bigness for an invertible sheaf ${\mathscr{L}}$ on some proper algebraic scheme $Z$, which is not necessarily reduced or irreducible, is defined in terms of the *Iitaka dimension*, which itself is given, up to a shift, by the Krull dimension of the ring $R(Z,{\mathscr{L}})=\bigoplus_{n\geq 0} H^0(Z,{\mathscr{L}}^{\otimes n})$. This generalization from integral to arbitrary schemes was analyzed by Cutkosky [@Cutkosky; @2014], and his results on the multiplicity or volume ${\operatorname{mult}}({\mathfrak{a}}_\bullet)={\operatorname{vol}}({\mathfrak{a}}_\bullet)$ for *graded families of ideals*, together with Huneke’s version [@Huneke; @1992] of the Briançon–Skoda Theorem, play a crucial role for the above.
As explained in [@Schroeer; @2017], our main result on TSC schemes has immediate consequences for the *Nisnevich topology* of completely decomposed étale maps [@Nisnevich; @1989]. This is a variant of the étale topology, where the local rings are henselian local rings rather than strictly local rings. We get:
[**(See Theorem \[cech\])**]{} For each quasicompact separated scheme $X$ and every abelian Nisnevich sheaf $F$, the canonical maps $$\check{H}^p_{\text{\rm Nis}}(X,F){\longrightarrow}H^p_{\text{\rm Nis}}(X,F)$$ from Čech cohomology to sheaf cohomology are bijective in every degree $p\geq 0$.
The paper is organized as follows: Section 1 contains our results on contractions for proper schemes over excellent Dedekind domains. The main result about TSC schemes is given in Section 2. The final Section 3 gives the application to Nisnevich cohomology.
Contractions over Dedekind domains
==================================
[\[Contractions\]]{}
Let $S={\operatorname{Spec}}(R)$ be the spectrum of a Dedekind domain $R$, and $X$ be an proper $S$-scheme. We write $f:X{\rightarrow}S$ for the structure morphism. For simplicity, we assume that the scheme $X$ is integral and that the ring $R$ is excellent. Note that we do not assume flatness; in particular, the structure morphism may factor over some closed point $\sigma\in S$.
A closed subscheme $E\subset X$ is said to be *contractible to points* if there is a commutative diagram $$\xymatrix{
X\ar[rr]^r\ar[dr]_f & & Y\ar[dl]^g\\
& S &\\
}$$ where $Y$ is an algebraic space, the structure morphism $g:Y{\rightarrow}S$ is proper, and $r:X{\rightarrow}Y$ is a morphism with ${\mathscr{O}}_Y=r_*({\mathscr{O}}_X)$ that is an open embedding on $X\smallsetminus E$ such that the image $Z=r(E)$ consists of finitely many closed points. Their images in $S$ are closed as well, and it follows that the connected components of $E$ are contained in closed fibers for the structure morphism $f:X{\rightarrow}S$. The morphism $r:X{\rightarrow}Y$ is unique up to unique isomorphism, and depends only on the underlying closed set for $E\subset X$, which follows from [@EGA; @II], Lemma 8.11.1.
Algebraic spaces can be glued along open subsets in the same way as ringed spaces (a consequence from [@Olsson; @2016], Proposition 5.2.5). In particular, the closed subset $E\subset X$ is contractible to points if and only if each connected component is *contractible to a single point*. Using Stein factorization, one easily sees the following permanence property:
[\[pullback\]]{} Assume $X'$ is another proper $S$-scheme that is integral, and let $X'{\rightarrow}X$ be a morphism. Suppose that a closed subset $E\subset X$ is contractible to points, and that the morphism $X'{\rightarrow}X$ is finite over $X\smallsetminus E$. Then the preimage $E'=E\times_XX'$ is contractible to points. Indeed, if $X{\rightarrow}Y$ is the contraction of $E\subset Y$, then the Stein factorization $Y'$ for the composition $X'{\rightarrow}Y$ is the contraction of $E'\subset X'$.
For general closed subsets $E\subset X$ it is often difficult to verify contractibility. However, by applying the previous result to the blowing-up $X'{\rightarrow}X$ with center $E$ one reduces to the case of effective Cartier divisors. Then more can be said:
[\[contractible\]]{} Suppose $E\subset X$ is an effective Cartier divisor contained in some closed fiber $X_\sigma=f^{-1}(\sigma)$. Let ${\mathscr{L}}={\mathscr{O}}_X(-E)$. If the restriction ${\mathscr{L}}|E$ is ample then the closed subset $E\subset X$ is contractible to points.
This immediately follows from Corollary 6.10 in Artin’s work [@Artin; @1970] on algebraic stacks: It suffices to treat the case that $E$ is connected, and we need to check two conditions. The first condition is straightforward: for every coherent sheaf ${\mathscr{F}}$ on $E$ the cohomology group $H^1(E,{\mathscr{F}}\otimes{\mathscr{L}}^{\otimes n})$ vanishes for all $n\gg 0$, because ${\mathscr{L}}|E$ is ample. The second conditions is somewhat more intricate: Since $E$ is proper, connected and contained in a closed fiber $X_\sigma=f^{-1}(\sigma)$, the rings $R_n=H^0(X,{\mathscr{O}}_{nE})$ are finite local artinian $R$-algebras. Write $k_n=R_n/{\mathfrak{m}}_{R_n}$ for their residue fields. The inclusions $E\subset 2E\subset\ldots$ induce an an inverse system $k_1\supset k_2\supset\ldots$ of fields, all of which contain the residue field $\kappa=\kappa(\sigma)$ and have finite degree. Let $k=\bigcap k_n$ be their intersection, and choose an index $n_0$ so that the inclusions $k_{n+1}\subset k_n$ are equalities for all $n\geq n_0$. Consider the resulting morphism $E{\rightarrow}{\operatorname{Spec}}(k)$ and, for each $n\geq 0$, the cartesian diagram $$\begin{CD}
R_n\times_{k_1} k @>>> k\\
@VVV @VVV\\
R_n @>>> k_1.
\end{CD}$$ Artin’s second condition stipulates that the upper vertical arrows must be surjective. To see this, choose some index $m\geq \max(n_0,n)$. Then $k_m=k$ and the residue class map $R_m{\rightarrow}k_m=k$ factors over the fiber product $R_n\times_{k_1}k$, so the projection in question is surjective.
The converse does not hold: For example, if $X$ is regular of dimension $d=2$, and $E=E_1+E_2$ is a curve with two irreducible components, having intersection matrix $\Phi=(\begin{smallmatrix}-2&1\\1&-2 \end{smallmatrix})$. The latter is negative-definite, so the curve $E$ is contractible. The linear combination $D=3E_1+E_2$ is contractible as well, yet ${\mathscr{O}}_D(-D)$ is not ample, because $(D\cdot E_2)=-1$.
Nevertheless, it is natural to ask for some form of converse. Indeed, we shall establish such a result based on the notion of bigness rather then ampleness. Let us recall the relevant definitions: Suppose $Z$ is a proper scheme over some ground field $k$. Given an invertible sheaf ${\mathscr{L}}$ on $Z$ we get a graded ring $R(Z,{\mathscr{L}})=\bigoplus_{n\geq 0} H^0(Z,{\mathscr{L}}^{\otimes n})$, which is is not necessarily of finite type or noetherian. Let $d=\dim(R)$ be its Krull dimension. The *Iitaka dimension* or *Kodaira–Iitaka dimension* is defined as $$\kappa({\mathscr{L}})=\begin{cases}
d-1 & \text{if $d\geq 1$;}\\
-\infty & \text{else.}
\end{cases}$$ Note that it will be crucial to allow reducible and non-reduced schemes $Z$ for what we have in mind. For integral normal schemes $Z$, the Iitaka dimension is a classical notion from birational geometry: If some ${\mathscr{L}}^{\otimes n_0}$ with $n_0\geq 1$ has a non-zero global section, the number $\kappa({\mathscr{L}})$ can also be seen as the maximal dimension of the images for the rational maps $X\dashrightarrow {\mathbb{P}}^m$ defined by ${\mathscr{L}}^{\otimes n}$, where $m=h^0({\mathscr{L}}^{\otimes n})-1$ and $n\geq 1$ runs over the positive multiplies of $n_0$. We refer to the monograph of Lazarsfeld ([@Lazarsfeld; @2004], Section 2.2) for more details. Iitaka dimension was only recently extended to arbitrary proper schemes, by the work of Cutkosky on asymptotics of ideals and linear series. In fact, in [@Cutkosky; @2014], Section 7 he defined it in the more general context of *graded linear series*, which can be seen as graded subrings $L=\bigoplus_{n\geq 0} L_n$ inside $R(Z,{\mathscr{L}})$.
According to [@Cutkosky; @2014], Lemma 7.1 we have $\kappa({\mathscr{L}})\leq \dim(Z)$ for arbitrary proper schemes $Z$. In case of equality $\kappa({\mathscr{L}})=\dim(Z)$ one says that the invertible sheaf ${\mathscr{L}}$ is *big*. We need the following observation on invertible sheaves that are not big:
[\[devissage\]]{} Set $d=\dim(Z)$. If ${\mathscr{L}}$ is not big, then for each coherent sheaf ${\mathscr{F}}$ on $Z$ there is a real constant $\beta\geq 0$ so that $h^0({\mathscr{F}}\otimes{\mathscr{L}}^{\otimes n})\leq \beta n^{d-1}$ for all integers $n\geq 0$.
This is a devissage argument similar to [@EGA; @IIIa], Theorem 3.1.2. Let ${\operatorname{Coh}}(Z)$ be the abelian category of all coherent sheaves ${\mathscr{F}}$ on $Z$, and ${\mathcal{C}}\subset{\operatorname{Coh}}(Z)$ be the subcategory of all sheaves for which the assertion holds. If $0{\rightarrow}{\mathscr{F}}'{\rightarrow}{\mathscr{F}}{\rightarrow}{\mathscr{F}}''{\rightarrow}0$ is a short exact sequence, the resulting long exact sequence immediately gives the following implications: $$\label{implications}
{\mathscr{F}}\in{\mathcal{C}}\Longrightarrow {\mathscr{F}}'\in{\mathcal{C}}{\quad\text{and}\quad}{\mathscr{F}}',{\mathscr{F}}''\in{\mathcal{C}}\Longrightarrow{\mathscr{F}}\in{\mathcal{C}}.$$ Furthermore, ${\mathcal{C}}$ contains all coherent sheaves with $\dim({\mathscr{F}})\leq d-1$, according to [@Cutkosky; @2014], Lemma 7.1. Let $Z_1,\ldots,Z_r\subset Z$ be the irreducible components, endowed with the reduced scheme structure. By [@Cutkosky; @2014], Lemma 10.1 combined with Lemma 9.1 we have $$\kappa({\mathscr{L}})=\max\{\kappa({\mathscr{L}}|Z_1),\ldots,\kappa({\mathscr{L}}|Z_r)\}.$$ In light of [@Cutkosky; @2014], Corollary 9.3 it follows that ${\mathscr{O}}_{Z_i}\in{\mathcal{C}}$, and hence ${\mathscr{O}}_{Z'}\in{\mathcal{C}}$ for every reduced closed subscheme $Z'\subset Z$. Note that the cited Corollary was formulated for projective rather than proper schemes, but the proof holds true without changes in the more general setting.
To proceed, we first suppose that ${\mathscr{F}}$ is a torsion-free coherent ${\mathscr{O}}_{Z_i}$-module, say of rank $r\geq 0$. Let $\eta\in Z_i$ be the generic point, choose a bijection ${\mathscr{F}}_{\eta}\simeq\kappa(\eta)^{\oplus r}$, and let ${\mathscr{F}}'$ be the resulting intersection ${\mathscr{F}}\cap{\mathscr{O}}_{Z_i}^{\oplus r}$ inside the sheaf ${\mathscr{M}}_{Z_i}^{\oplus r}$, where ${\mathscr{M}}_{Z_i}$ denotes the quasicoherent sheaf of meromorphic functions. Then ${\mathscr{F}}'$ is coherent and contained in both ${\mathscr{O}}_{Z_i}^{\oplus r}$ and ${\mathscr{F}}$. The quotient ${\mathscr{F}}''={\mathscr{F}}/{\mathscr{F}}'$ has dimension $\leq d-1$, thus ${\mathscr{F}}''\in{\mathcal{C}}$. Using ${\mathscr{F}}'\subset{\mathscr{O}}_{Z_i}^{\oplus r}$ we infer with that ${\mathscr{F}}'$ and thus ${\mathscr{F}}$ is contained in ${\mathcal{C}}$.
Next let ${\mathscr{F}}$ be an an ${\mathscr{O}}_{Z_{{\operatorname{red}}}}$-module, and write ${\mathscr{F}}_i$ for the restriction ${\mathscr{F}}|Z_i$ modulo its torsion subsheaf. Then we have a short exact sequence $$0{\longrightarrow}{\mathscr{F}}'{\longrightarrow}{\mathscr{F}}{\longrightarrow}\bigoplus_{i=1}^r{\mathscr{F}}_i.$$ The term on the right lies in ${\mathcal{C}}$, by the preceding paragraph, and thus also the subsheaf ${\mathscr{F}}''={\mathscr{F}}/{\mathscr{F}}'$. The term on the left has $\dim({\mathscr{F}}')\leq d-1$, thus lies in ${\mathcal{C}}$, which again by gives ${\mathscr{F}}\in {\mathcal{C}}$.
Finally, let ${\mathscr{F}}$ be arbitrary, and ${\mathscr{I}}={\operatorname{Nil}}({\mathscr{O}}_Z)$ be the nilradical, say with ${\mathscr{I}}^m=0$. In the short exact sequences $0{\rightarrow}{\mathscr{I}}^n{\mathscr{F}}/{\mathscr{I}}^{n+1}{\mathscr{F}}{\rightarrow}{\mathscr{F}}/{\mathscr{I}}^{n+1}{\mathscr{F}}{\rightarrow}{\mathscr{F}}/{\mathscr{I}}^n{\mathscr{F}}{\rightarrow}0$, the term on the left is annihilated by ${\mathscr{I}}$, whence lies in ${\mathcal{C}}$. Using induction on $n\geq 0$, one sees that the ${\mathscr{F}}/{\mathscr{I}}^n{\mathscr{F}}\in{\mathcal{C}}$. The case $n=m$ yields ${\mathscr{F}}\in{\mathcal{C}}$.
It is easy to characterize bigness in dimension one:
[\[big on curves\]]{} Suppose the proper scheme $Z$ is equidimensional, of dimension $d=1$. Then the invertible sheaf ${\mathscr{L}}$ is big if and only if $({\mathscr{L}}\cdot Z')>0$ for some irreducible component $Z'\subset Z$. In particular, this holds if $\deg({\mathscr{L}})>0$.
We have $h^0({\mathscr{L}}^{\otimes n})\geq \chi({\mathscr{L}}^{\otimes n}) = \deg({\mathscr{L}}) n + \chi({\mathscr{O}}_Z)$, where the degree is by definition $\deg({\mathscr{L}})=\chi({\mathscr{L}})-\chi({\mathscr{O}}_Z)$. If ${\mathscr{L}}$ is not big, Lemma \[devissage\] implies $\deg({\mathscr{L}})\leq 0$. Now suppose that $({\mathscr{L}}\cdot Z')>0$ for some irreducible component $Z'\subset Z$. By the above, the restriction ${\mathscr{L}}|Z'$ is big. According to [@Cutkosky; @2014], Lemma 9.1 combined with Lemma 10.1, the invertible sheaf ${\mathscr{L}}$ must be big. Conversely, assume ${\mathscr{L}}$ is big. Then there is some irreducible component $Z'\subset Z$ such that ${\mathscr{L}}|Z'$ is big. By [@Cutkosky; @2014], Lemma 7.1 we have $h^0({\mathscr{L}}^{\otimes n}|Z')\geq \alpha n$ for some real constant $\alpha>0$, and it follows $({\mathscr{L}}\cdot Z')>0$.
We now come to our converse for Proposition \[contractible\]:
[\[big\]]{} Suppose that $X$ is normal, and let $E\subset X$ be an effective Weil divisor that is contractible to points, with irreducible components $E_1,\ldots,E_r\subset E$. Suppose $D=\sum m_iE_i$ is a non-zero effective Cartier divisor supported on $E$, and let ${\mathscr{L}}={\mathscr{O}}_X(-D)$. Then the restriction ${\mathscr{L}}|D$ is big.
It suffices to treat the case that $E$ is connected. Let $r:X{\rightarrow}Y$ be the contraction, and $y=r(E)$ be the resulting closed point. Write $k=\kappa(y)$ for the residue field, choose a separable closure $k^{{\operatorname{sep}}}$ and consider the resulting geometric point $\bar{y}:{\operatorname{Spec}}(k^{{\operatorname{sep}}}){\rightarrow}Y$ and the ensuing strictly local ring ${\mathscr{O}}_{Y,\bar{y}}$. We now replace the scheme $Y$ by the spectrum of ${\mathscr{O}}_{Y,\bar{y}}$, and $X$ by the fiber product $X\times_Y{\operatorname{Spec}}({\mathscr{O}}_{Y,\bar{y}})$. This brings us into the situation that the scheme $Y$ is the spectrum of a strictly local excellent ring $R$, and $r:X{\rightarrow}Y$ is a proper morphism with $R=H^0(X,{\mathscr{O}}_X)$ that is an open embedding on $X\smallsetminus E$ and maps $E$ to the closed point $y\in Y$. Since the formal fibers of the excellent scheme ${\operatorname{Spec}}(R)$ are geometrically regular, me may base-change to the formal completion and assume that the local noetherian ring $R$ is complete.
Consider the short exact sequences $0{\rightarrow}{\mathscr{O}}_D(-nD){\rightarrow}{\mathscr{O}}_{(n+1)D}{\rightarrow}{\mathscr{O}}_{nD}{\rightarrow}0$, for each integer $n\geq 0$. The term on the left is ${\mathscr{L}}^{\otimes n}|D$, and we get a short exact sequence $$\label{infinitesimal}
0{\longrightarrow}H^0(D,{\mathscr{L}}^{\otimes n}|D){\longrightarrow}H^0(X,{\mathscr{O}}_{(n+1)D}) {\longrightarrow}H^0(X,{\mathscr{O}}_{nD}).$$ The schematic images for the morphisms $nD{\rightarrow}{\operatorname{Spec}}(R)$ are of the form ${\operatorname{Spec}}(R/{\mathfrak{a}}_n)$, for some inverse system of local Artin rings $R/{\mathfrak{a}}_n$. It yields a descending chain ${\mathfrak{a}}_1\supset{\mathfrak{a}}_2\supset\ldots$ of ${\mathfrak{m}}_R$-primary ideals given by $${\mathfrak{a}}_n=\{g\in R\mid g{\mathscr{O}}_X\subset{\mathscr{O}}_X(-nD)\}.$$ From this description we see that these ideals form a *graded family of ideals* in the sense of [@Cutkosky; @2014], that is, ${\mathfrak{a}}_m\cdot{\mathfrak{a}}_n\subset{\mathfrak{a}}_{m+n}$ for all $m,n\geq 0$. In other words, the subset $\bigoplus {\mathfrak{a}}_nT^n\subset R[T]$ is a subring, which one may call the *Rees ring* for the graded family of ideals. Since the complete local ring $R$ is reduced, the limit $$\alpha={\operatorname{mult}}({\mathfrak{a}}_\bullet)={\operatorname{vol}}({\mathfrak{a}}_\bullet)=\lim_{n{\rightarrow}\infty}\frac{{\operatorname{length}}(R/{\mathfrak{a}}_n)}{n^d}$$ exists as a real number by [@Cutkosky; @2014], Theorem 4.7. Here $d=\dim(R)$, and the number $\alpha$ is called the *multiplicity* or *volume* of the graded family of ideals. One should think of it as a generalization of the classical *Hilbert–Samuel multiplicities* $e({\mathfrak{b}},R)$, which is defined in terms of the graded family of ideal powers ${\mathfrak{b}}_n={\mathfrak{b}}^n$.
We now compute this number in two ways. For the first computation, we describe the ideals ${\mathfrak{a}}_n$ in terms of valuations: Let $x_i\in E_i$ be the generic points. Since $X$ is normal, the local rings ${\mathscr{O}}_{X,x_i}$ are discrete valuation rings. Let $v_i:F^\times{\rightarrow}{\mathbb{Z}}$ be the corresponding normalized valuation on the field of fractions $F={\operatorname{Frac}}(R)=\kappa(\eta)$, where $\eta\in X$ is the generic point. Then $${\mathfrak{a}}_n=\{g\in R\mid v_1(g)\geq nm_1,\ldots, v_r(g)\geq nm_r \}.$$ This reveals that the ideals ${\mathfrak{a}}_n$ are integrally closed: Indeed, the codimension one points $x_1,\ldots,x_r\in X$ admit a common affine neighborhood $U={\operatorname{Spec}}(A)$, according to [@Gross; @2012], Theorem 1.5. Write ${\mathfrak{p}}_1,\ldots,{\mathfrak{p}}_r\subset A$ for the corresponding prime ideals of height one. Then the localization $A'=S^{-1}A$ is a semilocal Dedekind domain, for the multiplicative system $S=A\smallsetminus({\mathfrak{p}}_1\cup\ldots\cup{\mathfrak{p}}_r)$. We see $${\mathfrak{a}}_n = R\cap ({\mathfrak{p}}_1^{nm_1}A'\cap\ldots\cap{\mathfrak{p}}_r^{nm_r}A'),$$ and this is integrally closed according to [@Swanson;; @Huneke; @2006] Proposition 6.8.1 together with Remark 1.1.3 (8). Setting ${\mathfrak{b}}={\mathfrak{a}}_1$, we moreover have ${\mathfrak{b}}^n\subset {\mathfrak{a}}_n$, and infer that the ideal ${\mathfrak{a}}_n$ is the integral closure of the ideal ${\mathfrak{b}}^n$.
According to the Briançon–Skoda Theorem in Huneke’s form [@Huneke; @1992], Theorem 4.13, there is an integer $l\geq 0$ so that ${\mathfrak{a}}_n\subset {\mathfrak{b}}^{n-l}$ for all $n\geq l$. Note that this is already a consequence from Izumi’s Theorem as given by Hübl and Swanson [@Huebl;; @Swanson; @2001], Theorem 1.2. It follows that ${\operatorname{length}}(R/{\mathfrak{a}}_n)\geq {\operatorname{length}}(R/{\mathfrak{b}}^{n-l})$. Passing to the limit, we obtain $$\alpha = \lim_{n{\rightarrow}\infty}\frac{{\operatorname{length}}(R/{\mathfrak{a}}_n)}{n^d} \geq \lim_{n{\rightarrow}\infty}\left(\frac{{\operatorname{length}}(R/{\mathfrak{b}}^{n-l})}{(n-l)^d} \cdot \frac{(n-l)^d}{n^d}\right).$$ Indeed, both factors in the sequence on the right converge. The second factor converges to one, whereas the first factor tends to the Hilbert–Samuel multiplicity $e({\mathfrak{b}},R)$. But such Hilbert–Samuel multiplicities are always integers $e\geq 1$, according to [@AC; @8-9], Chapter VIII, §4, No. 3. The upshot is that $\alpha\geq 1$.
Seeking a contradiction, we now assume that the restriction ${\mathscr{L}}|D$ is not big, and compute the number $\alpha$ in another way. Recall that $\dim(X)=\dim(R)=d$, such that $\dim(E)=d-1$. According to Lemma \[big\], we have $h^0({\mathscr{L}}^{\otimes n}|D)\leq \beta n^{d-2}$ for some real constant $\beta>0$. By definition of the ideals ${\mathfrak{a}}_n$, there are commutative diagrams $$\begin{CD}
0 @>>> {\mathfrak{a}}_{n+1} @>>> R @>>> H^0(X,{\mathscr{O}}_{(n+1)D})\\
@. @VVV @VVV @VVV\\
0 @>>> {\mathfrak{a}}_n @>>> R @>>> H^0(X,{\mathscr{O}}_{nD})
\end{CD}$$ with exact rows. Combining with the exact sequence , we see that the kernels for the surjection $R/{\mathfrak{a}}_{n+1}{\rightarrow}R/{\mathfrak{a}}_n$ are vector subspaces $V_n\subset H^0(D,{\mathscr{L}}^{\otimes n}|D)$. Inductively, we infer that $${\operatorname{length}}(R/{\mathfrak{a}}_n)\leq \sum_{i=0}^{n-1}\dim(V_i)\leq \sum_{i=0}^{n-1} h^0({\mathscr{L}}^{\otimes i}|E) \leq \beta\sum_{i=0}^{n-1}i^{d-2} \leq \gamma n^{d-1}$$ for some real constant $\gamma\geq 0$. This in turn gives $$\alpha = \lim_{n{\rightarrow}\infty}\frac{{\operatorname{length}}(R/{\mathfrak{a}}_n)}{n^d} \leq \lim_{n{\rightarrow}\infty} \frac{\gamma n^{d-1}}{n^d} =0,$$ contradiction.
If $X$ is a regular 2-dimensional scheme, with a curve $E\subset X$ that is contractible to a point, the intersection matrix $\Phi=(E_i\cdot E_j)$ is negative-definite, according to Mumford [@Mumford; @1961] in the complex case, Artin [@Artin; @1962] for algebraic surfaces, and Deligne [@SGA; @7b], Exposé X, Corollary 1.8 in the arithmetic situation. So for every non-zero effective Cartier divisor $D=\sum m_iE_i$, we have $D^2<0$, and thus the restriction ${\mathscr{L}}|D$ of the invertible sheaf ${\mathscr{L}}={\mathscr{O}}_X(-D)$ is big, according to Proposition \[big on curves\]. From this point of view, the preceding result can be seen as a generalization from dimension $d=2$ to higher dimensions.
Now back to our general setting $f:X{\rightarrow}S={\operatorname{Spec}}(R)$. Let $E\subset X$ be a closed subset that is contractible to points. If the proper algebraic space $Y$ resulting from the contraction $X{\rightarrow}Y$ admits an ample invertible sheaf, that is, comes from a projective scheme, we say that a closed subset $E\subset X$ is *projectively contractible to points*. This is a rather delicate condition that cannot be determined numerically in general.
The following is a variant of [@Schroeer; @2017], Theorem 10.2. The new feature is that we have a *ground ring* rather than a ground field, and that the contraction is *projective*. This extension will be essential for the application in the next section.
[\[modification\]]{} Let $E\subset X$ be an effective Cartier divisor that is contained in some closed fiber $X_\sigma=f^{-1}(\sigma)$. Furthermore, suppose that the structure morphism $f:X{\rightarrow}S$ is projective. Then there is an effective Cartier divisor $Z\subset E$ with the following property: On the blowing-up $g:X'{\rightarrow}X$ with center $Z\subset X$, the strict transform $E'\subset X'$ of $E\subset X$ becomes projectively contractible points. Moreover, we could choose $Z$ disjoint from any given finite subset $\{x_1,\ldots,x_m\}\subset E$.
Choose a very ample invertible sheaf ${\mathscr{L}}$ on the projective scheme $X$ so that there is a non-zero global section $s_0\in H^0(X,{\mathscr{L}})$ that does not vanish at any of the finitely many points in ${\operatorname{Ass}}({\mathscr{O}}_E)\cup\{x_1,\ldots,x_m\}$. Then the map $s_0:{\mathscr{O}}_E{\rightarrow}{\mathscr{L}}|E$ is injective, and bijective at the points $x_1,\ldots,x_m\in E$. The section $s_0$ defines an effective Cartier divisor $D\subset X$, and the intersection $Z=D\cap E$ remains Cartier in $E$. Replacing ${\mathscr{L}}$ and $s_0$ by suitable tensor powers, we may assume that ${\mathscr{L}}(-E)$ is very ample. Such a closed subscheme $Z\subset X$ is the desired center:
The exceptional divisor for the blowing-up $g:X'{\rightarrow}X$ is the effective Cartier divisor $g^{-1}(Z)$. Since $X'$ is integral and $g:X'{\rightarrow}X$ is dominant, the preimages $g^{-1}(D)$ and $g^{-1}(E)$ are Cartier as well. Write $ D',E'\subset X'$ for the strict transforms of $D$ and $E$, respectively. Since the center $Z$ is Cartier on $E$, the universal property of blowing-ups gives an $X$-morphism $E{\rightarrow}X'$ whose schematic image is the strict transform $E'$. In the same way, we have an $X$-morphism $D{\rightarrow}X'$ with image $D'$. Indeed, for each point $z\in Z$, let $f_1,f_2\in {\mathscr{O}}_{X,z}$ be generators for the respective stalks of the ideal sheaves ${\mathscr{O}}_X(-E),{\mathscr{O}}_X(-D)\subset{\mathscr{O}}_X$. By assumption, they form a regular sequence. According to [@Matsumura; @1980], Theorem 27 on page 98, they remain a regular sequence in the opposite order, which implies that the subscheme $Z$ is indeed Cartier in both $E$ and $D$.
According to [@Perling;; @Schroeer; @2017], Lemma 4.4 the strict transforms $D',E'\subset X'$ are Cartier, with $$g^{-1}(D) = D'+g^{-1}(Z){\quad\text{and}\quad}g^{-1}(E)= E'+g^{-1}(Z)$$ as Cartier divisors on $X'$. In particular, we have ${\mathscr{O}}_{E'}(-E')={\mathscr{O}}_E(Z-E)$ with respect to the identification $E'=E$. The latter sheaf is ample on $E$, because the sheaf ${\mathscr{L}}(-E)={\mathscr{O}}_X(D-E)$ is ample on $X$. By Theorem \[contractible\], the Cartier divisor $E'\subset X'$ is contractible to points. Let $r:X'{\rightarrow}{\tilde{X}}$ be the resulting contraction, where ${\tilde{X}}$ is a proper algebraic space.
It remains to construct an ample invertible sheaf on ${\tilde{X}}$. By the very definition of $X'={\operatorname{Proj}}(\bigoplus_{i\geq 0}{\mathscr{I}}^i)$ as a relative homogeneous spectrum, where ${\mathscr{I}}\subset{\mathscr{O}}_X$ is the ideal sheaf for the center $Z\subset X$, we have an invertible sheaf ${\mathscr{O}}_{X'}(1)={\mathscr{O}}_{X'}(-g^{-1}(Z))$ that is relatively ample for the blowing-up $g:X'{\rightarrow}X$. Consider the invertible sheaf $${\mathscr{L}}'=g^*({\mathscr{L}})(1)={\mathscr{O}}_{X'}(g^{-1}(D) - g^{-1}(Z)) ={\mathscr{O}}_{X'}(D').$$ We claim that $D'$ is disjoint from $E'$. This is well-known ([@Hartshorne; @1977], Chapter II, Exercise 7.12), but to fix ideas we provide an argument: Since $X'\smallsetminus g^{-1}(Z)=X\smallsetminus Z$, it suffices to check that $D'\cap g^{-1}(Z)$ and $E'\cap g^{-1}(Z)$ are disjoint. The inclusion $Z\subset D$ gives a commutative diagram $$\begin{CD}
0 @>>> {\mathscr{O}}_X(-D) @>>> {\mathscr{O}}_X @>>> {\mathscr{O}}_D @>>> 0\\
@. @VVV @VV{{\operatorname{id}}}V @VVV\\
0 @>>> {\mathscr{I}}@>>> {\mathscr{O}}_X @>>> {\mathscr{O}}_Z @>>> 0,\\
\end{CD}$$ and the Snake Lemma yields $0{\rightarrow}{\mathscr{O}}_X(-D){\rightarrow}{\mathscr{I}}{\rightarrow}{\mathscr{O}}_D(-Z){\rightarrow}0$. The term on the right is the ideal sheaf for the Cartier divisor $Z\subset D$, which coincides with ${\mathscr{O}}_D(-E)$. Restricting to $Z$ results in the short exact sequence $$0{\longrightarrow}{\mathscr{O}}_Z(-D){\longrightarrow}{\mathscr{I}}/{\mathscr{I}}^2{\longrightarrow}{\mathscr{O}}_Z(-E){\longrightarrow}0.$$ Applying this reasoning to the inclusion $Z\subset E$, we infer that the above sequence splits, and obtain a direct sum decomposition ${\mathscr{I}}/{\mathscr{I}}^2={\mathscr{O}}_Z(-D)\oplus{\mathscr{O}}_Z(-E)$. Following Grothendieck’s Convention, we regard sections $\Sigma=\sigma(Z)$ for the ${\mathbb{P}}^1$-bundle $${\mathbb{P}}({\mathscr{I}}/{\mathscr{I}}^2)={\operatorname{Proj}}{\operatorname{Sym}}({\mathscr{I}}/{\mathscr{I}}^2)=g^{-1}(Z){\longrightarrow}Z$$ as invertible quotients $\varphi:{\mathscr{I}}/{\mathscr{I}}^2{\rightarrow}{\mathscr{N}}$, via the identification $\Sigma ={\operatorname{Proj}}{\operatorname{Sym}}({\mathscr{N}})$. In the direct sum decomposition ${\mathscr{I}}/{\mathscr{I}}^2={\mathscr{O}}_Z(-D)\oplus{\mathscr{O}}_Z(-E)$, the first projection corresponds to the section $E'\cap g^{-1}(Z)$, whereas the second projection comes from $D'\cap g^{-1}(Z)$. It follows that the two sections are indeed disjoint. Hence ${\mathscr{L}}'$ is trivial in some open neighborhood of $E'$, and consequently ${\mathscr{L}}'=r^*({\tilde{{\mathscr{L}}}})$ for some invertible sheaf ${\tilde{{\mathscr{L}}}}$ on the algebraic space ${\tilde{X}}$.
Next, we verify that ${\tilde{{\mathscr{L}}}}$ is globally generated. Since the center $Z\subset X$ is locally of complete intersection, we may apply [@SGA; @6], Exposé VII, Lemma 3.5 together with the Projection Formula and obtain an identification $f_*({\mathscr{L}}')={\mathscr{I}}{\mathscr{L}}$. Since also $r_*({\mathscr{O}}_{X'})={\mathscr{O}}_{{\tilde{X}}}$, we arrive at the identifications $$\label{global sections}
H^0({\tilde{X}},{\tilde{{\mathscr{L}}}})=H^0(X',{\mathscr{L}}')=\{ \text{global sections $s$ of ${\mathscr{L}}$ with $s_Z=0$}\}.$$ The first identification reveals that the base-locus for the invertible sheaf ${\tilde{{\mathscr{L}}}}$ must be contained in the image $r(D')$. The exact sequence $0{\rightarrow}{\mathscr{L}}(-E){\rightarrow}{\mathscr{L}}{\rightarrow}{\mathscr{L}}_E{\rightarrow}0$ on the original scheme $X$ yields a long exact sequence $$\label{exact}
0{\longrightarrow}H^0(X,{\mathscr{L}}(-E))\stackrel{t_0}{{\longrightarrow}} H^0(X,{\mathscr{L}}){\longrightarrow}H^0(E,{\mathscr{L}}_E).$$ where $t_0\in H^0(X,{\mathscr{O}}_X(E))$ is the canonical section defining the inclusion $E\subset X$. Now recall that the sheaf ${\mathscr{L}}(-E)$ very ample and that the Cartier divisor $D\subset X$ is defined by a global section $s_0\in H^0(X,{\mathscr{L}})$. For each point $x\in D\smallsetminus Z=D\smallsetminus E$ we may choose a global section $t\in H^0(X,{\mathscr{L}}(-E))$ with $t(x)\neq 0$. In turn, the global section $s_1= tt_0$ of ${\mathscr{L}}$ also has $s_1(x)=t(x)t_0(x)\neq 0$, thus it defines an effective Cartier divisor $D_1\subset X$ with $x\not\in D_1$ and $Z\subset D_1$. Under the identification , the resulting section $s_1$ of ${\tilde{{\mathscr{L}}}}$ does not vanish at $r(x')\in{\tilde{X}}$, where $g(x')=x$.
Now suppose we have a point $z\in Z$. The corresponding point $z'\in g^{-1}(Z)\cap D'$ on $X'$ is an invertible quotient of ${\mathscr{I}}/{\mathscr{I}}^2\otimes\kappa(z)={\mathscr{I}}\otimes\kappa(z)$, whence defines a tangent vector at $z\in X$ not contained in $Z$, that is, a closed subscheme $T\subset{\operatorname{Spec}}({\mathscr{O}}_{X,z})$ of length two, satisfying $s_0|T=0$ and $t_0|T\neq 0$. Since ${\mathscr{L}}(-E)$ is very ample, we may choose a global section $t$ with $t(z)\neq 0$. As above, the global section $s_1= tt_0$ of ${\mathscr{L}}$ vanishes on $Z$ but not on $T$, thus defines a Cartier divisor $D_1'\subset X'$ that does not contain the point $z'$. Under the identification , the global section $s_1$ of ${\tilde{{\mathscr{L}}}}$ does not vanish at $r(z')\in{\tilde{X}}$. Summing up, we have shown that the sheaf ${\tilde{{\mathscr{L}}}}$ is globally generated.
The last step is to check that the globally generated invertible sheaf ${\tilde{{\mathscr{L}}}}$ is ample. Let ${\tilde{C}}\subset {\tilde{X}}$ be an integral curve contained in the fiber ${\tilde{X}}_\sigma$ for the structure morphism ${\tilde{X}}{\rightarrow}S$. We merely have to show that $({\tilde{{\mathscr{L}}}}\cdot {\tilde{C}})>0$. Let $C'\subset X'$ be the strict transform, such that $C'\not\subset E'$ and $({\tilde{{\mathscr{L}}}}\cdot {\tilde{C}})=({\mathscr{L}}'\cdot C')$. If $g(C')\subset X$ is a point, that is, $C'$ is a fiber of the ${\mathbb{P}}^1$-bundle $g^{-1}(Z)={\mathbb{P}}({\mathscr{I}}/{\mathscr{I}}^2)$, we have $({\mathscr{L}}'\cdot C')=({\mathscr{O}}_{X'}(1)\cdot C')\geq 1$.
Now suppose that the image $C=g(C')$ is a curve rather then a point. If $C$ is not contained in the center $Z\subset X$, fix a closed point $x\in C\smallsetminus Z=C\smallsetminus E$ and a global section $t$ for the very ample sheaf ${\mathscr{L}}(-E)$ that vanishes at $x\in C$ but not at the generic point $\eta\in C$. The resulting global section $s_1=tt_0$ of ${\mathscr{L}}$, via the exact sequence , vanishes along $\{x\}\cup Z$ but not at $\eta\in C$. From this we infer $({\tilde{{\mathscr{L}}}}\cdot {\tilde{C}})>0$.
It remains to treat the case that $ C\subset Z$. Let $\nu:B{\rightarrow}C'$ be the normalization map, and form the fiber product $$P=B\times_XX' = B\times_C{\mathbb{P}}({\mathscr{I}}/{\mathscr{I}}^2\otimes{\mathscr{O}}_C) = {\mathbb{P}}({\mathscr{E}}),$$ where the locally free sheaf ${\mathscr{E}}={\mathscr{L}}_1\oplus{\mathscr{L}}_2$ on $B$ is the sum of the two invertible sheafs ${\mathscr{L}}_1=q^*{\mathscr{O}}_C(-D)$ and ${\mathscr{L}}_2=q^*{\mathscr{O}}_C(-E)$. Here $q:B{\rightarrow}C$ is the composition of the normalization map $\nu:B{\rightarrow}C'$ and the induced map $g:C'{\rightarrow}C$. The two projections ${\operatorname{pr}}_i:{\mathscr{E}}{\rightarrow}{\mathscr{L}}_i$ correspond to disjoint sections $\Sigma _i\subset {\mathbb{P}}({\mathscr{E}})$, via $\Sigma _i={\operatorname{Proj}}{\operatorname{Sym}}({\mathscr{L}}_i)$. By functoriality of the construction, these $\Sigma _1$ and $\Sigma _2$ are the preimages of the sections $E'\cap g^{-1}(Z)$ and $D'\cap g^{-1}(Z)$ for $g^{-1}(Z)={\mathbb{P}}({\mathscr{I}}/{\mathscr{I}}^2){\rightarrow}Z$ with respect to the canonical morphism $B\times_XX'{\rightarrow}X'$. In particular, ${\mathscr{O}}_{P}(\Sigma_2)$ is the pullback of ${\mathscr{L}}'$.
The scheme $P={\mathbb{P}}({\mathscr{E}})$ is a ruled surface over the proper regular curve $B$, so its Picard group modulo numerical equivalence takes the form $N(P)={\mathbb{Z}}^2$. The pseudoeffective cone $$\overline{\operatorname{NE}}(P)\subset N(P)\otimes_{\mathbb{Z}}{\mathbb{R}}={\mathbb{R}}^2$$ must be generated by two extremal rays. Each fiber $F\subset P$ for the ruling has selfintersection number $F^2=0$. According to Proposition \[pullback\], the section $\Sigma_1\subset P$ is contractible, so its selfintersection is $(\Sigma_1)^2<0$. In light of [@Kollar; @1995], Lemma 4.12, it follows that the numerical classes of $F$ and $\Sigma_1$ are the two extremal rays for $\overline{\operatorname{NE}}(P)$. This in turn implies $(\Sigma_2)^2>0$. In particular, ${\mathscr{O}}_P(\Sigma_2)$ is ample on $\Sigma_2$. According to Fujita’s result ([@Fujita; @1983], see also [@Ein; @2000]), the invertible sheaf ${\mathscr{O}}_P(\Sigma_2)$ must be semiample. It follows that for some $n\geq 1$ the semiample sheaf ${\mathscr{O}}_P(n\Sigma_2)$ is the preimage of some ample sheaf on ${\tilde{P}}$, where $P{\rightarrow}{\tilde{P}}$ is the contraction of $\Sigma_1$. Consequently $(\Sigma_2\cdot \Sigma)>0$ for every integral curve $\Sigma\neq \Sigma_1$. In particular, this holds for the section $\Sigma\subset P$ arising from the diagonal map $B{\rightarrow}B\times_X X'=P$. By construction, the projection $P{\rightarrow}X'$ induces a surjection $B=\Sigma{\rightarrow}C'$. Since ${\mathscr{O}}_P(\Sigma_2)$ is the preimage of ${\mathscr{L}}'={\mathscr{O}}_{X'}(D')$, we must have $({\mathscr{L}}'\cdot C')>0$.
Totally separably closed schemes
================================
[\[TSC\]]{}
Recall that an integral scheme $X$ with generic point $\eta\in X$ is *totally separably closed* if it is normal and the function field $F={\mathscr{O}}_{X,\eta}=\kappa(\eta)$ is separably closed. A space or a scheme is called *local* if it contains exactly one closed point. The main result of this paper is:
[\[local\]]{} Let $X$ be an integral separated scheme that is totally separably closed, and $u,v\in X$ be two points. Then the intersection ${\operatorname{Spec}}({\mathscr{O}}_{X,u})\cap{\operatorname{Spec}}({\mathscr{O}}_{X,v})$ inside $X$ is local.
The intersection can be regarded as the underlying set of the schematic fiber product $$P={\operatorname{Spec}}({\mathscr{O}}_{X,u})\times_X{\operatorname{Spec}}({\mathscr{O}}_{X,v}).$$ Its image contains the generic point $\eta\in X$, in particular $P$ is non-empty. Furthermore, the scheme $P$ is affine, because $X$ is separated. Seeking a contradiction, we assume that the intersection is not local. Hence there are two closed points $\alpha\neq\beta$ inside $P$. Let $A,B\subset X$ be their closures in $X$. Both contain $u$ and $v$. In fact, the points $u,v\in A\cap B$ are generic points in the intersection. If the two points $u,v\in X$ admit a common affine open neighborhood, we immediately get a contradiction from [@Artin; @1971], Corollary 1.8. The idea of this proof is to construct, starting form $X$, another integral separated scheme $\tilde{X}$ that is totally separably closed and additionally enjoys the AF property, containing two points ${\tilde{u}}\neq{\tilde{v}}$ closely related to the original points $u\neq v$. Now the intersection $\tilde{P}={\operatorname{Spec}}({\mathscr{O}}_{{\tilde{X}},{\tilde{u}}})\times_{{\tilde{X}}}{\operatorname{Spec}}({\mathscr{O}}_{{\tilde{X}},{\tilde{v}}})$ is indeed local, and this will finally produced the desired contradiction.
*We reduce to the case that $X$ is the total separable closure of some proper ${\mathbb{Z}}$-scheme $X_0$.* First of all, we may assume that our scheme $X$ is quasicompact, simply by choosing affine open neighborhoods $U,V\subset X$ of $u,v\in X$ and replacing $X$ by their union. Next, we write $X={\varprojlim}X_\lambda$ as a filtered inverse limit of schemes $X_\lambda$, $\lambda\in L$ that are separated and of finite type over the ring $R={\mathbb{Z}}$, with affine transition maps $X_\mu{\rightarrow}X_\lambda$, $\lambda\leq\mu$. This is possible according to [@Thomason;; @Trobaugh; @1990], Appendix C, Proposition 7. Replacing $X_\lambda$ by the schematic images of the projection $X{\rightarrow}X_\lambda$, we may assume that the $X_\lambda$ are integral, and that the transition maps $X_\mu{\rightarrow}X_\lambda$ and the projections $X{\rightarrow}X_\lambda$ are dominant. Let $\eta_\lambda\in X_\lambda$ be the generic points, such that $\eta=(\eta_\lambda)_{\lambda\in L}$. The function fields $F_\lambda={\mathscr{O}}_{X_\lambda,\eta_\lambda}=\kappa(\eta_\lambda)$ form a filtered direct system of subfields inside $F={\mathscr{O}}_{X,\eta}=\kappa(\eta)$, with $F=\bigcup_{\lambda\in L} F_\lambda$. Let $F^{{\operatorname{sep}}}_\lambda$ be the relative separable algebraic closure of $F_\lambda\subset F$, and ${\operatorname{TSC}}(X_\lambda)$ the be the resulting integral closure of $X_\lambda$ with respect to the field extension $F_\lambda\subset F^{{\operatorname{sep}}}_\lambda$. Then the filtered inverse system $X_\lambda$ induces a filtered inverse system $Y_\lambda={\operatorname{TSC}}(X_\lambda)$. The morphism $X{\rightarrow}X_\lambda$ induces compatible morphisms $X{\rightarrow}Y_\lambda$, giving and identification $X={\varprojlim}Y_\lambda$.
Now suppose that the theorem is valid for all $Y_\lambda={\operatorname{TSC}}(X_\lambda)$. Let $u_\lambda,v_\lambda\in Y_\lambda$ be the images of $u,v\in X$. Then the schemes $$P_\lambda={\operatorname{Spec}}( {\mathscr{O}}_{Y_\lambda,u})\times_{Y_\lambda} {\operatorname{Spec}}({\mathscr{O}}_{Y_\lambda,v})$$ are local henselian. According to [@Artin; @1971], Lemma 2.6 the inverse limit $P={\varprojlim}P_\lambda$ is local henselian, contradiction.
This reduces us to the case that $X$ is the total separable closure of some integral scheme $X_0$ that is separated and of finite type over the ring $R={\mathbb{Z}}$. In light of Nagata’s Compactification Theorem in the relative version obtained by Lütkebohmert [@Luetkebohmert; @1993], we may additionally assume that the structure morphism $X_0{\rightarrow}{\operatorname{Spec}}({\mathbb{Z}})$ is proper. This concludes Step 1.
*We may assume that the point $u\in X$ lies in a closed fiber $X\otimes{\mathbb{F}}_p$.* If both points $u,v\in X$ lie in the generic fiber $X\otimes{\mathbb{Q}}$, we could replace $X_0$ and $X$ by their generic fibers. Now $X_0$ is proper over the field $k={\mathbb{Q}}$, and we immediately get a contradiction to [@Schroeer; @2017], Theorem 12.1. So we may assume without restriction that $u\in U$ lies in a closed fiber $X\otimes{\mathbb{F}}_p$, for some prime $p>0$.
*Here we write $X$ as a filtered inverse system $X_\lambda$, $\lambda\in L$ of proper ${\mathbb{Z}}$-schemes so that the geometry of $A,B\subset X$ is captured by their images $A_\lambda,B_\lambda\subset X_\lambda$.* Let $F_0\subset F$ be the inclusion of function fields coming from the canonical morphism $X{\rightarrow}X_0$. Changing the notation from Step 1, we now write $F_\lambda\subset F$, $\lambda\in L$ for the filtered direct system of subfields with $[F_\lambda:F_0]<\infty$, and let $X_\lambda{\rightarrow}X$ be the normalization of $X_0$ with respect to the field extension $F_0\subset F_\lambda$. This gives a filtered inverse system $X_\lambda$ of finite $X_0$-schemes, with $X={\varprojlim}X_\lambda$, where the transition maps $X_\mu{\rightarrow}X_\lambda$, $\lambda\leq \mu$ are finite. Note that all structure morphisms $X_\lambda{\rightarrow}{\operatorname{Spec}}({\mathbb{Z}})$ are proper, that $X{\rightarrow}{\operatorname{Spec}}({\mathbb{Z}})$ is separated and universally closed, and that the projections $X{\rightarrow}X_\lambda$ are universally closed. Write $$\label{induced points}
u_\lambda, v_\lambda, \alpha_\lambda,\beta_\lambda, \eta_\lambda\in X_\lambda$$ for the respective images of the points $u,v,\alpha,\beta,\eta\in X$. Let $A_\lambda,B_\lambda\subset X_\lambda$ be the closures of $\alpha_\lambda,\beta_\lambda\in X_\lambda$, which are also the images of $A,B\subset X$. In turn, we have $u_\lambda,v_\lambda\in A_\lambda\cap B_\lambda$. Replacing $L$ by some cofinal subset, we may assume that the points in are pairwise different.
Since inverse limits commute with inverse limits, we have $A\cap B={\varprojlim}(A_\lambda\cap B_\lambda)$. For the local rings this means ${\mathscr{O}}_{A\cap B,u} = {\varinjlim}{\mathscr{O}}_{{\mathscr{O}}_{A_\lambda\cap B_\lambda, u_\lambda}}$. Setting $C={\operatorname{Spec}}({\mathscr{O}}_{A\cap B,u})$ and $C_\lambda = {\operatorname{Spec}}({\mathscr{O}}_{{\mathscr{O}}_{A_\lambda\cap B_\lambda, u_\lambda}})$, we get $C\smallsetminus\{u\}={\varprojlim}(C_\lambda\smallsetminus\{u_\lambda\})$. But the left-hand side is empty, because $u\in A\cap B$ is a generic point. Thus $1=0$ already holds as global sections on some $C_\lambda$. By symmetry, the same applies for the point $v\in A\cap B$. Replacing $L$ by some cofinal subset, we thus may assume that $u_\lambda,v_\lambda\in A_\lambda\cap B_\lambda$ are generic points.
*Reduction to the case that the connected components of $A_0\cap B_0$ are irreducible.* The proper ${\mathbb{Z}}$-scheme $A_0\cap B_0$ has only finitely many irreducible components. Fix an irreducible component $C_0\subset A_0\cap B_0$, let $C'_0\subset A_0\cap B_0$ be the union of the other irreducible components, and consider the normalized blowing-up $Y_0{\rightarrow}X_0$ with center $Z=C\cap C'$. This morphism is an isomorphism over some open neighborhood of the set of generic points in $A_0\cap B_0$, and the strict transforms of $C_0$ and $C'_0$ become disjoint on $Y_0$. By induction on the number of irreducible components in $A_0\cap B_0$, we find a normalized blowing-up $Y_0'{\rightarrow}X_0$ that is an isomorphism over the set of generic points of $A_0\cap B_0$ so that the strict transform $A_0', B'_0$ have the property that the connected components of $A'_0\cap B'_0$ are irreducible. Replacing $X_0$ by $Y'_0$ and $X$ by ${\operatorname{TSC}}(Y'_0)$ we may assume that the connected components of $A_0\cap B_0$ are irreducible. In particular, $$\label{intersection empty}
\overline{\{u_0\}}\cap \overline{\{v_0\}}={\varnothing}.$$ Note that his property will later produce the desired contradiction.
*Construction of two auxiliary filtered inverse systems $X'_\lambda$ and $X''_\lambda$ consisting of projective schemes.* Choose an affine open neighborhood $V_0\subset X_0$ of $v_0$ not containing $u_0$. By Chow’s Lemma, there is a blowing-up $g_0:X'_0{\rightarrow}X_0$ with center $Z_0$ contained in $X_0\smallsetminus V_0$, such that $X'_0$ is projective. Set $Z'_0=g_0^{-1}(\overline{\{u_0\}})$. Replacing $X'_0$ by some further normalized blowing-up of $X'_0$ with center $Z_0'$, we may also assume that the the closed set $g_0^{-1}(\overline{\{u_0\}})$ is the support of an effective Cartier divisor $E'_0\subset X'_0$ contained in the closed fiber $X'_0\otimes{\mathbb{F}}_p$.
Let $X'_\lambda{\rightarrow}X'_0$ be its normalization of $X'_0$ with respect to the finite field extension $F_0\subset F_\lambda$. This gives a filtered direct system with finite transition maps, and we obtain a totally separably closed scheme $X'={\varprojlim}X'_\lambda$. The projection $X'{\rightarrow}X'_0$ is affine and the scheme $X'_0$ satisfies the the AF property, so the same holds for $X'$. The ensuing projective birational morphisms $X'_\lambda{\rightarrow}X_\lambda$ induces a birational morphism $X'{\rightarrow}X$ between integral schemes. The $X'_\lambda{\rightarrow}X_\lambda$ are isomorphisms over the open subsets $V_\lambda=V_0\times_{X_0}X_\lambda$, which contains $v_\lambda,\alpha_\lambda,\beta_\lambda$, hence $X'{\rightarrow}X$ is an isomorphism over $V=V_0\times_{X_0}X$, which contains $v,\alpha,\beta$. So we may regard the latter also as points $v',\alpha',\beta'\in X'$.
Let $A',B'\subset X'$ be the strict transforms of $A,B\subset X$, that is, the closures of $\alpha',\beta'\in X'$. According to [@Artin; @1971], Corollary 1.8 together with [@Schroeer; @2017], Theorem 7.6 the intersection $A'\cap B'$ is irreducible. Since ${\operatorname{Spec}}({\mathscr{O}}_{A'\cap B',v'})={\operatorname{Spec}}({\mathscr{O}}_{A\cap B,v})$, the point $v'\in A'\cap B'$ is generic, whence this must be the unique generic point. Consider the canonical morphism $\varphi:X'{\rightarrow}X_0$, which is a closed map. Suppose there would be a point $u'$ in the intersection $$\label{empty fiber}
A'\cap B'\cap \varphi^{-1}(u_0)=\overline{\{v'\}} \cap \varphi^{-1}(u_0).$$ Then $u_0\in \varphi(\overline{\{v'\}}) = \overline{ \{\varphi(v')\}}=\overline{ \{v_0\}}$, contradicting . Thus the intersection is empty. Passing to a cofinal subset of $L$, we may assume that already the fiber of $A'_0\cap B'_0$ over $u_0\in A_0\cap B_0$ is empty. On the other hand, the morphisms $A'_0{\rightarrow}A_0$ and $B'_0{\rightarrow}B_0$ are proper and dominant, so there are points $r'_0\in A'_0$ and $s'_0\in B'_0$ lying in $\varphi^{-1}(u_0)$.
Now recall that the fiber $g_0^{-1}(\overline{\{u_0\}})$ is a Cartier divisor $E'_0\subset X'_0$ lying inside the closed fiber $X'\otimes{\mathbb{F}}_p$. Write $n\geq 2$ for the common dimension of the schemes $X,X_\lambda, X',X'_\lambda$. Since $A_0,B_0\subsetneq X_0$, the irreducible schemes $A'_0,B'_0$ have dimension $\leq n-1$, and the intersections $A'_0\cap E'_0$ and $B'_0\cap E'_0$ have dimension $\leq n-2$. On the other hand, each irreducible component of $E'_0$ has dimension $n-1$, by Krull’s Principal Theorem. According to Proposition \[modification\], there is an effective Cartier divisor $Z_0'\subset E'_0$ not containing $\{r'_0,s'_0\}$ so that on the blowing-up $X''_0{\rightarrow}X'_0$ with center $Z'_0\subset X'_0$ the strict transform $E''_0$ of $E'_0$ becomes projectively contractible to points. As above, we get a filtered inverse system of projective schemes $X''_\lambda$ with finite transition maps, and the resulting $X''={\varprojlim}(X''_\lambda)={\operatorname{TSC}}(X''_0)$ comes with a birational morphism $X''{\rightarrow}X'$ of integral schemes. Consider the composite morphism $h: X''{\rightarrow}X$ and the corresponding $h_\lambda:X''_\lambda{\rightarrow}X_\lambda$. By construction, we have $h_0(E''_0)=\overline{\{u_0\}}$. This finishes Step 5.
*Construction of the contractions $r_\lambda:X''_\lambda{\rightarrow}{\tilde{X}}_\lambda$ resulting in a contradiction.* Let $E''_\lambda\subset X''_\lambda$ be the preimages of $E''_0\subset X''_0$. These subschemes are effective Cartier divisors, because $X''_\lambda$ is integral and $X''_\lambda{\rightarrow}X''_0$ is dominant. According to Proposition \[pullback\], the $E''_\lambda$ are projectively contractible to points. Moreover, the resulting contractions $r_\lambda:X''_\lambda{\rightarrow}{\tilde{X}}_\lambda$ coincide with the Stein factorization of $X''_\lambda{\rightarrow}{\tilde{X}}_0$, and yield yet another filtered inverse system ${\tilde{X}}_\lambda$ of projective schemes with finite transition maps. We have a commutative diagram $$\begin{CD}
{\tilde{X}}_\lambda @<r_\lambda << X''_\lambda @>h_\lambda>> X_\lambda\\
@VVV @VVV @VVt_{0\lambda} V\\
{\tilde{X}}_0 @<<r_0< X''_0 @>>h_0> X_0.
\end{CD}$$ Clearly, $u_\lambda$ is contained in the finite set $t_{0\lambda}^{-1}(u_0)$, where $t_{0\lambda}:X_\lambda{\rightarrow}X_0$ denotes the transition map. Using the above commutative diagram, we infer that the $h_\lambda^{-1}(u_\lambda)$ is contained in the effective Cartier divisor $E''_\lambda$ whose connected components are mapped to closed points in ${\tilde{X}}_\lambda$. But the fiber $h_\lambda^{-1}(u_\lambda)$ is connected, by Zariski’s Main Theorem. It follows that the set $h_\lambda^{-1}(u_\lambda)$ and in particular the elements $r_\lambda,s_\lambda\in h_\lambda^{-1}(u_\lambda)$ are mapped to the same point ${\tilde{u}}_\lambda\in{\tilde{X}}_\lambda$.
Now consider the resulting filtered inverse system ${\tilde{X}}_\lambda$, $\lambda\in L$ of projective schemes with finite transition maps. The inverse limit ${\tilde{X}}={\varprojlim}{\tilde{X}}_\lambda$ is another totally separably closed scheme. Since the scheme ${\tilde{X}}_0$ is projective and the morphisms ${\tilde{X}}{\rightarrow}{\tilde{X}}_0$ is integral, ${\tilde{X}}$ enjoys the AF property. The points ${\tilde{u}}_\lambda\in {\tilde{X}}_\lambda$ are compatible and yield a point ${\tilde{u}}\in {\tilde{X}}$. By construction, $X_\lambda{\rightarrow}{\tilde{X}}_\lambda$ are isomorphisms on an open neighborhood of $v_\lambda,\alpha_\lambda,\beta_\lambda$. So we may regard the latter as points on ${\tilde{X}}$, denoted by ${\tilde{v}}_\lambda,\tilde{\alpha}_\lambda,\tilde{\beta}_\lambda\in{\tilde{X}}_\lambda$. In turn, we get points ${\tilde{v}},\tilde{\alpha},\tilde{\beta}$ on the inverse limit ${\tilde{X}}$.
Since the points ${\tilde{u}},{\tilde{v}}\in {\tilde{X}}$ admit a common affine open neighborhood, [@Artin; @1971], Corollary 8.1 applies and we infer with [@Schroeer; @2017], Theorem 7.6 that the intersection ${\tilde{A}}\cap {\tilde{B}}$ is irreducible. Arguing as above, we see that ${\tilde{v}}\in {\tilde{A}}\cap {\tilde{B}}$ must be the generic point, in particular ${\tilde{u}}\in\overline{\{{\tilde{v}}\}}$. Since the projection ${\tilde{X}}{\rightarrow}{\tilde{X}}_0$ is closed, we also have ${\tilde{u}}_0\in\overline{ \{{\tilde{v}}_0\} }$. The contraction $r_0:X''_0{\rightarrow}{\tilde{X}}_0$ is closed as well, hence $$r_0(E''_0\cap \overline{\{v''_0\}}) \supset
r_0( r_0^{-1}({\tilde{u}}_0 ) \cap \overline{\{v''_0\}}) =
\{{\tilde{u}}_0\} \cap r_0( \overline{ \{v''_0\} }) =
\{{\tilde{u}}_0\} \cap \overline{ \{{\tilde{v}}_0\} }.$$ We see that $ E''_0\cap \overline{ \{v''_0\} }\neq {\varnothing}$. Finally, the blowing-up $h_0:X''_0{\rightarrow}X_0$ is a closed map with $h_0(E''_0)=\overline{\{u_0\}}$ and $h_0(v''_0)=v_0$, hence the sets $$h_0(E''_0\cap \overline{\{v''_0\}}) \subset
h_0(E''_0) \cap h_0(\overline{\{v''_0\}})=
\overline{\{u_0\}}\cap \overline{\{v_0\}}$$ are non-empty. But this contradicts .
Note that in Theorem \[local\] some assumption about separatedness is inevitable : For example, let $k$ be an algebraically closed field, $R_0$ be the henselization of $k[x,y]$ at the maximal ideal ${\mathfrak{m}}=(x,y)$, and $R$ be its total separable closure. Then $R$ is a local integral domain of dimension two that is TSC. Let $U\subset {\operatorname{Spec}}(R)$ be the complement of the closed point. Then $U$ has dimension one and contains infinitely many closed points. Let $R_1$ and $R_2$ be two copies of $R$, and $X={\operatorname{Spec}}(R_1)\cup{\operatorname{Spec}}(R_2)$ be the non-separated integral TSC scheme obtained by gluing along $U\subset{\operatorname{Spec}}(R_i)$. For the two closed points $u,v\in X$ we have ${\operatorname{Spec}}({\mathscr{O}}_{X,u})\cap{\operatorname{Spec}}({\mathscr{O}}_{X,v})=U$, which is not local.
In this example, the diagonal $\Delta:X{\rightarrow}X\times X$ is not affine. It is conceivable that Theorem \[local\] holds true under the weaker assumption that the diagonal is merely affine rather than a closed embedding.
Application to Nisnevich cohomology
===================================
[\[Nisnevich\]]{}
Let $X$ be a scheme, and write $({{\text{\rm Et}}}/X)$ for the category of étale $X$-schemes. The *Nisnevich topology* on this category is the Grothendieck topology defined by the pretopology whose covering families $(U_i{\rightarrow}U)_{i\in I}$ are those where for each $x\in U$ there is some index $i\in I$ and some $x_i\in U_i$ mapping to $x$, such that the residue field extension $\kappa(x)\subset\kappa(x_i)$ is trivial [@Nisnevich; @1989]. We write $X_{\text{\rm Nis}}$ for the ensuing topos of presheaves on $({{\text{\rm Et}}}/X)$ that satisfy the sheaf axiom for the Nisnevich topology. We refer to such sheaves as *Nisnevich sheaves*. Each point $x\in X$ yields a point $(P_*,P^*,\psi):({{\text{\rm Set}}}){\rightarrow}X_{\text{\rm Nis}}$ in the sense of topos-theory, and the corresponding local ring of the structure sheaf with respect to the Nisnevich topology is the henselization of the local ring ${\mathscr{O}}_{X,x}$ with respect to the Zariski topology. Every abelian Nisnevich sheaf $F$ comes with a spectral sequence $$E_2^{pq}={\check{H}}^p(X_{\text{\rm Nis}},\underline{H}^q(F)) \Longrightarrow H^{p+q}(X_{\text{\rm Nis}},F)$$ from Čech cohomology to sheaf cohomology (see for example [@Schroeer; @2017], Appendix B).
[\[cech\]]{} If $X$ is quasicompact and separated, then ${\check{H}}^p(X_{\text{\rm Nis}},\underline{H}^q(F)) =0$ for all $p\geq 0$, $q\geq 1$ and all abelian Nisnevich sheaves $F$. In particular, the canonical maps $${\check{H}}^p(X_{\text{\rm Nis}},F){\longrightarrow}H^p(X_{\text{\rm Nis}},F)$$ from Čech cohomology to sheaf cohomology are bijective for all $p\geq 0$.
The result was already established in [@Schroeer; @2017], Theorem 13.1 for schemes where the structure morphism $X{\rightarrow}{\operatorname{Spec}}({\mathbb{Z}})$ factors over the spectrum of a prime field. In other words, $X$ is a $k$-scheme for some ground field $k$. This assumption entered only via [@Schroeer; @2017], Theorem 12.1. But the latter holds true without the superfluous assumption of a ground field, by Theorem \[local\].
[ccccc]{}
M. Artin: Some numerical criteria for contractability of curves on algebraic surfaces. Am. J. Math. 84 (1962), 485–496.
M. Artin: Algebraization of formal moduli II: Existence of modifications. Ann. Math. 91 (1970), 88–135.
M. Artin: On the joins of Hensel rings. Advances in Math. 7 (1971), 282–296.
O. Benoist: Quasi-projectivity of normal varieties. Int. Math. Res. Not. IMRN 17 (2013), 3878–3885.
P. Berthelot, A. Grothendieck, L. Illusie (eds.): Théorie des intersections et théorème de Riemann–Roch (SGA 6). Springer, Berlin, 1971.
N. Bourbaki: Algèbre commutative. Chapitre 8–9. Masson, Paris, 1983.
S. Cutkosky: Asymptotic multiplicities of graded families of ideals and linear series. Adv. Math. 264 (2014), 55–113.
P. Deligne, N. Katz: Groupe de monodromie en géométrie algébrique (SGA 7 II). Springer, Berlin, 1973.
L. Ein: Linear systems with removable base loci. Comm. Algebra 28 (2000),5931–5934.
A. Engler, A. Prestel: Valued fields. Springer, Berlin, 2005.
T. Fujita: Semipositive line bundles. J. Fac. Sci. Univ. Tokyo 30 (1983), 353–378.
P. Gross: The resolution property of algebraic surfaces. Compos. Math. 148 (2012), 209–226.
A. Grothendieck: Éléments de géométrie algébrique II: Étude globale élémentaire de quelques classes de morphismes. Publ. Math., Inst. Hautes Étud. Sci. 8 (1961).
A. Grothendieck: Éléments de géométrie algébrique III: Étude cohomologique des faiscaux cohérent. Publ. Math., Inst. Hautes Étud. Sci. 11 (1961).
C. Huneke: Uniform bounds in Noetherian rings. Invent. Math. 107 (1992), 203–223.
C. Huneke: Absolute integral closure. In: A. Corso and C. Polini (eds.), Commutative algebra and its connections to geometry, pp. 119–135. Amer. Math. Soc., Providence, RI, 2011.
R. Hartshorne: Algebraic geometry. Springer, Berlin, 1977.
R. Hübl, I. Swanson: Discrete valuations centered on local domains. J. Pure Appl. Algebra 161 (2001), 145–166.
J. Kollár: Rational curves on algebraic varieties. Springer, Berlin, 1995.
R. Lazarsfeld: Positivity in algebraic geometry. I. Springer, Berlin, 2004.
W. Lütkebohmert: On compactification of schemes. Manuscr. Math. 80 (1993), 95–111.
H. Matsumura: Commutative algebra. Second edition. Benjamin/Cummings, Reading, Mass., 1980.
D. Mumford: The topology of normal singularities of an algebraic surface and a criterion for simplicity. Publ. Math., Inst. Hautes Étud. Sci. 9 (1961), 5–22.
Y. Nisnevich: The completely decomposed topology on schemes and associated descent spectral sequences in algebraic K-theory. In: J. Jardine, V. Snaith (eds.), Algebraic K-theory: connections with geometry and topology, pp. 241–342. Kluwer, Dordrecht, 1989.
M. Olsson: Algebraic spaces and stacks. American Mathematical Society, Providence, RI, 2016.
M. Perling, S. Schröer: Vector bundles on proper toric 3-folds and certain other schemes. Trans. Amer. Math. Soc. 369 (2017), 4787–4815.
F. Schmidt: Körper, über denen jede Gleichung durch Radikale auflösbar ist. Sitzungsber. Heidelberger Akad. Wiss. 2 (1933), 37–47.
S. Schröer: Geometry on totally separably closed schemes. Algebra Number Theory 11 (2017), 537–582.
I. Swanson, C. Hueke: Integral Closure of Ideals, Rings, and Modules. Cambridge University Press, Cambridge, 2006.
R. Thomason, T. Trobaugh: Higher algebraic $K$-theory of schemes and of derived categories. In: P. Cartier et al. (eds.), The Grothendieck Festschrift III, 247–435. Birkhäuser, Boston, MA, 1990.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Entanglement is an important phenomenon that enables quantum information processing algorithms and quantum communications protocols. Although entangled quantum states are often described in radix-2, higher-radix qudits can become entangled as well. In this work, we both introduce partial entanglement, a concept that does not exist for radix-2 quantum systems, and differentiate between partial and maximal entanglement within non-binary quantum information processing systems. We also develop and present higher-radix maximal entanglement generator circuits that are analogous to the well-known Bell state generator for binary quantum systems. Because higher-dimensioned qudits can be subjected to entangling processes that result in either partially or maximally entangled states, we demonstrate how higher-radix qudit circuits can be composed to generate these varying degrees of partial quantum entanglement. Theoretical results are provided for the general case where the radix is greater than two, and specific results based on a pair of radix-4 qudits are described.'
author:
-
title: 'Entanglement in Higher-Radix Quantum Systems'
---
qudit, entanglement, partial entanglement, maximal entanglement, quantum information science, QIS, quantum information processing, QIP
Introduction
============
Entanglement is an important aspect in quantum information science (QIS) that is unseen in conventional computing. For example, if a programmer writes a bit sequence to a register `reg0`, it is certainly not expected that this write operation would cause the content in another register, such as `reg1`, to also change. However, this is precisely the action that would occur in a quantum information processing (QIP) system if `reg0` and `reg1` collectively contained a set of qudits in a mutually entangled state. While the benefits of entanglement may not be readily apparent, it is exploited in numerous algorithms such as Shor’s factoring [@Sh:94], Grover’s search [@Gr:96], and others in QIS communications systems and protocols like BB84 and its derivatives. Additionally, entanglement over a long distance was recently demonstrated when the Micius satellite produced an entangled pair of photons in space and transmitted one of the pair to Earth while the entangled state with the remaining photon on the satellite was maintained [@YC+:17].
The aforementioned QIS algorithms and the Chinese experiment in [@YC+:17] were based on the use of quantum bits or “qubits” that have a binary basis, typically the so-called “computational basis,” of $\{\Ket{0},\Ket{1}\}$. There has been considerably less research and development on QIS systems based upon higher-dimensional quantum digits, or “qudits.” Qudits are simply a generalization of qubits that can be mathematically expressed as a linear combination of an extended computational basis set. For example, an $R$-dimensional qudit would have a corresponding computational basis set of $\{\Ket{0},\Ket{1}, \cdots, \Ket{r^n-1}\}$ where $R=r^n$, $r>2$, and $(r,n)\in \mathbb{Z}$.
Entangled states have been studied in the past for radix-3, -4, and -5 QIP systems [@enriq_2016], but the required operators for generating higher-radix entangled states have not been clearly outlined. Additionally, higher-radix QIP demonstrates the unique ability to generate partially entangled states, and achieving these variations of partial entanglement through a set of quantum operators can also be generalized. In this paper, methods for generating higher-radix partial and maximal entanglement will be discussed. A radix-4 system will be used to demonstrate the described techniques.
Background QIP Concepts
=======================
Information and Operations
--------------------------
Due to the postulates and axioms of quantum electrodynamic theory, the quantum state of a wavefunction is modeled as a vector over a discrete Hilbert vector space. The most common unit of quantum information is thus represented as a two-dimensional vector in $\mathbb{H}_2$ and is referred to as the qubit. Qubits are expressed as a linear combination of two basis vectors that span $\mathbb{H}$. The most common basis is the so-called computational basis represented as $\Ket{0_2}$ and $\Ket{1_2}$. We note that subscripts are used to refer to the value $r$ to avoid confusion while discussing systems of different values of $r$. A qubit is expressed as a linear combination of the two orthonormal basis states, $\Ket{0_2}=\left[ \begin{matrix}
1 & 0
\end{matrix} \right]^{\rm T}$ and $\Ket{1_2}=\left[ \begin{matrix}
0 & 1
\end{matrix} \right]^{\rm T}$. In general, a qubit is represented as
$$\label{eq:qubit_state}
\Ket{\phi_2} = \alpha\Ket{0_2}+\beta\Ket{1_2}$$
where $\alpha$ and $\beta$ are complex-valued probability amplitudes. For the qubit $\Ket{\phi_2}$, the probability that $\Ket{\phi_2}=\Ket{0_2}$ after measurement is $\alpha^*\alpha= |\alpha|^2$ and the probability that $\Ket{\phi_2}=\Ket{1_2}$ after measurement is $\beta^*\beta= |\beta|^2$ where $|\alpha|^2+|\beta|^2=1$.
For higher-dimensional systems, the quantum digit is modeled as a vector in $\mathbb{H}_r$ where $r>2$. The dimension of the overall QIP system is $R=r^n$ where $n$ indicates the amount of included qudits. A radix-$r$ qudit is a linear combination of $r$ basis states in the form of
$$\label{eq:gen_qudit}
\Ket{\phi_r} = \sum_{i=0}^{r-1} a_i \Ket{i_r}.$$
Using Eqn. \[eq:gen\_qudit\] to generate a radix-4 qudit results in
$$\label{eq:qudit_state}
\Ket{\phi_4} = a_0\Ket{0_4}+ a_1\Ket{1_4} + a_2\Ket{2_4}+ a_3\Ket{3_4} .$$
In Eqn. \[eq:qudit\_state\], the basis states used are $\Ket{0_4}=\left[ \begin{matrix}
1 & 0 & 0 & 0
\end{matrix} \right]^{\rm T}$, $\Ket{1_4}=\left[ \begin{matrix}
0 & 1 & 0 & 0
\end{matrix} \right]^{\rm T}$, $\Ket{2_4}=\left[ \begin{matrix}
0 & 0 & 1 & 0
\end{matrix} \right]^{\rm T}$, and $\Ket{3_4}=\left[ \begin{matrix}
0 & 0 & 0 & 1
\end{matrix} \right]^{\rm T}$, and $a_0$, $a_1$, $a_2$, and $a_3$ are probability amplitudes. The technological reasons prohibiting the widespread adoption of higher-radix classical computation are not present in the realization of QIP systems. Electronic information processing systems utilize the binary digit, or “bit,” as a basic unit of information, and a bit is typically represented by a voltage range. Due to the rapid advances in transistor scaling versus allowable voltage levels, the bit has dominated as the preferred choice in classical computing. Within QIS, there are no clear preferences with regard to the choice of information carriers. While many current QIP implementations are binary, or based upon qubits, higher-dimensional systems have certainly not been ruled out. Some examples of qudit-based photonic QIP systems that are found in the literature encode information with orbital angular momentum (OAM), time-energy, frequency, time-phase, and location.
For any radix-$r$ quantum system, computation is accomplished with quantum operations. These operators, also referred to as gates, are each represented by a unitary transfer function matrix, $\mathbf{U}$. The matrix $\mathbf{U}$ is of size $r^n\times r^n$ where $n$ indicates the number of radix-$r$ qudits being transformed.
Superposition
-------------
A qubit or qudit can comprise non-zero probability amplitudes for all of its basis states simultaneously. This characteristic is known as quantum “superposition.” Superposition allows quantum computing algorithms to parallelize computations since a state holds multiple values at once. In quantum systems, parallelized computation is achieved through parallelism in information representation rather than the spatial or temporal (*i.e.* pipelining) parallelism employed in classical systems.
*Maximal Superposition*\
\[def:max-super\] A state is is said to be “maximally superimposed” when the state vector can be expressed as a linear combination of all basis vectors such that each probability amplitude has an identical magnitude that is equal to $\frac{1}{r}$. In other words, the quantum state takes the value of $ \Ket{\phi_r} = \sum_{i=0}^{r-1} a_i \Ket{i_r}$ where each $|a_i|=a_i^*a_i=a_i a_i^*=\frac{1}{r}$.$\Box$
Entanglement
------------
When two or more quantum entities are entangled, they interact and behave as a single composite system. Operations and measurements performed on one portion of the entangled group directly impacts the state of the other entangled elements. Entanglement is a significant phenomenon as it powers quantum communication techniques. For example, entanglement enables quantum teleportation, a highly secure communication protocol where quantum states can be obtained in a near simultaneous manner as demonstrated in [@YC+:17].
Entanglement Generation
=======================
Two quantum units of any radix can only become entangled if the proper operators are applied. First, the pair is initialized into a basis state. Next, one of the units is transformed into a state of maximal superposition. Finally, the pair must interact with each other using one or more two-input operators.
The Chrestenson Gate
--------------------
The Hadamard gate is represented with the transfer matrix
$$\label{eq:Hadamard}
{\scriptsize
\mathbf{H}=\frac{1}{\sqrt{2}}
\arraycolsep=1.4pt\def\arraystretch{1.4}
\begin{bmatrix*}[r]
1 & 1 \\[-4pt]
1 & -1
\end{bmatrix*}}.$$
This gate is an important operator used to evolve a qubit into a maximally superimposed state. When a qubit is initially in a basis state, $\mathbf{H}$ transforms the qubit so that it has an equal probability of being measured as either $\Ket{0_2}$ or $\Ket{1_2}$. Quantum operators exist for many different computation bases, such as radix-3 and above, that achieve equal, and therefore maximal, superposition among the corresponding basis states. These operators are derived using the discrete Fourier transform on Abelian groups. General theory of Fourier transforms on Abelian groups is outlined in the literature [@vilenkin; @b6]. The multiple-valued generalization of the radix-2 Hadamard gate and its transfer matrix is composed of discretized versions of the orthogonal Chrestenson basis function set [@b6]. This QIP gate is generally referred to as the Chrestenson gate [@b7]. Examples of useful applications of the Chrestenson transform in QIP can be found in reference [@Chrest_gate_examples].
The Chrestenson operator, as the generalized version of the Hadamard operator, can be defined by a matrix that is parameterized depending upon the on the radix of computation. The resulting radix-$r$ Chrestenson transformation matrix for a single qudit has a size of $r \times r$, with column (or row) vectors that are orthonormal to one another since the matrix is unitary. Traditionally, the Chrestenson transformation matrix is expressed as being normalized with a scalar factor, $\frac{1}{\sqrt{r}}$, thus permitting each of the components within the Chrestenson transform matrix to take the form of one of the $r^{th}$ roots of unity raised to some integral power [@b6; @b7]. The $r^{th}$ roots of unity can be visualized as $r$ points that are evenly-spaced on the unit circle in the complex plane with one of the roots always being the real-valued unity value or +1. The roots of unity are indicated as $w_k$ where $k=0, 1, ..., (r-1)$, and the point (1,0), denoted as $w_0$, is always included in this set. Each root of unity satisfies $(w_k)^r=1$. A closed-form representation of the $r^{th}$ roots of unity is $w_k= e^{i\frac{2\pi}{r}\times k}$.
The structure of the Chrestenson transform matrix takes the form of a Vandermonde matrix where each row vector consists of a $r^{th}$ root of unity, $w_k$, raised to an integral power $j$. Each element of the matrix is $w_k^j$ where $j$ is the column index and $k$ is the row index. In this indexing system, $j=0$ represents the leftmost column vector and $j=(r-1)$ represents the rightmost column vector. Similarly, $k=0$ represents the topmost row vector and $k=(r-1)$ represents the bottommost row vector. The Hadamard matrix results from the Chrestenson transform matrix when $r = 2$, confirming that the Chrestenson transform is a generalization of the Hadamard transform. The generalized radix-$r$ Chrestenson transform matrix, $\mathbf{C}_r$, is represented with the matrix
$$\label{eq:gen_Chrestenson}
{\scriptsize
\mathbf{C}_r =
\frac{1}{\sqrt{r}}
\arraycolsep=1.4pt\def\arraystretch{1.6}
\begin{bmatrix*}[c]
w^{0}_{0} & w^{1}_{0} & \dots & w^{(r-1)}_{0} \\[-2pt]
w^{0}_{1} & w^{1}_{1} & \dots & w^{(r-1)}_{1} \\[-2pt]
\vdots & \vdots & \ddots & \vdots \\[-2pt]
w^{0}_{(r-1)} & w^{1}_{(r-1)} & \dots & w^{(r-1)}_{(r-1)}
\end{bmatrix*}}.$$
Using the fourth roots of unity, $w_0 = \exp[(i2\pi/4)*0] = 1$, $w_1 =\exp[(i2\pi/4)*1] = i$, $w_2 = \exp[(i2\pi/4)*2] = -1,$ and $w_3 = \exp[(i2\pi/4)*3] = -i$, in Eq. \[eq:gen\_Chrestenson\], the radix-4 Chrestenson gate transfer matrix becomes
$$\label{eq:rad-4_Chrestenson}
{\scriptsize
\mathbf{C_4}=
\frac{1}{\sqrt{4}}
\arraycolsep=1.4pt\def\arraystretch{1.2}
\begin{bmatrix*}[r]
1 & 1 & 1 & 1 \\[-2pt]
1 & i & -1 & -i \\[-2pt]
1 & -1 & 1 & -1\\[-2pt]
1 & -i & -1 & i
\end{bmatrix*}}.$$
The radix-4 Chrestenson gate ($\mathbf{C}_4$), allows a radix-4 qudit originally in a basis to evolve into a quantum state of equal superposition. More information about the radix-4 Chrestenson gate as well as a proposed physical implementation can be found in [@rad4chrest]. The following example shows how the radix-4 qudit $\Ket{a_4} = \Ket{0_4}$ evolves to $\Ket{b_4} = \frac{1}{2}\Ket{0_4} + \frac{1}{2}\Ket{1_4} + \frac{1}{2}\Ket{2_4} + \frac{1}{2}\Ket{3_4}$, taking the value of the first column of the radix-4 Chrestenson matrix, after it is applied to the $\mathbf{C_4}$ transform
$$\mathbf{C}_4\Ket{a_4} = \ket{b_4},$$
$${\scriptsize
\mathbf{C}_4\Ket{0_4} =
\frac{1}{\sqrt{4}}
\arraycolsep=1.4pt\def\arraystretch{1.2}
\begin{bmatrix*}[r]
1 & 1 & 1 & 1\\
1 & i & -1 & -i \\
1 & -1 & 1 & -1\\
1 & -i & -1 & i \\
\end{bmatrix*}\begin{bmatrix}
1 \\
0 \\
0 \\
0 \\
\end{bmatrix} =\frac{1}{2}\begin{bmatrix}
1 \\
1 \\
1 \\
1 \\
\end{bmatrix}},$$ $${\scriptsize
\mathbf{C}_4\Ket{0_4} = \frac{1}{2}[\Ket{0_4} + \Ket{1_4} + \Ket{2_4} + \Ket{3_4}].
}$$
The schematic symbol of the $\mathbf{C_4}$ gate is pictured in Fig. \[fig:C4\].
![Symbol of the radix-4 Chrestenson gate, $\mathbf{C_4}$.[]{data-label="fig:C4"}](Figures/c4_subscript.png){height=".45in"}
The Controlled Modulo-Add Gate
-------------------------------
The radix-2 $\mathbf{X}$ operator, or the $\mathbf{NOT}$ operator, performs a Pauli-$\mathbf{X}$ rotation on a qubit. Mathematically, the Pauli-$\mathbf{X}$ can be considered a modulo-2 addition-by-one operator since it evolves a qubit $\Ket{0_2}$ to be $\Ket{((0+1) \text{mod} \; 2)_2}=\Ket{1_2}$ and likewise evolves a qubit $\Ket{1_2}$ to $\Ket{((1+1) \text{mod} \; 2)_2}=\Ket{0_2}$. In the case where $\Ket{\phi_2}$ is in a state of superposition, $\Ket{\phi_2(t_0)}=a_0 \Ket{0_2} + a_1 \Ket{1_2}$, the $\mathbf{X}$ operation exchanges the probability amplitude coefficients of the quantum state yielding the qubit $\Ket{\phi_2(t_1)}=a_1\Ket{0_2}+a_0\Ket{1_2}$. The quantum gate or operator for the Pauli-$\mathbf{X}$ is represented with the transfer matrix
$$\label{eq:NOT}
{\scriptsize
\mathbf{X} = \arraycolsep=1.2pt\def\arraystretch{1.4}
\begin{bmatrix*}[r]
0 & 1 \\[-5pt]
1 & 0
\end{bmatrix*}}.$$
The controlled version of the $\mathbf{X}$ gate is the “controlled-$\mathbf{X}$” or “controlled-$\mathbf{NOT}$” gate denoted as $\mathbf{C}_{NOT}$. The controlled-$\mathbf{NOT}$ gate may also be referred to by the somewhat unconventional name of “controlled-modulo-add by one” gate. The $\mathbf{C}_{NOT}$ gate is defined as
$$\label{eq:CNOT}
{\scriptsize
\mathbf{C}_{NOT} =
\arraycolsep=1.2pt\def\arraystretch{1.4}
\begin{bmatrix*}[r]
1& 0 & 0 & 0 \\[-5pt]
0 & 1 & 0 & 0 \\[-5pt]
0& 0 & 0 & 1 \\[-5pt]
0 & 0 & 1 & 0
\end{bmatrix*}},$$
and it causes a target qubit to undergo a Pauli-$\mathbf{X}$ operation if the control qubit has a probability amplitude for $\Ket{1_2}$.
In the case of radix-2 qubit systems, only two different modulo-$k$ additions are possible, modulo-0 and modulo-1, since there are only two computational basis vectors. Furthermore, one of these is the trivial case of modulo-2 addition-by-zero that results in the identity transfer matrix.
We denote the single qudit modulo-addition operations as $\mathbf{M}_k$ for operators that cause a modulo-$k$ addition with respect to modulus-$r$ as is used in [@qlog_unit_ops]. As is the case with qubits (*i.e.*, $r=2$), the modulo-0 operation is equal to the identity transfer function, or $\mathbf{M}_0=\mathbf{I}_r$ where $\mathbf{I}_r$ is the $r \times r$ identity matrix. As an example, the non-trivial radix-4 $\mathbf{M}_k$ operators are
For higher-dimensional systems with radix-$r$, $r>2$, there are $r-1$ different non-trivial single-qudit modulo-$k$ additions. Considering all available control values for radix-$r$, as well as the different modulo-$k$ additions, there are a total of $r^2-r$ different and non-trivial controlled-modulo-addition operators. The radix-$r$ controlled Modulo-addition-$k$ matrix, $\mathbf{A}_{h,k}$ where $h$ and $k$ each contain a single value, takes the form of
$$\label{eq:gen-cont-mod-add}
{\scriptsize
\arraycolsep=1.4pt\def\arraystretch{1.0}
\begin{array}{c}
\mathbf{A}_{h,k}=
\left[
\arraycolsep=1.4pt\def\arraystretch{1.0}
\begin{array}{ccccccc}
\mathbf{D}_0 & \mathbf{0}_r & \cdots & \cdots & \cdots & \cdots & \mathbf{0}_r \\
\mathbf{0}_r & \mathbf{D}_1 & \mathbf{0}_r & \cdots & \cdots& \cdots & \mathbf{0}_r \\
\vdots & \mathbf{0}_r & \ddots & \mathbf{0}_r & \cdots & \cdots & \mathbf{0}_r \\
\vdots & \vdots & \mathbf{0}_r & \mathbf{D}_j & \mathbf{0}_r & \cdots & \mathbf{0}_r \\
\vdots & \vdots & \vdots & \mathbf{0}_r & \ddots & \mathbf{0}_r & \vdots \\
\vdots & \vdots & \vdots & \vdots & \mathbf{0}_r & \ddots & \mathbf{0}_r \\
\mathbf{0}_r & \mathbf{0}_r & \mathbf{0}_r & \mathbf{0}_r & \cdots & \mathbf{0}_r & \mathbf{D}_{(r-1)}
\end{array}
\right],
\\
\text{where, } \mathbf{D}_i = \begin{cases}
\mathbf{M}_0 = \mathbf{I}_r, & i \neq h \\
\mathbf{M}_k, & i=h.
\end{cases}
\end{array}}$$
In Eqn. \[eq:gen-cont-mod-add\], each submatrix along the diagonal is denoted as $\mathbf{D}_i$ and is of dimension $r \times r$. The two-qudit controlled variation of the modulo-add gate, $\mathbf{A}_{h,k}$, only allows the modulo-addition by $k$ operation to occur on the target whenever the control qudit is in state, $\Ket{h_r}$. For example,
$$\label{eq:r4_A_3,1}
{\scriptsize
\begin{array}{c}
\mathbf{A}_{3,1}=
\left[
\arraycolsep=1.4pt\def\arraystretch{1.0}
\begin{array}{rrrr;{2pt/2pt}rrrr;{2pt/2pt}rrrr;{2pt/2pt}rrrr}
1& 0 & 0 & 0 & 0 &0 &0 &0 &0 & 0 & 0 &0 &0 &0 &0 &0\\
0& 1 & 0 & 0 & 0 &0 &0 &0 &0 & 0 & 0 &0 &0 &0 &0 &0\\
0& 0 & 1 & 0 & 0 &0 &0 &0 &0 & 0 & 0 &0 &0 &0 &0 &0\\
0& 0 & 0 & 1 & 0 &0 &0 &0 &0 & 0 & 0 &0 &0 &0 &0 &0\\\hdashline[2pt/2pt]
0& 0 & 0 & 0 & 1 &0 &0 &0 &0 & 0 & 0 &0 &0 &0 &0 &0\\
0& 0 & 0 & 0 & 0 &1&0 &0 &0 & 0 & 0 &0 &0 &0 &0 &0\\
0& 0 & 0& 0 & 0 &0 &1 &0 &0 & 0 & 0 &0 &0 &0 &0 &0\\
0& 0 & 0 & 0& 0 &0 &0 &1 &0 & 0 & 0 &0 &0 &0 &0 &0\\\hdashline[2pt/2pt]
0& 0 & 0 & 0 & 0 &0 &0 &0 &1 & 0 & 0 &0 &0 &0 &0 &0\\
0& 0 & 0 & 0 & 0 &0 &0 &0 &0 & 1 & 0 &0 &0 &0 &0 &0\\
0& 0 & 0& 0 & 0 &0 &0 &0 &0 & 0 & 1 &0 &0 &0 &0 &0\\
0& 0 & 0 & 0 & 0 &0 &0 &0 &0 & 0 & 0 &1 &0 &0 &0 &0\\\hdashline[2pt/2pt]
0& 0 & 0 & 0 & 0 &0 &0 &0 &0 & 0 & 0 &0 &0 &0 &0 &1\\
0& 0 & 0 & 0 & 0 &0 &0 &0 &0 & 0 & 0 &0 &1 &0 &0 &0\\
0& 0 & 0 & 0 & 0 &0 &0 &0 &0 & 0 & 0 &0 &0 &1 &0 &0\\
0& 0 & 0 & 0& 0 &0 &0 &0 &0 & 0 & 0 &0 &0 &0 &1 &0\\
\end{array}\right]
\end{array}
=
\left[
\arraycolsep=1.4pt\def\arraystretch{1.0}
\begin{array}{cccc}
\mathbf{D}_0 & \mathbf{0}_4 & \mathbf{0}_4 & \mathbf{0}_4 \\
\mathbf{0}_4 & \mathbf{D}_1 & \mathbf{0}_4 & \mathbf{0}_4 \\
\mathbf{0}_4 & \mathbf{0}_4 & \mathbf{D}_2 & \mathbf{0}_4 \\
\mathbf{0}_4 & \mathbf{0}_4 & \mathbf{0}_4 & \mathbf{D}_3
\end{array}
\right]
}$$
only allows the $\mathbf{D}_3=\mathbf{M}_1$ operation to execute on the target qudit if the control qudit has a value of $\Ket{3_4}$, or at least a non-zero probability amplitude for $\Ket{3_4}$. The control qudit probability amplitudes for $\Ket{0_4}$, $\Ket{1_4}$, and $\Ket{2_4}$ do not evolve the target because $\mathbf{D}_0=\mathbf{D}_1=\mathbf{D}_2=\mathbf{M}_0 = \mathbf{I}_4$. In Eqn. \[eq:r4\_A\_3,1\], the dashed lines separate submatrices so the $\mathbf{D}_i$ values are apparent. The general symbol of the controlled Modulo-add gate is shown in Fig. \[fig:mod-add\]. The total amount of available controlled-modulo add operations, $\mathbf{A}_{h,k}$, varies depending on the radix. There are $r$ possible control values, $h$, in the range $\{0,\cdots,(r-1)\}$, and $r-1$ meaningful values in the range $\{1, \cdots, (r-1)\}$ for the $k$ value in the modulo-addition by $k$ operation. As an example, in a radix-4 QIP system, the controlled-modulo add operations are $\mathbf{A}_{0,1}$, $\mathbf{A}_{0,2}$, $\mathbf{A}_{0,3}$, $\mathbf{A}_{1,1}$, $\mathbf{A}_{1,2}$, $\mathbf{A}_{1,3}$, $\mathbf{A}_{2,1}$, $\mathbf{A}_{2,2}$, $\mathbf{A}_{2,3}$, $\mathbf{A}_{3,1}$, $\mathbf{A}_{3,2}$, and $\mathbf{A}_{3,3}$.
![Symbol of the controlled Modulo-add gate, $\mathbf{A}_{h,k}$.[]{data-label="fig:mod-add"}](Figures/controlled_modulo-add2_new.png){height="0.55in"}
The Bell State Generator
------------------------
The Bell state generator, the inspiration for this work, entangles qubit pairs. An entangled qubit pair, $\Ket{\alpha \beta_2}$ is characterized as a single QIS system wherein the quantum state formed from the qubit pair cannot be factored into a product of $\Ket{\alpha_2} \otimes\Ket{\beta_2}$.
*Entangled Qubit Pair* \[def:qubit-entang\]\
$\Ket{\alpha \beta_2}$ is considered entangled when it consists of two unfactorable basis states with equal and non-zero probability amplitudes. All other basis states in the state vector for the pair have zero-valued probability amplitudes. This leads to four possible entangled two-qubit pairs that are commonly referred to as the “Bell states.” The Bell states are
$$\label{eq:Bell}
{\small
\begin{split}
\begin{array}{rrr}
\Ket{B_{00}} = \Ket{\Phi^+} = \frac{\Ket{00_2} + \Ket{11_2}}{\sqrt{2}},&
\mbox{} &
\Ket{B_{01}} = \Ket{\Psi^+} = \frac{\Ket{01_2} + \Ket{10_2}}{\sqrt{2}},
\end{array}\\
\begin{array}{rrr}
\Ket{B_{10}} = \Ket{\Phi^-} = \frac{\Ket{00_2} - \Ket{11_2}}{\sqrt{2}},&
\mbox{} &
\Ket{B_{11}} = \Ket{\Psi^-} = \frac{\Ket{01_2} - \Ket{10_2}}{\sqrt{2}}.
\end{array}
\end{split}
}$$
The Bell states are created when a pair of qubits, each initialized to a basis state, are operated upon by a quantum algorithm referred to as a Bell state generator. The algorithm may be considered a quantum logic circuit or as a program on a quantum computer, but in either case, it is denoted by the sequence comprised of a Hadamard and $\mathbf{C}_{NOT}$ gate as shown in Fig. \[fig:bell\_state\_gen\].
![Bell State Generator.[]{data-label="fig:bell_state_gen"}](Figures/bell_state_gen2_new.png){height="0.55in"}
As a note, qubit pair entanglement is not limited to just the Bell states. Other arbitrary entangled pairs are possible. These qubit pairs are still entangled, but they contain non-equal magnitudes with respect to the non-zero probability amplitudes in the quantum state vector.
Demonstration of Partial Entanglement
=====================================
Qudits are maximally entangled whenever each of the possible basis state outcomes of an observation are equally likely. Additionally, measurement of one part of the entangled group causes others within the entangled group to collapse without direct observation. In contrast to maximal entanglement, a partially entangled state contains a subset of entangled basis states along with a remaining set that is not entangled. States of partial entanglement should not be confused with non-maximal entanglement where probability amplitudes are imbalanced. Partial entanglement does not exist in binary QIP systems. The concept of partial entanglement is unique for QIP where $r > 2$.
*Partial Entanglement* \[def:part-ent\]\
Consider a general radix = $r$ qudit expressed in terms of the the computational bases as $\Ket{\phi_r}=a_0\Ket{0}+a_1\Ket{1}+ \cdots + a_i\Ket{i}+ \cdots + a_{r-1}\Ket{r-1}$. When $\Ket{\phi_r}$ is entangled with another radix-$r$ qudit, $\Ket{\theta_r}$, the entangled quantum state can be expressed as $\Ket{\phi \theta_r}$. If the probability amplitudes of the entangled quantum state are such that a measurement of $\Ket{\phi_r}$ in a given state, $\Ket{i}$, implies that $\Ket{\theta_r}$ is correspondingly in some other given state, $\Ket{j}$ then it is verified that $\Ket{\phi \theta_r}$ are entangled. However, if the measurement of $\Ket{\phi_r}$ results in an observation of another distinct state, $\Ket{k}$ and furthermore, this measurement does not imply that $\Ket{\theta_r}$ is correspondingly in some distinct state $\Ket{h}$, then the pair $\Ket{\phi \theta_r}$ are said to be partially entangled. $\Box$
Using the structure of the Bell state generator as inspiration, a higher-radix partial entanglement generator can be created with a Chrestenson operator followed by a controlled-modulo-addition operation. An example circuit for radix-4 is seen in Fig. \[fig:part-ent-gen\]a. Here, the quantum state $\Ket{00_4}$ is transformed via the partial entanglement generator and evolves to
$$\label{eq:T_5a}
\scriptsize{
\begin{split}
\mathbf{T}_{Fig.~\ref{fig:part-ent-gen}a}\Ket{\phi \theta_4}&=\mathbf{T}_{Fig.~\ref{fig:part-ent-gen}a}\Ket{00_4} \\
&=\mathbf{A}_{3,1}\times (\mathbf{C}_4 \otimes \mathbf{I}_4)\Ket{00_4} \\
&=\frac{1}{2}\left[\Ket{00_4} + \Ket{10_4} + \Ket{20_4} + \Ket{31_4} \right] \\
&= \frac{1}{2}\left[ \left(\Ket{0_4} + \Ket{1_4} + \Ket{2_4}\right) \otimes\Ket{0_4}\right] + \frac{1}{2}\left[ \Ket{31_4} \right].
\end{split}}$$
The result of Eqn. \[eq:T\_5a\] exhibits some unique characteristics. It is observed that the qudit pair is entangled with respect to state $\Ket{31_4}$ since the measurement of $\Ket{\theta_4}$ resulting in $\Ket{\theta_4}=\Ket{1_4}$ implies that $\Ket{\phi_4}=\Ket{3_4}$. However, if $\Ket{\theta_4}$ is measured and results in $\Ket{\theta_4}=\Ket{0_4}$, there remains an equally likely chance that $\Ket{\phi_4}$ has assumed a basis value of $\Ket{0_4}$, $\Ket{1_4}$, or $\Ket{2_4}$. This is an example of a partially entangled pair. The output states for all possible input combinations of Fig. \[fig:part-ent-gen\]a are found in the center column of the leftmost partial entanglement results in Table \[tb:par\_full\_ent\_data\].
![Example radix-4 partial entanglement generators[]{data-label="fig:part-ent-gen"}](Figures/r4_part_ent_2examples.png){height="0.75in"}
The radix-4 partial entanglement generator of Fig. \[fig:part-ent-gen\]a can be expanded to create Fig. \[fig:part-ent-gen\]b. This second circuit includes an additional $\mathbf{A}_{h,k}$ operator, and a different type of partially entangled radix-4 state results:
$$\label{eq:T_5b}
\scriptsize{
\begin{split}
\mathbf{T}_{Fig.~\ref{fig:part-ent-gen}b}\Ket{00_4}&=\mathbf{A}_{2,2}\times\mathbf{A}_{3,1}\times (\mathbf{C}_4 \otimes \mathbf{I}_4)\Ket{00_4} \\
&=\mathbf{A}_{(2,3),(2,1)}\times (\mathbf{C}_4 \otimes \mathbf{I}_4)\Ket{00_4} \\
&=\frac{1}{2}\left[\Ket{00_4} + \Ket{10_4} + \Ket{22_4} + \Ket{31_4} \right] \\
&= \frac{1}{2}\left[ \left(\Ket{0_4} + \Ket{1_4}\right) \otimes\Ket{0_4}\right] + \frac{1}{2}\left[\Ket{22_4}+ \Ket{31_4} \right].
\end{split}}$$
In Eqn. \[eq:T\_5b\], the generated state shows an increase in the number of unique, fully-entangled basis state values and a decrease in the number of terms that have a factorable form and are indicative of the presence of partial entanglement. The output states for all possible input combinations of Fig. \[fig:part-ent-gen\]b are found in the third column of leftmost set of partial entanglement results in Table \[tb:par\_full\_ent\_data\].
When the radix-4 QIP system contains fewer than $r-1 = 4-1 = 3$ $\mathbf{A}_{h,k}$ gates, partially entangled states are generated. The degree of partial entanglement, or the number of basis states with a common factor, is directly related to the total number of $\mathbf{A}_{h,k}$ operations in the partial entanglement circuit. When including multiple $\mathbf{A}_{h,k}$ gates in a partial-entanglement circuit, each value of $h$ and $k$ must differ.
Although partially entangled states may not become as commonly implemented as fully entangled states in higher-radix systems, that does not mean they are simply an interesting phenomena without a use case. Possible implementations could involve generating either entangled states or random values in a controlled manner for higher-radix QIP communication systems. For instance, observing one of the entangled basis states results in known second qudit’s value, but observing an unentangled qudit that is factorable leaves an even distribution with respect to the possible observed state of the second qudit. In addition, since there are many combinations of partial entanglement generators, the output quantum state can be customized for a particular application.
[lr]{}
-------------- --------------------------------------------------------------------------------- ---------------------------------------------------------------------------------
**Input** $\mathbf{A}_{3,1}$ $\mathbf{A}_{2,2}\times\mathbf{A}_{3,1}=\mathbf{A}_{(2,3),(2,1)}$
$\Ket{00_4}$ $ \frac{1}{2}\left[\Ket{00_4} + \Ket{10_4} + \Ket{20_4} + \Ket{31_4} \right]$ $ \frac{1}{2}\left[\Ket{00_4} + \Ket{10_4} + \Ket{22_4} + \Ket{31_4} \right]$
$\Ket{01_4}$ $ \frac{1}{2}\left[\Ket{01_4} + \Ket{11_4} + \Ket{21_4} + \Ket{32_4} \right]$ $ \frac{1}{2}\left[\Ket{01_4} + \Ket{11_4} + \Ket{23_4} + \Ket{32_4} \right]$
$\Ket{02_4}$ $ \frac{1}{2}\left[\Ket{02_4} + \Ket{12_4} + \Ket{22_4} + \Ket{33_4} \right]$ $ \frac{1}{2}\left[\Ket{02_4} + \Ket{12_4} + \Ket{20_4} + \Ket{33_4} \right]$
$\Ket{03_4}$ $ \frac{1}{2}\left[\Ket{03_4} + \Ket{13_4} + \Ket{23_4} + \Ket{30_4} \right]$ $ \frac{1}{2}\left[\Ket{03_4} + \Ket{13_4} + \Ket{21_4} + \Ket{30_4} \right]$
$\Ket{10_4}$ $ \frac{1}{2}\left[\Ket{00_4} + i\Ket{10_4} - \Ket{20_4} -i \Ket{31_4} \right]$ $ \frac{1}{2}\left[\Ket{00_4} + i\Ket{10_4} -\Ket{22_4} - i\Ket{31_4} \right]$
$\Ket{11_4}$ $ \frac{1}{2}\left[\Ket{01_4} + i\Ket{11_4} - \Ket{21_4} - i\Ket{32_4} \right]$ $ \frac{1}{2}\left[\Ket{01_4} + i\Ket{11_4} - \Ket{23_4} -i \Ket{32_4} \right]$
$\Ket{12_4}$ $ \frac{1}{2}\left[\Ket{02_4} + i\Ket{12_4} - \Ket{22_4} -i \Ket{33_4} \right]$ $ \frac{1}{2}\left[\Ket{02_4} + i\Ket{12_4} - \Ket{20_4} -i \Ket{33_4} \right]$
$\Ket{13_4}$ $ \frac{1}{2}\left[\Ket{03_4} + i\Ket{13_4}- \Ket{23_4} -i \Ket{30_4} \right]$ $ \frac{1}{2}\left[\Ket{03_4} + i\Ket{13_4} - \Ket{21_4} -i \Ket{30_4} \right]$
$\Ket{20_4}$ $ \frac{1}{2}\left[\Ket{00_4} - \Ket{10_4} + \Ket{20_4} - \Ket{31_4} \right]$ $ \frac{1}{2}\left[\Ket{00_4} - \Ket{10_4} + \Ket{22_4} - \Ket{31_4} \right]$
$\Ket{21_4}$ $ \frac{1}{2}\left[\Ket{01_4} - \Ket{11_4} + \Ket{21_4} - \Ket{32_4} \right]$ $ \frac{1}{2}\left[\Ket{01_4} - \Ket{11_4} + \Ket{23_4} - \Ket{32_4} \right]$
$\Ket{22_4}$ $ \frac{1}{2}\left[\Ket{02_4} - \Ket{12_4} + \Ket{22_4} - \Ket{33_4} \right]$ $ \frac{1}{2}\left[\Ket{02_4} - \Ket{12_4} + \Ket{20_4} - \Ket{33_4} \right]$
$\Ket{23_4}$ $ \frac{1}{2}\left[\Ket{03_4} - \Ket{13_4} + \Ket{23_4} - \Ket{30_4} \right]$ $ \frac{1}{2}\left[\Ket{03_4} - \Ket{13_4} + \Ket{21_4} - \Ket{30_4} \right]$
$\Ket{30_4}$ $ \frac{1}{2}\left[\Ket{00_4} -i \Ket{10_4} - \Ket{20_4} +i \Ket{31_4} \right]$ $ \frac{1}{2}\left[\Ket{00_4} -i \Ket{10_4} - \Ket{22_4}+i \Ket{31_4} \right]$
$\Ket{31_4}$ $ \frac{1}{2}\left[\Ket{01_4} -i\Ket{11_4} - \Ket{21_4} +i \Ket{32_4} \right]$ $ \frac{1}{2}\left[\Ket{01_4} -i \Ket{11_4} - \Ket{23_4} +i \Ket{32_4} \right]$
$\Ket{32_4}$ $ \frac{1}{2}\left[\Ket{02_4} -i \Ket{12_4} - \Ket{22_4} +i \Ket{33_4} \right]$ $ \frac{1}{2}\left[\Ket{02_4} -i\Ket{12_4} - \Ket{20_4} +i \Ket{33_4} \right]$
$\Ket{33_4}$ $ \frac{1}{2}\left[\Ket{03_4} -i \Ket{13_4} - \Ket{23_4} +i \Ket{30_4} \right]$ $ \frac{1}{2}\left[\Ket{03_4} -i \Ket{13_4} - \Ket{21_4} +i \Ket{30_4} \right]$
-------------- --------------------------------------------------------------------------------- ---------------------------------------------------------------------------------
&
[|c|c|]{}\
**Input** & $\mathbf{A}_{1,3}\times\mathbf{A}_{2,2}\times\mathbf{A}_{3,1}=\mathbf{A}_{(1,2,3),(3,2,1)}$\
$\Ket{00_4}$ &$\frac{1}{2}\left[\Ket{00_4} + \Ket{13_4} + \Ket{22_4} + \Ket{31_4}\right]$\
$\Ket{01_4}$ & $\frac{1}{2}\left[\Ket{01_4} + \Ket{10_4} + \Ket{23_4} + \Ket{32_4}\right]$\
$\Ket{02_4}$ &$\frac{1}{2}\left[\Ket{02_4} + \Ket{11_4} + \Ket{20_4} + \Ket{33_4}\right]$\
$\Ket{03_4}$ & $\frac{1}{2}\left[\Ket{03_4} + \Ket{12_4} + \Ket{21_4} + \Ket{30_4}\right]$\
$\Ket{10_4}$ & $\frac{1}{2}\left[\Ket{00_4} + i\Ket{13_4} - \Ket{22_4} -i \Ket{31_4}\right]$\
$\Ket{11_4}$ &$\frac{1}{2}\left[\Ket{01_4} + i\Ket{10_4} -\Ket{23_4} -i \Ket{32_4}\right]$\
$\Ket{12_4}$ &$\frac{1}{2}\left[\Ket{02_4} +i \Ket{11_4} - \Ket{20_4} -i \Ket{33_4}\right]$\
$\Ket{13_4}$ &$\frac{1}{2}\left[\Ket{03_4} +i \Ket{12_4} - \Ket{21_4} -i \Ket{30_4}\right]$\
$\Ket{20_4}$ & $\frac{1}{2}\left[\Ket{00_4} -\Ket{13_4} + \Ket{22_4} - \Ket{31_4}\right]$\
$\Ket{21_4}$ &$\frac{1}{2}\left[\Ket{01_4} - \Ket{10_4} + \Ket{23_4} - \Ket{32_4}\right]$\
$\Ket{22_4}$ &$\frac{1}{2}\left[\Ket{02_4} - \Ket{11_4} + \Ket{20_4} - \Ket{33_4}\right]$\
$\Ket{23_4}$ & $\frac{1}{2}\left[\Ket{03_4} - \Ket{12_4} + \Ket{21_4} - \Ket{30_4}\right]$\
$\Ket{30_4}$ & $\frac{1}{2}\left[\Ket{00_4} -i\Ket{13_4} - \Ket{22_4} +i \Ket{31_4}\right]$\
$\Ket{31_4}$ & $\frac{1}{2}\left[\Ket{01_4} -i \Ket{10_4} - \Ket{23_4} +i \Ket{32_4}\right]$\
$\Ket{32_4}$ &$\frac{1}{2}\left[\Ket{02_4} -i \Ket{11_4} - \Ket{20_4} +i \Ket{33_4}\right]$\
$\Ket{33_4}$ & $\frac{1}{2}\left[\Ket{03_4} -i \Ket{12_4} - \Ket{21_4} +i \Ket{30_4}\right]$\
\[tb:par\_full\_ent\_data\]
Demonstration of Full Entanglement
==================================
The previous section described how partially entangled qudit pair generators can be created with a quantum circuit that includes a radix-$r$ Chrestenson gate and fewer than $r-1$ single-controlled-modulo-add-by-$k$ gates. In accordance with Theorem \[thm:max-ent-thm\], a total of appropriate $r-1$ controlled Modulo-add operations are required for full entanglement to occur.
*Maximal Entanglement Generator* \[thm:max-ent-thm\] A maximally entangled radix-$r$ qudit pair can be generated when the the pair is initialized to a basis state and one of the qudits evolves to a state of maximal superposition through the application of a Chrestenson gate, $\mathbf{C}_r$. This superimposed qudit is then applied to the control inputs of $r-1$ controlled modulo-add-by-$k$ gates, $\mathbf{A}_{h,k}$, and the other non-superimposed qudit is applied to the target. Each of the control values, $h$, of the $r-1$ $\mathbf{A}_{h,k}$ gates has a separate and distinct value from the set $\{0,1,\cdots,(r-1)\}$ and each of the modulo-add-by-$k$ target operations, $k$, of the $r-1$ $\mathbf{A}_{h,k}$ gates, takes on a separate and distinct value from the set $\{1,\cdots,(r-1)\}$.
Consider Eqn. \[eq:T\_5a\] wherein a single controlled modulo-add-by-$k$ gate, $\mathbf{A}_{3,1}$ results in a single fully entangled state and $r-1$ partially entangled states. Likewise, in Eqn. \[eq:T\_5b\] another controlled modulo-add-by-$k$ gate, $\mathbf{A}_{2,2}$ is applied resulting in $r-2$ fully entangled states and $r-2$ partially entangled states. Hence, by induction it is observed that the application of $r-1$ controlled modulo-add-by-$k$ gates, where all control values $h$ are unique from the set $\{0,1,\cdots,(r-1)\}$ and all modulo-addition constants $k$ are unique from the set $\{1,\cdots,(r-1)\}$, results in $r$ fully-entangled qudit basis states. $\Box$
In the case of the example radix-4 QIP system, three distinct $\mathbf{A}_{h,k}$ gates are required to generate full entanglement where each set of $h$ and $k$ values must contain unique integers. An illustration of a radix-4 full entanglement generator is given in Fig. \[fig:full-ent-gen\]. In this circuit, the $\mathbf{C}_4$ gate creates maximal control qudit superposition and the two-qudit operation $\mathbf{A}_{1,3}\times\mathbf{A}_{2,2}\times\mathbf{A}_{3,1}=\mathbf{A}_{(1,2,3),(3,2,1)}$ generates maximal entanglement among the qudit pair. When the initialized qudit state $\Ket{00_4}$ is evolved via the circuit in Fig. \[fig:full-ent-gen\], a fully entangled qudit pair results.
$$\label{eq:T_6}
\scriptsize{
\begin{split}
\mathbf{T}_{Fig.~\ref{fig:full-ent-gen}}\Ket{00_4}&= \mathbf{A}_{1,3}\times\mathbf{A}_{2,2}\times\mathbf{A}_{3,1} \times (\mathbf{C}_4 \otimes \mathbf{I}_4)\Ket{00_4} \\
&=\mathbf{A}_{(1,2,3),(3,2,1)}\times (\mathbf{C}_4 \otimes \mathbf{I}_4)\Ket{00_4} \\
&= \frac{1}{2}\left[\Ket{00_4} + \Ket{13_4} + \Ket{22_4} + \Ket{31_4} \right]. \\
\end{split}}$$
In Eqn. \[eq:T\_6\], all of the output value basis states have unique values for both qudits. Thus, the qudits are fully entangled. When other initial basis states of qudit pairs evolve through the example entanglement generator of Fig. \[fig:full-ent-gen\], alternative fully entangled qudit states result. These different fully entangled states resulting from the Fig. \[fig:full-ent-gen\] circuit are given in rightmost set of fully entangled results in Table \[tb:par\_full\_ent\_data\]. We note that this set of states is analogous to the radix-2 Bell states when the radix is extended to $r=4$. Therefore, one result of this paper is the generalization of the Bell state generator to a radix-$r$ system.
The qudit entanglement generator pictured in Fig. \[fig:full-ent-gen\] is not the only structure that creates maximally entangled radix-4 qudits. A $\mathbf{C}_4$ gate followed by any group of three different $\mathbf{A}_{h,k}$ gates, each with unique values for both $h$ and $k$, will create entangled states from qudits originally in a basis state. Since there are $r=4$ options for the control level, $h$, the three $\mathbf{A}_{h,k}$ gates needed for the full entanglement generator have a total of four different combinations for the implemented control values: $(\Ket{1_4}, \Ket{2_4}, \Ket{3_4})$, $(\Ket{0_4}, \Ket{1_4}, \Ket{2_4})$, $(\Ket{0_4}, \Ket{1_4}, \Ket{3_4})$, and $(\Ket{0_4}, \Ket{2_4}, \Ket{3_4})$. Considering the three $k$ values for the $r=4$ modulo-add-by-$k$ operations, there are six permutations for each control group. This gives a total of 24 radix-4 full entanglement generators. The order in which the $\mathbf{A}_{(h,k)}$ gates appear after the $\mathbf{C}_4$ gate is irrelevant, thus there are six orders for each of the 24 different full entanglement generators, resulting in 144 different and distinct $r=4$ full entanglement generators. In general, there are
$$\label{eq:total_max_ent_gen}
\prod_{i=2}^{r} (i^2-i)$$
different full entanglement circuits of the form described here for a pair of radix-$r$ qudits. Since combining $\mathbf{A}_{(h,k)}$ gates is commutative, $\mathbf{A}_{(h_1,k_1)} \times \mathbf{A}_{(h_2,k_2)} =\mathbf{A}_{(h_2,k_2)} \times \mathbf{A}_{(h_2,k_2)} $, the function
$$\label{eq:total_unique_max_ent_gen}
\prod_{i=2}^{r} \frac{(i^2-i)}{(i-1)} = \prod_{i=2}^{r} \frac{i(i-1)}{(i-1)} = \prod_{i=2}^{r} i = r!$$
is used to determine the number of maximal entanglement circuit configurations that produce unique transfer functions.
![Radix-4 full entanglement generator[]{data-label="fig:full-ent-gen"}](Figures/r4_full_ent_ex.png){height="0.75in"}
Conclusion
==========
We have introduced and defined the concept of partial entanglement for higher-dimensional QIS systems. In binary QIS, entanglement is either present or not. There is no concept of partial entanglement. In a higher-dimensioned QIS system, entanglement can be present in varying degrees. We have furthermore developed entanglement generators for higher-dimensional QIS systems. These generators, when they are configured to produce maximal entanglement, are directly analogous to the well known Bell state generators in binary qubit-based QIS. We have also shown how entanglement generators can be configured to produce partially entangled states. The partial and maximal entanglement state generators developed and described here use single-qudit Chrestenson operators for maximal superposition and controlled modulo-add-by-$k$ operators for entangling. These results will enable algorithms in binary QIS to be generalized and extended to higher-dimensional qudit-based systems.
[00]{}
P. Shor, “Algorithms for quantum computation: discrete logarithms and factoring,” in proc. *IEEE Ann. Symp. on Foundations of Computer Science*, pp. 124 – 134, 1994.
L. Grover, “A fast quantum mechanical algorithm for database search,” in proc. *ACM Symp. on Theory of Computing*, pp. 212 – 219, 1996.
Yin et al., “Satellite-based entanglement distribution over 1200 kilometers,” *Science* **356**, pp. 1140 – 1144, June 2017.
M. Enríquez, I. Wintrowicz, and K. Życzkowski, “Maximally entangled multipartite states: a brief survey," *J. Phys.: Conf. Ser.* vol. 698, no. 1, pp. 012003, 2016.
H. E. Chrestenson, “A class of generalized Walsh functions," *Pacific Journal of Mathematics*, vol. 5, no. 1, pp. 17 – 31, 1955.
N. Y. Vilenkin, “Concerning a class of complete orthogonal systems," *Dokl. Akad. Nauk SSSR, Ser. Math*, no. 11, 1947.
Z. Zilic and K. Radecka, “Scaling and better approximating quantum Fourier transform by higher radices," *IEEE Trans. on Computers*, vol. 56, no. 2, pp. 202 – 207, 2007.
Z. Zilic and K. Radecka, “The Role of Super-fast Transforms in Speeding up Quantum Computations,” in proc. *IEEE International Symposium on Multiple Valued Logic*, pp. 129 – 135, 2002.
K. N. Smith, T. P. LaFave, Jr., D. L. MacFarlane, and M. A. Thornton, “A Radix-4 Chrestenson Gate for Optical Quantum Computation,” in proc. *IEEE International Symposium on Multiple-Valued Logic*, pp. 260 – 265, 2018.
M. A. Thornton, D. W. Matula, L. Spenner, and D. M. Miller,“Quantum Logic Implementation of Unary Arithmetic Operations,” in proc. *IEEE International Symposium on Multiple-Valued Logic*, pp. 202 – 207, 2008.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A quantum deformation of the adjoint action of the special linear group on the variety of nilpotent matrices is introduced. New non-embedded quantum homogeneous spaces are obtained related to certain maximal coadjoint orbits, and known quantum homogeneous spaces are revisited.'
author:
- |
M. Domokos [^1]\
\
[Rényi Institute of Mathematics, Hungarian Academy of Sciences,]{}\
[P.O. Box 127, 1364 Budapest, Hungary,]{} [E-mail: [email protected] ]{}
title: A quantum homogeneous space of nilpotent matrices
---
MSC: 16W35; 20G42; 17B37; 81R50
Keywords: coadjoint orbit; quantum group; quantum homogeneous space; deformation; nilpotent matrices; comodule algebra
Introduction {#sec:intro}
============
The general linear group $GL(n,\mc)$ acts (from the right) on the space $M=M(n,\mc)$ of $n\times n$ matrices with complex entries by conjugation: $$\ad:M\times GL(n,\mc)\to M,\qquad (\xi,g)\mapsto g^{-1}\xi g.$$ This is an instance of the adjoint (coadjoint) action of a reductive Lie group on (the dual of) its Lie algebra. For the relevance of coadjoint orbits in representation theory and mathematical physics see [@kirillov].
Throughout this paper $G$ stands for the special linear group $SL(n,\mc)$. Note that two points of $M$ belong to the same $GL(n,\mc)$–orbit if and only if they belong to the same $G$–orbit. For the sake of technical convenience we shall restrict $\ad$ to $G$, and look for quantum deformations of the orbits. The precise formulation of the problem can be given in the language of affine algebraic geometry. We shall write $\calf(V)$ for the coordinate ring of an affine algebraic variety $V$. The action $\ad$ is encoded in its comorphism $$\ad^*:\calf(M)\to\calf(M)\otimes_{\mc}\fung,$$ which (in addition to being an algebra homomorphism) is a right coaction of the Hopf algebra $\fung$ on the algebra $\calf(M)$. Take $\xi\in M$ and its $G$–orbit $\orbxi$ in $M$. Recall that $\orbxi$ is closed if and only if $\xi$ is semisimple (i.e. diagonalizable). The Zariski closure $\orbclxi$ is an affine $G$–subvariety of $M$; equivalently, its coordinate ring $\calf(\orbclxi)$ is a quotient $\fung$–comodule algebra of $\calf(M)$. In other words, we have a commutative diagram $$\begin{array}{ccc}
\calf(M) &\stackrel{\ad^*}\longrightarrow &\calf(M)\otimes_{\mc}\fung\\
\downarrow & &\downarrow \\
\calf(\orbclxi)&\longrightarrow &\calf(\orbclxi)\otimes_{\mc}\fung
\end{array}$$ of algebra homomorphisms, where the vertical arrows are induced by restricting functions on $M$ to $\orbclxi$.
Our base ring to define quantum algebras will be $\mk=\mc[q]_{(q-1)}$, the localization of the polynomial ring $\mc[q]$ in the indeterminate $q$ at the maximal ideal $(q-1)$. In other words, $\mk$ is the subring of the field $\mc(q)$ of rational functions, consisting of the functions with no pole at $1$. Denote by $\funqg$ the quantum coordinate ring of $G$; this is a $\mk$–Hopf algebra defined in terms of generators and relations (see Section \[sec:main\]). Moreover, $\funqg$ is a deformation of $\fung$ in the following sense: $\funqg$ is a free $\mk$–module, and its quotient Hopf algebra modulo the ideal $(q-1)\funqg$ is isomorphic to $\fung$. (It is more standard to work over the ring $\mc[[h]]$ of formal power series when studying deformations of algebras, see for example [@bfgp]. However, all our constructions can be performed over $\mk$, and this choice of the base ring allows to relate our objects naturally to their classical counterpart in two different ways. The specialization $q\mapsto 1$ makes sense on the one hand, and on the other hand, it is also possible to extend scalars to $\mc(q)$ and formulate the results in a representation theoretic flavour.)
\[def:qdef\]
By a [*quantum deformation of an affine $G$–variety*]{} $V$ we mean a (right) $\funqg$–comodule algebra $\calf_q(V)$ with the following properties:
- $\calf_q(V)$ is a free $\mk$–module;
- $\calf_q(V)/(q-1)\calf_q(V)\cong \calf(V)$;
- The diagram $$\begin{array}{ccc}
\calf_q(V) &\stackrel{\varphi}\longrightarrow &\calf_q(V)\otimes_{\mk}\funqg\\
\downarrow & &\downarrow \\
\calf(V)&\longrightarrow &\calf(V)\otimes_{\mc}\fung
\end{array}$$ commutes, where the horizontal arrows are the coaction maps, and the vertical arrows are the natural surjections with kernel generated by $(q-1)$. When $V$ contains a dense $G$–orbit, we say that $(\calf_q(V),\varphi)$ is a [*quantum homogeneous $G$–space*]{}.
Denote by $P$ the subset of $M$ consisting of the matrices $\xi$ which have both of the following two properties:
- all non-zero eigenvalues of $\xi$ have multiplicity one;
- $\xi$ has only one nilpotent Jordan block.
Note that for $\xi\in P$ the $G$–orbit of $\xi$ is maximal; that is, it has maximal possible dimension. For all $\xi\in P$ we shall construct an $\funqg$–comodule algebra $\funqorb$, which is a quantum deformation of $\orbclxi$ in the above sense. In particular, we obtain quantum homogeneous spaces of nilpotent matrices. The algebra $\funqorb$ is defined as a quotient of the reflection equation algebra $\lqm$ modulo an ideal generated by adjoint invariants.
Quantizations of coadjoint orbits of $G$ given as quotients of the reflection equation algebra were considered in several papers, see for example [@dm1], [@dm2], [@dm3], [@gur-sap]. The case of semisimple orbits is settled in [@dm3]. Non-semisimple orbits of some special type are dealt with in [@dm1], [@dm2]; all of them are orbits of matrices with at most two eigenvalues, and with Jordan blocks of size one or two. The main reason that the methods of [@dm1], [@dm2] are restricted to such orbits is that the reflection equation algebra has few characters. In other words, all complex solutions of the reflection equations are matrices of this special type by [@mudrov].
There is a passage between the reflection equation algebra and the quantized coordinate ring $\funqm$ developed in [@bm], [@majid]. Exploiting this method, formulae from [@dl] provide a candidate for the ideal in the reflection equation algebra that should define $\funqorb$ for $\xi\in P$. The main technical problem is to prove that the resulting quotient is a free $\mk$–module. This is handled by applying results of [@dfl], that are based on the fact that sufficiently many characters of $\funqm$ are available to hit the orbit of each $\xi\in P$.
Main results {#sec:main}
============
First we recall two types of non-commutative algebras associated with the R-matrix of the vector representation of the Drinfeld-Jimbo algebra $U_q({\mathfrak{sl}}_n)$ (cf. [@drinfeld], [@jimbo]). Let $k$ be a commutative integral domain, $q$ an invertible element in $k$, and $R\in{\mathrm{End}}_k(k^n\otimes k^n)$ given by $$R^{is}_{jt}=\left\{
\begin{array}{cc}
q, &\mbox{ if }i=j=s=t;\\
1, &\mbox{ if }i=j,s=t,i\neq s;\\
q-q^{-1}, &\mbox{ if }i>j,i=t,j=s; \\
0, &\mbox{ otherwise}
\end{array}\right.$$ (see for example Section 9.2 of [@ks]). Set $\widehat{R}:=\flip\circ R$, where $\flip$ is the flip endomorphism $a\otimes b\mapsto b\otimes a$.
The [*quantum coordinate ring of $n\times n$ matrices*]{} is the $k$–algebra $\funkm$ generated by the entries of $X=[x^i_j]_{i,j=1}^n$, subject to the relations $$\widehat{R} (X\otimes I)(I\otimes X) =
(X\otimes I)(I\otimes X)\widehat{R}$$ (with the notation of [@rtf]). It has a bialgebra structure: the comultiplication and the counit maps are given on the generators by $\Delta(x^i_j)=x^i_s\otimes x^s_j$ (we use the convention of summing over repeated indices), and $\varepsilon(x^i_j)=\delta^i_j$. The [*quantum determinant*]{} $\qdet=\sum_{\sigma\in{\mathrm{Sym}}_n}
(-q)^{{\mathrm{length}}(\sigma)}
x^1_{\sigma(1)}x^2_{\sigma(2)}\cdots x^n_{\sigma(n)}$ is a central group-like element in $\funkm$. The quotient bialgebra of $\funkm$ modulo the ideal generated by $(\qdet-1)$ has an antipode $S$ that makes it a $k$–Hopf algebra $\funkg$, called the [*quantum coordinate ring of the special linear group*]{}. We shall denote by $t^i_j$ the image of $x^i_j$ under the natural homomorphism $\pi:\funkm\to\funkg$. The right adjoint coaction of $\funkg$ on itself can be lifted to a coaction $\beta:\funkm\to\funkm\otimes\funkg$: for $x\in\funkm$ with $\Delta^{(2)}(x)=\sum x_1\otimes x_2\otimes x_3$, we have $\beta(x)=\sum x_2\otimes S(\pi(x_1))\pi(x_3)$. We shall be interested in these algebras in the case when $(k,q)$ is $(\mk,q)$ or $(\mc(q),q)$. Moreover, to simplify notation we write $\funqm$ and $\funqg$ instead of $\funkkm$ and $\funkkg$.
The second algebra $\lkqm$ is the $k$–algebra generated by the entries of $L=[l^i_j]_{i,j=1}^n$, subject to the relations $$\label{eq:reflection}
\widehat{R} (I\otimes L)\widehat{R} (I\otimes L) =
(I\otimes L)\widehat{R}(I\otimes L)\widehat{R}.$$ This algebra is called the [*reflection equation algebra*]{} in [@kulish-sk], and the algebra of [*braided matrices*]{} in [@majid2]. An important property of $\lkqm$ is that the map $l^i_j\mapsto l^a_b\otimes S(t^i_a)t^b_j$, $i,j=1,\ldots,n$, extends to an algebra homomorphism $\beta:\lkqm\to\lkqm\otimes_k\funkg$, yielding a right coaction of $\funkg$ on $\lkqm$. In fact, the $\funqg$–comodule algebra $\lqm$ is a quantum deformation (in the sense of Definition \[def:qdef\]) of the $SL(n,\mc)$–variety $M$ endowed with the adjoint action, where $\lqm$ stands for ${\cal{L}}_{\mk,q}(M)$.
Both $\funqm$ and $\lqm$ are graded, the generators having degree one. A crucial fact for us is that there is a $\mk$–module isomorphism $\Phi:\funqm\to\lqm$, intertwining the right $\funqg$–coactions $\beta$ on $\funqm$ and $\lqm$. This is explained in Section 7.4 of [@majid], explicit formulae are given in (7.37); see also Section 10.3 in [@ks]. In particular, $\Phi(x^i_j)=l^i_j$ for all $i,j$, and $\Phi$ is homogeneous. We shall denote by $\Psi:\lqm\to\funqm$ the inverse $\mk$–module map to $\Phi$.
The [*quantum trace*]{} of an arbitrary $n\times n$ matrix $M$ with entries $m^i_j$ in a $\mk$–module is defined as $$\qtr(M)=\sum_{i=1}^n q^{n+1-2i}m^i_i.$$ The elements $\qtr(L^i)$, $i=1,2,\ldots$, are coinvariants for the coaction of $\funqg$ on $\lqm$; that is, $\beta(\qtr(L^i))=\qtr(L^i)\otimes 1$. Moreover, they are central in $\lqm$, see for example Theorem 10.3.8 in [@majid].
We need to recall the basic $\funqg$–coinvariants $\tau_1,\ldots,\tau_n$ in $\funqm$ with respect to the coaction $\beta$, introduced in [@dl]. The first of them, $\tau_1$, is the quantum trace $\qtr(X)$, the last of them, $\tau_n$ is the quantum determinant $\qdet$. For a $d$–element subset $I=\{i_1,\ldots,i_d\}$ of $\{1,\ldots,n\}$, the submatrix of $X$ consisting of the entries with row and column indices belonging to $I$ is a $d\times d$ quantum matrix, its quantum determinant will be denoted by $\qdet(I,I)$. For $d=1,\ldots,n$, we have $$\tau_d=\sum_{I=\{i_1,\ldots,i_d\}}
q^{(d(n+1)-2\sum_{i\in I}i)} \qdet(I,I),$$ where the summation ranges over the $d$–element subsets of $\{1,\ldots,n\}$.
Now take an $n\times n$ complex matrix $\xi$ from the set $P$ defined in Section \[sec:intro\]. In the $G$–orbit of $\xi$ choose a block diagonal Jordan normal form $$\jordxi={\mathrm{diag}}(J_r,\lambda_1,\ldots,\lambda_{n-r})
=\left(\begin{array}{cccccc}0 & 1 & & & & \\ & 0 & 1 & & & \\ & & 0 & & & \\ & & & \lambda_1 & & \\ & & & & \ddots& \\ & & & & & \lambda_{n-r}\end{array}\right),$$ where $J_r$ is an $r\times r$ nilpotent Jordan block ($r\in\{0,1,\ldots,n\}$), and $\lambda_1,\ldots,\lambda_{n-r}$ are pairwise different non-zero eigenvalues of $\xi$. The map $X\mapsto \jordxi$ extends to a $\mk$–algebra homomorphism $\evxi:\funqm\to\mk$; we write $\tau_d(\xi)$ for the image of $\tau_d$ under $\evxi$. Consider the elements $\Phi(\tau_d-\tau_d(\xi))=\Phi(\tau_d)-\tau_d(\xi)\in\lqm$, for $d=1,\ldots,n$. They are $\funqg$–coinvariants with respect to the coaction $\beta$, hence they are central in $\lqm$ by the results in [@majid]. Write $\idealxi$ for the ideal of $\lqm$ generated by them. Since this ideal is generated by coinvariants, it is a subcomodule with respect to the coaction $\beta$. We define $\funqorb$ as the quotient $\funqg$–comodule algebra $$\funqorb=\lqm/\idealxi$$ of $\lqm$. We shall keep the notation $\beta$ for the coaction of $\funqg$ on $\funqorb$.
\[thm:main1\] Let $\xi$ be an $n\times n$ complex matrix from $P$. Then the $\funqg$–comodule algebra $\funqorb$ is a quantum deformation (in the sense of Definition \[def:qdef\]) of $\orbclxi$, the closure of the $G$–orbit of $\xi$ in $M$.
Thus we obtained some new quantum homogeneous spaces. A particularly interesting extreme case is when $\xi=J_n$, the nilpotent $n\times n$ Jordan block. Then $\orbclxi$ is the set of all nilpotent $n\times n$ matrices, write $\nilp$ for this affine variety. It is the nullcone for the coadjoint action of $SL(n,\mc)$ on the dual of its Lie algebra, a relevant object both from the point of view of invariant theory and representation theory, see for example [@kostant], [@kraft].
\[thm:main2\] Let $\funqn$ denote the quotient $\funqg$–comodule algebra of $\lqm$ modulo the ideal generated by $\qtr(L^d)$, $d=1,\ldots,n$. Then $\funqn$ is a quantum deformation of the affine $G$–variety $\nilp$ of nilpotent $n\times n$ complex matrices.
Proofs {#sec:proofs}
======
\[lemma:qslfree\] Every $\mk$–submodule of $\funqg$ is free.
It is well known that every submodule of a free module over a principal ideal domain is free, see for example the remark after Theorem 5.3 in Chapter I of [@em]. Therefore it is sufficient to prove that $\funqg$ is a free $\mk$–module.
We write $\funkg^d$ for the $k$-submodule of $\funkg$ spanned by the products of degree at most $d$ in the generators $t^i_j$. Then $\funkg^d$ is the sum of the spaces of matrix elements of the tensor powers of exponent $\leq d$ of the fundamental corepresentation $e_j\mapsto e_i\otimes t^i_j$ $(j=1,\ldots,n)$ of $\funkg$ (cf. Section 11.2.3 in [@ks]). Since $\funkqg$ and $\fung$ have essentially the same corepresentation theory (see for example [@hayashi] or [@nym]), we have the equality $\dim_{\mc(q)}(\funkqg^d)=\dim_{\mc}(\fung^d)$ for $d=0,1,2,\ldots$.
By definition $\mk$ is a subring of $\mc(q)$; write $\cala$ for the $\mk$–subalgebra of $\funkqg$ generated by the $t^i_j$. Since these generators satisfy the defining relations of $\funqg$, the algebra $\cala$ is a homomorphic image of $\funqg$ (actually, they are isomorphic as we shall see below). Now we make use of the basis of the quantum coordinate ring of $n\times n$ matrices developed in [@gl]. It consists of certain (called [*preferred*]{} in loc. cit.) products of quantum minors. Although [@gl] works over a field, the arguments obviously show that as a $\mk$–module, $\funqg$ is spanned by the preferred products of quantum minors of size $\leq n-1$, and that the images of these elements in $\cala$ span $\funkqg$ as a $\mc(q)$–vector space. From the classical theory we know that the number of preferred products of (quantum) minors of size $\leq n-1$ and of total degree $\leq d$ is $\dim_{\mc}(\fung^d)$. Hence by the above dimension equality we obtain both that the homomorphism $\funqg\to\cala$ is an isomorphism and that $\funqg$ is a free $\mk$–module with basis consisting of the preferred products of quantum minors of size $\leq n-1$.
In particular, we may identify $\funqg$ with the $\mk$–subalgebra of $\funkqg$ generated by the elements $t^i_j$ $(i,j=1,\ldots,n)$, and we shall make this identification for the rest of the paper. Similarly, $\funqm$ is identified with a $\mk$–subalgebra of $\funkqm$ in the obvious way. Throughout this Section we shall use also the following convention. Given a $\mk$–submodule $W$ of $\funqm$, the symbol $\widehat{W}$ stands for the $\mc(q)$–vector subspace of $\funkqm$ spanned by $W$. Moreover, for any $\mk$–module $W$ we write $\overline{W}$ for the $\mc$–vector space obtained by taking the quotient of $W$ modulo the submodule $(q-1)W$, and write $\overline{w}\in\overline{W}$ for the image of $w\in W$ under the natural surjection. Note that by the defining relations of $\funqg$ it is clear that $\funqg/(q-1)\funqg$ is isomorphic to $\fung$, with an isomorphism mapping the generators $t^i_j$ to the corresponding coordinate functions on $G$.
\[lemma:right-ideal\] The right ideal $\ridealxi$ of $\funqm$ generated by $\tau_1-\tau_1(\xi),\ldots,\tau_n-\tau_n(\xi)$ is mapped by $\Phi$ onto the ideal $\idealxi$ in $\lqm$.
Note that $\Phi$ is not an algebra homomorphism. However, if $\beta(a)=a\otimes 1$ for some $a\in\funqm$, then $\Phi(ab)=\Phi(a)\Phi(b)$ in $\lqm$ for all $b\in\funqm$. This follows from the formula of Theorem 7.4.1 in [@majid] (see also the formula of Proposition 34 (i) in Section 10.3.2 of [@ks]), which expresses the multiplication in $\lqm$ in terms of the multiplication in $\funqm$. Therefore, $\Phi(\ridealxi)$ is the right ideal in $\lqm$ generated by $\Phi(\tau_1-\tau_1(\xi)),\ldots,\Phi(\tau_n-\tau_n(\xi))$. This latter right ideal is a two-sided ideal, since it is generated by central elements.
\[lemma:1\] We have the equality $\widehat{\ridealxi}\cap \funqm=\ridealxi$ for all $\xi\in P$.
Take sets of monomials $\Gamma_i\subset \funqm$ $(i=1,\ldots,n)$ in the variables $x^s_t$ such that $$\overline{\Lambda}=\bigcup_{i=1}^n\{\overline{(\tau_i-\tau_i(\xi))w}
\mid w\in\Gamma_i\}$$ is a $\mc$–basis of $\overline{\ridealxi}$ compatible with the filtration by degree. Using the corepresentation theory of $\funkqg$, classical results of Kostant [@kostant], and a commutative algebra lemma (cf. Lemma 5.2 in [@dfl]), it was proved in [@dfl] (see the last line of the proof of Proposition 5.3 in loc. cit.) that $$\Lambda=\bigcup_{i=1}^n\{(\tau_i-\tau_i(\xi))w
\mid w\in\Gamma_i\}$$ is a $\mc(q)$–basis of $\widehat{\ridealxi}$. An arbitrary element $f$ in the intersection $\widehat{\ridealxi}\cap \funqm$ can be uniquely written as a finite sum $f=\sum_{\lambda\in\Lambda}a_{\lambda}\lambda$, where the coefficients $a_{\lambda}$ are taken from $\mc(q)$. Assume that for some $\mu\in\Lambda$, the coefficient $a_{\mu}$ is not contained in $\mk$. Then there is a positive integer $d$ such that $(q-1)^da_{\lambda}\in\mk$ for all $\lambda\in\Lambda$, and $(q-1)^da_{\mu}$ is not contained in $(q-1)\mk$. Reducing the equality $(q-1)^df=\sum_{\lambda\in\Lambda}(q-1)^da_{\lambda}\lambda$ in $\funqm$ modulo the ideal $(q-1)$ we obtain that a non-trivial $\mc$–linear combination of the elements of $\overline{\Lambda}$ is zero. This contradiction shows that all the $a_{\lambda}$ are contained in $\mk$, hence $f$ is contained in $\ridealxi$.
\[thm:main1\]. The $SL(n,\mc)$–orbit of $\xi\in P$ is maximal (i.e. is not contained in the closure of another orbit). Therefore the vanishing ideal of $\orbclxi\subset M$ in $\calf(M)=\lqm/(q-1)\lqm$ is $\overline{\idealxi}$ by classical results of [@kostant]. Thus $\funqorb$ satisfies (ii) and (iii) from Definition \[def:qdef\] by construction. The only thing we have to prove is that (i) holds as well, namely that $\funqorb$ is a free $\mk$–module. By Lemma \[lemma:right-ideal\], the $\mk$–module isomorphism $\Psi:\lqm\to\funqm$ induces a $\mk$–module isomorphism between $\funqorb$ and $\funqm/\ridealxi$. Denote by $\evhatxi:\funkqm\to\mc(q)$ the $\mc(q)$–linear extension of $\evxi:\funqm\to\mk$ defined in Section \[sec:main\], and denote by $\widehat{\beta}$ the $\mc(q)$–linear extension of the coaction $\beta:\funqm\to\funqm\otimes_{\mk}\funqg$ to $\funkqm$. Consider the $\mc(q)$–linear map $\widehat{\gamma}=(\evhatxi\otimes\id)\circ\widehat{\beta}
:\funkqm\to\funkqg$, and its restriction $\gamma$ to $\funqm$. The kernel of $\widehat{\gamma}$ is $\widehat{\ridealxi}$ by Theorem 5.4 of [@dfl]. (To be more precise, the statement of [@dfl] is about $\calf_{\mc(q),q}(GL(n,\mc))$, instead of $\funkqg$. This does not make any difference, as the proof of the cited result works word by word if we replace $\calf_{\mc(q),q}(GL(n,\mc))$ by $\funkqg$; alternatively, one can refer to the results of the last Section of [@dl2].) It follows by Lemma \[lemma:1\] that $\ker(\gamma)=\ridealxi$. The image of the $\mk$–module homomorphism $\gamma$ is contained in $\funqg$. Therefore $\funqm/\ridealxi$ is isomorphic to a $\mk$–submodule of $\funqg$, hence is free by Lemma \[lemma:qslfree\].
\[thm:main2\] The elements $\Phi(\tau_1),\ldots,\Phi(\tau_n)$ generate the same ideal in $\lqm$ as the elements $\qtr(L),\qtr(L^2),\ldots,\qtr(L^n)$ by Lemma \[lemma:newton\] below.
\[lemma:newton\] The elements $\Phi(\tau_1),\ldots,\Phi(\tau_n)$ generate the same $\mk$-subalgebra in $\lqm$ as the elements $\qtr(L),\ldots,\qtr(L^n)$.
Denote by $A$, $B$, and $C$, respectively, the $\mk$–subalgebra of $\lqm$ generated by the first set, the second set, and their union, respectively. These are graded subalgebras of $\lqm$. Their degree $d$ homogeneous components $A_d$, $B_d$, $C_d$ are finitely generated $\mk$–modules. By the classical Newton formulae relating the characteristic coefficients of a matrix with the traces of its powers, their images $\overline{A}_d$, $\overline{B}_d$, $\overline{C}_d$ in $\calf(M)$ coincide. Applying Nakayama’s Lemma for finitely generated modules over the local principal ideal domain $\mk$ we obtain $C_d=A_d$ and $C_d=B_d$. This holds for all $d$, hence $A=C=B$.
Note that Lemma \[lemma:newton\] implies the existence of Newton formulae relating $\qtr(L),\ldots,\qtr(L^n)$ and $\Phi(\tau_1),\ldots,\Phi(\tau_n)$. A version of Newton formulae in $\lqm$ is given in [@iop].
Non-embedded quantum homogeneous spaces {#sec:non-embedded}
=======================================
\[def:non-embedded\]
A quantum homogeneous $G$–space $(\calf_q(V),\varphi)$ is [*embedded*]{} if there is a right coideal subalgebra $\cala$ in $\funqg$ such that
- $\calf_q(V)$ is isomorphic to $\cala$ as an $\funqg$–comodule algebra, where the coaction on the latter is the restriction to $\cala$ of the comultiplication $\Delta$.
- The quotient $\mk$–module $\funqg/\cala$ is torsion free.
Condition (i) above is standard, see for example Section 11.6.1 in [@ks]. Condition (ii) ensures that the natural surjection of $\calf_q(V)$ onto its ‘classical limit’ $\calf_q(V)/(q-1)\calf_q(V)$ can be identified with the restriction to $\cala$ of the natural surjection $\funqg\to\fung$.
The following result shows that our $\funqorb$ is a non-embedded quantum homogeneous space for most $\xi\in P$. This indicates that the use of the interplay between $\lqm$ and $\funqm$ due to [@bm], [@majid] combined with the ‘orbit map’ of [@dfl] can not be replaced by an orbit map going directly from $\lqm$ to $\funqg$ (like in [@dm2]).
\[prop:embedded\] Assume that for $\xi\in P$, one of the following holds:
- the size of the nilpotent Jordan block of $\xi$ is at least three;
- $n\geq 3$, and the size of the nilpotent Jordan block of $\xi$ is two;
- $\xi$ has at least three (different) non-zero eigenvalues.
Then $\funqorb$ is not an embedded quantum homogeneous space.
Assume in the contrary that for a $\xi$ satisfying one of the conditions (i), (ii), (iii), there exists a $\funqg$–comodule algebra injection $\iota: \funqorb\to\funqg$ such that $\funqg/\image(\iota)$ is torsion free as a $\mk$–module. Then we have a commutative diagram of algebra homomorphisms $$\begin{array}{ccccc}
\funqorb &\stackrel{\iota}\longrightarrow &\funqg
&\stackrel{\varepsilon}\longrightarrow &\mk\\
\downarrow & &\downarrow & &\downarrow\\
\calf(\orbclxi)&\stackrel{\bar\iota}\longrightarrow &\calf(G)
&\stackrel{\bar\varepsilon}\longrightarrow&\mc
\end{array}$$ where $\varepsilon$, $\bar\varepsilon$ denote the counit map of the corresponding Hopf algebra, and the vertical arrows are the natural surjections with kernel generated by $q-1$. Since $\funqg/\image(\iota)$ is torsion free as a $\mk$–module, the map $\bar\iota$ is also injective. Recall that $\funqorb$ is defined as a quotient of $\lqm$. Keep the notation $l^i_j$ for the images of the generators of $\lqm$ in $\funqorb$, as well as in $\calf(\orbclxi)$. The fact that $\bar\iota$ is injective implies that the complex $n\times n$ matrix $\overline{B}=[\bar\varepsilon\bar\iota(l^i_j)]_{n\times n}$ belongs to the same $G$–orbit as $\xi$. On the other hand, this matrix is obtained by reducing modulo $(q-1)$ the $n\times n$ matrix $B=[\varepsilon\iota(l^i_j)]_{n\times n}$ with entries in $\mk$. Now $B$ satisfies the reflection equation (\[eq:reflection\]). From the classification of solutions of the reflection equation given in [@mudrov] we see that it is impossible for a solution $B$ that $\overline{B}$ has the same Jordan normal form as $\xi$ (the paper [@mudrov] classifies complex solutions, but the proofs and the result can easily be extended to the case when $\mc$ is replaced by the domain $\mk$). This contradiction finishes the proof.
Multiplicities of irreducible corepresentations {#sec:multiplicities}
===============================================
Let us recall that the corepresentation theory of the $\mc(q)$–Hopf algebra $\funkqg$ is completely analogous to its classical counterpart for $\fung$ (i.e. the representation theory of $G$ expressed in a dual language). Namely, $\funkqg$ is cosemisimple, and its irreducible corepresentations are indexed by $\Omega$, the set of dominant integral weights for $G$; see for example the book [@ks]. For $\lambda\in\Omega$ denote $\varphi_{\lambda}:V_{\lambda}\to V_{\lambda}\otimes_{\mc(q)}\funkqg$ the irreducible corepresentation indexed by $\lambda$. Here $V_{\lambda}$ has a decomposition into a direct sum of weight subspaces having the same dimension as in the classical case of $\fung$. Given $\lambda\in\Omega$ denote by $c(\lambda)$ the dimension of the trivial weight subspace in the dual of the irreducible $SL(n,\mc)$–representation with highest weight $\lambda$. It is proved in [@kostant] that for any $\xi\in M(n,\mc)$ whose orbit is maximal, the multiplicity of the irreducible $SL(n,\mc)$–module associated to $\lambda$ in the coordinate ring $\calf(\orbclxi)$ is $c(\lambda)$ for all $\lambda\in\Omega$. The fact that $\funqorb$ is an appropriate deformation of $\calf(\orbclxi)$ can be expressed in terms of corepresentation theory as follows. Extend scalars and set $\funkqorb=\mc(q)\otimes_{\mk}\funqorb$. Since $\funqorb$ is a flat $\mk$–module, the $\mc(q)$–algebra $\funkqorb$ has the same presentation in terms of generators and relations over $\mc(q)$ as the algebra $\funqorb$ over $\mk$. Now $\funkqorb$ is a right comodule algebra over $\funkqg$ and we have the following decomposition of this corepresentation.
\[thm:multiplicity\] The corepresentation $\beta:\funkqorb\to\funkqorb\otimes_{\mc(q)}\funkqg$ for $\xi\in P$ decomposes into a direct sum of irreducible corepresentations as $$\beta\cong\sum^\oplus_{\lambda\in\Omega}c(\lambda)\varphi_{\lambda}.$$
For $d=0,1,2,\ldots$, denote by $\funkqorb^d$ the $\mc(q)$–subspace spanned by the products of length $\leq d$ in the generators $l^i_j$ of $\funkqorb$. This is a finite dimensional subcomodule, containing the free $\mk$–submodule $\funqorb^d$. A weight subspace of $\funkqorb^d$ of dimension $r$ intersects $\funqorb^d$ in a free $\mk$–submodule of rank $r$. Under the natural surjection $\funqorb^d/(q-1)\funqorb^d\cong \calf(\orbclxi)^d$ this $\mk$–submodule is mapped onto an $r$–dimensional classical weight subspace belonging to the same weight of $G$. Therefore the dimensions of the weight subspaces in $\funkqorb^d$ agree with the dimensions of the corresponding weight subspaces in $\calf(\orbclxi)^d$. The multiplicities of the irreducible summands can be calculated from the weight multiplicities by the same rule for $\funkqg$ as for $\fung$, hence the statement follows from the corresponding classical result for $G$.
Quantum spheres {#sec:spheres}
===============
Podleś [@p] introduced a family of quantum $2$–spheres, which became frequently cited examples in the literature on quantum homogeneous spaces. Here we point out that they can be obtained as coadjoint orbits in the case $n=2$. In this Section we change the setup and take $k=\mc$ as our base ring, so $q\in\mc^{\times}$ is a non-zero complex number. The algebra $\lcqm$ in the special case $n=2$ is generated by $l_{11}$, $l_{12}$, $l_{21}$, $l_{22}$, subject to the relations
$$\begin{aligned}
\label{eq:2rea}
l_{22}l_{12}=q^2l_{12}l_{22};&\qquad
l_{11}l_{12}=l_{12}l_{11}+(q^{-2}-1)l_{12}l_{22};&\\
l_{11}l_{22}=l_{22}l_{11};&\qquad
l_{21}l_{12}=l_{12}l_{21}+(q^{-2}-1)l_{22}(l_{22}-l_{11});&\\
l_{21}l_{22}=q^2l_{22}l_{21};&\qquad
l_{21}l_{11}=l_{11}l_{21}+(q^{-2}-1)l_{22}l_{21}.&\end{aligned}$$
The basic coinvariants in $\lcqm$ are $$\Phi(\tau_1)=ql_{11}+q^{-1}l_{22}=\qtr(L)$$ and (after an easy calculation) $$\Phi(\tau_2)=l_{11}l_{22}-q^2l_{12}l_{21}
=(q+q^{-1})^{-1}(q\qtr(L)^2-q^2\qtr(L^2)).$$
Classically a maximal $2\times 2$ coadjoint orbit is determined by specifying the values of the trace and the determinant. Motivated by this (or by the previous Sections), for a pair of complex parameters $(t,d)$ we consider the quotient algebra $$\lcqtd=\lcqm/(\qtr(L)=t,\Phi(\tau_2)=d).$$
With the notation of [@ks], for a pair of complex parameters $\alpha,\beta$, the Podleś sphere $\qsphere$ is generated by $x_{-1},x_0,x_1$, subject to the relations (71)–(74) on page 124 of [@ks].
\[prop:podles\] The $\funqg$–comodule algebra $\lcqtd$ is isomorphic to $\qsphere$, the Podleś sphere with parameters $\alpha=q^{-1}(q+q^{-1})^{-1/2}t$, $\beta=\alpha^2-(q^{-1}+q^{-3})d$.
Both algebras are defined in terms of generators and relations. Write $l_{ij}$ for the image in $\lcqtd$ of the corresponding generator of $\lcqm$. By the relation $\qtr(L)=t$ in $\lcqtd$ we can eliminate one of the generators. Thus $\lcqtd$ is generated by $il_{12},il_{21},(q+q^{-1})^{-1/2}(l_{11}-l_{22})$, where $i\in\mc$ is the imaginary unit with $i^2=-1$. Expressing the defining relations of $\lcqtd$ in terms of these three generators one gets the defining relations of $\qsphere$, so the map $il_{21}\mapsto x_{-1}$, $il_{12}\mapsto x_1$, $(q+q^{-1})^{-1/2}(l_{11}-l_{22})\mapsto x_0$ extends to an algebra isomorphism $\lcqtd\to\qsphere$.
[**Acknowledgement**]{}
The author thanks Pham Ngoc Ánh, Rita Fioresi, and Tom Lenagan for conversations on the topic of this paper.
[MMMM]{}
P. Bonneau, M. Flato, M. Gerstenhaber, and G. Pinczon, The hidden group structure of quantum groups: strong duality, rigidity and preferred deformations, Comm. Math. Phys. 161 (1994), 125-156.
T. Brzeziński and S. Majid, A class of bicovariant differential calculi on Hopf algebras, Lett. Math. Phys. 26 (1992), 67-78.
H. Cartan and S. Eilenberg, Homological Algebra, Princeton University Press, Princeton, 1956 (1999).
M. Domokos, R. Fioresi and T. H. Lenagan, Orbits for the adjoint coaction on quantum matrices, J. Geometry and Physics 47 (2003), 447-468.
M. Domokos and T. H. Lenagan, Conjugation coinvariants of quantum matrices, Bull. London Math. Soc. 35 (2003) 117-127.
M. Domokos and T. H. Lenagan, Weakly multiplicative coactions of quantized function algebras, J. Pure Appl. Alg. 183 (2003) 45-60.
J. Donin and A. Mudrov, $U_q({\rm sl}(n))$–covariant quantization of symmetric coadjoint orbits via reflection equation algebra, Quantization, Poisson brackets and beyond (Manchester, 2001), 61–79, Contemp. Math., 315, Amer. Math. Soc., Providence, RI, 2002.
J. Donin and A. Mudrov, Method of quantum characters in equivariant quantization, Comm. Math. Phys. 234 (2003), 533-555.
J. Donin and A. Mudrov, Explicit equivariant quantization on coadjoint orbits of ${\rm GL}(n,\mathbb C)$, Lett. Math. Phys. 62 (2002), no. 1, 17-32.
V. G. Drinfeld, Quantum groups. Proceedings of the International Congress of Mathematicians, Vol. 1, 2 (Berkeley, Calif., 1986), 798–820, Amer. Math. Soc., Providence, RI, 1987.
K. R. Goodearl and T. H. Lenagan, Quantum determinantal ideals, Duke Math. J. 103 (2000), 165-190.
D. Gurevich and P. Saponov, Quantum line bundles via Cayley-Hamilton identity, J. Phys. A: Math. Gen. 34 (2001), 4553-4569.
T. Hayashi, Quantum deformations of classical groups, Publ. RIMS Kyoto Univ. 28 (1992), 57-81.
A. Isaev, O. Ogievetsky, and P. Pyatov, Generalized Cayley-Hamilton-Newton identities, Czechoslovak J. Phys. 48 (1998), no. 11, 1369-1374.
Michio Jimbo, A $q$-analogue of $U({\mathfrak{gl}}(N+1))$, Hecke algebra, and the Yang-Baxter equation. Lett. Math. Phys. 11 (1986), 247-252.
A. A. Kirillov, Lectures on the Orbit Method, Amer. Math. Soc., Povidence RI, 2004.
A. Klimyk and K. Schmüdgen, Quantum Groups and Their Representations, Springer-Verlag, Berlin, Heidelberg, New York, 1997.
B. Kostant, Lie group representations on polynomial rings, Amer. J. Math. 85 (1963), 327-404.
H. Kraft, Geometrische Methoden in der Invariantentheorie, (German), Vieweg Verlag, Braunschweig-Wiesbaden, 1985.
P. P. Kulish and E. K. Sklyanin, Algebraic structures related to reflection equations, J. Phys. A: Math. Gen. 25 (1992), 5963-5975.
S. Majid, Foundations of Quantum Group Theory, Cambridge Univ. Press, Cambridge, 1995.
S. Majid, Examples of braided groups and braided matrices, J. Math. Phys. 31 (1991), 3246-3253.
A. Mudrov, Characters of $U_q(gl(n))$–reflection equation algebra, Lett. Math. Phys. 60 (2002), 283-291.
M. Noumi, H. Yamada, and K. Mimachi, Finite dimensional representations of the quantum group $GL_q(n;\mc)$ and the zonal spherical functions on $U_q(n-1)\backslash U_q(n)$, Japanese J. Math. [**19**]{} (1993), 31-80.
P. Podleś, Quantum spheres, Lett. Math. Phys. [**14**]{} (1987), 193-202.
N. Yu. Reshetikhin, L. A. Takhtadzhyan, and L. D. Faddeev, Quantization of Lie groups and Lie algebras, (Russian), Algebra i Analiz 1 (1989), 178–206.
[^1]: Partially supported by OTKA No. T034530, T046378, and the Bolyai Fellowship.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We introduce unbiased estimators for the Shannon entropy and the class number, in the situation that we are able to take sequences of independent samples of arbitrary length.'
address:
- 'Math Dept, University of Missouri, Columbia, MO 65211, U.S.A.'
- 'Jülich Supercomputing Centre, Jülich Research Centre, 52425 Jülich, Germany.'
author:
- 'Stephen Montgomery-Smith'
- Thomas Schürmann
title: Unbiased Estimators for Entropy and Class Number
---
Introduction {#introduction .unnumbered}
------------
This paper supposes that we may pick a sequence of arbitrary length of independent samples $w_1,w_2,\dots$ from an infinite population. Each sample belongs to one of $M$ classes $C_1,C_2,\dots,C_M$, and the probability that a sample belongs to class $C_i$ is $p_i$. So these probabilities satisfy the constraints $0\le p_i\le 1$ and $\sum_{i=1}^M\,p_i =1$.
The goal of this paper is to present methods to estimate the Shannon entropy $H=-\sum_{i=1}^M\,p_i\log(p_i)$ (see [@shann]). An obvious method is to take a sample of size $n$, and compute the estimators $\hat p_i = k_i/n$, where $k_i$ are the number of samples from the class $C_i$. However this is known to systematically underestimate the entropy, and it can be significantly biased [@grass88][@harris][@herzel][@miller]. Recently there have been more advanced estimators for the entropy which have smaller bias [@grass88][@grass03][@harris][@herzel][@miller][@schuer].
In this paper we introduce new entropy estimators that have bias identically zero. We also introduce an unbiased estimator for the class number $M$, a problem of interest to ecologists (see for example the review article [@chao05]). The disadvantage of all our methods is that there is no *a priori* estimate of the sample size. For this reason, we postpone rigorous analysis of variance and other measures of confidence until it becomes clear that these estimators are of more than theoretical value.
We will use the following power series. Define the harmonic number by $h_n = \sum_{k=1}^n 1/k$, $h_0 = 0$. Then for $|x|<1$ $$\begin{aligned}
\frac1{(1-x)^2} = \displaystyle \sum_{k=1}^\infty k x^{k-1} ,
\\
\log(1-x) = - \displaystyle \sum_{k=1}^\infty \frac{x^k}k ,
\\
\frac{\log(1-x)}{(1-x)} = - \displaystyle \sum_{k=1}^\infty h_k x^k .\end{aligned}$$
First Estimator for Entropy {#first-estimator-for-entropy .unnumbered}
---------------------------
For each $1 \le i \le M$, let $N_i$ denote the smallest $k \ge 1$ for which $w_k \in C_i$. Then $$\hat H_1 = \sum_{i=1}^M \frac{I_{N_i\ge 2}}{N_i-1}$$ is an unbiased estimator for the entropy. The proof is straightforward. The marginal distribution of $N_i$ satisfies the geometric distribution $\Pr(N_i = k) = p_i (1-p_i)^{k-1}$. Thus $$E\left(\frac{I_{N_i\ge 2}}{N_i-1}\right) =
\sum_{k=2}^\infty \frac{p_i(1-p_i)^{k-1}}{k-1} = - p_i \log(p_i) .$$
Depending upon the applications, a possible disadvantage of this estimator is that complete knowledge of all possible classes needs to be known in advance.
Second Estimator for Entropy {#second-estimator-for-entropy .unnumbered}
----------------------------
Let $N$ denote the smallest $k \ge 1$ such that $w_1$ and $w_{k+1}$ belong to the same class. Then $\hat H_2 = h_{N-1}$ is an unbiased estimator for the entropy.
This follows, since conditional upon $w_1 \in C_i$, the distribution of $N$ satisfies the geometric distribution $\Pr(N = k | w_1 \in C_i) = p_i (1-p_i)^{k-1}$. Thus $$E(h_{N-1} | w_1 \in C_i) = \sum_{k=1}^\infty h_{k-1} p_i (1-p_i)^{k-1}
= - \log(p_i) ,$$ and hence $$E(h_{N-1})
= \sum_{i=1}^M E(h_{N-1} | w_1 \in C_i) \Pr(w_1 \in C_i)
= - \sum_{i=1}^M p_i \log(p_i) .$$
While we don’t wish to focus on analysis of the variance, it is certainly clear that this single estimator by itself will have unusable confidence limits. While the variance can be reduced by taking the mean of $n$ of these estimators, we propose the following version. For each $1 \le j \le n$, let $N^{(j)}$ be the smallest $k \ge 1$ such that $w_j$ and $w_{k+j}$ belong to the same class. Then the unbiased estimator is $$\hat H_3 = \frac1n \sum_{j=1}^n h_{N^{(j)}-1} .$$
Estimator for Class Number {#estimator-for-class-number .unnumbered}
--------------------------
This is very similar to the second estimator for entropy. Define $N$ and $N^{(j)}$ as in the previous section. Then $\hat M_1 = N$ is an unbiased estimator for the class number $M$. The proof is almost identical to that provided in the previous section.
We may also produce an unbiased estimator $$\hat M_2 = \frac1n \sum_{j=1}^n N^{(j)} .$$ This last quantity can also be considered as a corrector to the naive estimator $\hat M_3$, which is defined as the number of classes observed in the first $n$ samples, or alternatively, as the cardinality of the set $A$, where $A$ is the set of $1 \le i \le M$ such that there exists $1 \le k \le n$ for which $w_k \in C_i$.
After picking $n$ samples, then continue picking samples until every class observed in the first $n$ samples is observed at least once more. For each $i \in A$, let $F_i$ denote the smallest $k \ge 1$ such that $w_k \in C_i$, and let $L_i$ denote the smallest $k \ge 1$ such that $w_{k+n} \in C_i$. Then $$\hat M_2 = \hat M_3 + \frac1n \sum_{i\in A} (L_i - F_i) .$$
If one doesn’t wish to record the order in which the samples are obtained, one can simply compute the expected value of this quantity over all possible rearrangements of obtaining this data, and derive the unbiased estimator $$\hat M_4 = \hat M_3 + \frac1n \sum_{i\in A}
\left(\frac{m+1}{s_i+1}-\frac{n+1}{r_i+1} \right) ,$$ where $r_i$ is the number of times the $i$th class appears in the first sample of size $n$, $m$ is the size of the subsequent sample, and $s_i$ is the number of times the $i$th class appears in the subsequent sample.
[99]{} A. Chao, Species richness estimation. In N. Balakrishnan, C. B. Read, and B. Vidakovic, eds., [*Encyclopedia of Statistical Sciences*]{} pp 7909-7916 (New York, Wiley, 2005). P. Grassberger, Phys. Lett. [**A 128**]{}, 369 (1988). P. Grassberger, www.arxiv.org, physics/0307138 (2003). B. Harris, The statistical estimation of entropy in the non-parametric case [*Topics in Information Theory*]{} ed I Csiszar (Amsterdam: North-Holland) pp 323-55 (1975). H. Herzel, Sys. Anal. Mod. Sim. [**5**]{}, 435 (1988). G. Miller, Note on the bias of information estimates. In H. Quastler, ed., [*Information theory in psychology II-B*]{}, pp 95-100 (Free Press, Glencoe, IL 1955). T. Schürmann, J. Phys. A: Math. Gen. [**37**]{}, L295-L301 (2004). C. E. Shannon and W. Weaver, [*The Mathematical Theory of Communication*]{}, (University of Illinois Press, Urbana, IL 1949).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Collective behaviour has been observed in hadronic measurements of high multiplicity proton+lead collisions at the Large Hadron Collider (LHC), as well as in (proton, deuteron, helium-3)+gold collisions at the Relativistic Heavy-Ion Collider (RHIC). To better understand the evolution dynamics and the properties of the matter created in these small systems, a systematic study of the soft hadronic observables together with electromagnetic radiation from these collisions is performed, using a hydrodynamic framework. Quantitative agreement is found between theoretical calculations and existing experimental hadronic observables. The validity of the fluid dynamical description is estimated by calculating Knudsen and inverse Reynolds numbers. Sizeable thermal yields are predicted for low $p_T$ photons. Further predictions of higher order charged hadron anisotropic flow coefficients and of thermal photon enhancement are proposed.'
author:
- Chun Shen
- 'Jean-François Paquet'
- 'Gabriel S. Denicol'
- Sangyong Jeon
- Charles Gale
bibliography:
- 'references.bib'
title: Collectivity and electromagnetic radiation in small systems
---
Introduction
============
High energy nucleus-nucleus collision experiments conducted at the Relativistic Heavy-ion Collider (RHIC) and the Large Hadron Collider (LHC) probe QCD (Quantum Chromodynamics) under extreme conditions and create a novel state of matter: the quark-gluon plasma (QGP). This QGP has been found to be strongly-coupled and to exhibit striking collective behaviour. Relativistic hydrodynamics has been a successful effective theory which provides a quantitative description of this collectivity. In fact, a program of quantitative comparison between theory and experiment now offers the genuine prospect of being able to even extract transport properties of the QGP, as well as to set strong constraints on properties of the initial state [@Gale:2013da; @*Heinz:2013th].
More specifically, some recent measurements have shown evidence of collectivity in the high multiplicity events of small collision systems such as p+Pb, (p, d, $^3$He)+Au collisions, at RHIC and LHC energies [@Chatrchyan:2013eya; @Chatrchyan:2013nka; @Aad:2013fja; @Adare:2014keg; @Adare:2015ctn; @Loizides:2016tew]. These measurements suggest the creation of hot and strongly-coupled QGP droplets within a reaction zone of merely a few fm in size, suggesting that the mechanism responsible for the rapid approach to local equilibrium of plasma produced in heavy ion collisions may also operate in such collisions. Because the hydrodynamic description relies on well-separated distance/time scales between microscopic and macroscopic physics, the dynamics of these small collision systems appears to challenge the very validity of the fluid dynamical approach. The origin of these collective features is therefore actively investigated, and one would like to elucidate whether they are inherited from properties of the initial state [@Dumitru:2010iy; @Dusling:2012cg; @Dusling:2013qoz; @Dumitru:2014yza], or appear during the collective expansion [@Nagle:2013lja; @Bozek:2011if; @Bozek:2012gr; @Schenke:2014zha; @Werner:2013ipa; @Kozlov:2014fqa].
In order to unveil the dynamics in these small systems, one must investigate multiple aspects of experimental observables within a consistent framework. A complete slate of measurements could include hadronic anisotropic flow [@Bozek:2013uha; @Kozlov:2014fqa], the mass ordering of identified particle $v_2$ [@Adare:2014keg; @ABELEV:2013wsa], particle interferometry [@Bozek:2013df; @Shapoval:2013jca], as well as penetrating probes such as QCD jets [@Shen:2016egw] and direct real and virtual photons [@Shen:2015qba]. This last observable, radiated throughout the dynamical evolution of the hot expanding medium and suffering from negligible final state interactions, is a particularly valuable probe since the local properties of the fluid at the photon’s production point can be directly carried to the detectors. Current experiments are able to isolate low-energy direct photons $p_T \lesssim 1$ GeV. Those photons are thus penetrating [*and*]{} soft: they enjoy a unique status among all the observables measured in hadronic collisions.
In this paper, we build on previous work [@Shen:2015qba] by performing detailed comparisons with various hadronic measurements of high multiplicity p+Pb collisions at 5.02 TeV. A microscopic transport stage is employed to describe the out-of-equilibrium dynamics of the collision systems in the dilute hadronic phase. Its effect on hadronic observables in small systems is quantified. Systematic studies of hadronic and direct photon observables in (p, d, $^3$He)+Au collisions at the top RHIC energy are also presented within the same theoretical framework.
The rest of the paper is organized as follows. A description of the components of the hydrodynamical modelling, together with the choice of parameters, appear in Sec. II. A variety of collective signals involving soft hadronic observables are compared with existing experimental measurements in Sec. III. In Sec. IV, direct photon production from the small collision systems is studied in detail. The electromagnetic tomography and the viscous corrections to photon observables are discussed. Final remarks and conclusions appear in Sec. V.
Hydrodynamic modelling
======================
In this work, all the collision systems are numerically simulated using an hydrodynamics + hadronic cascade framework [@Shen:2014vra]. Fluctuating initial conditions in the transverse plane are generated using the Monte-Carlo-Glauber (MC-Glauber) model. The nucleon spatial configurations inside the heavy nuclei are sampled considering realistic repulsive 2-body nucleon-nucleon correlations [@Alvioli:2009ab]. For the light nuclei, the spatial topography of the deuteron’s two-nucleon system is obtained from sampling the Hulthen wavefunction [@Adare:2013nff], and fluctuating $^3$He configurations come from results of Green’s function Monte Carlo calculations using the AV18+UIX model interaction [@Carlson:1997qn]. Multiplicity fluctuations in every binary collision are randomly-sampled from a $\Gamma$-distribution at the wounded nucleon positions. The shape and scale parameters in the $\Gamma$-distribution are chosen to reproduce the multiplicity distribution measured in pp collisions at the same collision energy [@Shen:2014vra]. The inclusion of such multiplicity fluctuation was shown to be essential to reproduce the measured high multiplicity tail of the charged hadron yield distributions in small collision systems [@Alver:2008aq; @Bozek:2013uha; @Kozlov:2014fqa; @Shen:2014vra]. In order to simulate efficiently high multiplicity events, centrality selection is determined by sorting minimum bias events according to their initial total entropy at mid-rapidity, $dS/d\eta_s\vert_{\eta_s = 0}$, where $\eta_s = \frac{1}{2} \log \frac{t+z}{t-z}$ is the space-time rapidity. Triggering on initial total entropy of the system was shown to be a good approximation to determining centrality according to final charged hadron multiplicity, as done in experiments [@Shen:2015qta].
Starting from $\tau_0 = 0.6$ fm/$c$, an individual fluctuating entropy density profile is then evolved using a (2+1)D viscous hydrodynamics, [VISH2+1]{}, with a lattice-QCD based equation of state (EoS), s95p-v1.2 [@Huovinen:2009yb]. The effect of the longitudinal dynamics of the system is also investigated using (3+1)D viscous hydrodynamics [@Schenke:2010nt]. Second order non-linear terms are included in the evolution of the shear stress tensor [@Denicol:2012cn], $$\begin{aligned}
\tau_\pi \dot{\pi}^{\langle \mu\nu \rangle} + \pi^{\mu\nu} = && 2\eta \sigma^{\mu\nu} - \delta_{\pi\pi} \pi^{\mu\nu} \theta + \phi_7 \pi_\alpha^{\langle \mu} \pi^{\nu \rangle \alpha} \notag \\
&& - \tau_{\pi \pi} \pi_\alpha^{\langle \mu} \sigma^{\nu \rangle \alpha}.\end{aligned}$$ The transport coefficients, $\tau_\pi$, $\delta_{\pi\pi}$, $\phi_7$, and $\tau_{\pi\pi}$ are fixed using formulae derived from the Boltzmann equation near the conformal limit [@Denicol:2014vaa]. An effective specific shear viscosity $\eta/s = 0.08$ is chosen for the collisions at the top RHIC energy and a slightly larger value $\eta/s = 0.10$ is used for p+Pb collisions at 5.02 TeV. With these choices of $\eta/s$, the simulated results can reproduce the measured charged hadron anisotropic flow coefficients, $v_{2,3}$, in central collisions fairly well, as will be shown in Figs. \[fig1\] and \[fig2\]. The hydrodynamic description is switched to a microscopic hadronic cascade, UrQMD v3.4 [@Bass:1998ca; @Bleicher:1999xi], at $T_\mathrm{sw} = 155$MeV as the collision system becomes more dilute and out-of-equilibrium.
The equation of state (EoS) s95p-v1.2, fitted to lattice calculations at high temperature is matched in the confinement region to a hadron resonance gas which comprises the same hadronic content implemented in UrQMD v3.4 [^1]. In this way, the total energy of system is ensured to remain the same during the conversion from fluid cells into hadrons. The value of $T_\mathrm{sw}$ is fixed to reproduce the correct $p$ to $\pi$ ratio measured in $0-20$% d+Au collisions [@Adare:2013esx].
At the end of hadronic scatterings and resonance decays, the particle spectra and flow observables of stable particles are analyzed.
The anisotropic flow coefficients of particle of interests are evaluated using the scalar-product method [@Luzum:2012da], $$v_n\{\mathrm{SP}\} = \frac{\langle v_n(p_T) v^\mathrm{ref}_n \cos[n(\Psi_n(p_T) - \Psi_n^\mathrm{ref})]\rangle}{\sqrt{\langle (v^\mathrm{ref}_n)^2 \rangle }}.
\label{eq1}$$ For hadronic flow, we choose $v_n$ of charged hadrons integrated from $p_T = 0.3$ to 3 GeV as the reference flow vector for p+Pb collisions at 5.02 TeV to compare with the CMS data [@Chatrchyan:2013nka]. The $v^\mathrm{ref}_n$ is set to $v^\mathrm{ch}_n$ integrated from $p_T = 0.2$ to 2 GeV for (p, d, $^3$He)+Au collisions at the top RHIC energy, in line with PHENIX measurements.
To accumulate enough statistics for the hadronic and direct photon observables, we evolve 300 fluctuating events through hydrodynamics for every centrality class presented in this paper. Then 2000 hadronic cascade runs are simulated for each hydrodynamic event. Finally, these 2000 “oversampled” events from the same hydrodynamic event are combined in the hadronic flow analysis.
Collectivity in small systems
=============================
In this section, the collective behavior of hadronic observables is explored, for small collision systems at RHIC and LHC energies.
Prelude: The effect of broken longitudinal boost-invariance
-----------------------------------------------------------
---------------------------------------------------------------- ---------------------------------------------------------------
{width="0.45\linewidth"} {width="0.45\linewidth"}
---------------------------------------------------------------- ---------------------------------------------------------------
---------------------------------------------------------------------------------------------------------------- --
{width="0.45\linewidth"} {width="0.45\linewidth"}
---------------------------------------------------------------------------------------------------------------- --
Light-heavy nuclei collisions are asymmetric: boost-invariance in the longitudinal direction is explicitly broken. This study starts by quantifying the effects of longitudinal boost-non-invariance in small systems on hadronic and photon flow observables, at mid-rapidity.
To model these collisions in three dimensions, we extend the MC-Glauber initial energy density profiles in the longitudinal direction with the following envelope functions [@Kozlov:2014fqa]: $$\begin{aligned}
e({\bf x_\perp}, \eta) =&& f_L(\eta)\left[\sum_{i=1}^{N_\mathrm{part}^\mathrm{left}} \exp\left(-\frac{({\bf x_\perp - x}_i)^2}{2\sigma^2} \right) \right] \notag \\
&+& f_R(\eta) \left[\sum_{i=1}^{N_\mathrm{part}^\mathrm{right}} \exp\left(-\frac{({\bf x_\perp - x}_i)^2}{2\sigma^2} \right) \right],\end{aligned}$$ where the envelope function $f_{L,R}(\eta)$ is given by $$\begin{aligned}
&&f_{L,R}(\eta) = \left(1 \pm \frac{\eta}{\eta_\mathrm{max}} \right) \notag \\
&&\quad\times \left[\theta(\vert\eta\vert - \eta_0)\exp\left(-\frac{(\eta - \eta_0)^2}{2\sigma_\eta^2} \right) + \theta(\eta_0 - \vert \eta \vert) \right]\end{aligned}$$ Here, $\eta_{\mathrm{max}}=y_\mathrm{beam}$, the beam rapidity. The parameters, $\eta_0$ and $\sigma_\eta$, are fixed such that the measured pseudo-rapidity dependence of charged hadron multiplicity is reproduced.
The full (3+1)D hydrodynamic equations are solved with [MUSIC]{} [@Schenke:2010nt]. The transport parameters in the hydrodynamic simulation and transition to the hadronic cascade are the same as those in the (2+1)D simulations.
-------------------------------------------------------------- -------------------------------------------------------------
{width="0.45\linewidth"} {width="0.45\linewidth"}
-------------------------------------------------------------- -------------------------------------------------------------
-------------------------------------------------------------------------- ---------------------------------------------------------------------
{width="0.45\linewidth"} {width="0.45\linewidth"}
-------------------------------------------------------------------------- ---------------------------------------------------------------------
Fig. \[figdNdeta\] shows the pseudo-rapidity dependence of charged hadrons multiplicity and their anisotropic flow coefficients in central p+Pb collisions at 5.02 TeV and central d+Au collisions at 200 GeV. The $\eta$-dependence in charged hadron yields is quite strong in these highly asymmetric collision systems. The results from boost-invariant simulations are indicated as the green dashed lines, within a pseudo-rapidity $\vert \eta \vert < 1$ interval. Although $dN^\mathrm{ch}/d\eta$ is a symmetric function in $\eta$ for the (2+1)D simulations, the integrated $dN^\mathrm{ch}/d\eta \vert_{\vert \eta \vert< 0.5}$ still agrees well with the full (3+1)D simulations.
Fig. \[figvnEta\] compares the pseudo-rapidity dependence of charged hadron $v_n$ of (2+1)D and (3+1)D simulations, in both central p+Pb collisions and central d+Au collisions. The momentum anisotropy is relatively flat in the mid-rapidity region, $\vert \eta \vert < 1$ and the (2+1)D results (indicated by the circle and triangle markers) agree quite well with the full (3+1)D simulations.
Fig. \[fig0.2\] features the $p_T$-differential particle spectra and charged hadron $v_n(p_T)$, as obtained with (2+1)D and (3+1)D simulations at mid-rapidity. The hadronic observables computed with (2+1)D simulations serve as a good approximation to the (3+1)D results at mid-rapidity, even in these highly asymmetric systems.
We postpone a discussion of photon observables until Section \[photon\_section\], but it suffices to say here that thermal photons are sensitive to the entire evolution history of the medium than hadronic observables. In Figs. \[fig0.3\], the thermal photon spectra and the anisotropic flow coefficients are shown, for (2+1)D and (3+1)D simulations, for photons at mid-rapidity. Calculations show that boost-invariance is still a very good approximation for thermal photon production at mid-rapidity in d+Au collisions at RHIC. We verified that similar results were found in p+Pb collisions at LHC energy.
Hadronic flow observables
-------------------------
As shown in the previous subsection, boost-invariance is still a good approximation for mid-rapidity observables in asymmetric systems. We shall thus henceforth conduct our systematic study using boost-invariant conditions. All hadronic flow observables are computed for pseudo-rapidity $\vert \eta \vert < 0.5$.
------------------------------------------------------------- --------------------------------------------------------
{width="0.4\linewidth"} {width="0.4\linewidth"}
------------------------------------------------------------- --------------------------------------------------------
Collision system $\frac{dN^\mathrm{ch}}{d\eta} \big\vert_{\vert \eta \vert < 0.5}$ $\langle p_T \rangle({\pi^+})$ (GeV) $\langle p_T \rangle({K^+})$ (GeV) $\langle p_T \rangle({p})$ (GeV) $v_2^\mathrm{ch}\{2\}$ $v_3^\mathrm{ch}\{2\}$
-------------------------- ------------------------------------------------------------------- -------------------------------------- ------------------------------------ ---------------------------------- ------------------------ ------------------------
0-5% p+Au @ 200 GeV 11.8(1) 0.52(1) 0.72(2) 0.98(3) 0.037(1) 0.0091(3)
0-5% d+Au @ 200 GeV 17.7(1) 0.50(1) 0.70(1) 0.95(2) 0.054(1) 0.0114(4)
0-5% $^3$He+Au @ 200 GeV 22.9(1) 0.49(1) 0.69(1) 0.93(2) 0.059(1) 0.0116(4)
We start our comparisons with integrated hadronic observables. Because it is difficult to estimate the number of particle tracks detected in the CMS experiments with our theoretical model, we use the conversion table (Table 1) in Ref. [@Chatrchyan:2013nka] to map our calculations in different centrality bins to the experimental measurements. As shown in Fig. \[fig1\], quantitative agreement is achieved between our hydrodynamic simulations, the measured identified particle averaged transverse momentum ($\langle p_T \rangle$) and charged hadron anisotropic flow coefficients, $v_{2,3}\{2\}$ in top 20% p+Pb collisions at 5.02 TeV. The proton $\langle p_T \rangle$ appears slightly underestimated but nevertheless falls within experimental uncertainties. The centrality dependence of these global observables is well reproduced by our approach.
The effect of the hadronic afterburner is also highlighted in the same figures, with the solid lines including the effect of hadronic rescattering with UrQMD, and the dashed line including only hadronic decays, but no rescattering. The proton $\langle p_T \rangle$ is found to have a small but visible increase with the additional hadronic scattering, but the actual increase is much smaller than in the case of Au+Au and Pb+Pb collisions [@Ryu:2015vwa]. The transport phase helps the small system to further develop a few percent of anisotropic flow. This effect is more pronounced in peripheral bins, where the fireball lifetime is too short to convert all the system’s spatial eccentricity into momentum anisotropy during the hydrodynamic stage.
Predictions of the integrated hadronic observables in $0-5$% (p, d, $^3$He)+Au collisions at 200 GeV are summarized in Table \[table1\] for future comparison. With respect to p+Pb collisions at 5.02 TeV, the identified particle mean $\langle p_T \rangle$ are about 20% smaller at the top RHIC energy. This is because the system lifetime is about 20% shorter compared to the collisions at LHC energy. This limits the development of radial flow during the hydrodynamic evolution. Moreover, the pressure gradients are also smaller at 200 GeV, which translates into a smaller expansion rate. These two factors yield anisotropic flow coefficients about 50% smaller in p+Au collisions at RHIC than in p+Pb collisions at 5.02 TeV. The elliptic flow coefficients in d+Au and $^3$He+Au collisions, on the other hand, are comparable with the $v_2\{2\}$ in p+Pb collisions because of larger initial eccentricities in these systems.
{width="0.95\linewidth"}
![Charged hadron anisotropic flow $v_{2,3}\{\mathrm{SP}\}$ compared with the CMS [@Chatrchyan:2013nka] and the ATLAS measurements [@Aad:2013fja] in 0-2% p+Pb collisions at 5.02 TeV. The shaded bands represent statistical uncertainty.[]{data-label="fig3"}](figs/pPb_vn_vs_CMS){width="0.9\linewidth"}
Now, we take a closer look at $p_T$-differential observables. The charged hadron anisotropic flow coefficients, $v_{2,3}\{\mathrm{SP}\}(p_T)$, are compared with experimental measurements in Fig. \[fig2\] and Fig. \[fig3\] for small collision systems at RHIC and LHC energies, respectively. At the top RHIC energy, our hybrid approach with $\eta/s = 0.08$ for $T > 155$ MeV successfully provides a consistent description of the PHENIX anisotropic flow measurements in $0-5$% p+Au, d+Au, and $^3$He+Au collisions. Hadronic rescattering from the transport phase is found to increases the high $p_T$ charged hadron $v_n\{\mathrm{SP}\}$ and improves the agreement with experimental data. Prediction of $p_T$-differential triangular flow $v_3\{\mathrm{SP}\}(p_T)$ in p+Au and d+Au collisions are shown for future comparison. In Fig. \[fig3\], a same level of agreement for charged hadron $v_{2,3}\{\mathrm{SP}\}(p_T)$ is achieved in the top 2% p+Pb collisions at 5.02 TeV with an effective $\eta/s = 0.10$ for $T > 155$ MeV. The effect of the hadronic cascade lessens as the collision energy is increased. This is because the larger pressure gradients at higher collision energy drive the system to develop hydrodynamic radial flow faster. Most of the spatial eccentricity has already been converted to momentum anisotropy before switching to hadronic transport.
![Identified particle spectra compared with experimental measurements in (a) 0-20% d+Au collisions at 200 GeV [@Adare:2013esx] and (b) minimum bias p+Pb collisions at 5.02 TeV [@Chatrchyan:2013nka]. The shaded bands represent statistical uncertainty.[]{data-label="fig4"}](figs/dAu_C0-20_SP_vs_PHENIX "fig:"){width="0.9\linewidth"}\
![Identified particle spectra compared with experimental measurements in (a) 0-20% d+Au collisions at 200 GeV [@Adare:2013esx] and (b) minimum bias p+Pb collisions at 5.02 TeV [@Chatrchyan:2013nka]. The shaded bands represent statistical uncertainty.[]{data-label="fig4"}](figs/pPb5020_MB_SP_vs_CMS "fig:"){width="0.9\linewidth"}
In Fig. \[fig4\], identified particle spectra are compared with experimental measurements for 0-20% d+Au collisions at 200 GeV and minimum bias p+Pb collisions at 5.02 TeV. In 0-20% d+Au collisions, our hybrid calculations provide a good description of the soft hadron spectra up to 1.5 GeV. Agreement with measurements at higher $p_T$ may require the contribution from recombination with jet shower partons, which is not considered here.
We note that the effect of hadronic rescattering, also shown in Fig. \[fig4\], is very small for pions and kaons. The proton spectrum at $p_T < 0.5$ GeV is reduced due to baryon anti-baryon annihilation in the transport phase. The fact that the high $p_T$ region of proton spectra remains the same suggests the hadronic rescatterings play a minor role in the dilute gas phase. Hence, the observed increase of the proton mean $\langle p_T \rangle$ observed earlier (Fig. \[fig1\]) owes mainly to baryon anti-baryon annihilation.
![Identified pion and proton elliptic flow coefficients compared with the PHENIX [@Adare:2014keg] and ALICE measurements [@ABELEV:2013wsa] in 0-5% d+Au collisions at 200 GeV (a) and 0-20% p+Pb collisions at 5.02 TeV (b), respectively. Predictions of pion and proton $v_2\{\mathrm{SP}\}$ in top 5% p+Au and $^3$He+Au collisions are shown in panels (c) and (d). The legends apply to all four panels. The shaded bands represent statistical uncertainty.[]{data-label="fig5"}](figs/pid_v2_exp){width="1.0\linewidth"}
The mass-ordering of hadronic flow coefficients has long been considered a hallmark of fluid-dynamical behavior [@Huovinen:2001cy]. Alternative interpretations have however recently appeared for asymmetric [@Zhou:2015iba; @Schenke:2016lrs] and even symmetric [@Li:2016ubw] heavy-ion collisions. In the current work, mass ordering in the identified particle elliptic flow is investigated in Fig. \[fig5\]. The hydrodynamic model quantitatively produced the mass splitting between pion and proton $v_2\{\mathrm{SP}\}(p_T)$ measured in 0-5% d+Au collisions at 200 GeV and 0-20% p+Pb collisions at 5.02 TeV. Within hydrodynamic framework, the larger difference between pion and proton elliptic flow in p+Pb collisions at 5.02 TeV can be understood as the consequence of a stronger radial flow blue shifts the proton $v_2$ to high $p_T$ regions at higher collision energy. The effect of hadronic rescattering is found to be small, implying that that most of the mass splitting is developed in the hydrodynamic phase. This suggests that a strongly-coupled QGP core in the small collision systems can be at the origin of the mass ordering in measured identified particle $v_2$. Predictions of pion and proton elliptic flow coefficients in p+Au and $^3$He+Au collisions are shown in Figs. \[fig5\]c and \[fig5\]d for future comparison. The amount of mass splitting is found to be similar for the three systems studied, at the top RHIC energy.
Photon radiation {#photon_section}
================
A hot and rapidly expanding QGP droplet radiates thermal photons. As shown in the previous section, the hydrodynamic medium has been well calibrated to reproduce various aspects of hadronic observables in the small collision systems. In this section, the significance of thermal photon enhancement in the final measurable direct photon signal is addressed. More specifically, only thermal photon radiation from fluid cells whose temperature is higher than the switching temperature, $T_\mathrm{sw} = 155$MeV is considered. Since the number of hadrons and average number of collisions per hadrons is small in p+A collisions, photon emission in the late stage of the medium is not expected to be large. An estimation of the contributions from temperature below $T_\mathrm{sw}$ will be presented in the next section.
The leading-order QGP photon emission rate [@Arnold:2001ms] is used for temperature larger than 180 MeV, and hadronic photon production rates below. In the hot hadronic phase, photon production through meson-meson scattering [@Turbide:2003si], from (the imaginary part of) many-body $\rho$-spectral function, from $\pi-\pi$ bremsstrahlung [@Rapp:1999ej; @Rapp:1999qu; @Liu:2007zzw; @Heffernan:2014mla], and from $\pi-\rho-\omega$ reaction channels [@Holt:2015cda] are considered. Shear viscous corrections to the 2 to 2 scattering processes in the QGP phase [@Shen:2014nfa] and to the meson-meson reactions in the hadronic phase [@Dion:2011pp; @Shen:2014thesis] are included. Using the fact that the shear stress tensor - $\pi^{\mu\nu}$ - is symmetric, traceless, and orthogonal to the flow velocity, the photon emission rates can be written as [@Shen:2014nfa], $$\begin{aligned}
&& \!\!\!\!\!\!\!\! E_q\frac{d \Gamma}{d^3 k}(E_q, T, \pi^{\mu\nu}) \notag \\
&& = \Gamma_0 (E_q, T) + \delta \Gamma (E_q, T, \pi^{\mu\nu}) \notag \\
&& = {\Gamma}_0(E_q, T) + \frac{\pi^{\mu\nu} \hat{q}_\mu \hat{q}_\nu}{2(e+P)} \chi\left(\frac{E_ q}{T}\right) {\Gamma}_\pi (E_q, T),
\label{eq3}\end{aligned}$$ where ${\Gamma}_0(E_q, T)$ and $\delta \Gamma (E_q, T)$ denote the equilibrium rate and first order shear viscous correction, respectively. Finally, decay photons from short-lived resonances that can not be subtracted in the experimental cocktail background are included when computing the direct photon observables [@Rapp:1999qu; @vanHees:2014ida; @Shen:2014thesis; @Paquet:2015lta]. A detailed list and a discussion of the non-cocktail decay channels that produce photons can be found in Chapter 21 of Ref. [@Shen:2014thesis][^2].
Prompt photons {#sec:prompt}
--------------
Collision system Centrality $N_\mathrm{coll}$ $N_\mathrm{part}$
--------------------- ------------ ------------------- -------------------
p+Pb @ 5.02 TeV 0-1% 15.92(4) 16.92(4)
0-5% 14.47(2) 15.47(2)
0-20% 12.51(1) 13.51(1)
0-100% 6.50(1) 7.50(1)
p+Au @ 200 GeV 0-5% 10.28(1) 11.28(1)
0-20% 8.62(1) 9.62(1)
0-100% 4.66(1) 5.66(1)
d+Au @ 200 GeV 0-5% 18.48(2) 18.19(2)
0-20% 15.75(2) 15.40(1)
0-100% 7.90(2) 8.31(1)
$^3$He+Au @ 200 GeV 0-5% 26.48(2) 25.35(2)
0-20% 22.67(1) 21.59(1)
0-100% 10.59(1) 10.88(1)
: The averaged number of binary collisions $N_\mathrm{coll}$ and participant nucleons $N_\mathrm{part}$ in p+Pb collisions at 5.02 TeV and (p, d, $^3$He)+Au collisions at 200 GeV. Statistical uncertainties of the last digits are in parentheses.[]{data-label="table2"}
The photons produced by the very first nucleon-nucleon collisions are the prompt photons. These are evaluated with perturbative QCD at next-to-leading order in $\alpha_s$ [@Aurenche:1987fs; @Aversa:1988vb; @incnlo], as in previous work [@Paquet:2015lta; @Paquet:2015Thesis]. The isospin effect is included to account for the different proton-to-neutron ratio of each colliding ion. Cold nuclear effects are taken into account with the nCTEQ15 nuclear parton distribution functions [@Kovarik:2015cma]. The perturbative calculation of photon production is scaled up from the nucleon-nucleon result, by the number of binary collisions $N_\mathrm{coll}$ which is summarized in Table \[table2\] for the different systems investigated in this work.
We note that the nCTEQ15 nuclear parton distribution functions were constrained using nuclear deep-inelastic scattering ($e A \to e + X$) and nuclear Drell-Yan ($p A \to l^+ l^- + X$). The default parametrization of nCTEQ15 was also constrained with pion production in d+Au collisions, although a separate parametrization that did not use these pion measurements was also made available. It is this second version of nCTEQ15 — referred to by the authors as nCTEQ15-np — that is used in this work. The rationale for this choice is that constraining nuclear parton distribution functions with hadronic measurements from small systems makes the explicit assumption that no QGP is formed, i.e. that no significant energy loss or thermal hadron production occurs in such collisions. Since the opposite assumption is made in the present work, it would not be consistent to use nuclear distribution functions partly constrained by hadronic measurements. This is also the reason why the widely used EPS09 nuclear parton distribution functions [@Eskola:2009uj] are not used in this work[^3].
--------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------
{width="0.4\linewidth"} {width="0.4\linewidth"}
--------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------
The isospin effect is only significant at high $p_T$, which is shown in Figure \[prompt\_photon\_RAA\]a by plotting $R_{\textrm{pAu}}$, $R_{\textrm{dAu}}$, $R_{\textrm{He$^3$Au}}$ and $R_{\textrm{AuAu}}$ at RHIC ($\sqrt{s_{NN}}=200$ GeV). The introduction of cold nuclear effects with nCTEQ15 differentiates $R_{\textrm{AuAu}}$ from that of the small systems (p+Au, d+Au, He$^3$+Au), as seen in Figure \[prompt\_photon\_RAA\]b.
As investigated in Refs. [@Arleo:2011gc] and [@Klasen:2013mga], uncertainties on prompt photon production in nuclear collisions originate from the factorization/renormalization/fragmentation scale dependence, photon fragmentation functions and nuclear parton distribution functions. The questionable reliability of perturbative QCD at low $p_T$ further increases the uncertainty in this region of momentum. To constrain the scale dependence of the calculation, the factorization, renormalization and fragmentation scales are taken to be proportional to the photon transverse momentum, and the proportionality constant is fixed [@Paquet:2015lta; @Paquet:2015Thesis] using proton-proton measurements.
The photon fragmentation function BFG-II [@Bourhis:1997yu] is used. It appears to be slightly better than other fragmentation functions at describing low momentum photon measurements in proton-proton collisions [@Klasen:2014xfa].
The nuclear distribution functions nCTEQ15 themselves have uncertainties [@Kovarik:2015cma] reflecting the limited constraining power of available nuclear data. They can be used to provide an uncertainty band on prompt photon predictions. This uncertainty was studied in Ref. [@Arleo:2011gc] for different nuclear parton distribution functions, EPS09. Given that the nuclear distribution function uncertainties of EPS09 and nCTEQ15 are of the same order, this previous work can be used as a guide for the size of the uncertainties due to the nuclear parton distribution functions. Thus, uncertainties from nCTEQ15 of order 10% are expected for the nuclear modification factor of prompt photons.
Combining all the above uncertainties in prompt photons, along with possible final state effects on prompt photon production (e.g. the effect of parton energy loss on fragmentation photon production), it is clear that work remains to be done to determine the precise contribution of this source of photons in small collision systems. Nevertheless, we believe that the calculation presented above provides a sufficiently good determination of prompt photons to establish if thermal photons can be observed above the prompt photon background.
------------------------------------------------------------ ------------------------------------------------------------------
{width="0.45\linewidth"} {width="0.45\linewidth"}
------------------------------------------------------------ ------------------------------------------------------------------
Direct photon spectra and $v_n$
-------------------------------
Fig. \[fig6\] presents the direct photon spectra and elliptic flow coefficients in 0-5% most central p+Pb collisions at 5.02 TeV. Contributions from individual production channels are shown. We find that the thermal sources are significant in the total direct photon signal for $p_T < 2$ GeV. Thermal radiation represents about 1.6 times that of the prompt contribution. Decay photons coming from the short-lived resonances contribute only about 10% to the direct photon yield. In the thermal photon signal, almost 80% of the photons come from $T > 180$ MeV region: indeed radiation from the QCD phase diagram at temperatures above that of the crossover is observed. Contribution of thermal photons from spacetime regions with $T<180$ MeV is as small as decay photons coming from the short-lived resonances. The relative importance of photon originating from the low and high temperature region of the plasma is an important feature of the calculation that we discuss in more details in the next section.
In Fig. \[fig6\]b, the net $p_T$-differential elliptic flow coefficient of direct photons is shown together with the contributions from individual channels. Prompt photons are assumed to carry zero anisotropy. The direct photon elliptic flow in $0-5$% p+Pb collisions reaches its maximum $\sim$ 0.05 at $p_T \sim 1.7$ GeV. The strength of this signal is comparable to that of the direct photon $v_2$ measured in $0-40$% Pb+Pb collisions [@Lohner:2012ct]. The rise and fall in the direct photon elliptic flow reflects the competition between thermal and prompt sources at different $p_T$ [@Shen:2013cca]. Although the elliptic flow of hadronic photons and decay photons from short-lived resonances are large, their contribution to the total direct photon elliptic flow is limited, owing to their small yield as shown in Fig. \[fig6\]a.
![(Color online) [*Panels (a-d):*]{} Thermal photon enhancement in 0-5%, 0-20%, and minimum bias p+Pb collisions at 5.02 TeV and (p, d, $^3$He collisions at 200 GeV. In Panel (c), the PHENIX measurement is from Ref. [@Adare:2012vn]. []{data-label="fig7"}](figs/photon_RpPb){width="0.9\linewidth"}
We summarize the thermal photon enhancement in all four small collision systems as the nuclear modification factor $R_{pA}^\gamma$. From panels (a) to (d), one sees that collision systems at RHIC and the LHC exhibit qualitatively similar behaviour. We find sizeable enhancements in the direct photon signals from thermal radiation with respect to the prompt production for $p_T < 3$ GeV. The prompt photon $R_{pA}^\gamma$ is below 1 in the low $p_T$ region because of nuclear shadowing effects (see Section \[sec:prompt\]). The thermal enhancement over this baseline is the largest in p+Pb collisions at 5.02 TeV. At the top RHIC energy, we find that direct photons $R_\mathrm{pAu}^\gamma > R_\mathrm{dAu}^\gamma > R_{^{3}\mathrm{HeAu}}^\gamma$. This is because the number of binary collisions from p+Au to $^{3}$He+Au collisions (see Table \[table2\]) increases more rapidly than the thermal radiation; this makes for a larger relative weight of the prompt photon production. As a function of centrality, direct photon $R_{pA}^\gamma$ is larger in more central collisions. In $0-5$% most central collisions, direct photon $R_{pA}^\gamma$ are predicted to reach up to 2 at $p_T \sim 1.2$ GeV for all four systems. Although this thermal enhancement is smaller than what it is in Au+Au or Pb+Pb collisions [@Adare:2008ab; @Adam:2015lda; @Shen:2016odt], it still can serve as a precious signature of the hot medium in small collision systems. In Fig. \[fig7\]c, our result in minimum bias d+Au collisions is compared to the available PHENIX measurement [@Adare:2012vn]. The current experimental uncertainty does not permit a distinction between the scenarios with, and without the thermal signal.
![(Color online) [*Panels (a-d):*]{} Direct photon $v_{2,3}$ in 0-5% and 0-20% centralities of the small collision systems. The shaded bands represent statistical uncertainty.[]{data-label="fig7b"}](figs/photon_vn){width="0.9\linewidth"}
The direct photon anisotropic flow coefficients $v_{2,3}\{\mathrm{SP}\}$ are shown in Figs. \[fig7b\] for all four systems investigated. Direct photons from p+Pb collisions at 5.02 TeV carry the largest anisotropic flow, owing to the combined effects of higher temperatures achieved and larger pressure gradients. The direct photon anisotropic coefficients at top RHIC energy are similar, for p+Au, d+Au, and $^3$He+Au collisions despite the initial eccentricities $\varepsilon_n$ being quite different [@Nagle:2013lja]. This similarity in the momentum anisotropy of these three small systems was observed previously in hadrons (see Fig. \[fig5\]), and the considerable damping effect of non-flowing prompt photons makes it even harder to distinguish differences in the direct photon $v_n$ of these systems.
Centrality, system size and center-of-mass energy dependence
------------------------------------------------------------
![(Color online) The centrality dependence of $p_T$-integrated thermal and prompt photon yields compared with charged pion multiplicity $N_{\pi^+}$ in p+Pb collisions at 5.02 TeV in the mid-rapidity region, $\vert y \vert < 0.5$. The meaning of the ordinate is specified in the legend.[]{data-label="fig7.1"}](figs/Ngamma_vs_Nch_withfit){width="0.95\linewidth"}
In this section, we take a closer look at the centrality, system size and center-of-mass energy dependence of photon production in relativistic nuclear collisions. All three of these parameters can be mapped reasonably well to the pion multiplicity $N_{\pi^+}$: more central collision, larger system size and higher center-of-mass energy all translate into larger pion production. The pion multiplicity is thus used as proxy for these quantities. To facilitate comparisons of different systems, the photon spectra is integrated in transverse momentum to obtain the photon multiplicity, which we evaluate in this section with the cuts $1<p_T^\gamma<3$ GeV.
As observed in Fig. \[fig7\], the thermal photon enhancement is more pronounced in central collisions than at minimum bias. This can also be seen in Fig. \[fig7.1\], which shows the centrality dependence of the thermal and prompt photon multiplicities in p+Pb collisions at 5.02 TeV: the thermal photon multiplicity increases approximately with $N_{\pi^+}^{1.45}$, while the prompt photon multiplicity goes much more slowly as $N^\gamma_\mathrm{prompt} \propto N_{\pi^+}^{0.67}$. The increase of thermal photons can be understood as a combination of changes in the systems’ space-time volume, average temperature and blueshift effect. First note that the thermal photon multiplicity shown in Fig. \[fig7.1\] is not independent of the blueshift, because of the limited integration range used in $p_T^\gamma$ ($1<p_T^\gamma<3$ GeV). This effect can be quantified by evaluating the photon multiplicity without including blueshift, shown in Fig. \[fig7.1\]. This results in a slightly smaller slope of $N_{\pi^+}^{1.40}$, implying that the contribution of the blueshift to the centrality dependence of the thermal photon multiplicity is small. The growth of the spacetime volume $V_4$ as a function of centrality is slower compared to $N^\gamma_\mathrm{th}$, with a scaling exponent 1.11 shown in Fig. \[fig7.1\]. The remaining difference between $N_{\pi^+}^{1.40}$ and $N_{\pi^+}^{1.11}$ can be attributed to the increase of systems’ average temperature, which is thus the dominant factor in the centrality dependence of thermal photons in small systems (as verified by a direct calculation). It is found that thermal and prompt photon production in small systems at RHIC show a similar centrality dependence as found above for p+Pb collisions at 5.02 TeV.
![(Color online) Multiplicity of thermal and prompt photons ($1<p_T^\gamma<3$ GeV) as a function of the multiplicity of pions for all collision systems and centralities investigated in this work, along with the results from Au+Au collisions at RHIC and Pb+Pb collisions at the LHC from Ref. [@Paquet:2015lta][]{data-label="mult_photons_pions"}](figs/Ngamma_vs_Nch_withfit_allsystems){width="0.95\linewidth"}
A slightly different picture emerges when investigating the overall centrality, system size and center-of-mass energy dependence of all systems studied in this work, along with calculation for Au+Au collisions at RHIC ($\sqrt{s_\mathrm{NN}}=200$ GeV) and Pb+Pb collisions at the LHC ($\sqrt{s_\mathrm{NN}}=2760$ GeV) [@Paquet:2015lta]. Figure \[mult\_photons\_pions\] provides an overview for all these systems. Thermal photons are found to grow as $N_{\pi^+}^{1.42}$, which is very similar to the power found above for the centrality dependence at a fixed center-of-mass energy (see Fig. \[fig7.1\]). The prompt photon multiplicity is found to go as $N_{\pi^+}^{1.12}$, which is slower than thermal photons but significantly larger than the $N^\gamma_\mathrm{prompt} \propto N_{\pi^+}^{0.67}$ observed in Fig. \[fig7.1\] for the centrality dependence in p+Pb collisions. This difference in the scaling of prompt photons is not expected: at a fixed center-of-mass energy, the centrality and system size dependence of prompt photons originates mainly from the number of binary collisions that multiply the perturbative QCD calculation of prompt photons (see Section \[sec:prompt\]). On the other hand, changes in the center-of-mass energy of the collisions have a significant effect on both the number of binary collisions and the perturbative QCD calculation. The larger exponent found in Fig. \[mult\_photons\_pions\] reflects more closely the combination of these two effects.
Note that the pion multiplicity is also a good proxy for the number of decay photons being produced in a collision, since $\pi^0\to\gamma \gamma$ is the dominant source of decay photons in heavy ion collisions. The observation in Figure \[mult\_photons\_pions\] that the thermal and prompt photon multiplicities grow faster than the pion multiplicity summarizes the fact that the direct photon signal grows faster than the decay photon background as the collision energy and system size increases, making its measurement easier. Inversely, direct photons tend to be increasingly difficult to measure in small system. This subject is addressed in the next section.
Inclusive and decay photons
---------------------------
In order to estimate the difficulty in measuring the proposed direct photon observables, we investigate the “photon decay cocktail” in the small systems, at top RHIC and LHC energies. For the cocktail content we include decays from $\pi^0$, $\eta$, $\omega$, $\eta^\prime$, $\phi$, $\Sigma_0$, and $\rho_0$: the same species included in the ALICE direct photon analysis in Pb+Pb collisions [@LohnerThesis; @Adam:2015lda]. These particles contribute over 99% of the decay photons in the calculation. The remaining less than 1% of the decay photons from other short-lived resonances are included in the direct photon signal.
------------------------------------------------------------------------------ ------------------------------------------------------------------------------
{width="0.35\linewidth"} {width="0.35\linewidth"}
------------------------------------------------------------------------------ ------------------------------------------------------------------------------
In Figs. \[fig8\], we show the direct, decay, and inclusive (direct + decay) photon spectra in $0-5$% p+Pb collisions at 5.02 TeV, and $^3$He+Au collisions at 200 GeV. Because the $\pi^0$ production in p+Pb collisions scales faster than number of binary collisions (see Figs. \[fig7.1\] and \[mult\_photons\_pions\]), the signal to background (direct/decay photons) ratio is worse than the one in Au+Au or Pb+Pb collisions. This translates in a big gap between the direct and decay photon spectra at low $p_T$ in Figs. \[fig8\](a) and (b). In panels (c) and (d), the ratio of inclusive to decay photons, $R_\gamma$, is shown as a function of $p_T$. At 5.02 TeV, the value of $R_\gamma$ is about 1.02$ - $1.03 for $p_T < 2$ GeV and increases up to $\sim 1.05$ at $p_T = 3$ GeV. $R_\gamma$ is larger at the top RHIC energy. In 0$-$5% $^3$He+Au collisions, $R_\gamma$ is about $4-5$% higher than unity for $p_T < 2$ GeV. It increases rapidly for $p_T > 2$ GeV, where the prompt photon signal becomes more and more dominant. The larger values of $R_\gamma$ in $^3$He+Au collisions compared to those in p+Pb collisions can be understood as a consequence of the faster growth of $N_\mathrm{coll}$ as a function of centrality in the less asymmetric collision systems. The prompt photon yield increases relatively faster in $^3$He+Au collisions than in p+Pb collisions. The faster increase of $R_\gamma$ as a function of $p_T$ in $^3$He+Au collisions is because a weaker hydrodynamic radial flow is generated at top RHIC energy. The hadron spectra who produce decay photons are less blueshifted at lower collision energy. The direct photon signal for $p_T > 2$ GeV shines out more easily at RHIC than at the LHC. The $R_\gamma$ values in the small collision systems are about factor of $4\sim5$ smaller compared to those measured in Au+Au and Pb+Pb collisions [@Adare:2014fwh; @Adam:2015lda]. This indicates that direct photon measurements in small collision systems are challenging, and that measurements at 200 GeV should be somewhat easier than at 5.02 TeV.
Thermal photons in pp collisions
--------------------------------
----------------------------------------------------------------------- ---------------------------------------------------------------------
{width="0.4\linewidth"} {width="0.4\linewidth"}
----------------------------------------------------------------------- ---------------------------------------------------------------------
Collective flow signatures were also found in high multiplicity pp collisions at LHC energies [@Khachatryan:2016txc; @Aad:2015gqa]. We estimate the amount of thermal photon radiation in pp collisions at 200 GeV and 5.02 TeV in Figs. \[fig8.1\]. Since the size of the medium created in the pp collisions is extremely small, systems may not achieve quasi-thermal equilibrium in every collision event. Hence, we explore the following three scenarios for thermal photon production: considering thermal radiation from only the top 5% highest multiplicity events, thermal radiation from only the top 20% events, and thermal radiation from all events. For all these three cases, the thermal photon yields are smaller compared to the prompt component in minimum bias pp collisions at both 200 and 5.02 TeV. At 200 GeV, our direct photon spectrum agrees reasonably with the PHENIX measurements [@Adare:2009qk]. The relative size of thermal radiation to the prompt production increases with the collision energy from 35% to 40% at $p_T \sim 1$ GeV.
Theoretical uncertainties {#secV}
=========================
In this section we estimate some of the theoretical uncertainties inherent in some of calculations presented in this work.
---------------------------------------------------------------- ----------------------------------------------------------------------
{width="0.4\linewidth"} {width="0.4\linewidth"}
---------------------------------------------------------------- ----------------------------------------------------------------------
Photon emission from the dilute hadronic phase
----------------------------------------------
The results presented up to this point included thermal photon radiation from fluid cells with temperatures higher than the switching temperature to the hadronic afterburner, $T_\mathrm{sw} = 155$ MeV. The system’s dynamical evolution in the dilute hadronic phase is modelled by microscopic transport simulations. However, electromagnetic radiation is not currently included in the transport phase. To estimate the additional thermal radiation from the dilute hadronic stage, it is possible to extend the hydrodynamic evolution below $T_\mathrm{sw}$ and evaluate thermal photon production from this hydrodynamic medium’s profile. The hydrodynamic evolution is stopped at $T = 105$MeV, as done in Ref. [@Paquet:2015lta].
Since inelastic scatterings that associated with species changing processes are expected to stop quickly, we use an equation of state that implements partial chemical freeze-out at 150 MeV in order to maintain correct hadronic chemical contents. Fugacity enhancements in hadronic photon emission rates are included in the calculations. The choice of 150 MeV as partial chemical freeze-out temperature, and 105 MeV as the temperature as which the additional hydrodynamic evolution is stopped are obviously not unique, but are chosen as reasonable parameters to be used to estimate late stage photon production.
In Fig. \[fig9\]a, one sees that the hadronic photon emission, defined as photon emitted at temperature below 180 MeV, roughly doubles if thermal radiation is included down to $T = 105$MeV. The final direct photon spectra increases by about 10%. The $v_2$ of photons emitted below $T=180$ MeV remains almost unchanged by the addition of lower temperature photon emission, as shown in Fig. \[fig9\]b. This indicates that the momentum anisotropy of the system is fully built up and saturated around 155 MeV. In the end, the addition of these large-$v_2$ late stage photons increase the thermal and direct photon $v_2$ by about 30% and 25%, respectively.
The bands in plotting the photon elliptic flow coefficients indicate our estimation of the theoretical uncertainties generated by the inclusion (or not) of the additional thermal contribution from the dilute hadronic phase.
We note that there are reasons to believe that production of photons in the late stage of the medium should be significantly lower in small systems than in heavy ion collisions. In the hadronic transport simulations, we find the collision rate in central p+Pb collisions is only $\sim$0.2 collisions per particle below 155 MeV, in contrast to $\gtrsim$ 0.5 collisions per particle in semi-peripheral Pb+Pb collisions. This smaller collision rate is in part a consequence of the large expansion rate found at low temperature in small systems, which is around four times larger than what it is in heavy-ion collisions in the same temperature range. Considering the smaller number of collisions per hadrons, and the overall smaller number of hadrons present in the final state of small systems, it is thus less likely that a significant number of photons are produced at late time. In this sense, we expect photon production in small systems to be more biased towards the earlier and hotter regions of the plasma.
------------------------------------------------------------------------------------ -------------------------------------------------------------------------------
{width="0.4\linewidth"} {width="0.4\linewidth"}
------------------------------------------------------------------------------------ -------------------------------------------------------------------------------
------------------------------------------------------------------ -------------------------------------------------------------
{width="0.4\linewidth"} {width="0.4\linewidth"}
------------------------------------------------------------------ -------------------------------------------------------------
Pre-equilibrium contributions
-----------------------------
Because the lifetime of the plasma produced in small systems is shorter ($\sim4$fm/$c$) than in heavy ion collisions ($\sim15-20$fm/$c$), the pre-equilibrium dynamics of the system may have a more sizeable influence on experimental observables. Non-trivial initial flow velocity profiles and early-stage electromagnetic probe production are two examples of the possible influence of pre-equilibrium dynamics on observables [@Vujanovic:2016anq]. A rigorous treatment of these effects would require a detailed model of pre-equilibrium dynamics, which still the subject of much active research and is currently an open question. Nevertheless, within the hydrodynamic framework used in this paper, it is possible to address the more modest question of the relative sensitivity of hadronic and photonic observables to the time $\tau_0$ at which thermalization is assumed to occur, which was fixed to $\tau_0 = 0.6$ fm/$c$ for the results presented up to this point.
In Figs. \[fig10a\], we investigate the effect of $\tau_0$ on direct photon observables by choosing smaller values of $\tau_0$ for starting the hydrodynamics. Calculations with different $\tau_0$ are tuned such that the final charged hadron multiplicity remains the same. We find about 15% more viscous entropy production if hydrodynamic evolutions are started at $\tau_0 = 0.2$ fm/$c$ compared to ones started at $\tau_0 = 0.6$ fm/$c$.
The thermal photon spectrum is flatter with a smaller $\tau_0$ in Fig. \[fig10a\]. This is a consequence of the following two main factors. Firstly, more high $p_T$ photons are emitted from high temperature hot spots at the early time. At $\tau_0 = 0.2$ fm/$c$, the peak temperature of the medium can reach up to 650 MeV. Secondly, with a smaller $\tau_0$ the system’s pressure gradients accelerate fluid cells earlier and develop more radial flow. It gives a stronger blueshift to the thermal photons emitted at the late stage. Both effects make the emitted thermal photon spectrum harder. Meanwhile, the large expansion rate shortens the fireball lifetime by $\sim 10\%$. The reduction of the space-time volume results a smaller thermal photon production with $\tau_0 = 0.2$ fm/$c$ compared to the collision events who started its transverse expansion at $\tau_0 = 0.6$ fm/$c$. Finally, the large pressure gradients at the early time also help the anisotropic flow to develops faster during the hydrodynamic evolution. Direct photons $v_2$ in Fig. \[fig10a\] is found to be considerably larger with a smaller $\tau_0$.
In Fig. \[fig10b\], we verify the $\tau_0$ sensitivities on hadronic observables. Both identified particle mean $p_T$ and charged hadron $v_{2,3}$ increase as $\tau_0$ gets smaller. This is consistent with the direct photon observables in Fig. \[fig10a\] that hydrodynamic flow is developed earlier and larger with a smaller $\tau_0$.
The effect of $\tau_0$ on photons and hadrons thus appear to be qualitatively similar, with photons being slightly more sensitive to the initial time than hadrons, as one can reasonably expect from probes that can be emitted at earlier times.
--------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------ ----------------------------------------------------------------------------------
{width="0.32\linewidth"} {width="0.32\linewidth"} {width="0.32\linewidth"}
--------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------ ----------------------------------------------------------------------------------
Out-of-equilibrium corrections {#section:viscousEffects}
------------------------------
Temperature and velocity gradients are larger in small systems than in heavy ion collisions. Consequently, the out-of-equilibrium corrections could also be expected to be larger. In the last part of this section, we investigate the shear viscous corrections to direct photon observables.
We explore the sensitivity of direct photon spectra and $v_2$ to the shear viscous corrections $\delta \Gamma$ by constraining it to be smaller than a certain fraction of its corresponding equilibrium contribution. Based on Eq. (\[eq3\]), we compute the following ratio for a photon with energy $E_q$ in the local rest frame of a fluid cell, $$r(E_q, T, \pi^{\mu\nu};a) = \frac{\vert \delta \Gamma (E_q, T, \pi^{\mu\nu}) \vert}{a \Gamma_0(E_q, T)},
\label{eq4}$$ where the coefficient $a$ is a constant parameter, which determines the maximum allowed fraction of equilibrium contribution for the $\delta f$. Then the constrained photon emission rate in a fluid cell is evaluated as, $$\begin{aligned}
&& \!\!\!\!\!\!\!\!\!\! E_q\frac{d \Gamma}{d^3 k}(E_q, T, \pi^{\mu\nu};a) = \notag \\
&& \Gamma_0 (E_q, T) + \frac{\delta \Gamma (E_q, T, \pi^{\mu\nu}) }{\mathrm{max}\{1, r(E_q, T, \pi^{\mu\nu}; a)\}}.
\label{eq5}\end{aligned}$$
Because shear viscous correction is only available for those photons emitted from the 2 to 2 scatterings in the QGP phase and meson-meson reactions, we will focus on the thermal photon flow observables from these two channels in Figs. \[fig11\](a) and (b). We compared the thermal photon spectra and their elliptic flow coefficients with the fraction parameter $a = 0, 0.5, 1$, and $\infty$ in Eq. (\[eq5\]). By choosing $a = 1$, we allow the maximum size of $\delta \Gamma$ to be equal to its equilibrium part. In this case, the thermal photon observables are very close to the no-constraint case ($a = \infty$) after integrated over all space-time volume. Some noticeable differences are present only for $p_T > 2.5$ GeV. If the constraint increases to $a = 0.5$ ($\vert \delta \Gamma \vert < 0.5 \Gamma_0$), thermal photon elliptic flow starts to show some sizeable variation. In Fig. \[fig11\](c), we show the sensitivity of total direct photon $v_2$ to different choices of the parameter $a$. Among the cases $a = 0.5, 1.0$, and $\infty$, the largest variation of the direct photon $v_2$ reaches up to $\sim 15\%$ in $2 < p_T < 3$ GeV.
Conclusion
==========
In this work, we have presented a systematic study of hadronic collective observables and of direct photon probes in small collisions systems at RHIC and LHC energies, within a consistent dynamical approach.
It was found that hydrodynamic simulations can provide a good description of the hadronic flow observables for both inclusive charged hadrons and identified particles. The effects of the hadronic transport description of the dilute hadronic phase on hadronic observables were quantified and found to be small in general. It was also found that, in the absence of longitudinal fluctuations, a boost-invariant assumption can provide a reasonable description of mid-rapidity photonic and hadronic observables.
A thermal photon enhancement - over the case where no thermal sources are present - of the differential spectra was predicted in high multiplicity collision events, in small systems at both RHIC and LHC energies. In addition, we found that direct photons carry sizeable anisotropic flow $v_{2,3}\{\mathrm{SP}\}$. These proposed signals can serves as independent tests of the hydrodynamic description of small hot and dense systems. An analysis of decay cocktail and inclusive photons was presented to provide guidance to future experimental measurements.
Several aspects of theoretical uncertainties in describing the dynamics of small systems were explored in the last section of this work. Photon emissions from the dilute hadronic phase were found to contribute up to $\sim10\%$ in the total direct photon signal. The direct photon elliptic flow was shown to have a larger sensitivity than hadrons to the choice of thermalization time $\tau_0$. The shear viscous corrections to the photon production calculations were examined and found to be under control.
Future work will include the study of more realistic sub-nucleon fluctuations in the initial state [@Mantysaari:2016ykx; @Mantysaari:2016jaz], of longitudinal fluctuations [@Khachatryan:2015oea; @Aaboud:2016jnr], as well as the interplay between soft-hard components in the intermediate $p_T$ region, and the inclusion of bulk viscosity [@Ryu:2015vwa].
It is worth re-emphasizing that measurements of low $p_T$ photons represent precious information which completes and complements what is learned with hadronic observables. These photons are both penetrating [*and*]{} soft: they are a unique characterization tool for systems of all sizes and shapes.
This work was supported in part by the Natural Sciences and Engineering Research Council of Canada. JFP was supported by the U.S. D.O.E. Office of Science, under Award No. DE-FG02-88ER40388. Computations were made in part on the supercomputer Guillimin from McGill University, managed by Calcul Québec and Compute Canada. The operation of this supercomputer is funded by the Canada Foundation for Innovation (CFI), NanoQuébec, RMGA and the Fonds de recherche du Québec - Nature et technologies (FRQ-NT).
Space-time structure of medium evolution and photon emissions in small systems
==============================================================================
In this appendix, the space-time evolution of the fluid-dynamical description of small systems is scrutinized.
----------------------------------------------------------------------------- -----------------------------------------------------------------------------
{width="0.42\linewidth"} {width="0.42\linewidth"}
{width="0.42\linewidth"} {width="0.42\linewidth"}
{width="0.42\linewidth"} {width="0.42\linewidth"}
----------------------------------------------------------------------------- -----------------------------------------------------------------------------
------------------------------------------------------- ------------------------------------------------------------------
{width="0.42\linewidth"} {width="0.42\linewidth"}
{width="0.42\linewidth"} {width="0.42\linewidth"}
{width="0.42\linewidth"} {width="0.42\linewidth"}
------------------------------------------------------- ------------------------------------------------------------------
Semi-quantitative estimators of the validity of fluid dynamics are Knudsen and Reynolds numbers in numerical simulations. The Knudsen number is estimated according to Ref. [@Niemi:2014wta], $${\rm Kn} = K_\theta = \tau_\pi \theta = \frac{5 \eta}{e + \mathcal{P}} \theta = \frac{5\eta}{sT} \theta,
\label{eqA1}$$ where $\tau_\pi$ is the shear relaxation time and $\theta = \partial_\mu u^\mu$ is system’s expansion rate. In order to have a realistic estimation of $K_\theta$ in the dilute hadronic phase, we adopt a temperature dependent $\eta/s(T)$ for $T < T_\mathrm{sw} = 155$ MeV [@Niemi:2011ix], $$\begin{aligned}
\frac{\eta}{s}(T) &=& \left(\frac{\eta}{s}\right)_\mathrm{min} + 0.0594 \left(1 - \frac{T}{T_\mathrm{sw}} \right) \notag \\
&& + 0.544 \left(1 - \frac{T}{T_\mathrm{sw}} \right)^2.
\label{eqA2}\end{aligned}$$ The smallness of the Knudsen number reflects how fast the system can evolve towards local thermal equilibrium. Complementarily, the distribution of the inverse Reynold’s number [@Niemi:2014wta], $$R_\pi^{-1} = \frac{\sqrt{\pi^{\mu\nu} \pi_{\mu\nu}}}{e + \mathcal{P}},
\label{eqA3}$$ describes how far the system is out-of-equilibrium at every space-time point. Small values of Kn and $R_\pi^{-1}$ are needed for viscous effects to remain perturbative.
Figs. \[fig12\] shows the space-time distribution of the temperature profiles, together with the evolution of the Knudsen and of the inverse Reynold’s number in two typical individual fluctuating events in central p+Pb and $^3$He+Au collisions. In spite of the small system size, the Knudsen numbers above the switching temperature, $T_\mathrm{sw} = 155$ MeV, remain $\sim 0.5$ in $^3$He+Au collisions at the top RHIC energy and reaches up to $0.6\sim0.7$ in the p+Pb collisions at 5.02 TeV. The result that $K_\theta < 1$ in the high temperature QGP phase suggests that a hydrodynamic description of these small systems is within the validity of the theory. In Figs. \[fig12\](c) and (f), the evolution of the inverse Reynold’s number is shown for the two events. The values of $R_\pi^{-1}$ remain small during the entire evolutions which means that the relative size of the shear stress tensor is small compared to its corresponding ideal part.
In Figs. \[fig13\], we study the space-time structure of the thermal photon emission in one p+Pb collisions at 5.02 TeV. In the left panels, the thermal photon emission is shown for three $p_T$ cuts. In the right panels, the relative size of shear viscous correction to the photon yield is shown for the same momentum cuts. The largest viscous corrections are commonly found at the early stage of the evolution for all three $p_T$ bins. For $0.4 < p_T < 1$GeV, the maximum size of viscous correction is less 1% compared to its equilibrium part. The relative size of the $\delta \Gamma$ correction increases quadratically as a function of $p_T$. For $2 < p_T < 3$ GeV, the shear viscous correction can reach up to 60% of its equilibrium part. However, the corresponding left-hand-side plot indicates that the region where viscous corrections are large, at early time and large radius, are also regions of low photon emissivity. Thus very few photons are produced from these regions where viscous correction are large, and viscous corrections are under control in space-time regions important for the thermal photon spectrum. This is in line with the findings of Section \[section:viscousEffects\].
[^1]: There are 305 hadronic species in total, up to 2.25 GeV in their mass, included in the hadron resonance gas phase in s95p-v1.2 and UrQMD v3.4. Please see [<http://urqmd.org/itypes.html>]{} for details.
[^2]: Compared to the list in Chapter 21 of Ref. [@Shen:2014thesis], we exclude photons from the decay channel $\rho_0 \rightarrow \pi^+ + \pi^- + \gamma$ in the short-lived resonance contribution. Photons from this channel are added in the decay photon cocktail.
[^3]: Since hadronic measurements represent only a small fraction of the measurements used to constrain nuclear parton distribution functions, it may appear that the use of hadronic data can only result in a small contamination of the extracted functions. It is important to note, however, that measurements are not necessarily given the same weight when constraining distribution functions. For example, hadronic measurements are given a large weight in EPS09 to provide better constraints on the gluon distribution, increasing the influence of this measurement on the nuclear parton distribution functions.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Given a metabolic network in terms of its metabolites and reactions, our goal is to efficiently compute the minimal knock out sets of reactions required to block a given behaviour. We describe an algorithm which improves the computation of these knock out sets when the elementary modes (minimal functional subsystems) of the network are given. We also describe an algorithm which computes both the knock out sets and the elementary modes containing the blocked reactions directly from the description of the network and whose worst-case computational complexity is better than the algorithms currently in use for these problems. Computational results are included.'
address:
- 'Institut für Mathematische Optimierung, Otto-von-Guericke-Universität, Universitätsplatz 2, D-39106 Magdeburg, Germany'
- 'Max-Planck-Institut für Dynamik komplexer technischer Systeme, Sandtorstr. 1, D-39106 Magdeburg, Germany'
- 'Department of Mathematics, Simon Fraser University, 8888 University Drive, Burnaby, British Columbia V5A 1S6 Canada'
author:
- 'Utz-Uwe Haus'
- Steffen Klamt
- Tamon Stephen
title: Computing knock out strategies in metabolic networks
---
Introduction
============
Systems biology studies the complex systems which occur at many levels of biology. Such systems involve large numbers of components and interactions. We consider [*metabolic networks*]{}, that is, a set of metabolites that can be interconverted by biochemical reactions. A fundamental question about metabolic networks is to find knock out strategies that block the operation of a given reaction or set of reactions. A target reaction is blocked if it cannot operate in a steady state. Some reactions can be easily knocked out, while others may be expensive or impossible to knock out directly. In this case, we consider the problem of blocking target reactions by inhibiting other reactions so that the targets cannot continue in a steady state. Some applications of this problem are outlined in [@KG04] and [@Kla06], and an implementation has been included as part of [@Klamt], a [MATLAB]{} package for analyzing cellular and biochemical networks.
In this paper, we consider methods of computing the minimal sets of reactions that need to be disabled to block a given target. We call such a knock out set a [*cut set*]{}. We remark that we focus only on the (inclusion-wise) [*minimal*]{} cut sets since these are the cheapest ways of blocking reactions in terms of effort and impact on the system. The list of minimal cut sets contains the same information as the full list of cut sets but is much shorter.
We consider two main approaches. The first is to build the hypergraph of elementary modes and then compute the minimal cut sets as the transversal of this hypergraph. This strategy has been employed successfully in [@KG04], however we observe we can substantially improve the computation of the transversal hypergraph. The second approach is to compute the minimal cut sets directly using the ideas of Gurvich and Khachiyan [@GK99] and others on generating monotone boolean formulae. This procedure also generates the set of elementary modes employing a given set of reactions as a by-product, which is a potentially useful feature.
We expect that these methods can be adapted to more of the complex systems that are typical at many levels of biology. Indeed, the question of finding minimal cut sets can be abstracted to finding the minimal failure modes of a network, which is a natural question arising in various contexts. Some suggestions for using these methods in other types of biochemical networks are presented in [@KSRL+06].
Preliminaries
=============
We model a metabolic network as a number $m$ of metabolites involved in a set $Q$ of $q$ reactions (where $q$ is typically between $m$ and $2m$). For our purposes, these reactions can be encoded in a $m \times q$ matrix $N$ whose columns encode the metabolites produced and consumed by a given reaction. The matrix $N$ is known as the [*stoichiometric*]{} matrix. The reactions may be divided into two types: [*reversible*]{} reactions that can either produce a given output from a given input or vice-versa; and [*irreversible*]{} reactions which cannot operate in reverse. Let $S$ be the index set of the reversible reactions and $U=Q \setminus S$ be the index set of the irreversible reactions. We call our set of target reactions $T$, for simplicity we will usually assume they are irreversible, i.e. $T \subseteq U$.
Given such a network, we are interested in its potential steady-state flux vectors. In steady state, the reaction rates balance the metabolites, i.e. for each metabolite it holds that the sum of the rates of all reactions consuming the metabolite equals the sum of the rates of the reactions producing it. We can represent such a steady state as a vector $x \in
\R^q$ s.t. $Nx=0$ and $x_i \ge 0$ for all $i \in U$. Then we can formally define a [*cut set*]{} $C \subset Q$ as a subset such that the system: $$\label{eq:cutset}
\{x \in \R^q | Nx=0, x_i \ge 0 ~ \forall i \in U, x_c=0 ~ \forall c \in C\}$$ has only solutions with $x_t=0$ for all $t \in T$. A [*minimal cut set*]{} (MCS) is simply a cut set none of whose proper subsets is a cut set. A concept closely related to minimal cuts sets is that of an elementary mode. An [*elementary mode*]{} (EM) is a minimal set of reactions that can exist in a steady state. The importance of EM’s in metabolic networks is discussed, for example, in [@SFD00] and [@SKB+02]. EM’s are, up to a scaling factor, support minimal solutions to the system: $$\{x \in \R^q | Nx=0, x_i \ge 0 ~ \forall i \in U\}$$ The problem of computing the EM’s of a given system has a nice geometric formulation: it reduces to finding the extreme rays of the cone $\{r | \hat{N}r=0, r \ge 0\}$, where $\hat{N}$ is $N$ modified to represent reversible reactions as opposite pairs of irreversible reactions. This is described in [@GK04]. The EM’s can then be computed by applying the double description method (see for example [@FP96]) to this cone. As observed in [@GK04], EM’s are characterized up to a constant by their binary support patterns. Hence we will also use the term EM to refer to this support pattern, which we can view as a set of reactions.
For the purposes of finding cut sets for a given target $T$, we consider only the EM’s that include at least one target reaction. Note that cut sets are exactly the sets of reactions that intersect each of these EM’s. The collection $\E$ of these EM’s defines a Sperner hypergraph $\H=(R,\E)$ on the set of reactions. (A hypergraph is Sperner if it has no nested edges.) The key observation is that cut sets are exactly the sets that intersect every edge of $\H$. In the terminology of hypergraphs such sets are known as [*hitting sets*]{} or [*vertex covers*]{}. The collection of all minimal hitting sets for $\H$ is itself a hypergraph $\H'=(R,\E')$ which is dual to $\H$ in the sense that its minimal hitting sets are the edges of $\H'$. The hypergraph $\H'$ is known as the [*transversal hypergraph*]{} of $\H$ and is denoted $\Tr(\H)$.
The approach to computing minimal cut sets presented in [@KG04] is to first compute the EM’s hypergraph $\H$ via the double description method and then compute $\Tr(\H)$. The computation of $\Tr(\H)$ is done through an enumeration scheme. This succeeds in solving four large scale networks arising in biomass synthesis in [*E.coli*]{}. The computation benefits substantially from effective preprocessing, but nevertheless consumes a lot of time and memory. A faster and more memory efficient algorithm is described in Section \[se:berge\].
The double description method and the algorithms of Section \[se:berge\] have uncertain complexity. For this reason, we give in Section \[se:joint\] an algorithm which generates both the EM’s and the MCS’s directly from the stoichiometric matrix which has a surprisingly good complexity bound of $m^{\poly(\log{m})}$ in the output size. See [@stougie:07] for an overview of the known complexity results regarding EM and MCS algorithms.
We could consider weighting the reactions and looking for a single minimum weight cut set. This assumes that the costs are independent, which is questionable - it may be possible to attain some economies when knocking out multiple reactions. Additionally, designing a weighting function for blocking a metabolic reaction requires quantifying the costs to disable various reactions, which is a labour intensive task. Since we can in some interesting cases produce the entire list of minimal cut sets, we view generating the entire list as a suitable goal.
If we do want to find a minimum weight cut set, this problem is the [*minimum set cover*]{} problem on the dual hypergraph produced by interchanging the roles of the vertices and edges of $\H$. Minimum set cover is a classical NP-complete problem, see for example [@ADP80].
Characteristics of metabolic networks {#se:struct}
-------------------------------------
In our model the stoichiometric matrix $N$ is a real $m \times q$ matrix. Typically, we would expect $N$ to have many zero entries as any particular reaction will only involve a few metabolites. Those non-zero entries will usually be small integers since chemical reactions are discrete rearrangements of molecules. Note that a column describes the same reaction when it is scaled by a positive factor. Recall that reactions may be reversible or irreversible. The hypergraphs of EM’s and MCS’s are 0-1 matrices $\H$ and $\H'$ determined by $N$. Each row of $\H$ ($\H'$) is an indicator function for a given EM (MCS), that is, a row indicates the complement of a maximal $C$ such that (\[eq:cutset\]) has a solution with $x_t>0$ for some $t \in T$ (indicates a minimal $C$ such that (\[eq:cutset\]) has only solutions with $x_t=0$ for all $t \in T$). Both $\H$ and $\H'$ have the same number of columns as $N$, but in our applications they will have many more rows. The orders of the rows and columns are arbitrary, but they can affect the performance of the algorithms.
### Typical behaviour
We understand from biological considerations that we would expect to get many small hitting sets. The intuition is that such networks have some important reactions whose loss quickly impairs the operation of the network. This can be quantified through “fragility coefficients” [@KG04], which are an average of MCS sizes. Their examples produced fragility coefficients in a narrow range.
### Test cases {#se:cases}
The motivating problems from [@KG04] are four networks obtained from studying biomass synthesis in [*E.coli*]{}. These are the growth modes for substrates acetate, succinate, glycerol and glucose from the network presented in [@SKB+02]. The objective is to block the single target reaction representing growth in each of these networks. For the purposes of this computation, a pair of reactions corresponding to the same multifunctional enzyme (transketolase) has been combined. This modifies the input hypergraph by merging a pair of vertices, creating some nested edges.
Computing minimal cut sets via elementary modes {#se:berge}
===============================================
The method proposed in [@KG04] to compute minimal cut sets involves two steps. The first step is to compute the set of EM’s via polyhedral methods, and the second step is to compute the transversal hypergraph of the EM’s hypergraph. Their method of computing the transversal hypergraph requires enumerating many possible partial solutions. As a result it consumes substantial time and memory, and is ripe for improvement.
Algorithms
----------
In this section we sketch algorithms for the transversal hypergraph problem, including the enumeration algorithm of [@KG04] and the algorithm described by Berge in [@Ber89]. For the latter, we describe several useful modifications.
### Enumeration algorithm
This is the original algorithm implemented in [@KG04], the predecessor to . Beginning at size 1, it tests for subsets of a given size. It maintains a list of unused partial cut sets to avoid full enumeration of the subsets.
The problem with this algorithm is that the list of partial cut sets can grow quite quickly. Nevertheless, it can solve large problems. A major reason for this is the abundance of small cut sets, which keeps the list of partial cut sets manageable.
### Berge’s algorithm {#se:ourberge}
This algorithm [@Ber89] orders the edges $e_1,e_2,\ldots,e_r$ of the hypergraph $\H$, and then computes in sequence the transversal of each hypergraph $\H_i$ consisting of edges $e_1$ through $e_i$. This can be done by taking all the edges created by adding a vertex from $e_i$ to an edge in $\H_{i-1}$ and keeping all the inclusionwise-minimal edges. $\H_1$ has an edge consisting of a single vertex for each vertex from $e_1$.
The performance of this algorithm depends on the size of the intermediate transversals generated, which in turn depends on the order of the vertices. Intuitively we do not expect the size of a transversal of a subgraph to substantially exceed the larger of the size of the initial graph and the size of the final transversal hypergraph. In practice this does often turn out to be the case, but Takata exhibited an example where an intermediate transversal will have size $\Theta(m^{\log(\log(m))})$ in the combined size of the input and output for any ordering of the edges, see [@Hage07]. We do not know of any non-trivial upper bounds for the size of intermediate transversals or how to produce favourable edge orderings when they exist.
A naive implementation of Berge’s algorithm will be slow, but there are a number of modifications that can make it more effective. The main bottleneck is the removal of superset rows from the list generated from $\H_{i-1}$ and $e_i$. We do this in time $O(n^2)$ in the length of the list using the simple algorithm described below. There are algorithms for superset removal that work in time $O(n^2/\log{n})$ or slightly better, see [@Pri95] and [@SE96]. It is known, that there is a lower bound of $O(n^2/\log^2 n)$, if the complete subset lattice is constructed by algorithm, see [@Pri99]. It appears that no non-trivial lower bounds are known for superset removal alone, and we remark that it is an interesting question.
When implementing Berge’s algorithm, we can avoid generating some edges that are clearly supersets before entering the removal phase. For instance whenever an edge $f$ of $\H_{i-1}$ intersects $e_i$, we can add only that edge to the list of candidates for $\H_i$ since $f$ will be generated as an edge using the common vertex with $e_i$ and all edges generated using the other vertices of $e_i$ will contain this edge. Additionally, we can see that the list of edges retained in this way will itself be superset free since it is a subhypergraph of $\H_{i-1}$. Hence we only need to check which of the new edges generated are supersets of these retained edges: newly generated edges cannot be subsets of edges from $\H_{i-1}$ since they are edges from $\H_{i-1}$ with an additional vertex, and if $f_1 \cup v_1 \subset f_2 \cup v_2$ in the new edge list, then $f_1$ must contain $v_2$ and hence would in fact be in the retained edges list.
Our implementation of Berge’s algorithm is now included in .
### Others {#se:others}
Several algorithms have recently been proposed which build on the Berge algorithm and are effective for some problems, see for example [@BMR03] and [@KS05]. A nice recent survey of methods for computing transversals is [@EMG06]. It focuses on work based on the ideas of Fredman and Khachiyan [@FK96] concerning generation of logical formulae. These methods provide better theoretical performance and we consider them in Section \[se:joint\].
Implementation {#se:impb}
--------------
While Berge’s algorithm is well known, naive implementations of it are quite slow. Modifications of Berge’s algorithm, such as the ones mentioned in Section \[se:others\] have apparently yielded good results, but we know of no public implementations of such an algorithm. Thus we implemented our own version of the algorithm described in Section \[se:ourberge\]. We used [MATLAB]{} because it is the platform for and . Our code is freely available for academic use [@BergeCode].
We tested our code on the examples of [@KG04]. These are obtained from the growth-related EM’s calculated in [@SKB+02] via simple modifications described in Section \[se:pre\]. Our results are presented in Section \[se:compb\]. Below we describe some details of our implementation, as well as the original implementation. We used as much of the setup as possible, including input data structures and pre- and post-processing code to facilitate comparison between the core algorithms.
### Preprocessing {#se:pre}
Beginning from the computed full set of EM’s, we first select only those containing at least one of the target reactions. As mentioned in Section \[se:struct\], some groups of reactions may be catalyzed by the same multifunctional enzyme. These reactions are cut simultaneously by disabling such an enzyme. We combine the reactions corresponding to such groups by a logical “or” operation as the corresponding elementary mode is disabled if any of its constituent reactions are disabled. This merges vertices in the input hypergraph; the merged vertices in the new modes represent the set of reactions blocked by the enzyme. In our examples there is only one target reaction (biomass synthesis) and one multifunctional enzyme (transketolase). This gives us our initial hypergraph $\H$ of modes that include the target reaction.
There are several further preprocessing steps applied to $\H$. The first is to scan for zero rows and columns, which should not occur. The second is to find columns of ones, which correspond to cut sets of size one and can be noted and removed until postprocessing. Next duplicate columns are identified. They are treated as a single column and then reexpanded during post-processing. In terms of the hypergraph, they represent vertices that are in exactly the same set of edges, and can thus be merged during the calculation.
Finally, for Berge’s algorithm, we remove rows that are supersets of other rows. Our current implementation does this in time $\Omega(m^2)$. Superset or duplicate rows may be introduced when merging vertices corresponding to multifunctional enzymes. In our examples, about 20% of the rows in $\H$ are supersets of other rows, these come from the pairs of EM’s that contain only one of the transketolase reactions, each pair contains one EM using transketolase1 and transaldo, and one EM instead using transketolase2. The first of these becomes a superset of the second upon merging the two columns. Removing these supersets in preprocessing cut the observed running time by about 30%. Removing superset rows does not speed up the algorithm in , whose bottleneck is generating the possible cut sets of a given size.
### Postprocessing {#se:post}
Following the application of either algorithm to the hypergraph produced in Section \[se:pre\], there is a small amount of postprocessing that needs to be done. Merged vertices are reexpanded: each cut set containing a merged vertex $v$ will be replaced by cut sets containing exactly one of the vertices that were merged into $v$. Size one cut sets will be introduced for each column of ones that was removed prior to the calculation. Finally, the cut sets involving reactions corresponding to a multifunctional enzyme are split into cut sets containing the constituent reactions.
### Coding issues
Because we are working with large matrices, memory use is a key consideration, it is essential to use a bit-level representation of the binary matrices describing the EM’s and MCS’s. This is done in the implementation of [@KG04].
We had to accommodate [MATLAB]{}’s strengths and weakness: we obtained substantial time savings through small changes in the main bottleneck routine. This routine removes the rows from one list that are supersets of rows from another. Since memory allocation in [MATLAB]{} is slow rather than resizing the matrix when superset rows are identified, we mark rows for removal in a pass through the matrix and then we generate a single new matrix containing those rows. It is also useful to take advantage of [MATLAB]{}’s internal parallelization. [MATLAB]{} can quickly check a single row for super- or sub-setness against an entire matrix using a couple of bitwise comparisons. We always cycle through the rows of the shorter of the two lists and check its rows against the longer list.
### Computational results {#se:compb}
In this section we compare the performance of the transversal hypergraph code written for to our implementation of Berge’s algorithm now included in . Our test base is the four EM problems from [@KG04] (see Section \[se:cases\]) and their hypergraph transversals. The transversals are denoted by primes.
We first compare the algorithms from the point of view of the largest intermediate lists generated. This tells us how much memory each algorithm uses, and gives an idea of how fast it can run. In the case of the algorithm, the measure is the largest list of partial cut sets generated. In the case of the Berge algorithm, the measure is the largest intermediate transversal generated before removing nested subsets. Due to the reductions of Section \[se:pre\], this sometimes turned out to be smaller than output it generated after postprocessing. Results are in Table \[ta:bsizes\]. We used as inputs both the EM’s found for the networks described in Section \[se:cases\] and the dual hypergraphs containing the MCS’s that we computed (denoted with a ’).
Problem: ’ ’ ’ ’
-------------------------- ------ ----- ------- ------ -------- ------- --------- -------
Input columns 104 104 104 104 105 105 105 105
Preprocessed columns 21 98 26 101 28 103 34 103
Input rows 363 245 3421 1255 9479 2970 21592 4225
Preprocessed rows 289 244 2722 1254 7472 2969 18481 4224
Raw output rows 54 280 159 2589 376 7047 918 18481
Final output rows 245 289 1255 2722 2970 7472 4225 18481
[FluxAnalyzer]{} largest 3563 – 69628 – 342025 – 902769 –
Berge largest 94 296 304 2669 657 7047 1714 18569
time 5.1 – 633.5 – 8696.2 – 54099.1 –
Berge time 0.7 1.1 7.1 35.8 29.6 215.0 206.5 727.4
: Sizes of intermediate lists generated in computing transversals and computation times.[]{data-label="ta:bsizes"}
\[ta:btimes\]
We record the number of columns and rows both before and after preprocessing. The preprocessing reduces the problem size substantially mainly by removing columns corresponding to reactions which form cut sets by themselves. The number of output rows is given before and after postprocessing.
The routine was not able to solve the dual problems due to memory limitations. This routine is much better at converting EM graphs to MCS graphs than vice versa because many of the MCS’s are small in these instances. For the given examples, all the EM’s are large, so the computation begins by building a long list of partial cut sets.
While Berge’s algorithm provides no complexity guarantees, it worked very well for these problems. Of particular note is that, unlike the algorithm, the information carried by Berge’s algorithm during its intermediate stages was never much larger than the size of the final output prior to preprocessing.
We also give running times in seconds in Table \[ta:btimes\]. Both codes are written in [MATLAB]{}, take identical inputs, and use the same preprocessing code as noted in Section \[se:pre\]. These comparisons were run on a Sun Fire V890 with 32 GB memory and 16 1200 MHz processors. Note that running times include some preprocessing, such as removing duplicate rows for Berge, but excludes postprocessing. This method of reporting was used in [@KG04].
Computing Minimal Cut Sets Directly {#se:joint}
===================================
In this section, we consider methods of generating the MCS’s directly from the stoichiometric matrix. The techniques outlined here are based on an algorithm of Fredman and Khachiyan [@FK96] for dualizing boolean functions. They offer better complexity guarantees than the algorithms of Section \[se:berge\]. They work directly from the stoichiometric matrix for a network and generate the hypergraph of EM’s containing the blocked reactions as a byproduct of the computation.
Algorithm
---------
The cut sets generated by a given stoichiometric matrix define a boolean function, that is a function that takes a binary pattern of included reactions as input, and yields 1 if this set of reactions is a cut set, and 0 if it is not. Further, this is a [*monotone*]{} function in the sense that if a given set is a cut set, then any superset of that set is also a cut set. Thus the problem of finding [*minimal*]{} cut sets can be viewed as a problem of representing such a boolean function via its minimal true assignments. The support of an EM is then the complement of a maximal false assignment.
A monotone boolean function can be represented uniquely both by its minimal true assignments and its maximal false assignments. The process of converting from one representation to the other is sometimes called dualization, and is equivalent to the hypergraph transversal problem. Fredman and Khachiyan [@FK96] proposed an algorithm that generates the transversal incrementally using an algorithm that is slightly superpolynomial in the size of the graph and the transversal. This key idea in this algorithm is to recurse on a variable that occurs with relatively high frequency.
As described in [@GK99], this algorithm can be implemented from a function evaluation oracle, producing both the hypergraph of minimal true assignments and its transversal in $m^{o(\log(m))}$ oracle calls, where $m$ is the combined size of the two hypergraphs. This fits very well with our problem: given the stoichiometric matrix, we want to generate both the MCS’s and the EM’s. Note that the boolean function characterizing the cut set is monotone because every superset of a cut set is again a cut set.
We remark that the amount of memory required to generate the next clause is bounded by $m^{\poly(\log{m})}$, while the worst-case memory blow up for Berge is unknown, but not polynomial.
Implementation {#se:impj}
--------------
The algorithm of [@FK96] gives remarkable theoretical results, but the only implementation we know of is that of [@BEGK06]. Their code is not public and uses hard-coded oracles different from the one for our problem. So we implemented a prototype of the Fredman-Khachiyan algorithm with a suitable oracle, again in [MATLAB]{} for easy comparison to and the results in Section \[se:impb\]. We emphasize, though, that this algorithm proceeds directly from the stoichiometric matrix, in contrast to the Berge algorithm, which requires the completed computation of the EM’s as input.
### Oracle {#se:oracle}
We test whether a given set $C$ is a cut set by checking if the system (\[eq:cutset\]) has any solutions with $x_t > 0$ for some $t \in T$. This is a linear programming feasibility problem, and thus can be solved in polynomial time. We do this via an external call to [CPLEX]{} [@ilog-cplex-uuh], which is known to have a fast and reliable LP solver. We can test for non-trivial solutions by maximizing the sum of the target variables $\sum_{i \in T} x_i$ subject to (\[eq:cutset\]). If this is greater than zero or unbounded we have a non-trivial solution.
### Duality checker {#se:checker}
Using this oracle, we implemented a version of the “Algorithm A” duality checker from [@FK96]. This checker either verifies that our current collections of EM’s and MCS’s form dual hypergraphs, in which case both sets are complete and we are done, or it finds a set of reactions that is not a superset of any current EM, and whose complement is not a superset of any current MCS. Given such a clause we use the oracle to check if it is a mode or if its complement is a cut set. If it is a mode, we test whether it remains a mode upon removing in turn each of its constituent reactions – if the resulting set is no longer a mode, we return the removed reaction to the set, otherwise it stays out. In this way we obtain an EM, which we add to our list. Similarly, given a cut set, we obtain an MCS.
The essence of the algorithm is to recurse on a frequently occurring reaction in one of the lists until we arrive at a trivial case. The recursion then tests separately if duality holds assuming that this variable is true and assuming that it is false. By taking a frequently occurring variable, Fredman and Khachiyan ensure that the sizes of the lists decrease fast enough to guarantee that the algorithm runs in time $m^{O(\log^2(m))}$, where $m$ is the current joint length of the two lists. As with Berge’s algorithm, we find that the bottleneck is removing supersets from lists: when we recurse on a reaction some of the elements of the reduced lists, which omit this reaction, will no longer be minimal. These must be removed, and this takes $O(m^2)$ time.
Fredman and Khachiyan also provide an “Algorithm B” which achieves further economy through observing some interdependencies in the two subproblems. This reduces the time guarantee to $m^{o(\log(m))}$, but the resulting algorithm is much more complicated, so we did not implement a prototype.
We did try several variations of the simpler Algorithm A. We found some useful corners to cut: it is helpful to short circuit the recursion by treating more base cases than suggested by the algorithm. We can substantially reduce the number of superset removal calls required by doing them only when necessary before recursing rather than at the start of the checking routine.
We ran our code on the stoichiometric matrices that are used to generate EM’s in [@SKB+02]. Our oracle tests if a set of reactions blocks the growth reaction, the two transketolase reactions are treated as a single reaction for the purposes of blocking. Thus the EM’s we generate use a single bit to indicate if either of the two reactions are active. As with the preprocessing of Section \[se:pre\] this merges certain EM’s.
Problem
----------------------- -------- ---------- ----------- -----------
EM’s 289 2722 7472 18481
MCS’s 245 1255 2970 4225
Total recursive calls 107781 11129110 122136668 764239195
Time to generate 194.8 10672.2 103511.2 677599.3
: Call counts for the Fredman-Khachiyan algorithms.[]{data-label="ta:jsizes"}
In Table \[ta:jsizes\] we record the number of calls to the (recursive) duality checker used in our implementation of Fredman and Khachiyan’s algorithm. Each stoichiometric matrix has 89 metabolites and 105 or 106 reactions.
The algorithm is written in [MATLAB]{} and the oracle uses external calls to [CPLEX]{} when necessary to solve linear programs. Running times are included in Table \[ta:jsizes\] above. We used the same computer as in Section \[se:compb\].
We remark that if our objective is to produce the EM’s containing the target reactions rather than the MCS’s, we can compress the network as described in [@KGv06]. This speeds up the computation, since it now needs to generate only the few cut sets for the compressed modes (which are hard to interpret), see Table \[ta:compress\_gen\]. In this case, since our objective is to produce the EM’s, we did not merge the two transketolase reactions.
Problem
----------------------- ------- --------- ---------- ----------- -- --
Compressed reactions 40 40 42 42
EM’s 363 3421 9479 21592
Total recursive calls 38503 3487200 20971005 217252316
Time to generate 45.8 2707.0 17202.0 210749.8
: Data for generating the EM’s from the compressed network[]{data-label="ta:compress_gen"}
In fact it is possible to do most of this compression in such a way that the cut sets generated can be expanded to yield the MCS’s for the original system. Working with these partially compressed networks takes less than 10% longer than the fully compressed networks reported in Table \[ta:compress\_gen\].
We can also produce the full set of EM’s using this method by blocking the full set of reactions. This requires modifying the oracle to check for non-trivial solutions to (\[eq:cutset\]) via a rank check. If we generate them from the compressed matrices, this is somewhat tractable, but slower than generating only those containing the target reactions.
Conclusions
===========
As has been often seen in recent years, biological problems are ripe for the application of mathematics. In Section \[se:berge\], we use a non-trivial, but simple algorithm to compute MCS’s from EM’s in minutes rather than hours in the context of a large metabolic network problem. With such large data sets a careful implementation of the algorithm was as essential to make it useful.
While Berge’s algorithm is successful in practice, it is poorly understood in theory, and thus must be considered suspect. Additionally, it requires the precomputation of the EM’s via an algorithm which also has uncertain worst-case complexity. Thus we also considered algorithms based on the dual generation framework of Fredman and Khachiyan [@FK96]. These offer a guaranteed theoretical performance which is close to polynomial, i.e. $m^{\poly(\log{m})}$ where $m$ is the joint size of the EM’s and MCS’s.
The Fredman-Khachiyan oracle-based algorithms also have the advantage that, unlike Berge’s algorithm, the lists are generated [*incrementally*]{} - at each step a new EM or MCS is added to the current collection, and in time and memory $m^{\poly(\log{m})}$ in the current joint size of the lists. Thus if we only have the resources to do a partial computation, we will get a partial answer. Indeed, much larger systems exists for which the full sets of EM’s and MCS’s are too large to store. In this situation, even if we could compute a partial list of EM’s, its dual is meaningless. In contrast, the oracle-based algorithm can produce a sampling of EM’s and MCS’s as resources permit.
The oracle-based algorithms also provide a method of computing the EM’s containing target reactions without computing the full set of EM’s. This is desirable since there may be an enormous number of EM’s, only a small fraction of which contain the target reactions. The double description algorithm can be modified to compute only the EM’s containing a target, however as noted in [@KGv06], unless implemented carefully, this may be slower than computing the full set of reactions.
However, there are several clear drawbacks to oracle-based algorithms. The obvious ones are that they are more difficult to implement and slower. The high number of recursive calls used to solve the small problem (see Table \[ta:jsizes\]) suggests that it will be difficult to make such an algorithm competitive with Berge, especially in the case where the EM’s are given. It is encouraging that our simple implementation is able to solve even the largest problem (), although the time required was long. This gives us some hope that a more advanced implementation of the oracle-based algorithm could be competitive in this application. One place to start would be to implement Fredman and Khachiyan’s more intricate, but theoretically faster $m^{o(\log(m))}$ algorithm.
The oracle-based methods could be improved by additional preprocessing as is done when obtaining the EM’s from the MCS’s. For example, single reaction cut sets are treated separately in the algorithms of Section \[se:berge\], this could also be implemented in the Fredman-Khachiyan framework. We expect this would yield a mild improvement in the running time.
Acknowledgments
---------------
We are grateful for fruitful discussions with Annegret Wagler and Robert Weismantel This work was partially supported by DFG FG-468, the Research Focus Program Dynamic Systems funded by the Kultusministerium of Saxony-Anhalt, and by a President’s Research Grant at Simon Fraser University.
[CLMS[[$^{+}$]{}]{}07]{}
G. Ausiello, A. D’Atri, and M. Protasi, *[Structure preserving reductions among convex optimization problems]{}*, J. Comput. Syst. Sci. **21** (1980), 136–153.
Endre Boros, Khaled Elbassioni, Vladimir Gurvich, and Leonid Khachiyan, *[An Efficient Implementation of a Quasi-Polynomial Algorithm for Generating Hypergraph Transversals and its application in Joint Generation]{}*, Discrete Applied Math **154** (2006), no. 16, 2350–2372.
Claude Berge, *[Hypergraphs. Combinatorics of finite sets. Transl. from the French]{}*, [North-Holland Mathematical Library, 43. Amsterdam etc.: North-Holland. x, 255 p.]{}, 1989.
James Bailey, Thomas Manoukian, and Kotagiri Ramamohanarao, *[A Fast Algorithm for Computing Hypergraph Transversals and its Application in Mining Emerging Patterns]{}*, [Proceedings of the 3rd IEEE International Conference on Data Mining (ICDM)]{}, [[IEEE]{} Computer Society]{}, 2003, pp. 485–488.
Flavio Chierichetti, Vincent Lacroix, Alberto Marchetti-Spaccamela, Marie-France Sagot, and Leen Stougie, *Modes and cuts in metabolic networks: [C]{}omplexity and algorithms*, Tech. Report Ext. rep. 2007-01, Technische Universiteit Eindhoven, Eindhoven, 2007.
Thomas Eiter, Kazuhisa Makino, and Georg Gottlob, *[Computational aspects of monotone dualization: a brief survey]{}*, Discrete Applied Math, doi:10.1016/j.dam.2006.04.017, in press., 2007.
Michael L. Fredman and Leonid Khachiyan, *On the complexity of dualization of monotone disjunctive normal forms*, J. Algorithms **21** (1996), no. 3, 618–628.
Komei Fukuda and Alain Prodon, *[Double Description Method Revisited]{}*, [Deza, M., Euler, R. and Manoussakis, I. (eds.) Combinatorics and Computer Science, Springer. Lect. Notes Comput. Sci. 1120]{}, Springer, 1996, pp. 91–111.
V. Gurvich and L. Khachiyan, *[On generating the irredundant conjunctive and disjunctive normal forms of monotone Boolean functions.]{}*, Discrete Applied Math. **96-97** (1999), 363–373.
Julien Gagneur and Steffen Klamt, *[Computation of elementary modes: a unifying framework and the new binary appraoch]{}*, BMC Bioinformatics **5** (2004), no. 175.
Matthias Hagen, *[Lower bounds for three algorithms for the transversal hypergraph problem]{}*, 33rd International Workshop on Graph-Theoretic Concepts in Computer Science (WG 2007), 2007, to appear.
Utz-Uwe Haus, Steffen Klamt, and Tamon Stephen, *[MATLAB code for hypergraph transversal]{}*, Available at: [ http://www.math.sfu.ca/[\~]{}tamon/Software/Berge/index.html]{}, 2007.
ILOG, *[CPLEX]{}*, 1997–2007, For information see: .
Steffen Klamt and Ernst Dieter Gilles, *[Minimal cut sets in biochemical reaction networks]{}*, Bioinformatics **20** (2004), no. 2, 226–234.
Steffen Klamt, Julien Gagneur, and Axel von Kamp, *Algorithmic approaches for computing elementary modes in large biochemical reaction networks*, IEE Proceedings Systems Biology **154** (2006), no. 4, 249–255.
Steffen Klamt, *[Generalized concept of minimal cut sets in biochemical networks]{}*, Biosystems **83** (2006), no. 2-3, 233–247.
Dimitris Kavvadias and Elias C. Stavropoulos, *[An Efficient Algorithm for the Transversal Hypergraph Generation]{}*, Journal of Graph Algorithms and Applications **9** (2005), no. 2, 239–264.
Steffen Klamt, Julio Saez-Rodriguez, and Ernst Dieter Gilles, *[Structural and functional analysis of cellular networks with [CellNetAnalyzer]{}]{}*, BMC Systems Biology **1:2** (2007), CellNetAnalyzer can be downloaded for free (academic use) via the following web-site:\
[ http://www.mpi-magdeburg.mpg.de/projects/cna/cna.html]{}.
Steffen Klamt, Julio Saez-Rodriguez, Jonathan Lindquist, Luca Simeoni, and Ernst Dieter Gilles, *A methodology for the structural and functional analysis of signaling and regulatory networks*, BMC Bioinformatics **7** (2006), no. 56.
Paul Pritchard, *A simple sub-quadratic algorithm for computing the subset partial order*, Inform. Process. Lett. **56** (1995), no. 6, 337–341.
[to3em]{}, *On computing the subset graph of a collection of sets*, J. Algorithms **33** (1999), no. 2, 187–203.
Hong Shen and D. J. Evans, *Fast sequential and parallel algorithms for finding extremal sets*, Int. J. Comput. Math. **61** (1996), no. 3-4, 195–211.
Stefen Schuster, David A. Fell, and Thomas Dandekar, *[A general definition of metabolic pathways useful for systematic organization and analysis of complex metabolic network]{}*, [Nat. Biotechnol.]{} **18** (2000), 326–332.
J[ö]{}rg Stelling, Steffen Klamt, Katja Bettenbrock, Stefan Schuster, and Ernst Dieter Gilles, *Metabolic network structure determines key aspects of functionality and regulations*, Nature **420** (2002), 190–193.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Recent measurements of the $\pi^+\pi^-$ invariant mass distribution at RHIC show a shifted peak for the $\rho$ meson in 100$A$ GeV in peripheral Au + Au and even in $p$ + $p$ collisions. A recent theoretical study based on a picture of in-medium production rates of pions, showed that a large shift could result from a combination of the Boltzmann factor and the collisional broadening of the $\rho$. Here we argue that the two-pion density of states is the appropriate quantity if one assumes a sudden break-up of the system. Methods for calculating the density of states which include Bose effects are derived. The resulting invariant mass distributions are significantly enhanced at lower masses and the $\rho$ peak is shifted downward by $\sim$ 35 MeV.'
author:
- Scott Pratt
- Wolfgang Bauer
title: 'Thermal production of the $\rho$ meson in the $\pi^+\pi^-$ channel '
---
Introduction {#sec:intro}
============
One of the most compelling motivations for studying heavy ion collisions is the prospect for observing the restoration of chiral symmetry. The spontaneous breaking of chiral symmetry is accompanied by the creation of a quark-antiquark condensate whose coupling to nucleons is responsible for the great bulk of the nucleon mass, and is therefore responsible for most of the mass of the universe. The transient nature of the heavy-ion reaction precludes a detailed investigation of all the quasi-particle modes in the highly excited collision volume. However, the $\rho$ meson is unique for it typically decays inside the spatial region where the vacuum structure might undergo novel changes. A neutral $\rho$ decays with 99% probability into a $\pi^+\pi^-$ pair and decays with a small probability into $e^+e^-$ or $\mu^+\mu^-$ pairs. The electromagnetic channels are especially useful because dilepton pairs will largely leave the collision volume unscathed by interactions with the thousands of other constituents. Since the $\rho$ has the same quantum numbers as the photon, the invariant mass spectrum of dileptons is dominated by the $\rho$ for masses between 600 and 800 MeV. Experiments at the CERN SPS for $e^+e^-$ [@ceres] and $\mu^+\mu^-$ [@na50; @na38] suggest that the $\rho$ has either dissolved [@siemenschin] (as would be expected in a quark-gluon plasma), has moved down a few hundred MeV [@brownrho] (due to chiral symmetry restoration), or has been broadened via collisions by many hundreds of MeV [@rappwambaugh].
Recently, the possibility of studying in-medium properties of the $\rho$ meson through the $\pi^+\pi^-$ channel has been discussed [@kolbprakash; @rapp]. Unlike dileptons, pions are not penetrating probes and are likely to re-interact before they escape. Since temperatures fall to near 100 MeV at breakup, where the $\rho/\pi$ ratio falls to a few percent, the chance that a $\pi^+$ is accompanied by a $\pi^-$ that originated from the same $\rho$, rather than a charged pion from a different source, is only a few percent. Thus, a background subtracted invariant mass distribution should have a $\rho$ peak that comprises only a few percent of the integrated distribution.
The STAR collaboration at RHIC has measured such a peak in $pp$ collisions, and for the first time, in peripheral relativistic heavy ion collisions [@star_result]. A surprisingly significant downward shift of the mass was observed even in $pp$ collisions, especially at low $p_t$, and an even larger shift was observed in peripheral Au + Au collisions. Results are not yet available for central collisions where it is more difficult to observe the peak since the $\rho/\pi$ ratio falls. Eventually, the $\rho$ peak should also be measured for central collisions given sufficient statistics.
In references [@kolbprakash] and [@rapp], the mass distribution was predicted by considering the in-medium rate of $\rho$ decays into $\pi^+\pi^-$ pairs, $dN/dM d^3x dt$. This is the same approach as has been applied for dilepton studies. In reference [@rapp], these rates were corrected for collision broadening and for Bose effects. Collision broadening was shown to be particularly important in moving strength to lower-lying masses. However, emission of pions is of a fundamentally different character than that of dileptons. First, the final-state distribution is not necessarily proportional to the decay rate since the decay rate is often balanced by a formation rate of similar magnitude. Secondly, collisional broadening can not be applied in the same manner since measurements are made in the asymptotic state. Finally, the presence of the $\rho$ alters the two-pion scattering partial waves at non-zero separations which should affect the mass distribution. As we will demonstrate, the production-rate calculations of [@kolbprakash] and [@rapp] provide different results than a freeze-out prescription which is governed by the available phase space.
If the last strong interactions felt by the two pions used in the distribution can be considered sufficiently hard to statistically sample the outgoing phase space, the two-pion density of states should govern the invariant mass distribution. Other non-randomizing interactions, identical particle symmetrization and mean field interactions, would then serve to modify the density of states. Although these $\rho$s probably decayed during the breakup stage, which is well below the critical density, the decaying $\rho$ mesons might still sample a region where mean-field effects, i.e., in-medium mass shifts, are not negligible. Since pions are Goldstone bosons, they probably leave the region with their energy and momenta unchanged during their exiting trajectory, and one expects that a modification of the 2-pion invariant mass distribution would reflect the in-medium modifications of the $\rho$ rather than those of the pion. Whereas, this mass shift may be on the order of 100 MeV at high temperature, it is unlikely to be much more than 25 MeV at breakup when densities have fallen well below nuclear density.
It is not the aim of the current paper to model the in-medium mass shift of the $\rho$, but rather to investigate how the invariant-mass distribution would look in a thermal description based entirely on the two-pion density of states, and the associated Bose effects. In the next section, methods for calculating the two-particle density of states are presented along with a comparison with the functional forms one would expect from rate calculations. After convoluting with the Boltzmann weighting, we find that the $\rho$ peak is shifted downward by $\sim 30$ MeV relative to the nominal $\rho$ mass. The shift is due to three factors, the Boltzmann weighting [@barz], the fact that the density of states peaks below the $\rho$ mass, and the inclusion of other partial waves. Bose-Einstein effects also enhance the distribution at lower masses [@lafferty; @opal; @rapp], especially for heavy ion collisions where the pionic phase space filling factors are approaching unity [@bertsch; @prattqm2002]. In section \[sec:bose\] methods are presented for including Bose effects into the two-pion density of states. The resulting mass distribution is strengthened at lower invariant masses, but the peak did not shift appreciably.
Invariant mass distributions from the two-pion density of states {#sec:densityofstates}
================================================================
Since the first measurements of the $\rho$ meson [@abolins; @erwin], the masses and widths have fluctuated by several MeV depending on the analysis. Currently, the Particle Data Group assigns a nominal mass of 771.1 MeV and a width of 149.2 MeV [@pdg], with uncertainties for each number being near 1 MeV. The $\rho$ mass has been determined from a number of means, $e^+e^-\rightarrow \pi^+\pi^-$ reactions, $pp$ collisions, and $\pi p\rightarrow\pi\pi p$ reactions [@protopopescu]. Electro-production of the $\rho$ is complicated by the interference with the $\omega\rightarrow 2\pi$ channel [@omegainterference] which constructively interferes with the $\rho^0$ channel since the electromagnetic coupling violates isospin conservation. Since $pp$ collisions are typically highly inelastic, extracting the $\rho$ mass is complicated by the same factors that complicate the study in a heavy-ion environment. In the $\pi p\rightarrow \pi\pi p$ reaction, the proton is treated as a source of pions which are assumed to scatter elastically with the incoming pions. In fact, $\pi^+\pi^-$ phase shift analyses have been successfully performed. The cross section for $\pi^+\pi^-$ reactions should have a Breit-Wigner form, $$\label{eq:sigma}
\sigma(M)=3\frac{4\pi}{q^2}\frac{(\Pi_I)^2}
{(M^2-M_0^2)^2+(\Pi_I)^2},$$ where $q$ is the momentum of either pion in the center-of-mass frame ($M=2\sqrt{m_\pi^2+q^2}$), and $\Pi_I$ is the imaginary part of the one-loop self energy of the relativistic propagator. $$\begin{aligned}
\label{eq:piself}
\Pi_I&=&\Pi_{I,0}\frac{M_0}{M}\left(\frac{q}{q_0}\right)^3F(q,q_0),\\
\Pi_{I,0}&=&\Gamma_0 M_0.\\\end{aligned}$$ Here, $\Gamma_0$ is the nominal width, and $q_0$ is the momentum required to provide the nominal mass, $M_0$. The last term, $F(q/q_0)$ is a form factor whose exact form is in doubt[@lafferty]. The product $q^2\sigma$ peaks precisely at the nominal mass irrespective of the form factor.
The spectral function of the $\rho$ is related to the imaginary part of the propagator, $$\begin{aligned}
\label{eq:rhospectral}
S_\rho(M)&=&\frac{2M}{\pi}\Im \frac{1}{(M^2-M_0^2)+i\Pi_I},\\
&=&\left(\frac{2M}{\pi\Pi_I}\right)BW(M)\\
BW(M)&=&\frac{(\Pi_I)^2}{(M^2-M_0^2)^2+(\Pi_I)^2}.\end{aligned}$$ Here, $S_\rho$ is usually associated with the number of states available to the $\rho$ with a given mass. The real part of $\Pi$ is being ignored for the current discussion. Since the Breit-Wigner function, $BW(M)$, always peaks at $M=M_0$, and since $\Pi_I/M$ is rapidly growing with $M$ near the $\rho$ mass, the spectral function always peaks below the $\rho$ mass. Setting the form factor in Eq. (\[eq:piself\]) to unity, the peak of the $\rho$ spectral function shifts downward by 5 MeV. Applying some of the different expressions for $\Pi_I$ discussed in [@lafferty] may result in the peak being shifted further downward, perhaps as much as an additional 5 MeV.
The change in the total density of states can be expressed in terms of phase shifts, [@pratt87; @landaulifshitz]: $$\Delta\rho(M)=\frac{1}{\pi}\sum_\ell(2\ell +1)\frac{d\delta_\ell}{dM}.$$ Given the relation between the phase shift and the self energy, one can express $\Delta\rho$ in terms of the self energy, $$\begin{aligned}
\label{eq:rhoofdelta}
\tan\delta&=&\frac{\Pi_I}{M_0^2-M^2},\\
\Delta\rho_{\pi\pi}(M)&=&\frac{3}{\pi}\frac{2M \Pi_I}{(M_0^2-M^2)^2-\Pi_I^2}
\left(1+\frac{M_0^2-M^2}{2\Pi_I M}\frac{d\Pi_I}{dM}\right).\end{aligned}$$ The first term is the spectral function of the $\rho$, which is often associated with the probability of having a $\rho$ meson of mass $M$. Together, the two terms describe the entire correction to the density of states, including the effects of modifying the outgoing partial waves.
![\[fig:spectral\] The spectral function of the $\rho$ (dotted line) is broader than the Breit-Wigner form (dashed line). The total density of states, including effects of modifying the outgoing partial waves, is noticeably shifted to the left relative to the other forms. The difference is especially noticeable at small invariant masses, where the three forms rise proportional to $q^6$, $q^3$ and $q$ respectively.](prattbauer_fig1.eps){width="50.00000%"}
Figure \[fig:spectral\] illustrates the importance of using the correct expression for the density of states. The spectral function of the $\rho$ is peaked below the Breit-Wigner function, and the total density of states is peaked even lower. The difference is especially strong at low invariant masses, as the Breit-Wigner function rises as $q^6$, the $\rho$ spectral function rises as $q^3$ and the pionic density of states rises as $q$. This relative scaling with $q$ would hold for any $p$-wave interaction.
Thus far, the distribution of masses has not incorporated the Boltzmann factor, which should push the peak even lower with the thermal weight, $e^{-M/T}$ [@barz]. More precisely, one needs to integrate over the modes in momentum space due to relativistic effects, $$\Delta \frac{dN_{\pi\pi}}{dMd^3x}=\int \frac{d^3P_\rho}{(2\pi)^3}
e^{-\sqrt{P_\rho^2+M^2}/T} \Delta\rho_{\pi\pi}(M).$$ This should represent the background-subtracted 2-pion invariant-mass distribution. As can be seen in Fig. \[fig:dndm\], the Boltzmann weight pushes the distribution increasingly downward for lower temperatures. The upper panel shows the mass distribution assuming a temperature of 170 MeV, which is a reasonable temperature for thermal models of $pp$ collisions, while the lower panel shows the result for a temperature of 110 MeV, which may be reasonable for the breakup temperature in central heavy ion collisions. Calculations using both the $\rho$ spectral function and the two-pion density of states are displayed to illustrate the importance of choosing the appropriate form for the density of states. The Boltzmann factor greatly magnifies the enhancements at low $M$, to the point that a second peak appears for lower temperature.
![\[fig:dndm\] The thermal mass distribution of $\pi^+\pi^-$ pairs is shown for three calculations, both at $T=110$ MeV and $T=170$ MeV. Using the full density of states as calculated from taking derivatives of phase shifts results in a broader distribution for the $\ell=1$ channel (full line) than using the $\rho$ spectral function (dotted line). Including all $s$, $p$ and $d$ channels (dashed line) provides significant strength at low invariant masses due to the $s$ wave channels and moderate strength at higher masses from the $d$ wave channels.](prattbauer_fig2.eps){width="50.00000%"}
The $\pi^+\pi^-$ density of states is also affected by phase shifts in other channels. For $\pi^+\pi^-$, the $s$-wave channel is split into two isospin components, 2/3 weight for $I=0$ and 1/3 weight for $I=2$. The $I=0$ channel is particularly important as it corresponds to the mythical $\sigma$ meson. Although phase shift analyses do not reveal a sharp peak as in a resonance [@pichowsky; @kaminski; @grayer; @rosselet; @shrinivasan], the phase shifts are considerable, rising steadily from zero at threshold to approximately 90 degrees at $M=2M_K\sim 1$ GeV, where the kaon channel opens. At the two-kaon threshold, the behavior of the phase shifts becomes complicated and an inelastic treatment becomes warranted. Since one uses derivatives of the phase shifts to find the density of states, interpolating data for phase shifts can be dangerous due to noise in the experimentally determined phase shifts. Thus, we apply a simple form that describes the general behavior, $$\delta_{I=0,S=0}=aq+b(M-2m_\pi).$$ The first coefficient $a$ is the scattering length, which is small due to constraints from chiral symmetry. The number varies throughout the literature by several tens of percent. We use the value, $a=0.204/m_\pi$ [@kermani]. The second term does not contribute to the scattering length, as $(M-2m_\pi)\sim q^2$ at low $q$. Choosing $b=9.1\times 10^{-4}$ GeV$^{-1}$ crudely reproduces experimental phase shifts, which are reviewed in [@pichowsky]. Since these phase shifts rise half as far as those in the delta channel, have one third the spin degeneracy, and have a 2/3 weight in the $\pi^+\pi^-$ channel, they are noticeably less important than the $\rho$ channel in affecting the overall density of states, unless one is near the two-pion threshold where $p$-wave interactions vanish.
Other phase shifts also contribute: $(I=2,\ell=0)$, $(I=0,\ell=2)$ and $(I=2,\ell=2)$. Since none of these phase shifts exceed more than a few degrees, they make nearly negligible contributions to the density of states. For the $(I=2,\ell=0)$ channel, we apply an effective range expansion [@losty], $$\cot\delta=\frac{1}{qa}+\frac{1}{2}Rq,$$ where $a=-0.13$ MeV$^{-1}$ and $R=1.0$ MeV$^{-1}$. The $d$ wave is also composed of $I=0$ and $I=2$ pieces. For the $(I=0,\ell=2)$ piece, the data [@estabrooks] are rough, and we make a simple expansion, $$\delta_{I=0,\ell=2}=cq^5,$$ where $c=6.2$ GeV$^{-1}$. The parameter $a$ is uncertain to the 50% level. For the $(I=2,\ell=2)$ partial wave, we use an expansion [@losty], $$\delta_{I=2,\ell=2}=-8.4 q^5+12.5q^6 {\rm ~GeV}^{-1}.$$ None of the these three channels are well understood, but none have a substantial impact at or below the $\rho$ region of invariant mass.
Figure \[fig:dndm\] also shows the invariant mass distribution of a thermal ensemble with $T=110$ MeV using all the $s$, $p$ and $d$ channels. The $s$-wave contributions are non-negligible near the $\rho$ mass, and dominate near the two-pion threshold. The $d$-wave contributions matter only for masses near or greater than 1.0 GeV.
Bose Einstein Corrections {#sec:bose}
=========================
Bose Einstein corrections should preferentially enhance low-mass pairs since low-mass pairs are more likely to include a low-momentum pion. This has been investigated within the context of the $\rho$ peak as well as the influence on $Z$ boson decay modes [@lafferty]. In this section, we present a means to include Bose enhancement effects which are consistent with the statistical picture described in the previous section.
In order to demonstrate Bose enhancement effects, we revert to the fundamental definition of the two-particle density of states. $$\begin{aligned}
\label{eq:rhodef}
\rho(M)&=&\frac{1}{2\pi}\Im {\rm Tr}\frac{1}{M-H+i\epsilon}\\
&=&\frac{1}{2\pi}\Im {\rm Tr}\sum_{n=0}\frac{1}{M-H_0+i\epsilon}
\left(V\frac{1}{M-H_0+i\epsilon}\right)^n\end{aligned}$$ We will work in the two-pion rest frame, so the trace would cover all two-pion states that have total momentum zero. When including Bose effects, one would sum all such two-pion states, plus average over the distribution of other identical particles whose probability of being populated is $$f({\bf q})=\frac{f_0({\bf q})}{1-f_0({\bf q})}=f_0({\bf q})(1+f({\bf q})).$$ Thus, $(1+f)$ can be considered as an Bose enhancement factor while $f_0$ is the phase space filling factor if Bose statistics were neglected.
Using the cyclic property of the trace, Eq. (\[eq:rhodef\]) can be written in terms of a derivative with respect to $M$, $$\begin{aligned}
\rho(M)&=&\rho_0(M)+\frac{1}{2\pi}\Im \frac{d}{dM} {\rm Tr}
\sum_{n=1}\frac{1}{n}\left(\frac{V}{M-H_0+i\epsilon}\right)^n\\
&=&\rho_0(M)+\frac{1}{\pi}\Im \frac{d}{dM} {\rm Tr}
\sum_{n=1}\frac{1}{n}\left(\frac{{\cal P}}{M-H_0}V+i\pi\delta(M-H_0)V\right)^n.\end{aligned}$$ One can expand the $n$ terms and note that the sum includes all possible orderings of $n$ factors where each factor is either the principal value piece, which is real, or the imaginary part which is proportional to the density of states. One could restrict this sum to cover only those terms where the first factor is the imaginary part and multiply by a factor of $n/N_\rho$, where $N_\rho$ is the number of times that $i\pi\delta(M-H_0)$ appears in the term. The sum over $n$ can then be transformed into a sum of all possible numbers of appearances of the real part. $$\begin{aligned}
\rho(M)&=&\rho_0(M)+\frac{1}{2\pi}\Im \frac{d}{dM} {\rm Tr}
\sum_{N_\rho=1}\frac{1}{N_\rho}
\left(i\pi\delta(M-H_0){\mathcal R}\right)^{N_\rho}\\
{\mathcal R}&\equiv&V+V\frac{{\cal P}}{E-H_0}{\mathcal R}.\end{aligned}$$ Here, ${\mathcal R}$ is often referred to as the $R$-matrix. This can be written in terms of a logarithm, $$\Delta\rho(M)=\frac{1}{\pi}\Im\frac{d}{dM} {\rm Tr}
\log\left(\frac{1+i\pi\delta(M-H_0){\mathcal R}}{1-i\pi\delta(M-E){\mathcal R}}
\right).$$ Thus, the density of states is determined completely by a single matrix, $$\tau\equiv \pi\rho_0(M){\mathcal R},$$ which is evaluated only for those states whose energy equals $M$. In a partial-wave basis, $\tau$ is related to the phase shift, $\tau=\tan\delta$. In a plane-wave basis, the matrix $\tau$ links one direction of the relative momentum with another, i.e., the matrix should be written with indices, $\tau_{\Omega_1,\Omega_2}$.
The presence of other particles alters $\tau$. Each matrix element $V$ used to construct $\tau$ is modified by the presence of other particles by the Bose enhancement factor, $$V({\bf q}_1,-{\bf q}_1;{\bf q}_2,-{\bf q}_2)\rightarrow
V({\bf q}_1,-{\bf q}_1;{\bf q}_2,-{\bf q}_2)
\sqrt{(1+f({\bf q}_1))(1+f(-{\bf q}_1))
(1+f({\bf q}_2))(1+f(-{\bf q}_2))}.$$ If the intermediate states contained in the definition of ${\mathcal R}$ are not affected by the phase space density, one can scale $\tau$ in the same manner as $V$. Then, given the fact that each state appears in both the bra and ket, one can modify $\tau$ in a simple manner to account for Bose effects, $$\tau({\bf q}_1,-{\bf q}_1;
{\bf q}_2,-{\bf q}_2)=
\tau_0({\bf q}_1,-{\bf q}_1;
{\bf q}_2,-{\bf q}_2)
(1+f({\bf q}_1))(1+f(-{\bf q}_1)).$$
The density of states is comprised of integrals of a cyclic nature, $I_n$, $$\begin{aligned}
\Delta\rho(M)&=&\frac{1}{\pi}\Im\frac{d}{dM}\sum_{n=1,3,5\cdots} I_n/n\\
\label{eq:indef}
I_n(M)&=&\int \frac{d\Omega_1}{4\pi} \frac{d\Omega_2}{4\pi}\cdots
\frac{d\Omega_n}{4\pi} \tau_0(\Omega_1,\Omega_2)\tau_0(\Omega_2,\Omega_3)
\cdots\tau_0(\Omega_n,\Omega_1)\\
\nonumber
&&(1+f({\bf q}_1))(1+f(-{\bf q}_1))\cdots (1+f({\bf q}_n))(1+f(-{\bf q}_n)).\end{aligned}$$ Unless the momentum of the pair $P=0$, the phase space densities will be sensitive to the direction of the relative momentum $\Omega$.
For a purely $s$-wave interaction, $\tau_0$ has no angular dependence and $I_n$ easily incorporates Bose effects, $$I_n=\left(\tau_0 \int \frac{d\Omega}{4\pi} (1+f({\bf q}))
(1+f(-{\bf q}))\right)^n.$$ The correction to the density of states is then $$\begin{aligned}
\Delta\rho(M)&=&\frac{1}{\pi}\frac{d\tau/dM}{1+\tau^2},\\
\tau&=&\tan\delta\int\frac{d\Omega}{4\pi} (1+f({\bf q}))
(1+f(-{\bf q})),\end{aligned}$$ where $\delta$ is the phase shift as measured in the absence of Bose modifications.
For a $p$ wave interaction, $\tau_0$ has the angular dependence, $$\tau_0(\Omega_1,\Omega_2)=\bar{\tau}_0 \hat{q}_1\cdot\hat{q}_2.$$ By choosing a coordinate system where the $z$ axis is parallel to the total pair momentum, there is reflection symmetry about the $x$, $y$ and $z$ planes. By making use of the identity, $$\begin{aligned}
\int \frac{d\Omega_b}{4\pi} (\vec{A}\cdot\hat{b})
(\hat{b}\cdot\vec{C}) F(\Omega_b)&=&\vec{A}'\cdot\vec{C}\\
A_i^\prime&=&A_iF_i,\\
F_x&=& \int \frac{d\Omega}{4\pi} F(\Omega) \cos^2\phi\sin^2\theta,\\
F_y&=& \int \frac{d\Omega}{4\pi} F(\Omega) \sin^2\phi\sin^2\theta,\\
F_z&=& \int \frac{d\Omega}{4\pi} F(\Omega) \cos^2\theta,\end{aligned}$$ one can iteratively perform the integral in Eq. (\[eq:indef\]). $$\begin{aligned}
I_n(M)&=&\left(\bar{\tau}_0F_x\right)^n
+\left(\bar{\tau}_0F_z\right)^n+\left(\bar{\tau}_0F_z\right)^n,\\\end{aligned}$$
Using $F(\Omega)=(1+f({\bf q}))(1+f(-{\bf q}))$, one can calculate $\Delta\rho$, $$\begin{aligned}
\Delta\rho(M)&=&\frac{1}{\pi}\left(\frac{d\tau_x/dM}{1+\tau_x^2}
+\frac{d\tau_y/dM}{1+\tau_y^2}+\frac{d\tau_z/dM}{1+\tau_z^2}\right),\\
\tau_x&=&\tan\delta\int \frac{d\Omega}{4\pi} (1+f({\bf q}))
(1+f(-{\bf q})) 3\sin^2\phi\sin^2\theta,\\
\tau_y&=&\tan\delta\int \frac{d\Omega}{4\pi} (1+f({\bf q}))
(1+f(-{\bf q})) 3\cos^2\phi\sin^2\theta,\\
\tau_z&=&\tan\delta\int \frac{d\Omega}{4\pi} (1+f({\bf q}))
(1+f(-{\bf q})) 3\cos^2\theta.\end{aligned}$$ The calculation of $\Delta\rho(M)$ must be repeated for each value of the total momentum since $f({\bf q})$, which is defined in the two-pion rest frame, changes when the total momentum is changed.
The $p$-wave and $s$-wave corrections to the density of states do not interfere with one another since they have opposite parities and $(1+f({\bf
q}))(1+f(-{\bf q}))$ has even parity. However, calculation of the $\ell=2$ contributions would be complicated by the fact that the elliptical distortion of the Bose enhancement factors would mix the $\ell=0$ and $\ell=2$ contributions. For the calculations here, the $\ell=2$ contributions were calculated by assuming that the Bose enhancement factors were independent of $\Omega$, then using enhancement factors which had been averaged over all directions of $\Omega$.
The mean Bose enhancement, $\langle (1+f_1)(1+f_2)\rangle$, is shown as a function of the invariant mass and momentum of the decaying $\rho^0$ in Fig. \[fig:meanenhancement\] assuming a breakup temperature of 110 MeV and an effective chemical potential of 90 MeV. The enhancement has been averaged over the directions of the relative momentum. The enhancement is largest for low-momentum, low-mass pairs since these pions most strongly sample the region of high phase space density. For higher invariant masses, the Bose enhancement is actually stronger for higher pair momenta, as it allows one of the outgoing pions to have low $p_t$ and sample the high phase-space density region. From viewing Fig. \[fig:meanenhancement\], it is clear that the Bose modifications to the invariant mass distribution would be more acute if experiments were to focus on pion pairs with low total momentum.
![\[fig:meanenhancement\] The mean values of $\langle (1+f_1)(1+f_2)\rangle$ are shown as a function of the invariant mass and total momentum of the outgoing pion pair. The values have been averaged over all directions of the relative momentum. The enhancement factors exceed 2.0 for a low values of the pair momentum and invariant mass.](prattbauer_fig3.eps){width="50.00000%"}
![\[fig:dndm\_bose\] The mass distribution is shown with (solid line) and without (dashed line) Bose effects for $T=110$ MeV, $\mu=90$ MeV. Bose effects enhance the probability of producing low invariant-mass pairs since they are more likely to have low momenta and stronger Bose enhancement factors.](prattbauer_fig4.eps){width="50.00000%"}
Bose corrected densities of states are shown in Fig. \[fig:dndm\_bose\] for $T=110$ MeV and $\mu=90$ MeV. A non-zero chemical potential was used to account for the relative overpopulation of pionic phase space which may result from rapid cooling [@gong] and might be magnified by the effects of chiral symmetry restoration [@haglin]. Analyses of $\pi\pi$ correlations from RHIC indeed point to high phase space densities [@bertsch; @prattqm2002], especially for central collisions of heavy ions. As expected, lower-mass states were more enhanced by Bose effects. Since the density of states was proportional to the derivative of $(\tan\delta\langle(1+f)(1+f')\rangle$, and since the averaged phase space filling factors generally fall as $M$ increases, the density of states was less enhanced for intermediate masses as compared to the no-Bose case. The peak of the distribution shifted downward by only one MeV after the inclusion of Bose effects.
Although the position of the peak was not much affected by Bose effects shown in Fig. \[fig:dndm\_bose\], Bose effects led to a near doubling of the distribution at low masses. These effects are most important at low $p_t$ where the phase space densities are higher. In $pp$ collisions, a movement of the $\rho$ peak was observed for low $p_t$ pairs[@aguilarbenitez; @star_result] which is suggestive of Bose effects. However, at the $\rho$ peak each pion has a relative momentum of $\sim 300$ MeV/c and will largely sample phase-space regions with moderate to low phase space densities. Although the invariant mass distribution is mainly altered at invariant masses below the $\rho$ peak, Bose effects should contribute to washing out the peak by increasing the declining background to the peak structure in Fig. \[fig:dndm\_bose\].
summary
=======
Our principal finding is that the $\rho$ peak in the $\pi^+\pi^-$ invariant mass distribution should be approximately 35 MeV lower than the nominal $\rho$ mass, if one accepts the scenario of a sudden breakup that thermally samples the two-pion density of states. The shift was the result of convoluting the density of states which is shifted by $\sim$ 10 MeV below the $\rho$ mass with the Boltzmann factor. Given the extra cooling inherent to heavy ion collisions, the breakup temperature is probably near 110 MeV, well below the characteristic temperatures used to describe $pp$ collisions. This low temperature is responsible for the additional downward shift of the peak in heavy ion collisions. In addition to the shift of the peak, the distribution showed significant additional strength at invariant masses near the two-pion threshold. This additional strength hinged on using the correct expressions for the density of states, especially in the $\ell=0$ channels. Although the position of the peak was not much affected by Bose effects, Bose effects led to a near doubling of the distribution at low masses.
The thermal model presented here rests critically on a pair of assumptions. First, we have assumed that the breakup is sudden, i.e., the last strong interaction experienced by the particles samples the outgoing two-particle phase space. Indeed, interferometric measurements do suggest a sudden breakup [@starhbt; @phenixhbt; @prattqm2002]. If emission were gradual, e.g., surface evaporation, this picture would be invalid. An appropriate treatment of the surface would include the dynamics of surface penetration and absorption and might include collisional broadening. For instance, spectral lines in stars are affected by collisional broadening. The “truth” of the breakup at RHIC probably has elements of both volume-like breakup and surface-like evaporation. Thus, the effect of collisions, which played a pivotal role in moving the distribution downward in [@rapp], requires more study.
The second assumption inherent to these calculations is related to the neglect of finite-size effects. The enhancement factors applied to small-angle correlation studies are usually based on the outgoing wave function, $|\phi({\bf q},{\bf r})|^2$ [@bauergelbkepratt; @heinzjacak]. For large sources, there is a straight-forward correspondence between the integrated wave functions and the phase shifts [@boal], [@corrtail], $$\frac{1}{\pi}\sum_\ell (2\ell+1)\frac{d\delta_\ell}{dE}
=\frac{qM}{8\pi^2}\int d^3r
\left( |\phi({\bf q},{\bf r})|^2-|\phi_0({\bf q},{\bf r})|^2 \right).$$ Since the modification of the wave function is mostly confined to a region where $qR<\pi$, one expects that the treatments shown here should work well for $q>100$ MeV/c, or for masses greater than 400 MeV. For masses near threshold, a different approach, based on the actual scattered wave functions, would be warranted. Such an approach could also incorporate the effects of the Coulomb interaction between pions.
Finally, it should be emphasized that other correlations, besides those resulting from the change in the two-pion density of states, will play a role in any experimental measurement. Experimental analyses are typically based on a like-sign subtraction. This should eliminate global correlations such as elliptic flow which correlate same-sign and opposite-sign pairs equally. However, any correlation based on charge conservation should survive the subtraction [@bassdanpratt; @starbalance]. For every $\pi^+$, there is a $\sim 75$% chance that local charge conservation will result in an extra $\pi^-$ being emitted with a similar rapidity. This should provide a bump in the like-sign-subtracted invariant-mass distribution that peaks for masses near 400 MeV. The ratio of the $\rho$ peak in the like-sign subtracted distribution to the bump from charge correlation is approximately determined by the chance that a given $\pi^+$ had its last interaction with other hadrons through the decay of a $\rho^0$. This ratio should be smaller for central collisions since the breakup temperature is lower which reduces the $\rho/\pi$ ratio.
The in-medium mass of the $\rho$ might be altered by $\sim$20 MeV at breakup. Given that this peak is also spread out and distorted as shown in the calculations presented here, it is certainly challenging to isolate the contribution from the $\rho$ and to quote a peak height to a better accuracy than 20 MeV. Upcoming runs at RHIC may increase the statistics by more than an order of magnitude. Thus, we believe that there remains a good chance that the $\rho$ can be studied in detail, even in the central collision Au+Au environment.
Finally, we compare the experimentally observed mass shifts to results of our model. In reference [@star_result] it was reported that the $\rho$ shift downward in $pp$ collisions by $\sim 20$ MeV at higher $p_t$ while shifting downward by $\sim 45 MeV$ at low $p_t$. The shift appeared to be 5 to 10 MeV larger for high-muliplicity $pp$ collisions and perhaps another 5 MeV lower for peripheral Au+Au reactions. Similar behavior for $pp$ collisions had been reported for $\sqrt{s}=27$ GeV $pp$ collisions [@aguilarbenitez]. The shifts that we extracted were as large 35 MeV, but these calculations assumed a lower temperature, 110 MeV, and a higher effective chemical potential, 90 MeV, than would be appropriate for $pp$ phenomenology. For a temperature of 170 MeV, and zero chemical potential, the shift was in the range of 20 MeV, a somewhat smaller shift than what was observed by STAR. It appears that the experimental mass shift is 10 to 20 MeV stronger than what we would expect from our approach. But, before this discrepancy can be attributed to novel in-media phenomena, i.e., a mass shift of the $\rho$, it should be stressed that systematic uncertainties described in [@star_result] are of the order of 10 MeV. This problem would be served well by both a higher statistics experimental analysis and a more detailed theoretical modeling. An improved calculation would consider finite-size effects, the influence of other resonances, and the effects of experimental acceptances and efficiencies.
This work was supported by the U.S. National Science Foundation, Grant No. PHY-02-45009 and by the U.S. Department of Energy, Grant No. DE-FG02-03ER41259.
[99]{} D. Adamova et al., http://arXiv.org, nucl-ex/0209024 (2002). NA50 Collab., M.C. Abreu et al., European Phys. J. C[**13**]{}, 69 (2000). NA38 Collab., M.C. Abreu et al., Phys. Lett B[**368**]{}, 239 (1996). P.J. Siemens and S.A. Chin, Phys. Rev. Lett. [**55**]{}, 1266 (1985). G.E. Brown and M. Rho, Phys. Rep. [**363**]{}, 85 (2002). R. Rapp and J. Wambach, Adv. Nucl. Phys. [**25**]{} (2000). P.F. Kolb and M. Prakash, Phys. Rev. C[**67**]{}, 044902 (2003). R. Rapp, http://arXiv.org, hep-ph/0305011 (2003). J. Adams et al., www.arXiv.org, nucl-ex/0307023 (2003). H.W. Barz, G. Bertsch, B.L. Friman, H. Schultz and S. Boggs, Phys. Lett. B[**265**]{}, 219 (1991). G.D. Lafferty, Z. Phys. C[**60**]{}, 659 (1993). P.D. Acton et al., Phys. Lett. B[**267**]{}, 143 (1991). G.F. Bertsch, Phys. Rev. Lett. [**72**]{}, 2349 (1994); [*ibid.*]{} [**77**]{}, 789(E) (1996). S. Pratt, Nucl. Phys. A[**715**]{}, 389c (2003). M. Abolins, R.L. Lander, W. Mehlhop, N. Xuong and P.M. Yager , Phys. Rev. Lett. [**11**]{}, 381 (1963). A.R. Erwin, R. March, W.D. Walker and E. West, Phys. Rev. Lett. [**6**]{}, 628 (1961). Particle Data Group, K. Hagiwara et al., Phys. Rev. D[**66**]{}, 010001 (2002). S.D. Protoposecu, M. Alston-Garnjost, A. Barbaro-Galtieri, S.M. Flatt, J.H. Friedman, T.A. Lasinski, G.R. Lynch, M.S. Rabin, and F.T. Solmitz, Phys. Rev. D.[**7**]{}, 1279 (1973). P.S. Biggs et al., Phys. Rev. Lett. [**24**]{}, 1201 (1970). S. Pratt, P.J. Siemens and Q.N. Usmani, Phys. Lett. [**B189**]{}, 1 (1987). L.D. Landau and E.M. Lifshitz, , Reed Educational and Professional Publishing, Oxford, 230 (1999). M.A. Pichowsky, A. Szczepaniak and J.T. Londergan, Phys.Rev. D[**64**]{}, 036009 (2001). R. Kaminiski, L. Lesniak and K. Rybicki, Acta. Phys. Polon. B[**31**]{}, 895 (2000). G. Grayer, et al., Nucl. Phys. B[**75**]{}, 189 (1974). L. Rosselet, et al., Phys. Rev. D[**15**]{}, 574 (1977). V. Shrinivasan et al., Phys. Rev. D[**12**]{}, 681 (1975). M. Kermani et al., Phys. Rev. C[**58**]{}, 3431 (1998). M.J. Losty, V. Chaloupka, A. Ferrando, L. Montanet, E. Paul, D. Yaffe, A. Zieminski, J. Alitti, B. Gandois and J. Louie, Nucl. Phys. B[**69**]{}, 185 (1974). P. Estabrooks and A.D. Martin, Nucl. Phys. B[**79**]{}, 301 (1974). C. Greiner, C. Gong and B. Müller, Phys. Lett. B[**16**]{}, 226 (1993). S. Pratt and K. Haglin, Phys. Rev. C[**59**]{}, 3304 (1999). M. Aguilar-Benitez et al., Z. Phys. C [**50**]{}, 405 (1991). C. Adler et al. Phys. Rev. Lett. 87 , 082301 (2001). K. Adcox, et al., Phys. Rev. Lett. [**88**]{}, 192302 (2002). W. Bauer, C.K. Gelbke and S. Pratt, Ann. Rev. of Nucl. and Part. Science [**42**]{}, 77 (1992). U. Heinz and B. Jacak, Ann. Rev. Nucl. Sci. [**49**]{}, 529 (1999). B. Jennings, D. Boal and J. Shillcock, Phys. Rev. C[**33**]{}, 1303 (1986). S. Pratt and S. Petriconi, http://arXiv.org, nucl-th/0305018 (2003). S.A. Bass, P. Danielewicz, and S. Pratt, Phys. Rev. Lett. [**85**]{}, 2689 (2000). J. Adams et al., Phys. Rev. Lett. [**90**]{}, 172301 (2003).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Xiaodong Cao and Wenguang Zhai
title: 'Conditional Results for a Class of Arithmetic Functions: a variant of H. L. Montgomery and R. C. Vaughan’s method '
---
[**Abstract.**]{} Let $a, b ,c $ and $k$ be positive integers such that $1\leq a\leq
b,a<c<2(a+b), c\ne b$ and $(a,b,c)=1$. Define the arithmetic function $f_k(a,b;c;n)$ by $$\sum_{n=1}^{\infty}\frac{f_k(a,b;c;n)}{n^s}=\frac{\zeta (as)\zeta
(bs)}{\zeta^k(cs)}, \Re s >1.$$ Let $\Delta_k(a,b;c;x)$ denote the error term of the summatory function of the function $f_k(a,b;c;n).$ IN this paper we shall give two expressions of $\Delta_k(a,b;c;x)$. As applications, we study the so-called $(l,r)$-integers, the generalized square-full integers, the $e-r$-free integers, the divisor problem over $r$-free integers, the $e$-square-free integers. An important tool is a generalization of a method of H. L. Montgomery and R. C. Vaughan.
Introduction and main results
=============================
W. G. Nowak[@nowak], M. Küleitner and W. G. Nowak [@kn] studied a class of very general arithmetic function $a(n),$ which possess a generating Dirichlet series $$\sum_{n=1}^{\infty}\frac {a(n)}{n^s}=\frac{f_1(m_1s)\cdots f_K(m_Ks)}{g_1(n_1s)\cdots g_J(n_Js)} h(s),$$ where $f_k$ and $g_j$ are certain generalizations of Riemann zeta-function, $m_1\le \cdots \le m_K$ and $n_1\le \cdots \le n_J$ are natural numbers, and $h(s)$ is a good function which is regular and bounded in a sufficiently large half-plane. People are usually concerned with the summatory function $\sum_{n\leq x}a(n),$ especially sharp upper and lower bounds of its error term. The above two papers give an upper bound and a lower bound for $a(n)$ in a very general sense. Some special cases are also studied, see for example, [@ba1; @ba2; @ku; @wu2].
The aim of this paper is to study a special case of $a(n),$ in which case we can get better upper results. Let $a, b ,c $ and $k$ be positive integers such that $1\leq a\leq
b,a<c<2(a+b), c\ne b$ and $(a,b,c)=1$. Let $\zeta(s)$ denote the Riemann zeta-function. The arithmetic function $f_k(a,b;c;n)$ is defined by $$\begin{aligned}
\sum_{n=1}^{\infty}\frac{f_k(a,b;c;n)}{n^s}=\frac{\zeta (as)\zeta
(bs)}{\zeta^k(cs)}, \Re s >1.\end{aligned}$$ In this paper we are concern with the summatory function $$\begin{aligned}
A_k(a,b;c;x):=\sum_{n\le x}f_k(a,b;c;n),\ x\geq 2.\end{aligned}$$ The expected asymptotic formula of $A_k(a,b;c;x)$ is of the form $$\begin{aligned}
A_k(a,b;c;x) =\frac {\zeta (\frac ba)}{\zeta^k
(\frac ca)}x^{\frac 1a}+\frac {\zeta (\frac ab)}{\zeta^k (\frac
cb)}x^{\frac 1b}+\Delta_k(a,b;c;x)\end{aligned}$$ when $a\ne b.$ When $a = b$, then an appropriate limit should be taken in the above formula. As usual, $\Delta_k(a,b;c;x)$ is called the error term of the function $A_k(a,b;c;x).$ For convenience, we also use notations $A(a,b;c;x)$, $\Delta(a,b;c;x)$ to denote $A_1(a,b;c;x)$, $\Delta_1(a,b;c;x)$, respectively.
By Theorem 2 of M. Küleitner and W. G. Nowak[@kn] or Theorem 3 of W. G. Nowak[@nowak], it is easy to prove that $$\begin{aligned}
\Delta_k(a,b;c;x):=\Omega\left(x^{\max (\frac
{1}{2(a+b)},\frac{1}{2c})}\right).\end{aligned}$$ Hence one may conjecture that $$\begin{aligned}
\Delta_k(a,b;c;x):=O\left(x^{\max (\frac
{1}{2(a+b)},\frac{1}{2c})+\varepsilon}\right).\end{aligned}$$
Many special cases of the function $f_k(a,b;c;n)$ have been extensively studied in number theory. We take some examples.
\(1) The case $(a,b,c,k)=(2,3,6,1)$ is the well-known square-full number problem(see [@ca; @co2; @cd; @ns; @wu2]). Suppose $a\nmid b,$ the cases $(a,b,c,k)=(a,b,2b,1)$ or $(a,b,c,k)=(b,a,2b,1)$ are studied for the the generalized square-full number problem(see [@co1; @su2] ).
\(2) Suppose $r\geq 2$ is a fixed integer, the case $(a,b,c,k)=(1,1,r,1)$ is the $r$-free divisor problem(see [@ba1; @ba2; @fz1; @fz2; @ku]).
\(3) Suppose $1<r<l$ are fixed integers, the case $(a,b,c,k)=(1,r,l,1)$ corresponds to the distribution of the so-called $(r,l)$-integers(see [@co4; @sh; @ss1; @ss2; @su3; @ya]).
\(4) Suppose $r\geq 1$ is a fixed integer, the case $(a,b,c,k)=(1,2^r+1,2^r,1)$ corresponds to the the distribution of the so-called e-$r$-free integers(see [@cz2; @su1; @to1; @to2; @wu1]).
\(5) Suppose $r\geq 2$ is a fixed integer, the case $(a,b,c,k)=(1,1,r,r+1)$ corresponds to the Dirichlet divisor problem over the set of $r$-free integers(see [@fen; @re]).
From the right-hand side of (1.1) it is easily seen that the unconditional asymptotic formula we could possibly prove at present is at most $$\begin{aligned}
A_k(a,b;c;x) =\frac {\zeta (\frac ba)}{\zeta^k
(\frac ca)}x^{\frac 1a}+\frac {\zeta (\frac ab)}{\zeta^k (\frac
cb)}x^{\frac 1b}+O(x^{\frac 1c}\exp(-A (\log x)^{\frac 35}(\log
\log x)^{-\frac 15})),\end{aligned}$$ where $A>0$ is some absolute constant. Now define $\theta_k(a,b;c)$ denote the infimum of $\alpha_k(a,b;c)$ such that the estimate $$\Delta_k(a,b;c;x)\ll x^{\alpha_k(a,b;c)+\varepsilon}$$ holds.
From (1.1) we also see that the function $f_k(a,b;c;n)$ is related to the divisor function $d(a,b;n):=\sum_{n=m_1^am_2^b}1,$ which satisfies $$\sum_{n=1}^\infty \frac{d(a,b;n)}{n^s}=\zeta(as)\zeta(bs), \Re
s>1.$$ We write $$\begin{aligned}
D(a,b;x):= \sum_{n\leq x}d(a,b;n)=\zeta (\frac ba)x^{\frac
1a}+\zeta (\frac ab)x^{\frac 1b} +\Delta(a,b;x)\end{aligned}$$ for $a\ne b$, and let $0<\alpha(a,b)<1/(a+b)$ be a real number such that the estimate $$\Delta(a,b;x)\ll x^{\alpha(a,b)+\varepsilon}$$ holds.
As usual, $\Delta(a,b;x)$ is called the error term of the asymmetric two-dimensional divisor problems. For the history and classical results of $\Delta(a,b;x),$ see for example[@iv; @ikkn; @kr; @mn; @ri; @zc1].
If $\alpha(a,b)\geq 1/c,$ then by the convolution approach we get easily $\theta_k(a,b;c)\leq \alpha(a,b).$ Thus the difficulty of the evaluation of the function $A_k(a,b;c;x)$ is basically the difficulty of the evaluation of the function $D(a,b;x).$ Without the loss of generality, we always suppose later that $\alpha(a,b)< 1/c.$
The exponent $1/c$ in the error term in (1.6) is closely related to the distribution of the non-trivial zeros of $\zeta(s).$ People usually assume the Riemann-hypothesis (RH) to reduce the constant $1/c.$ See for example, [@ba1; @ns; @ss2; @su2]. From now on, we always suppose that RH holds.
In 1981, Montgomery and Vaughan[@mv] developed a new ingenious method to treat the distribution of $r$-free integers, which was also used by many other authors, see for example, Baker[@ba1], Nowak and Schmeier[@ns], Nowak[@nowak] etc. However, as W. G. Nowak and M. Schmeier[@ns] observed in subsection (The Divisor Problem For $(l,r)$-Integers) that: The $r=2, l=3$ is some exceptional. That is, in some cases, by Montgomery-Vaughan’s method one could not get directly better estimates than the usual approach.
The main aim of this paper is to find a suitable expression of the error term in (1.3) for every case. We have to consider two different cases: $c>b$ and $c<b$. For these two cases, we have to use different convolution approaches.
Consider first $c>b.$ We define the function $\mu_k$ by $$\begin{aligned}
\sum_{n=1}^{\infty}\frac {\mu_k(n)}{n^s}=\frac {1}{\zeta ^k(s)},
\Re s>1.\end{aligned}$$ (Also see Titchmarsh[@ti], page 165-166.) Clearly $\mu_1$ is the well-known Möbius function $\mu$. Then Theorem 2 of Nowak[@nowak] essentially implies the following theorem.
(W. G. Nowak) Let $x\ge 2,$ $a\le b<c<2(a+b)$ and $\Delta_k(a,b;c;x)$ be defined by (1.3). If the RH is true, then for any $ 1\le y<
x^{\frac 1c}$ $$\begin{aligned}
\Delta_k(a,b;c;x)=\sum_{l\le y}\mu_k (l)\Delta\left(a,b;\frac
{x}{l^c}\right)+O\left(x^{\frac {1}{2a}+\varepsilon}y^{\frac
12-\frac {c}{2a}}+x^\varepsilon\right).\end{aligned}$$
Taking $y=x^{\frac {1-2a\alpha (a,b)}{a+c-2ac\alpha (a,b)}}$ and noting $\mu_{k} (l)\ll l^\varepsilon$, we get
Suppose RH. If $a\leq b<c<2(a+b)$ and $\alpha{(a,b)}<\frac 1c$, then $$\begin{aligned}
\theta_k(a,b;c)\leq \frac {1-a\alpha (a,b)}{a+c-2ac\alpha (a,b)}.\end{aligned}$$
[**Remark 1.1.**]{} The proof of Theorem 1 is based on a classical idea of Montgomery and Vaughan[@mv] and the Dirichlet convolution $$f_k(a,b;c;n)=\sum_{n=m_1m_2^c}d(a,b;m_1) \mu_k(m_2).$$
[**Remark 1.2.**]{} Very fortunately, for arithmetical function $\mu_k $ we have an analogue of the well-known Vaughan’s identity of Möbius function $\mu$(see Lemma 4.1 below, in fact this is the third useful Vaughan-type’s identity except the well-known von Manlgoldt function $\Lambda$ and Möbius function $\mu$ ). By the method of exponential sums, one could improve the result in Corollary 1.1. For the related works, we refer to papers [@ba1; @ba2; @ca; @ku; @wu2]. However, this is not the main aim of the present paper.
We now turn to the case $c<b.$ Some examples of this type can be found in [@ss1; @ss2; @su2; @su3]. In this case, we hope to find an estimate of the form $\Delta_k(a,b;c;x)\ll x^{\alpha}\ (\alpha
<1/b)$ such that the second main term $\frac {\zeta (\frac
ab)}{\zeta^k (\frac cb)}x^{\frac 1b}$ becomes a real main term. We can also use the convolution (1.13) as our first choice to study $A_k(a,b;c;x)$. Actually it is easy to check that Theorem 1 also holds for $b/2<c<b.$ But when $a+c\le b,$ we have checked that it is very difficult to prove $\theta_k(a,b;c)<1/b$ via Theorem 1 directly(Also see page 9, section 5 in [@cz2]). In order to overcome this difficulty, we choose another convolution approach.
Let the arithmetic function $u_k(a;c;n)$ be defined by $$\begin{aligned}
u_k(a;c;n):=\sum_{n=l^ad^c}\mu_k(d),\end{aligned}$$ which satisfies $$\begin{aligned}
\sum_{n=1}^{\infty}\frac{u_k(a;c;n)}{n^s}=\frac
{\zeta(as)}{\zeta^k(cs)},\Re s>1.\end{aligned}$$ We note that when $a=k=1,$ the function $u_k(a;c;n)$ is just the characteristic function of the set of the $c$-free integers. Hence we can write $$f_k(a,b;c;n)=\sum_{n=m_1m_2^b}u_k(a;c;m_1).$$
The function $u_k(a;c;n)$ plays an important role in this case. Define $$\begin{aligned}
\Delta_k(a;c;x):=\sum_{n\le x}u_k(a;c;n)-\frac {x^{\frac
1a}}{\zeta^k(\frac ca)}:=A_k(a;c;x)-\frac {x^{\frac
1a}}{\zeta^k(\frac ca)}.\end{aligned}$$
For $\Delta_k(a;c;x)$, similar to Theorem 1 we also have
Let $x\ge 2$, $a<c$ and $\Delta_k(a;c;x)$ be defined by (1.17). If the RH is true, then for any $ 1\le y< x^{\frac 1c}$ we have $$\begin{aligned}
\Delta_k(a;c;x)=\sum_{l\le y}\mu_k (l)\psi\left((\frac
{x}{l^c})^{\frac 1a}\right)+O\left(x^{\frac
{1}{2a}+\varepsilon}y^{\frac 12-\frac {c}{2a}}+y^{\frac
12+\varepsilon}\right).\end{aligned}$$
On taking $y=x^{\frac {1}{a+c}}$ in Theorem 2 we get immediately the following
Under the conditions of Theorem 2, we have $$\begin{aligned}
\Delta_k(a;c;x)\ll x^{\frac {1}{a+c}+\varepsilon}.\end{aligned}$$
Now we state our main result for the case $c<b$, which improves Theorem 1 in the case $a<c<b<2c$.
Suppose RH is true. Let $x\ge 2$, $a<c<b<2c$, $\Delta_k(a,b;c;x)$ and $\Delta_k(a,c;x)$ be defined by (1.3) and (1.17) respectively. Suppose $\Delta_k(a;c;x)\ll
x^{\alpha_k{(a;c)}+\varepsilon}$ such that $\alpha_k{(a;c)}<1/b$(a natural restriction). Then for any $1\le y<x^{\frac 1b}$ we have $$\begin{aligned}
\Delta_k(a,b;c;x)&=\sum_{d\le y}\Delta_k\left(a;c;\frac
{x}{d^b}\right)- \sum_{m\le \frac
{x}{y^b}}u_k(a;c;m)\psi\left((\frac x m)^{\frac
1b}\right)\\
&\ \ \ +O\left(x^{\frac {1}{2c}}y^{1-\frac
{b}{2c}}+(xy^{-b})^{\alpha_k{(a;c)}}+x^{\frac 1a}y^{-1-\frac
ba}\right).\nonumber\end{aligned}$$
Under the conditions of Theorem 3, we have $$\begin{aligned}
\Delta_k(a,b;c;x)\ll x^{\frac {1}{a+b-ab\alpha_k
(a;c)}+\varepsilon}.\end{aligned}$$
[**Remark 1.3.**]{} Since we use different convolution approaches in Theorem 1 ($b<c$) and Theorem 3($b>c$), the exponential sums appeared in these two theorems are also different. Hence we have to use different ways to estimate exponential sums in these two theorems. We note that Corollary 3 implies $\frac {1}{a+b-ab\alpha_k
(a;c)} <\frac 1b$, hence in the asymptotic formula (1.3), the second main term becomes a real main term.
[**Remark 1.4.**]{} All corollaries above can be further improved by more precise estimate for the exponential sums involved(e.g., see [@ba2; @bp; @cz1; @fi; @hu1; @mn; @rs; @sw1; @wu2]).
The organization of this paper is as follows. In section 2 we shall give short proofs of Theorem 1 and Theorem 2. The proof of Theorem 3 will be given in section 3. In section 4, by the well-known Heath-Brown’s method we shall further improve Corollary 2 and obtain a non-trivial estimate for $\Delta_k(a;c;x) (k=1,2)$, and then give some of its applications to problems related to the exponential convolution. In section 5 we discuss some applications of Corollary 1.1 and Corollary 1.3. Finally, in section 6 we give an example to explain how to get a sharper upper bound by Theorem 3.
[**Notation.**]{} Throughout this paper $ \varepsilon$ denotes a fixed positive constant, not necessarily the same in all occurrences. As usual, let $\tau(n)$ and $\omega(n)$ denote the divisor function , and the number of prime factors of $n$, respectively. We also use $\tau_k(n)$ to denote the number of decompositions of $n$ into $k$ factors, and let $\tau_1(n)=1$. Let $q_r(n)$ denote the characteristic function of the set of $r$-free integers. $x> 1$ is real, $\mathcal{L}=\log x$, $\{t\}$ denotes the fractional part of $t, \psi(t)=\{t\}-1/2, \Vert t\Vert=\min
(\{t\},1-\{t\})$. We let $e(t)=\exp (2\pi it)$ and $\delta
(x)=\exp(-A (\log x)^{\frac 35}(\log \log x)^{-\frac 15})$ for some fixed constant $A>0$. $m\sim M$ means that $cM<m<CM$ for some constants $0<c<C$.
The proof of Theorem 1 and 2
============================
Let $A(s)=\sum_{n=1}^{\infty}a(n)n^{-s}$ converge absolutely for $\Re s=\sigma>\sigma_a$, and let functions $H(u)$ and $B(u)$ be monotonically increasing such that $$|a(n)|\le H(n),\ \ (n\ge
1),$$ $$\sum_{n=1}^{\infty}|a(n)|n^{-\sigma}\le B(\sigma),\ \ \sigma>\sigma_a.$$ If $s_0=\sigma_0+it_0, b_0>\sigma_a, b_0\ge b>0,
b_0\ge\sigma_0+b>\sigma_a, T\ge
1$ and $x\ge 1$, then for $x\notin \mathbb{N}$ $$\begin{aligned}
\sum_{n\le x}a(n)n^{-s_0}=&\frac {1}{2\pi
i}\int_{b-iT}^{b+iT}A(s_0+s)\frac{x^s}{s}\mathrm{d}s+O\left(\frac{x^bB(b+\sigma_0)}{T}\right)\\
&+O\left(x^{1-\sigma_0}H(2x)\min (1,\frac {\log
x}{T})\right)+O\left( x^{-\sigma_0}H(N)\min (1,\frac {x}{T\Vert
x\Vert})\right).\end{aligned}$$
This lemma is the well-known Perron’s formula, for example, see Theorem 2 of page 98 in Pan[@pa].
Let $y\ge 1$ and define $$\begin{aligned}
g_y(s):=\sum_{n>y}\frac {\mu_k(n)}{n^s}, \ \ \ \Re s>1.\end{aligned}$$
Suppose RH is true, then $g_y(s)$ can be continued analytically to $\Re s=\sigma>\frac
12+\varepsilon$, and we have uniformly for $\sigma$ that $$\begin{aligned}
g_y(s)\ll y^{\frac 12 -\sigma + \varepsilon}(|t| +1)^{\varepsilon}, \sigma\ge
\frac 12+ \varepsilon.\end{aligned}$$
This lemma follows from Lemma 3 of Nowak[@nowak] immediately. .
[**The Proof of Theorem 1 and 2.**]{} Let $\delta=\frac{\varepsilon}{10}$ and $1\le y< x^{\frac 1c}$. We begin the proof of Theorem 1 in the same way as that of Theorem 1 in Montgomery and Vaughan[@mv]. Here we only give the details of our proof for the case $a<b$. The proof for the case $a=b$ is similar.
Define $$\begin{aligned}
f_{1,y}(n):=\sum_{ \stackrel { l^cm=n}{ l\le y}
} \mu_k(l)d(a,b;m),
f_{2,y}(n):=\sum_{ \stackrel { l^cm=n}{ l> y}
} \mu_k(l)d(a,b;m),\end{aligned}$$ hence $$\begin{aligned}
f_k(a,b;c;n)=\sum_{ l^cm=n}
\mu_k(l)d(a,b;m)=f_{1,y}(n)+f_{2,y}(n).\end{aligned}$$ We now write $A_k(a,b;x)$ in the form $$\begin{aligned}
A_k(a,b;c;x):=\sum_{n\le x}f_k(a,b;c;n)=S_1(x)+S_2(x),\end{aligned}$$ where $$\begin{aligned}
S_1(x)=\sum_{n\le x}f_{1,y}(n),\end{aligned}$$ and $$\begin{aligned}
S_2(x)=\sum_{n\le x}f_{2,y}(n).\end{aligned}$$ We first evaluate $S_1(x)$. From (1.8) we get
$$\begin{aligned}
S_1(x)&= \sum_{ \stackrel { l^cm\le x}{ l\le y}
} \mu_k(l)d(a,b;m)=\sum_{ l\le y
} \mu_k(l)D\left(a,b;\frac {x}{l^c}\right)\\
&=\sum_{l\le y}\mu_k (l)\left(\frac {\zeta(\frac ba)}{l^{\frac
ca}}x^{\frac 1a}+\frac
{\zeta(\frac ab)}{l^{\frac cb}}x^{\frac 1b}+\Delta\left(a,b;\frac {x}{l^c}\right)\right)\nonumber\\
&=\zeta (\frac ba)x^{\frac 1a}\sum_{l\le y}\frac
{\mu_k(l)}{l^{\frac ca}}+\zeta (\frac ab)x^{\frac 1b}\sum_{l\le
y}\frac {\mu_k(l)}{l^{\frac cb}}+\sum_{l\le
y}\mu_k (l)\Delta\left(a,b;\frac {x}{l^c}\right).\nonumber\\
&=\operatorname*{Res}\left( \zeta (as)\zeta (bs)x^ss^{-1}\sum_{l\le y}\frac
{\mu_k(l)}{l^{cs}},\frac 1a \right)\nonumber \\
&\ \ \ +\operatorname*{Res}\left( \zeta (as)\zeta (bs)x^ss^{-1}\sum_{l\le
y}\frac {\mu_k(l)}{l^{cs}},\frac 1b \right) +\sum_{l\le y}\mu_k
(l)\Delta\left(a,b;\frac {x}{l^c}\right).\nonumber\end{aligned}$$
From (2.1) we have $$\begin{aligned}
\sum_{n=1}^{\infty}\frac {f_{2,y}(n)}{n^s}=g_y(cs)\zeta(as)\zeta
(bs).\end{aligned}$$
From (2.3), (2.7), (2.9) and Lemma 2.1 we obtain that $$\begin{aligned}
S_2(x)=\frac {1}{2\pi i}\int_{\frac 1a +\varepsilon-ix^2}^{\frac
1a+\varepsilon+ix^2}g_y(cs)\zeta (as)\zeta(bs)x^ss^{-1}\mathrm{d}s
+O(x^\delta),\end{aligned}$$ since $f_{2,y}(n)\ll n^\delta$ by a divisor argument.
[**Case (i).**]{} $a<b<2a$. When we move the line of integration to $\Re s=\sigma =\frac {1}{2a}+\delta_0$ with $\delta_0=\min
\{\delta, \frac {1}{2b}(1-\frac {b}{2a})\}$, then by the residue theorem $$\begin{aligned}
&\frac {1}{2\pi i}\int_{\frac 1a +\varepsilon-ix^2}^{\frac
1a+\varepsilon-ix^2}g_y(cs)\zeta (as)\zeta(bs)x^ss^{-1}\mathrm{d}s\\
&=\operatorname*{Res}\left( g_y(cs)\zeta (as)\zeta (bs)x^ss^{-1},\frac 1a
\right)+\operatorname*{Res}\left( g_y(cs)\zeta (as)\zeta (bs)x^ss^{-1},\frac 1b
\right)\nonumber\\
&\ \ \ +I_1+I_2-I_3,\nonumber\end{aligned}$$ where $$\begin{aligned}
&I_1=\frac {1}{2\pi i}\int_{\frac {1}{2a}+\delta_0+ix^2}^{\frac
1a+\varepsilon+ix^2}g_y(cs)\zeta (as)\zeta
(bs)x^ss^{-1}\mathrm{d}s, I_2=\frac {1}{2\pi i}\int_{\frac
{1}{2a}+\delta_0-ix^2}^{\frac {1}{2a}+\delta_0 +ix^2}g_y(cs)\zeta
(as)\zeta (bs)x^ss^{-1}\mathrm{d}s,\\
&I_3=\frac {1}{2\pi i}\int_{\frac {1}{2a}+\delta_0 -ix^2}^{\frac
1a+\varepsilon-ix^2}g_y(cs)\zeta (as)\zeta
(bs)x^ss^{-1}\mathrm{d}s.\end{aligned}$$
From Lemma 2.2, we have $$g_y(cs)\ll y^{\frac 12-\frac {c}{2a}}(|t|^\delta +1),( \sigma \ge \frac {1}{2a} +\delta)$$ Thus $$\begin{aligned}
g_y(cs)\zeta (as)\zeta (bs)\ll
y^{\frac 12-\frac {c}{2a}}(|t|^{3\delta} +1),( \sigma \ge \frac
{1}{2a} +\delta).\end{aligned}$$ From (2.12) it is not difficult to see that $$\begin{aligned}
I_j\ll y^{\frac 12-\frac {c}{2a}}x^{\frac
{1}{2a}+8\delta},(j=1,2,3).
\end{aligned}$$
Now combining (1.3), (2.5), (2.8), (2.10), (2.11) and (2.13) completes the proof of Theorem 1 in this case.
[**Case (ii).**]{} $b\ge 2a$. In this case, moving the line of integration in (2.10) to $\Re s=\sigma =\frac {1}{2a}+\delta$, we can treat $S_2(x)$ as in the above case except the second residue in relation (2.11) vanishes. In addition, applying Abel summation formula and the estimate $\sum_{n\le x}\mu_k(n)\ll x^{\frac 12
+\varepsilon}$(this can be proved in the same way as that of Theorem 14.25(C) in Titchmarsh[@ti], also see (2.10) in Nowak[@nowak]), it is easy to check that $$\begin{aligned}
x^{\frac 1b}\sum_{l>y}\frac {\mu_k(l)}{l^{\frac cb}}\ll x^{\frac
1b+\delta}y^{\frac 12-\frac cb}\ll x^{\frac
{1}{2a}+\delta}y^{\frac {1}{2}-\frac {c}{2a}}.\end{aligned}$$ Hence $$\begin{aligned}
x^{\frac 1b}\sum_{l\le y}\frac {\mu_k(l)}{l^{\frac cb}}&=x^{\frac
1b}\sum_{l=1}^{\infty}\frac {\mu_k(l)}{l^{\frac cb}}-x^{\frac
1b}\sum_{l> y}\frac {\mu_k(l)}{l^{\frac
cb}}\\
&=\frac {1}{\zeta^k(\frac cb)}x^{\frac 1b}+O\left( x^{\frac
{1}{2a}+\delta}y^{\frac {1}{2}-\frac {c}{2a}}\right)\nonumber.\end{aligned}$$
Therefore, Theorem 1 also holds in this case. This completes the proof of Theorem 1. The proof of Theorem 2 is very similar to that of Theorem 1, we omit the details here.
**The proof of Theorem 3**
==========================
Let $a<c<b$ and $\Delta_k(a;c;x)$ be defined by (1.17). If $\Delta_k(a;c;x)\ll x^{\alpha{(a;c)}+\varepsilon}$ such that $\alpha{(a;c)}< 1/b$, then for $s>\alpha(a;c)$ we have $$\begin{aligned}
\sum_{m\le x}u_k(a;c;m)m^{-s}=&\frac {x^{\frac
1a-s}}{(1-as)\zeta^k(\frac ca)}+\frac {\zeta
(as)}{\zeta^k(cs)}\\
&+\Delta_k(a;c;x)x^{-s}-s\int_{x}^{\infty}\Delta_k(a;c;t)t^{-s-1}\mathrm{d}t.\nonumber\end{aligned}$$
By partial summation formula and (1.17) we get $$\begin{aligned}
&\sum_{m\le
x}u_k(a;c;m)m^{-s}\\
&=A_k(a;c;x)x^{-s}+s\int_{1}^{x}A_k(a;c;t)t^{-s-1}\mathrm{d}t\nonumber\\
&=\frac {x^{\frac 1a-s}}{\zeta^k(\frac
ca)}+\Delta_k(a;c;x)x^{-s}+s\int_{1}^{x}\left(\frac {t^{\frac
1a}}{\zeta^k(\frac
ca)}+\Delta_k(a;c;t)\right)t^{-s-1}\mathrm{d}t\nonumber\\
&=\frac {x^{\frac 1a-s}}{(1-as)\zeta^k(\frac
ca)}+\Delta_k(a;c;x)x^{-s}-\frac {as}{(1-as)\zeta^k(\frac
ca)}+s\int_{1}^{x}\Delta_k(a;c;t)t^{-s-1}\mathrm{d}t.\nonumber\end{aligned}$$
Suppose that $s>1$, we have from (1.15) and the condition $\alpha{(a;c)}<\frac 1b$, when $x\rightarrow \infty$ $$\begin{aligned}
\frac {\zeta (as)}{\zeta^k(cs)}=-\frac {as}{(1-as)\zeta^k(\frac
ca)}+s\int_{1}^{\infty}\Delta_k(a;c;t)t^{-s-1}\mathrm{d}t.\end{aligned}$$ By analytic continuation this equation also holds for $s>\alpha(a;c)$. Substituting (3.3) into (3.2) completes the proof of Lemma 3.1.
Let $x\ge 2$, $a<c<b$, and $\Delta_k(a,b;c;x)$ be defined by (1.3). If $\Delta_k(a;c;x)\ll
x^{\alpha{(a;c)}+\varepsilon}$ such that $\alpha{(a;c)}< 1/b$, then for any $1\le y<x^{ 1/b}$ we have
$$\begin{aligned}
\Delta_k(a,b;c;x)=&\sum_{d\le y}\Delta_k\left(a;c;\frac
{x}{d^b}\right)- \sum_{m\le \frac
{x}{y^b}}u_k(a;c;m)\psi\left((\frac x m)^{\frac
1b}\right)\nonumber\\
& -\frac {x^{\frac 1b}}{b} \int_{\frac {x}{y^b}}^{\infty}\frac
{\Delta_k(a;c;t)}{t^{1+\frac 1b }}\mathrm{d}t+\psi
(y)\Delta_k\left(a;c; \frac {x}{y^b}\right)+O(x^{\frac
1a}y^{-1-\frac ba}).\nonumber\end{aligned}$$
Let $1\le y\le x^{1/b}$. Applying (1.1),(1.15)-(1.17) and Dirichlet’s hyperbolic argument, we get $$\begin{aligned}
&A_k(a,b;c;x)=\sum_{n\le x}f_k(a,b;c;n)=\sum_{md^b\le x}u_k(a;c;m)\\
&=\sum_{d\le y}\sum_{m\le \frac {x}{d^b}}u_k(a;c;m)+\sum_{m\le
\frac
{x}{y^b}}u_k(a;c;m)\sum_{y<d\le (\frac xm)^{\frac 1b}}1\nonumber\\
&=\frac {x^{\frac 1 a}}{\zeta^k(\frac ca)}\sum_{d\le y}\frac
{1}{d^{\frac ba}}+x^{\frac 1b}\sum_{m\le \frac {x}{y^b}}\frac
{u_k(a;c;m)}{m^{\frac 1b}}-y\sum_{m\le \frac
{x}{y^b}}u_k(a;c;m)\nonumber\\
&\quad+\sum_{d\le y}\Delta_k\left(a;c;\frac {x}{d^b}\right)-
\sum_{m\le \frac {x}{y^b}}u_k(a;c;m)\psi\left((\frac x m)^{\frac
1b}\right)+ \psi (y)\sum_{m\le \frac {x}{y^b}}u_k(a;c;m).\nonumber\end{aligned}$$
Applying Lemma 3.1 with $s=\frac 1b$, we get $$\begin{aligned}
\sum_{m\le \frac {x}{y^b}}\frac {u_k(a;c;m)}{m^{\frac 1b}} =&\frac
{\left( \frac {x}{y^b}\right)^{\frac 1a-\frac 1b}}{(1-\frac
ab)\zeta^k(\frac ca)}+\frac {\zeta (\frac ab)}{\zeta^k(\frac
ca)}\\
&+\Delta_k(a;c; \frac {x}{y^b})\left( \frac
{x}{y^b}\right)^{-\frac 1b}-\frac 1b \int_{\frac
{x}{y^b}}^{\infty}\frac {\Delta_k(a;c;t)}{t^{1+\frac 1b
}}\mathrm{d}t.\nonumber\end{aligned}$$
In addition, we have from the Euler-Maclaurin formula that $$\begin{aligned}
\sum_{d\le y}\frac {1}{d^{\frac ba}}=\zeta (\frac ba)+\frac
{y^{1-\frac ba}}{(1-\frac ba)}-\psi(y)y^{-\frac ba}+O(y^{-1-\frac
ba}).\end{aligned}$$
Applying (1.17) again we also have $$\begin{aligned}
\sum_{m\le \frac
{x}{y^b}}u_k(a;c;m)=\frac {\left( \frac {x}{y^b}\right)^{\frac
1a}}{\zeta^k(\frac ca)}+\Delta_k(a;c; \frac {x}{y^b}).\end{aligned}$$ Substituting (3.5)-(3.7) into (3.4), we get $$\begin{aligned}
A_k(a,b;c;x)=&\frac {\zeta(\frac ba)}{\zeta^k(\frac ca)}x^{\frac 1
a}+\frac {\zeta(\frac ab)}{\zeta^k(\frac cb)}x^{\frac 1b}-
\sum_{m\le \frac {x}{y^b}}u_k(a;c;m)\psi\left((\frac x m)^{\frac
1b}\right)\\
&+\sum_{d\le y}\Delta_k\left(a;c;\frac {x}{d^b}\right) -\frac
{x^{\frac 1b}}{b} \int_{\frac {x}{y^b}}^{\infty}\frac
{\Delta_k(a;c;t)}{t^{1+\frac 1b }}\mathrm{d}t\nonumber\\
&+\psi (y)\Delta_k\left(a;c; \frac {x}{y^b}\right)+O(x^{\frac
1a}y^{-1-\frac ba}).\nonumber\end{aligned}$$ Now Lemma 3.2 follows from (1.3) and (3.8) at once.
Let $\Delta_k(a;c;x) $ be defined by (1.17). If RH is true, then for any fixed $\delta>0$ we have $$\begin{aligned}
\int_{1}^{T}\Delta_k(a;c;u)\mathrm{d}u\ll T^{1+\frac {1}{2c}+
\delta }.\end{aligned}$$
It suffices to prove that for any $M>2$, we have $$\begin{aligned}
\int_{M}^{2M}\Delta_k(a;c;u)\mathrm{d}u\ll M^{1+\frac {1}{2c}+
\delta}.\end{aligned}$$
Taking in Lemma 2.1 $H(n)=n^\varepsilon, B(\sigma)=(\sigma-1)^{-k}, b=1+1/\log M, T=M^5$, we get $$\begin{aligned}
\sum_{n\leq x}u_k(a;c;u)=\frac {1}{2\pi i}\int_{b-iT}^{b+iT}\frac
{\zeta (as)}{\zeta^k(cs)}\frac{u^s}{s}\mathrm{d}s+O(M^{-4}).\end{aligned}$$
It is well-known that if RH is true, then for any fixed $0<\eta<1/2,$ we have $$\begin{aligned}
\zeta (s)\ll (|t|+1)^{\eta},\
\zeta^{-1}(s)\ll (|t| +1)^\eta,\ \ \sigma
>\frac 12 + \eta.\end{aligned}$$ Moving the line of integration to $\Re s=\frac {1}{2c}+\delta$, we have by (3.11) and the estimate $\zeta(s)\ll (1+|t|)^{1/2}\
(\sigma\geq 0)$ that $$\begin{aligned}
\Delta_k(a;c;u)=\frac {1}{2\pi i}\int_{\frac
{1}{2c}+\delta-iT}^{\frac {1}{2c}+\delta+iT}\frac {\zeta
(as)}{\zeta^k(cs)}\frac{u^s}{s}\mathrm{d}s+O(M^{-1}).\end{aligned}$$
Thus we have $$\begin{aligned}
\int_{M}^{2M}\Delta_k(a;c;u)\mathrm{d}u & = \int_{\frac
{1}{2c}+\delta-iT}^{\frac {1}{2c}+\delta+iT}\frac {\zeta
(as)}{\zeta^k(cs)}\frac {\mathrm{d}s}{s}\int_{M}^{2M}u^s \mathrm{d}u+O(1)\\
&=\int_{\frac {1}{2c}+\delta-iT}^{\frac
{1}{2c}+\delta+iT}\frac{\zeta
(as)\left((2M)^{1+s}-M^{1+s}\right)}{\zeta^k(cs)s(1+s)}\mathrm{d}s+O(1).\nonumber\\
&\ll M^{1+\frac
{1}{2c}+\delta}\int_{-T}^{T}\left|\frac{\zeta(a(\frac
{1}{2c}+\delta+it))}{\zeta^k(c(\frac
{1}{2c}+\delta+it))}\right|\frac{\mathrm{d}t}{(1+|t|)^2}+1\nonumber\\
&\ll M^{1+\frac {1}{2c}+ \delta }.\nonumber\end{aligned}$$ Namely (3.10) holds. This completes the proof of Lemma 3.3.
[**The Proof of Theorem 3.**]{} Theorem 3 follows immediately from Lemma 3.2 and Lemma 3.3 with $\delta =\frac
{\varepsilon}{10}$.
[**The Proof of Corollary 1.3.**]{} It is easy to check $$\begin{aligned}
\sum_{d\le y}\Delta_k\left(a;c;\frac {x}{d^b}\right)\ll \sum_{d\le
y}\left(\frac {x}{d^b}\right)^{\alpha_k(a;c)}\ll
x^{\alpha_k(a;c)}y^{1-b\alpha_k(a;c)}\end{aligned}$$ and $$\begin{aligned}
\sum_{m\le \frac {x}{y^b}}u_k(a;c;m)\psi\left((\frac x m)^{\frac
1b}\right)&\ll \sum_{m\le \frac {x}{y^b}}|u_k(a;c;m)|\\
&\ll \sum_{m\le \frac {x}{y^b}}d(a,c;m)\ll \left(\frac
{x}{y^b}\right)^{\frac 1 a}.\nonumber\end{aligned}$$ Taking $y=x^{\frac {1-a\alpha_k (a;c)}{a+b-ab\alpha_k (a;c)}}$ we find that Corollary 1.3 is an immediate consequence of Theorem 3 and the above two estimates.
**Estimates for $\Delta_k (a;c;x) (k=1,2)$ and an application of Theorem 2**
=============================================================================
**Some preliminary lemmas**
---------------------------
To treat the exponential sums appeared in Theorem 1 and 2, for the arithmetic function $\mu_k (k\ge 2)$ one needs an analogue of the well-known Vaughan’s identity of Möbius function $\mu$. First we shall prove such an identity.
(Vaughan’s identity). Let $1\le N_1<N$. Suppose that $U, V $ be two parameters with $1\leq U, V\le N_1$. Then for any arithmetic function $f$ we have $$\begin{aligned}
\sum_{N_1<n\le N}\mu_k(n)f(n)={\sum}_1-{\sum}_2-{\sum}_3,\end{aligned}$$ where $$\begin{aligned}
&{\sum}_1=\sum_{U<m\le N/V}A(m)\sum_{\stackrel{N_1/m<n\le N/m}{V<n}}\mu_k(n)f(mn),\\
&{\sum}_2=\sum_{U<m\le UV}B(m)\sum_{N_1/m<n\le N/m}\tau_k(n)f(mn),\\
&{\sum}_3=\sum_{m\le U}B(m)\sum_{N_1/m<n\le N/m}\tau_k(n)f(mn),\\
&A(m)=\sum_{\stackrel{ed_1=m}{e\le U}}\mu_k(e)\tau_{k}(d_1),\ \
B(m)=\sum_{\stackrel{d_1d_2=m}{d_1\le U,d_2\le
V}}\mu_k(d_1)\mu_k(d_2).\end{aligned}$$
Let $$\begin{aligned}
F(U,s):=\sum_{d\le U}\frac{\mu_k(d)}{d^s},\Re s>1,\end{aligned}$$ then $$\begin{aligned}
\frac {1}{\zeta^k(s)}&=\left(\frac
{1}{\zeta^k(s)}-F(V,s)\right)\left(1-\zeta^k(s)F(U,s)\right)-F(U,s)F(V,s)\zeta^k(s)\\
&\ \ +F(U,s)+F(V,s),\Re s>1.\nonumber\end{aligned}$$ Equating coefficients from both sides of (4.3) gives the following identity $$\begin{aligned}
\mu_k(n)=b_1(n)+b_2(n)+b_3(n)+b_4(n),\end{aligned}$$ where $$\begin{aligned}
b_1(n)&=-\sum_{\stackrel{dm=n}{d>V,m>1}}\mu_k(d)\left(\sum_{\stackrel{ed_1=m}{e\le
U}}\mu_k(e)\tau_{k}(d_1)\right),\\
b_2(n)&=-\sum_{\stackrel{d_1d_2m=n}{d_1\le U,d_2\le
V}}\mu_k(d_1)\mu_k(d_2)\tau_{k}(m),\\
b_3(n)& =\begin{cases}
{\displaystyle}\mu_k(n)
& \mbox{if $n\le U$,}\\[1em]
{\displaystyle}0 & \mbox{if $n>U,$}
\end{cases}
\ \ b_4(n) =\begin{cases}
{\displaystyle}\mu_k(n)
& \mbox{if $n\le V$,}\\[1em]
{\displaystyle}0 & \mbox{if $n>V.$}
\end{cases}\end{aligned}$$
From (1.10) we have $$\sum_{md=n}\tau_{k}(m)\mu_k(d) =\begin{cases}
{\displaystyle}1
& \mbox{if $n=1$,}\\[1em]
{\displaystyle}0 & \mbox{if $n>1.$}
\end{cases}$$ In the sum for $b_1(n)$ we can replace the condition $m>1$ by $m>U$, since the sum over $m$ vanishes by (4.5) when $1<m\le U$. Now multiplying the above identity (4.4) by $f(n)$ we get (4.1).
We will also exploit the following several lemmas. Lemma 4.2 is Lemma 1 of Graham and Pintz[@gp](also see Theorem 18 of Vaaler[@va]), Lemma 4.3 is well-known, Lemma 4.4 is Lemma 6 of Fouvry and Iwaniec[@fi], Lemma 4.5 is Lemma 4 of the second paper in[@ca](also see (2.1) in Wu[@wu2]), Lemma 4.6 is Lemma 12 of Cao[@ca].
Suppose $H>0$. There is a function $\psi^*(x)$ such that $$\begin{aligned}
\psi^*(x)&=\sum_{1\le |h|\le H}\gamma (h)e(hx), \ \ \gamma(h)\ll
\frac {1}{|h|}, \\
|\psi^*(x)-\psi (x)|&\le \frac {1}{2H+2}\sum_{|h|\le
H}\left(1-\frac {|h|}{H}\right)e(hx).\end{aligned}$$
Let $X\ne 0$ and $\nu\ne 0,1$. If $(\kappa,\lambda)$ is an exponent pair, then $$\begin{aligned}
\sum_{n\sim N}e(Xn^{\nu})\ll \left(XN^{\nu-1}\right)^\kappa
N^\lambda+X^{-1}N^{-\nu+1}.\end{aligned}$$
Let $0<M\le N<\gamma N\le \lambda M$, and $|a_n|\le 1$. Then we have $$\begin{aligned}
\sum_{N<n\le \gamma N}a_n=\frac
{1}{2\pi}\int_{-M}^{M}\left(\sum_{M<n\le \lambda
M}a_nn^{-it}\right)N^{it}(\gamma^{it}-1)t^{-1}\mathrm{d}t+O\left(\log
(2+M)\right).\end{aligned}$$
Let $x\ge 2$, $\alpha, \beta, \gamma$ be given real numbers with $\alpha (\alpha -1)\beta\gamma\neq 0$, $|a(m)|\le 1$, $b(n_1,n_2)\le 1$. Suppose $G=xM^\alpha N_1^\beta
N_2^\gamma$, $(\kappa, \lambda)$ is an exponent pair and $$\begin{aligned}
T(M,N_1,N_2)=\sum_{m\sim M}\sum_{n_1\sim N_1}\sum_{n_2\sim
N_2}a(m)b(n_1,n_2)e(xm^\alpha n_1^\beta n_2^\gamma).\end{aligned}$$ Then $$\begin{aligned}
T(M,N_1,N_2)\mathcal{L}^{-2}\ll
\left(G^{\kappa}M^{1+\lambda+\kappa}(N_1N_2)^{2+\kappa}\right)^{\frac
{1}{2+2\kappa}}+M^{\frac 12}N_1N_2+M(N_1N_2)^{\frac 12}+G^{-\frac
12}MN_1N_2.\end{aligned}$$
Let $x\ge 2$, $\beta, \gamma$ be given real numbers with $\beta\gamma\neq 0$,$|a(m)|\le 1$, $|b(n)|\le 1$, $(\kappa, \lambda)$ is an exponent pair. Suppose $D$ is a subdomain of $\{(m,n):m\sim M, n\sim N\}$ bounded by finite algebraic curves , $G=xM^\beta N^\gamma$ and $$\begin{aligned}
T_1(M,N)=\sum_{(m,n)\in D}a(m)b(n)\psi(xm^\beta n^\gamma).\end{aligned}$$ Then $$\begin{aligned}
T_1(M,N)\mathcal{L}^{-6}\ll
\left(G^{\kappa}M^{1+\lambda+\kappa}N^{2+\kappa}\right)^{\frac
{1}{2+2\kappa}}+M^{\frac 12}N+MN^{\frac 12}+G^{-\frac 12}MN.\end{aligned}$$
**An estimate of $\Delta (a;c;x)$**
-------------------------------------
Suppose $1\leq a<c$ are two fixed integers. In this subsection, we shall estimate the error term $\Delta (a;c;x)$ defined by (1.17) with $k=1.$
Let $r>1$ be a fixed real number. The function $\Delta_k(1;r;x)$ is defined on $[1,\infty)$ such that for any $1\leq y\leq x^{1/r}$ one has $$\begin{aligned}
\Delta_k(1;r;x) =\sum_{l\le y}\mu_k (l)\psi\left(\frac
{x}{l^r}\right)+O\left(x^{\frac {1}{2}+\varepsilon}y^{\frac 12-\frac
{r}{2}}+y^{\frac 12+\varepsilon}\right).\end{aligned}$$
It follows easily from Theorem 2 that $$\begin{aligned}
\Delta_k(a;c;x)=\Delta_k(1;\frac ca;x^{\frac 1a}).\end{aligned}$$ Hence we only need to estimate $\Delta_k(1;r;x)$ for real $r>1$.
Now we define $$\begin{aligned}
\alpha (r) =\begin{cases}
{\displaystyle}\frac{7}{8r+6}
& \mbox{if $1< r\le 5$ and $r\ne 2$,}\\[1em]
{\displaystyle}\frac{17}{54}
& \mbox{if $r=2$,}\\[1em]
{\displaystyle}\frac{67}{514} & \mbox{if $5<r\leq 6,$}\\[1em]
{\displaystyle}\frac {11(r-4)}{12r^2-37r-41} & \mbox{if $6<r\le 12,$}\\[1em]
{\displaystyle}\frac {23(r-1)}{24r^2+13r-37} & \mbox{if $12< r\le 20.$}
\end{cases}\end{aligned}$$ For $r>20$, $\alpha (r)$ is defined by the following procedure. Let $q\ge 2$ and $Q=2^q$. For every $r>20$, there is a unique integer $q$ such that $\frac {12Q-q-5}{2}<r\le \frac
{24Q-q-6}{2}$. With this value of $q$, define $$\begin{aligned}
\alpha (r)=\frac {(12Q-1)r-12Q+1}{12Qr^2+(6Qq+1)r-(6Qq+12Q+1)}.\end{aligned}$$
In this subsection we shall prove that
Let $\alpha (r)$ be defined by (4.8) and (4.9), respectively. If RH holds, then $$\Delta(1;r;x)=O\left(x^{\alpha(r)+\varepsilon}\right).$$
[**Remark 4.1.**]{} Certainly one can improve the exponent $\alpha
(r)$ further for some special values of $r$. For example, R. C. Baker and K. Powell[@bp] obtained recently that $\alpha
(3)=\frac {17}{74}$, $\alpha (4)=\frac {17}{94}$ and $\alpha
(5)=\frac {3}{20}$. In addition, for large values of $r$, one can take $\alpha (r)=\frac {1}{r+c^*r^{1/3}}$ for some constant $c^*>0$ (see Theorem 2, [@gp].)
From (4.7) and Theorem 4 we get
Let $1\leq a<c$ be two fixed integers. If RH holds, then $$\begin{aligned}
\Delta(a;c;x)=O\left(x^{\frac {1}{a}\alpha(\frac
ca)+\varepsilon}\right).\end{aligned}$$
Theorem 4 is proved in Jia[@ji] for the case $r=2$. S. W. Graham and J. Pintz[@gp] showed that Theorem 4 holds for any integer $r>3$. However it is easily seen that the argument of [@gp] can be applied to any $r\ge 2.$ So we only give a proof of Theorem 4 for $1<r<2$.
Taking $y=x^{\frac {4}{4r+3}}$ , by Theorem 2, (4.6) and a simple splitting argument, an estimate for $x^{\alpha(r)+\varepsilon}\ll
Y\le y$ $$\begin{aligned}
\sum_{Y<l\le 2Y}\mu (l)\psi\left(\frac {x}{l^r}\right)\ll
x^{\alpha(r)+\varepsilon}\end{aligned}$$ would suffice to complete the proof of Theorem 4.
Choose $U=Y^{\frac 12}, V=Y^{\frac 14}$. Let $|a(m)|\le 1$ and $|b(n)|\le 1$ be any complex-valued arithmetic functions. If we can show the estimates $$\begin{aligned}
\sum_{U< m\le Y/V}a(m)\sum_{Y< mn \le 2Y}b(n)\psi\left(\frac
{x}{m^rn^r}\right)\ll x^{\alpha(r)+\varepsilon}\end{aligned}$$ and $$\begin{aligned}
\sum_{ m\le U}a(m)\sum_{Y< mn \le 2Y}\psi\left(\frac
{x}{m^rn^r}\right)\ll x^{\alpha(r)+\varepsilon},\end{aligned}$$ then (4.11) follows from Lemma 4.1.
We first estimate the type II sum (4.12). Assume $N\ll M$, applying Lemma 4.6 with $(\kappa,\lambda)=(\frac 12,\frac 12)$, we get $$\begin{aligned}
&\mathcal{L}^{-6}\sum_{ m\sim M}a(m)\sum_{\stackrel {n\sim N}{Y<
mn \le
2Y}}b(n)\psi\left(\frac {x}{m^rn^r}\right)\\
&\ll \left(\left(\frac{x}{(MN)^r}\right)^{\frac 12}M^2N^{\frac
52}\right)^{\frac 13}+M^{\frac 12}N+MN^{\frac 12}+\left(\frac
{x}{(MN)^r}\right)^{-\frac 12}MN\nonumber\\
&\ll \left(x(MN)^{4-r}(MN)^{\frac 12}\right)^{\frac
{1}{6}}+(MN)N^{-\frac 12}+x^{-\frac 12}(MN)^{1+\frac r2}\nonumber\\
&\ll x^{\frac 16}Y^{\frac{9-2r}{12}}+YN^{-\frac 12}+x^{-\frac
12}Y^{1+\frac r2}\ll x^{\alpha(r)}.\nonumber\end{aligned}$$ Hence (4.14) holds under the condition $N\ll M$. If $N\gg M$, using Lemma 4.4 to separate the dependence between the variables $n$ and $m$, then interchanging the roles of $m$ and $n$, we can show that (4.14) also holds in this case. The estimate (4.12) follows from (4.14) by a simple splitting argument.
Now we turn to estimate the type I sum (4.13). If $M\ge V$, by the same the argument as that of (4.14), we get $$\begin{aligned}
\mathcal{L}^{-6}\sum_{ m\sim M}a(m)\sum_{\stackrel {n\sim N}{Y< mn
\le 2Y}}\psi\left(\frac {x}{m^rn^r}\right)\ll
x^{\alpha(r)}.\end{aligned}$$
If $M\le V$, applying Lemma 4.2 with $H=Yx^{-\alpha (r)}$ and Lemma 4.3 with $(\kappa,\lambda)=(\frac 12,\frac 12)$, we get that $$\begin{aligned}
&\sum_{ m\le V}a(m)\sum_{Y< mn \le 2Y}\psi\left(\frac
{x}{m^rn^r}\right)\\
& \ll \sum_{ m\le V}\left(\frac 1H\frac Ym+\sum_{1\le |h|\le
H}\frac {1}{|h|}\left|\sum_{Y/m< n \le 2Y/m} e\left(\frac {hx}{m^rn^r}\right)\right|\right)\nonumber \\
& \ll x^{\alpha (r)}\mathcal{L}+\sum_{ m\le V}\sum_{1\le h\le
H}\frac 1h\left( (hx)^{\frac 12}Y^{-\frac r2}+(hxm)^{-1}Y^{1+r}\right)\nonumber\\
& \ll x^{\alpha (r)}\mathcal{L}+x^{\frac 12}Y^{-\frac r2}H^{\frac
12}V+x^{-1}Y^{1+r}\mathcal{L}\nonumber\\
& \ll x^{\alpha (r)}\mathcal{L}+x^{\frac 12-\frac {\alpha (r)}{2}}Y^{\frac 34-\frac r2}+x^{-1}y^{1+r}\mathcal{L}\nonumber\\
&\ll x^{\alpha (r)}\mathcal{L}.\nonumber\end{aligned}$$ (Here note that if $1.5\le r<2$, we use the bound $Y\gg
x^{\alpha(r)}$, otherwise we use $Y\le y$) Finally, it follows from (4.15) and (4.16) that (4.13) always holds. This completes the proof of Theorem 4.
**An estimate of $\Delta_2(a;c;x)$**
-------------------------------------
Let $a,c$ be two fixed integers such that $1\leq a<c\leq 9a/2$ and $\Delta_2(a;c;x)$ be defined by (1.17) with $k=2$. Assume that RH holds, then $$\begin{aligned}
\Delta_2(a;c;x)\ll x^{ \frac{7}{8c+6a}+\varepsilon}.\end{aligned}$$
Similar to the proof of Theorem 4, we only need to show that for $1< r \le 9/2$ one has $$\begin{aligned}
\Delta_2(1;r;x)=O\left(x^{\beta(r)+\varepsilon}\right),\
\beta(r)=7/(8r+6).\end{aligned}$$ Taking $y=x^{\frac {4}{4r+3}}$ , by Theorem 2, (4.6) and a simple splitting argument, an estimate $$\begin{aligned}
\sum_{Y<l\le 2Y}\mu_2 (l)\psi\left(\frac {x}{l^r}\right)\ll
x^{\beta(r)+\varepsilon}\ \ (x^{\beta(r)+\varepsilon}\ll Y\le y)\end{aligned}$$ would suffice to complete the proof of Theorem 5.
Choose $U=Y^{\frac 12}, V=Y^{\frac 14}$. Let $|a_1(m)|\le 1$ and $|b_1(n)|\le 1$ be any complex-valued arithmetic functions. If we can show $$\begin{aligned}
\sum_{U< m\le Y/V}a_1(m)\sum_{Y< mn \le 2Y}b_1(n)\psi\left(\frac
{x}{m^rn^r}\right)\ll x^{\beta(r)+\varepsilon}\end{aligned}$$ and $$\begin{aligned}
\sum_{ m\le U}a_1(m)\sum_{Y< mn \le 2Y}\tau(n)\psi\left(\frac
{x}{m^rn^r}\right)\ll x^{\beta(r)+\varepsilon},\end{aligned}$$ then (4.19) follows from Lemma 4.1.
The estimate (4.20) can be proved by the same approach of (4.12), so we omit its detals. Hence we only need to prove (4.21). From (4.20) we get easily that $$\sum_{Y^{1/4}<m\le U}a_1(m)\sum_{Y< mn \le 2Y}\tau(n)\psi\left(\frac
{x}{m^rn^r}\right)\ll x^{\beta(r)+\varepsilon}.$$ So it suffices for us to prove $$\begin{aligned}
\sum_{ m\le V}a_1(m)\sum_{Y< mn \le 2Y}\tau(n)\psi\left(\frac
{x}{m^rn^r}\right)\ll x^{\beta(r)+\varepsilon}.\end{aligned}$$ Let $1\le M\le V$. Now we are in a position to estimate the exponential sum $$\begin{aligned}
S_r(M,Y):=\sum_{ M<m\le 2M}a_1(m)\sum_{Y< mn_1n_2 \le
2Y}\psi\left(\frac {x}{m^rn_1^rn_2^r}\right).\end{aligned}$$
Without the loss of generality, we suppose $n_1\ll n_2$, hence $n_1\ll (YM^{-1})^{\frac 12}$. Applying a simple splitting argument, we have for some $N_1\ll (YM^{-1})^{\frac 12}$ $$\begin{aligned}
\mathcal{L}^{-1}S_r(M,Y)\ll \sum_{M< m\le
2M}a_1(m)\sum_{N_1<n_1\le 2N_1}\sum_{Y< mn_1n_2 \le
2Y}\psi\left(\frac {x}{m^rn_1^rn_2^r}\right).\end{aligned}$$ Now we consider two cases.
(Case i): $\frac 12\le N_1\ll VM^{-1}$. In this case, applying Lemma 4.2 with $H=Yx^{-\beta (r)}$ and Lemma 4.3 with $(\kappa,\lambda)=(\frac 12,\frac 12)$, similar to the estimate of (4.16), we can obtain $$\begin{aligned}
S_r(M,Y) \ll x^{\beta(r)}\mathcal{L}^3.\end{aligned}$$
(Case ii): $ VM^{-1}\le N_1\ll (YM^{-1})^{\frac 12}$. Applying Lemma 4.4 to separate the dependence between the variable $n_2$ and the variables $m, n_1$, we get for $N_2=\frac {Y}{MN_1}$ that $$\begin{aligned}
&\sum_{M< m\le 2M}a_1(m)\sum_{N_1<n_1\le 2N_1}\sum_{Y< mn_1n_2 \le
2Y}\psi\left(\frac {x}{m^rn_1^rn_2^r}\right)\\
&= \frac {1}{2\pi}\int_{-N_2}^{N_2}\left(\sum_{M< m\le
2M}a_1(m)\sum_{N_1<n_1\le 2N_1}\sum_{N_2<n_2\le
8N_2}n_2^{-it}\psi\left(\frac {x}{m^rn_1^rn_2^r}\right)\right)
N_2^{it}(8^{it}-1)t^{-1}\mathrm{d}t+O\left(MN_1\mathcal{L}\right)\nonumber\\
&=\frac {1}{2\pi}\int_{-N_2}^{N_2}\left(\sum_{MN_1<d\le
4MN_1}\sum_{N_2<n_2\le 8N_2}c(d)n_2^{-it}\psi\left(\frac
{x}{d^rn_2^r}\right)\right)N_2^{it}\frac{8^{it}-1}{t}\mathrm{d}t+O\left(MN_1\mathcal{L}\right),\nonumber\end{aligned}$$ where $$c(d)=\sum_{d=mn_1, M<m\le 2M, N_1<n_1\le 2N_1}a_1(m)\ll
d^\varepsilon.$$
If $N_2\gg MN_1$, applying Lemma 4.6 with $(M,N)=(N_2,MN_1)$ and $(\kappa,\lambda)=(\frac 12,\frac 12)$ to estimate the inner sum in the above expression, we get(similar to (4.14)) $$\begin{aligned}
&\ \ \ \ x^{-\frac {\varepsilon}{2}}\sum_{MN_1<d\le
4MN_1}\sum_{N_2<n_2\le
8N_2}c(d)n_2^{-it}\psi\left(\frac {x}{d^rn_2^r}\right)\\
&\ll \left(\left(\frac {x}{Y^r}\right)^{\frac
12}N_2^2(MN_1)^{\frac 52}\right)^{\frac 13}+N_2^{\frac
12}(MN_1)+N_2(MN_1)^{\frac 12}+\left(\frac {x}{Y^r}\right)^{-\frac
12}MN_1N_2\nonumber\\
&\ll \left(x^{\frac 12}Y^{-\frac r2+2}(MN_1N_2)^{\frac
14}\right)^{\frac 13}+(MN_1N_2)^{\frac 12}(YM)^{\frac
14}+(MN_1N_2)(MN_1)^{-\frac 12}+x^{-\frac 12}Y^{\frac
r2+1}\nonumber\\
&\ll \left(x^{\frac 12}Y^{\frac {9-2r}{4}}\right)^{\frac
13}+Y^{\frac 12}(YV)^{\frac 14}+YV^{-\frac 12}+x^{-\frac
12}Y^{\frac r2+1}\ll x^{\beta(r)}.\nonumber\end{aligned}$$ If $N_2\ll MN_1$, using the same approach but with $(M,N)=(MN_1,
N_2)$ in Lemma 4.6 we get that (4.27) still holds.
Combining (4.26) and (4.27), we obtain that $$\begin{aligned}
S_r(M,Y) \ll x^{\beta(r)+\varepsilon}\end{aligned}$$ holds in the Case ii.
The estimate (4.22) now follows from the proofs of the above two cases. .
**An application of Theorem 2**
-------------------------------
The exponential convolution(e-convolution) was introduced by M. V. Subbarao[@su1]. Let $n>1$ be an integer of canonical form $n=p^{a_1}_{1}\cdots
p^{a_s}_{s}$. An integer $d$ is called an exponential divisor (e-divisor) of $n$ if $d=p^{b_1}_{1}\cdots p^{b_s}_{s}$, where $b_1|a_{1},\cdots,b_s|a_{s}$. Let $\tau^{(e)}(n)$ denote the number of exponential divisors of $n$, which is called the exponential divisor function. Let $r\geq 2 $ be a fixed integer. The integer $n>1$ is called exponentially $r$-free (e-$r$-free) if all the exponents $a_1,\cdots, a_s$ are $r$-free. Let $q_r^{(e)}$ denote the characteristic function of the set of e-r-free integers. The e-unitary convolution was introduced by N. Minculete and L. Tóth [@mt]. The function $I(n)=1 (n\ge 1)$ has inverses with respect to e-convolution and e-unitary convolution denoted by $\mu^{(e)}(n)$ and $\mu^{(e)*}(n)$, respectively. These are the unitary and exponential analogues of the Möbius function. These arithmetic functions attract the interests of many authors, see for example [@cz2; @ha; @ks1; @ks2; @lu; @mt; @pw; @sa1; @sa2; @sw2; @ss; @to1; @to2; @to3].
L. Tóth [@to2] showed that the Dirichlet series of $\mu^{(e)}$ is of the form $$\begin{aligned}
\sum_{n=1}^{\infty}\frac{\mu^{(e)}(n)}{n^s}=\frac{\zeta
(s)}{\zeta^2(2s)}W_1(s), \Re s >1,\end{aligned}$$ where $W_1(s):=\sum_{n=1}^{\infty}\frac{w_1(n)}{n^s}$ is absolutely convergent for $\Re s>\frac {1}{5}$.
Let $$\begin{aligned}
\Delta_{\mu^{(e)}} (x):=\sum_{n\le x}\mu^{(e)}(n)-A_1x,\end{aligned}$$ where $$\begin{aligned}
A_1:=m(\mu^{(e)})=\frac{W_1(1)}{\zeta^2(2)}=\prod_{p}\left(1+\sum_{n=2}^{\infty}\frac
{\mu(n)-\mu(n-1)}{p^n}\right).\end{aligned}$$
L. Tóth [@to2] showed $ \Delta_{\mu^{(e)}} (x)=O\left(
x^{\frac{91}{202}+\varepsilon} \right)$ under RH. The exponent $\frac{91}{202}$ was improved to $\frac {37}{94}$ by X. Cao and W. Zhai[@cz2].
Similarly, N. Minculete and L. Tóth[@mt] showed that the Dirichlet series of $\mu^{(e)*}$ is of the form $$\begin{aligned}
\sum_{n=1}^{\infty}\frac{\mu^{(e)*}(n)}{n^s}=\frac{\zeta
(s)}{\zeta^2(2s)}W_2(s), \Re s >1,\end{aligned}$$ where $W_2(s):=\sum_{n=1}^{\infty}\frac{w_2(n)}{n^s}$ is absolutely convergent for $\Re s>\frac {1}{4}$. Let $$\begin{aligned}
\Delta_{\mu^{(e)*}} (x):=\sum_{n\le x}\mu^{(e)*}(n)-A_2x,\end{aligned}$$ where $$\begin{aligned}
A_2:=m(\mu^{(e)*})=\frac{W_2(1)}{\zeta^2(2)}=\prod_{p}\left(1+\sum_{n=2}^{\infty}\frac
{(-1)^{\omega(n)}-(-1)^{\omega(n-1)}}{p^n}\right).\end{aligned}$$
Under RH, the estimate $ \Delta_{\mu^{(e)*}} (x)=O\left(
x^{\frac{91}{202}+\varepsilon} \right)$ was proved in [@mt].
As an application of Theorem 2 and Theorem 5, from (4.29) and (4.31) we get the following
Let $ \Delta_{\mu^{(e)}} (x)$ and $\Delta_{\mu^{(e)*}} (x)$ be defined by (4.30) and (4.32), respectively. If RH is true, then we have $$\begin{aligned}
\Delta_{\mu^{(e)}} (x)=O\left( x^{\frac{7}{22}+\varepsilon}
\right)\end{aligned}$$ and $$\begin{aligned}
\Delta_{\mu^{(e)*}} (x)=O\left(
x^{\frac{7}{22}+\varepsilon} \right).\end{aligned}$$
For comparison, we have numerically that $$\frac{91}{202}=0.45049\cdots,\ \ \frac
{37}{94}=0.39361\cdots,\ \ \frac
{7}{22}=0.31818\cdots.$$
**Some applications of Corollary 1.1 and Corollary 1.3**
=========================================================
**The distribution of generalized square-full integers**
--------------------------------------------------------
In 1963, E. Cohen [@co1] generalized square-full integers in the following way: Let $a $ and $b $ are fixed positive integers. Let $n>1$ be an integer of canonical form $n=p^{a_1}_{1}\cdots
p^{a_r}_{r}$ and $R_{a,b}$ denote the set of all $n$ such that each exponent $a_i (1\le i\le r)$ is either a multiple of $a$ or is contained in the progression $at+b(t\ge 0).$ Obviously $R_{2,3}$ is the set of square-full integers. Let $a\nmid b$ , $f_{(a,b)}(n)$ denote the characteristic function of the set $R_{a,b}$. By Lemma 2.1 in E. Cohen[@co1] one has $$\begin{aligned}
\sum_{n=1}^{\infty}\frac{f_{(a,b)}(n)}{n^s}=\frac{\zeta (as)\zeta
(bs)}{\zeta(2bs)}, \Re s >1.\end{aligned}$$ We are interested in the summatory function of $f_{a,b}(n).$
First consider the case $a<b$. Suppose also that $a\nmid b$. In this case the problem is closely related to the estimate of $\Delta(a,b;x).$ We take $a=a, b=b, c=2b$ and $k=1$ in (1.1), then the estimate (1.4) implies $\Delta(a,b;2b;x)=\Omega(x^{\frac
{1}{2(a+b)}})$. Suppose $\Delta(a,b;x)\ll
x^{\alpha{(a,b)}+\varepsilon}$ such that $\alpha{(a,b)}<1/2b$. By Corollary 1.1 with $(a,b,c,k)=(a,b,2b,1)$, under the RH we have the asymptotic formula $$\begin{aligned}
\sum_{n\le x}f_{(a,b)}(n)=\frac {\zeta (\frac ba)}{\zeta (\frac
{2b}{a})}x^{\frac 1a}+\frac {\zeta (\frac ab)}{\zeta ( 2)}x^{\frac
1b}+ O\left(x^{\frac {1-a\alpha (a,b)}{a+2b-4ab\alpha
(a,b)}+\varepsilon}\right),\end{aligned}$$ which improves Theorem 3.2 of [@su2].
The distribution of square-full integers (the case $a=2, b=3$) has received special attention. In this special case, the error term in (5.2) becomes $(x^{11/72+\varepsilon})$, which was first proved in [@ns]. The exponent $11/72$ was improved by several authors. The best known result is duo to Wu[@wu2], who obtained exponent $\frac{12}{85}=0.1411\cdots$(also subject to the RH).
Now we suppose $b<a.$ From (1.4) we get $\Delta(b,a;2b;x)=\Omega(x^{\frac
{1}{4b}})$. So without the loss of generality, we always suppose $b<a<4b.$
If $b<a< 2b$. Applying Corollary 1.1 with $(a,b,c,k)=(b,a,2b,1)$, under the RH one has $$\begin{aligned}
\sum_{n\le x}f_{(a,b)}(n)=\frac {\zeta (\frac ab)}{\zeta (
2)}x^{\frac 1b}+\frac {\zeta (\frac ba)}{\zeta (\frac
{2b}{a})}x^{\frac 1a}+O\left(x^{\frac {1-b\alpha
(a,b)}{3b-4b^2\alpha (a,b)}+\varepsilon}\right),\end{aligned}$$ which improves Theorem 3.4 of [@su2].
Finally look at the case $2b<a<4b.$ In this case, D. Suryanarayana[@su2] proved that(see Remark 3.3 therein) $$\sum_{n\le x}f_{(a,b)}(n)=\frac {\zeta (\frac ab)}{\zeta (
2)}x^{\frac 1b}+O(x^{1/2b} \delta(x))$$ unconditionally. If RH is true, then Remark 3.4 of D. Suryanarayana[@su2] claimed that $$\begin{aligned}
\sum_{n\le x}f_{(a,b)}(n)=\frac {\zeta (\frac ab)}{\zeta (
2)}x^{\frac 1b}+O\left(x^{\frac {2a-b}{5ab-4b^2}}\omega (x)\right),\end{aligned}$$ where $\omega (x)=exp\{A\log x(\log\log x)^{-1}\}$, $A$ is a positive absolute constant. Here we note that on the right-hand side of (5.5) the second main term $\frac {\zeta (\frac ba)}{\zeta
(\frac {2b}{a})}x^{\frac 1a}$ is absorbed into the error term .
Applying Corollary 1.3 with $(a,b,c,k)=(b,a,2b,1)$ and $\Delta (b;2b;x)\ll x^{\frac
{17}{54b}+\varepsilon}$ in Corollary 4.1, we have under RH that $$\begin{aligned}
\sum_{n\le x}f_{(a,b)}(n)=\frac {\zeta (\frac ab)}{\zeta (
2)}x^{\frac 1b}+\frac {\zeta (\frac ba)}{\zeta (\frac
{2b}{a})}x^{\frac 1a}+O\left( x^{\frac
{54}{37a+54b}+\varepsilon}\right),\end{aligned}$$ which took the second main term $\frac {\zeta (\frac ba)}{\zeta
(\frac {2b}{a})}x^{\frac 1a}$ out of the error term in (5.4) when $2b<a<54b/17=3.176\cdots b.$
**On the order of the error function of the $(l, r)$-integers**
---------------------------------------------------------------
For given integers $l,r$ with $1<r<l$, we say an integer $n$ is a $(l,
r)$-integers if it has the form $m^ln$ where $m, n $ are integers and $n$ is $r$-free. The definition of the $(l, r)$-integers was introduced by M. V. Subbarao and V. C. Harris[@sh]. Let $g_{l,r}(n)$ denote the characteristic function of the set of $(l,
r)$-integers. By Lemma 2.6 in M. V. Subbarao and D. Suryanarayana[@ss1] we have $$\begin{aligned}
\sum_{n=1}^{\infty}\frac{g_{l,r}(n)}{n^s}=\frac{\zeta (s)\zeta
(ls)}{\zeta(rs)}, \Re s >1.\end{aligned}$$ Hence $g_{l,r}(n)=f_{1,l;r}(n).$ We define the error term by $$\begin{aligned}
\Delta(1,l;r;x):=\sum_{n\le x}g_{l,r}(n)-\frac {\zeta (l)}{\zeta
(r)}x-\frac {\zeta(\frac 1l)}{\zeta (\frac rl)}x^{\frac 1l}.\end{aligned}$$ From (1.4) we have $\Delta(1,l;r;x):=\Omega(x^{\frac {1}{2r}})$.
If $l\ge 2r$, the distribution of $(l,r)$-integers is almost the same as the distribution of $r$-free numbers. From Theorem 4 we get under RH that $$\begin{aligned}
\sum_{n\le x}g_{l,r}(n)=\frac {\zeta (l)}{\zeta
(r)}x+O\left(x^{\alpha(r)+\varepsilon}\right),\end{aligned}$$ where $\alpha(r)$ is defined by (4.8) and (4.9). In particular, if $l\ge 4$ we have $\Delta(1,l;2;x)=O(x^{\frac
{17}{54}+\varepsilon})$.
If $r<l< 2r$ and $\alpha(r)\geq 1/l$ we get that (5.8) holds too. However, if $\alpha(r)<1/l,$ by Corollary 1.3 with $(a,b,c,k)=(1,l,r,1)$, we get under RH that $$\begin{aligned}
\sum_{n\le x}g_{l,r}(n)=\frac {\zeta (l)}{\zeta (r)}x+\frac
{\zeta(\frac 1l)}{\zeta (\frac rl)}x^{\frac 1l}+O\left(x^{\frac
{1}{l+1-l\alpha(r)}+\varepsilon}\right).\end{aligned}$$ In particular, $\Delta(1,3;2;x)=O(x^{\frac {18}{55}+\varepsilon})$.
The previously best known error term is due to M. V. Subbarao and D. Suryanarayana[@ss2]. Since $\frac
{18}{55}=0.32727\cdots<\frac 13$, (5.8) and (5.9) answer a conjecture of M. V. Subbarao and D. Suryanarayana in [@ss1](see page 123) for the special case $r=2$. It should be noted that if $2\le r\le 10, l=r+1$, we get the second main term; if $r\ge 11,
r<l=r+2$, we also get the second main term. Hence (5.9) is an substantial improvement to theirs.
**The distribution of e-$r$-free integers**
-------------------------------------------
In this subsection we consider the distribution of e-$r$-free integers. For the distribution of e-square-free integers, J. Wu[@wu1] showed that $\Delta_{q_2^{(e)}} (x)=O\left(x^{\frac
14}\delta (x)\right)$, improving an earlier result of M. V. Subbarao[@su1]. In the general case, L. Tóth[@to1] obtained that $\Delta_{q_r^{(e)}} (x)=O\left(x^{\frac
{1}{2^r}}\delta(x)\right)$. Under RH, X. Cao and W. Zhai[@cz2] showed that $\Delta_{q_2^{(e)}} (x)=O(x^{\frac
{38}{193}+\varepsilon})$, improving the exponent $\frac 15$ of L. Tóth[@to2]. In this subsection we shall study this topic more carefully.
\(i) The Dirichlet series of $q_r^{(e)}$ is of form $$\begin{aligned}
Q_r^{(e)}(s):=\sum_{n=1}^{\infty}\frac{q_r^{(e)}(n)}{n^s}=\frac
{\zeta (s)\zeta((2^r+1)s)}{\zeta(2^rs)\zeta(2^{r+1}s)}U_r(s),\Re
s>1,\end{aligned}$$ where $U_r(s)$ is absolutely convergent for $\Re s>\frac
{1}{2^{r+1}+1}$.
\(ii) Let $$\begin{aligned}
\Delta_{q_r^{(e)}} (x):=\sum_{n\le
x}q_r^{(e)}(n)-C_1(r)x-C_2(r)x^{\frac{1}{2^r+1}},\end{aligned}$$ where $$\begin{aligned}
C_1(r):=\frac{\zeta(2^r+1)U_r(1)}{\zeta(2^r)\zeta(2^{r+1})},\ \
C_2(r):=\frac{\zeta(\frac{1}{2^r+1})U_r(\frac{1}{2^r+1})}{\zeta(\frac{2^r}{2^r+1})\zeta(\frac{2^{r+1}}{2^r+1})}.\end{aligned}$$ Then $$\begin{aligned}
\Delta_{q_r^{(e)}} (x)=\Omega\left(x^{\frac {1}{2^{r+1}}}\right).\end{aligned}$$
\(iii) If the RH is true, we have $$\begin{aligned}
\Delta_{q_r^{(e)}} (x)=O\left(x^{\frac
{1}{2^r+2-(2^r+1)\alpha(2^r)}+\varepsilon}\right),\end{aligned}$$ where $\alpha(r)$ is defined by (4.8) and (4.9). In particular, $\Delta_{q_2^{(e)}} (x)=O(x^{\frac {38}{193}+\varepsilon})$, here $\frac {38}{193}=0.1968\cdots<\frac 15$.
[**Remark 5.1.**]{} Note that we always have $\frac
{1}{2^r+2-(2^r+1)\alpha(2^r)}<\frac {1}{2^r+1}$. In addition when $r=2$, if we use the new estimate $\alpha (4)=\frac {17}{94}$ proved by R. C. Baker and K. Powell[@bp] recently, one can slightly improve the above result to $\Delta_{q_2^{(e)}}
(x)=O(x^{\frac {94}{479}+\varepsilon})$. In section 6 we shall improve the exponent $\frac {94}{479}$ further by the exponential sum method.
Since the function $q_r^{(e)}$ is multiplicative and $q_r^{(e)}(p^\alpha)=q_r(\alpha )$ for every prime power $p^\alpha$. For $r\ge 2$, it is easy to verify that $q_r^{(e)}(p)=q_r^{(e)}(p^2)=\cdots=q_r^{(e)}(p^{2^r-1})=1$, $q_r^{(e)}(p^{2^r})=0$, $q_r^{(e)}(p^{2^r+1})=\cdots=q_r^{(e)}(p^{2^{r+1}-1})=1$, and $q_r^{(e)}(p^{2^{r+1}})=0$. Hence for $\Re s>1$ $$\begin{aligned}
\sum_{n=1}^{\infty}\frac{q_r^{(e)}(n)}{n^s}=\prod_{p}\left(1+\sum_{m=1}^{\infty}\frac
{q_r(m)}{p^{ms}}\right)\end{aligned}$$ Applying the product representation of Riemann zeta-function $$\begin{aligned}
\zeta
(s)=\prod_{p}(1+p^{-s}+p^{-2s}+p^{-3s}+\cdots)=\prod_{p}(1-p^{-s})^{-1},\Re
s>1,\end{aligned}$$ we have for $\Re s>1$ $$\begin{aligned}
\zeta (s)\zeta
((2^r+1)s)=\prod_{p}\left((1-p^{-s})(1-p^{-(2^r+1)s})\right)^{-1}.\end{aligned}$$ Let $$\begin{aligned}
f_{q_r^{(e)}}(z):&=1+\sum_{m=1}^{\infty}q_r (m)z^m\\
&=1+z+\cdots+z^{2^r-1}+z^{2^r+1}+\cdots+z^{2^{r+1}-1}+
\sum_{m=2^{r+1}+1}^{\infty} q_r (m)z^m.\nonumber\end{aligned}$$
By a simple calculation one get for $|z|<1$ $$\begin{aligned}
(1-z)(1-z^{2^r+1})&= 1 - z - z^{2^r+1} + z^{2^r+2},\\
f_{q_r^{(e)}}(z)(1-z)(1-z^{2^r+1})&=1-z^{2^r}-z^{2^{r+1}}+\sum_{m=2^{r+1}+1}^{\infty}c_mz^m,\end{aligned}$$ and $$\begin{aligned}
&\left(1+z^{2^r}+z^{2(2^r)}+z^{3(2^r)}+\cdots\right)\left(1+z^{2^{r+1}}+z^{2(2^{r+1})}+z^{3(2^{r+1})}+\cdots\right)\\
&=1+z^{2^r}+2z^{2(2^r)}+2z^{3(2^r)}+\cdots.\end{aligned}$$
From the above two relations, we easily obtain for $ |z|<1$ $$\begin{aligned}
&&f_{q_r^{(e)}}(z)(1-z)(1-z^{2^r+1})(1-z^{2^r})^{-1}(1-z^{2^{r+1}})^{-1}\\
&&=1+\sum_{m=2^{r+1}+1}^{\infty}C_mz^m.\nonumber\end{aligned}$$
Taking $z=p^{-s}$ in (5.18), then combining (5.15), (5.16) and (5.17) completes the proof of (5.10) in Theorem 7.
Applying Theorem 2 of M. Küleitner and W. G. Nowak[@kn] and (5.10), we immediately get (5.13). By Corollary 1.3 with $(a,b,c,k)=(1,2^r+1,2^r,1)$ and Theorem 4, we obtain $$\begin{aligned}
\Delta (1,2^r+1;2^r;x)=O\left(x^{\frac
{1}{2^r+2-(2^r+1)\alpha(2^r)}+\varepsilon}\right).\end{aligned}$$ Now (5.14) follows from (5.10) , (5.19) and a simple convolution argument at once, and this completes the proof of Theorem 7.
**The divisor problem over the set of $r$-free numbers**
--------------------------------------------------------
Let $r\geq 2$ be a fixed integer. Winfried Recknagel[@re] and Hailiang Fen[@fen] investigated the divisor problem over the set of $r$-free numbers. Hailiang Fen[@fen] showed that $$\begin{aligned}
\Psi_r(s):=\sum_{n=1}^{\infty}\frac{\tau (n)q_r(n)}{n^s}=\frac
{\zeta^2 (s)}{\zeta^{r+1}(rs)}V_r(s),\Re s>1,\end{aligned}$$ where $V_r(s)$ is absolutely convergent for $\Re s>\frac
{1}{r+1}$. In particular, $V_2(s)=\zeta^2(3s)W_2(s)$, $W_2(s)$ is absolutely convergent for $\Re s>\frac {1}{4}$.
Let $$\begin{aligned}
\sum_{n\le x}\tau (n)q_r(n)=\operatorname*{Res}_{s=1}\Psi_r(s)\frac
{x^s}{s}+\Delta_{(r)}^\tau(x).\end{aligned}$$ Hence this problem is reduced to estimate the error term $\Delta_{r+1}(1,1,r;x)$. Hailiang Fen[@fen] showed that $$\begin{aligned}
\Delta_{r+1}(1,1,r;x)=\left\{\begin{array}{ll}
O(x^{1/r} \delta(x)),&\mbox{if $r=2,3$}\\
\Delta_{r+1}(1,1,r;x)=O(x^{\frac {131}{416}}(\log
x)^{\frac{26947}{8320}}) ,& \mbox{if $r\geq 4$},
\end{array}\right.\end{aligned}$$where the second estimate in (5.22) follows from M. N. Huxley’s bound(see [@hu2]) $$\Delta (1,1;x)\ll x^{\frac{131}{416}}(\log
x)^{\frac{26947}{8320}}.$$
Applying Corollary 1.1 with $(a,b,c,k)=(1,1,2,3)$ and $(a,b,c,k)=(1,1,3,4)$ respectively, and with the help of (5.23), we obtain under RH that $$\begin{aligned}
\Delta_{3}(1,1;2;x)=O(x^{\frac {285}{724}+\varepsilon}),\ \ \
\Delta_{4}(1,1;3;x)=O(x^{\frac {285}{878}+\varepsilon}).\end{aligned}$$ Finally, from (5.20) and (5.24) we get immediately the following
If RH is true, then $$\begin{aligned}
\Delta_{(2)}^\tau(x)=O(x^{\frac {285}{724}+\varepsilon}),\ \ \
\Delta_{(3)}^\tau(x)=O(x^{\frac {285}{878}+\varepsilon}).\end{aligned}$$
**The distribution of e-square-free integers**
==============================================
In this section we shall use the method of exponential sums, and $\alpha (4)=\frac {17}{94}$(recall Remark 4.1 in Section 4) proved by R. C. Baker and K. Powell[@bp] to prove
Let $\Delta_{q_2^{(e)}} (x)$ be defined by (5.11). Assume that RH holds, then $$\begin{aligned}
\Delta_{q_2^{(e)}} (x)=O\left(x^{\frac
{23}{124}+\varepsilon}\right).\end{aligned}$$
[**Remark 6.1.**]{} For comparison, we have $\frac
{23}{124}=0.18548\cdots$ and $\frac {94}{479}=0.19624\cdots$.
In the proof of Theorem 9 we need the following lemma (see Lemma 6.9 of Krätzel[@kr]).
Let $1\le Z<Z_1\le x^{\frac 1a}$, $a>0,b>0$. If $(\kappa,\lambda)$ is any exponent pair and if $$(2\lambda-1)a>2\kappa b,\ \ (2\lambda-1)b>2\kappa a,$$ then $$\begin{aligned}
\sum_{Z<n\le Z_1}\psi\left(\left(\frac {x}{n^a}\right)^{\frac
1b}\right)\ll x^{\frac {2(\kappa+\lambda-\frac 12)}{a+b}}\log
x+Z_1\left(\frac {Z_1^a}{x}\right)^{\frac 1b}\log x.\end{aligned}$$
[**Proof of Theorem 9.**]{} Similar to the proof of Theorem 7, we need only to prove $$\begin{aligned}
\Delta(1,5;4;x)=O\left(x^{\frac {23}{124}+\varepsilon}\right).\end{aligned}$$ Let $\alpha =\frac {23}{124}$. Applying Theorem 3 with $(a,b,c,k)=(1,5,4,1)$ and $y=x^{\frac {101}{744}}$, the following two estimates $$\begin{aligned}
\sum_{d\le y}\Delta (1,4;\frac {x}{d^5})\ll x^{\alpha+\varepsilon}\end{aligned}$$ $$\begin{aligned}
\sum_{m\le \frac {x}{y^5}}u_1(1;4;m)\psi\left((\frac x m)^{\frac
15}\right)\ll x^{\alpha+\varepsilon}\end{aligned}$$ would suffice to finish the proof of Theorem 9.
We first estimate the sum in (6.3). Let $y_1=x^{\frac {3}{62}}$, we split the sum in (6.3) into two parts and write
$$\begin{aligned}
\sum_{d\le y}\Delta (1,4;\frac {x}{d^5})=\sum_{d\le y_1}\Delta
(1,4;\frac {x}{d^5})+\sum_{y_1<d\le y}\Delta (1,4;\frac
{x}{d^5}):=S_1+S_2.\end{aligned}$$
Clearly, it follows from $\alpha (4)=\frac {17}{94}$ that $$\begin{aligned}
S_1\ll \sum_{d\le y_1}\left(\frac {x}{d^5}\right)^{\frac
{17}{94}+\varepsilon}\ll x^{\frac {17}{94}+\varepsilon}y_1^{\frac
{9}{94}}\ll x^{\alpha+\varepsilon}.\end{aligned}$$ To estimate $S_2$, we discuss two cases.
[**Case (i)**]{} $y_1\ll D\le y^{\frac {843}{10912}}=y_2$. Applying Theorem 2 with $Y_*=x^{\frac 15}D^{-\frac {4}{5}}$ and a simple splitting argument, we have for some $1\ll N\ll Y_*$ $$\begin{aligned}
\sum_{d\sim D}\Delta (1,4;\frac {x}{d^5})&=\sum_{d\sim
D}\left(\sum_{n\le Y_*}\mu (n)\psi\left(\frac
{x}{d^5n^4}\right)+O\left((\frac {x}{D^5})^{\frac
12+\varepsilon}{Y_*}^{-\frac
32}+{Y_*}^{\frac 12+\varepsilon}\right)\right)\\
&\ll \mathcal{L}\left|\sum_{d\sim D}\sum_{n\sim N}\mu
(n)\psi\left(\frac {x}{d^5n^4}\right)\right|+x^{\frac
15+\varepsilon}D^{-\frac {3}{10}}+x^{\frac
{1}{10}+\varepsilon}D^{\frac {3}{5}}\nonumber\end{aligned}$$ Applying Lemma 4.6 with $(\kappa,\lambda)=(\frac 12,\frac 12)$ and $(M,N)=(N,D)$, one get $$\begin{aligned}
&\mathcal{L}^{-6}\sum_{d\sim D}\sum_{n\sim N}\mu
(n)\psi\left(\frac
{x}{d^5n^4}\right)\\
&\ll \left(\left(\frac {x}{D^5N^4}\right)^\kappa
N^{1+\kappa+\lambda}D^{2+\kappa}\right)^{\frac{1}{2+2\kappa}}+D^{\frac
12}N+DN^{\frac 12}+x^{-\frac 12}D^{\frac 72}N^{3}\nonumber\\
&\ll x^{\frac 16}+x^{\frac 15}D^{-\frac
{3}{10}}+x^{\frac {1}{10}}D^{\frac {3}{5}}+x^{\frac {1}{10}}D^{\frac {11}{10}}\nonumber\\
& \ll x^{\frac 15}y_1^{-\frac {3}{10}}+x^{\frac {1}{10}}y_2^{\frac
{11}{10}} \ll x^{\alpha+\varepsilon}.\nonumber\end{aligned}$$ Combining (6.7) and (6.8), in this case one obtain $$\begin{aligned}
\sum_{d\sim D}\Delta (1,4;\frac {x}{d^5})\ll
x^{\alpha+\varepsilon}.\end{aligned}$$ [**Case (ii)**]{} $y_2=y^{\frac {843}{10912}}\ll D\le y$. Applying Theorem 2 with $Y_*=x^{\frac {13}{62}}D^{-1}$, and a simple splitting argument, we have for some $1\ll N\ll Y_*$ $$\begin{aligned}
\sum_{d\sim D}\Delta (1,4;\frac {x}{d^5})&=\sum_{d\sim
D}\left(\sum_{n\le Y_*}\mu (n)\psi\left(\frac
{x}{d^5n^4}\right)+O\left((\frac {x}{D^5})^{\frac
12+\varepsilon}{Y_*}^{-\frac
32}+{Y_*}^{\frac 12+\varepsilon}\right)\right)\\
&\ll \mathcal{L}\left|\sum_{d\sim D}\sum_{n\sim N}\mu
(n)\psi\left(\frac
{x}{d^5n^4}\right)\right|+x^{\alpha+\varepsilon}+x^{\frac
{13}{124}+\varepsilon}D^{\frac 12}.\nonumber\end{aligned}$$ Now applying lemma 4.2 with $H=DNx^{-\alpha }$ and Lemma 4.3 with $(\kappa,\lambda)=BA^2BA(\frac 16,\frac 46)=(\frac
{13}{40},\frac {22}{40})$, we easily obtain that $$\begin{aligned}
&\sum_{d\sim D}\sum_{n\sim N}\mu (n)\psi\left(\frac
{x}{d^5n^4}\right)\ll \sum_{n\sim N}\left| \sum_{d\sim
D}\psi\left(\frac {x}{d^5n^4}\right)\right|\\
&\ll \sum_{n\sim N}\left(\frac DH+\sum_{1\le |h|\le H}\frac
{1}{|h|}\left|\sum_{d\sim D}e\left(\frac
{hx}{d^5n^4}\right)\right|\right)\nonumber\\
&\ll x^{\alpha}+\sum_{n\sim N}\sum_{1\le h\le H}\frac 1h\left(
\left(\frac {hx}{n^4D^6}\right)^\kappa D^\lambda+\left(\frac
{hx}{n^4D^6}\right)^{-1}\right)\nonumber\\
&\ll x^{\alpha}+\left(\frac {x^{1-\alpha}}{N^3D^5}\right)^\kappa
D^\lambda N+x^{-1}N^5D^6\nonumber\\
&\ll x^{\alpha}+x^{\frac {13}{62}+(\frac
{23}{62}-\alpha)\kappa}D^{\lambda-2\kappa-1}+x^{\frac
{3}{62}}D\nonumber\\
&\ll x^{\alpha}+x^{\frac {1339}{4960}}y_2^{-\frac
{11}{10}}+x^{\frac {3}{62}}y \ll x^{\alpha}.\nonumber\end{aligned}$$ (Here we use $1-3\kappa>0$.) Combining (6.10) and (6.11), one also has $$\begin{aligned}
\sum_{d\sim D}\Delta (1,4;\frac {x}{d^5})\ll
x^{\alpha+\varepsilon}.\end{aligned}$$
Combining the above two cases, then using a simple splitting argument we obtain $$\begin{aligned}
S_2 \ll x^{\alpha+\varepsilon}.\end{aligned}$$
Hence (6.3) follows from (6.5), (6.6) and (6.13).
Now we turn to prove (6.4). From (1.14), applying the Drichlet’s hyperbolic argument, we have for any $1<Z< \frac {x}{y^5}=Y$ $$\begin{aligned}
&\sum_{m\le \frac {x}{y^5}}u_1(1;4;m)\psi\left((\frac x m)^{\frac
15}\right)=\sum_{m_1m_2^4\le Y}\mu
(m_2)\psi\left((\frac {x}{ m_1m_2^4})^{\frac 15}\right)\\
&=\sum_{m_1\le Z}\sum_{m_2^4\le \frac {Y}{m_1}}\mu
(m_2)\psi\left((\frac {x}{ m_1m_2^4})^{\frac 15}\right)
+\sum_{m_2\le (\frac{Y}{Z})^{\frac 14}}\mu (m_2)\sum_{Z<m_1\le
\frac {Y}{m_2^4}}\psi\left((\frac {x}{ m_1m_2^4})^{\frac
15}\right).\nonumber\end{aligned}$$
Applying Lemma 6.1 with $(\kappa,\lambda)=A^2BA(\frac 16,\frac
46)=(\frac {2}{40},\frac {33}{40})$ and $(a,b,Z_1)=(1,5,\frac
{Y}{m_2^4})$, we have $$\begin{aligned}
\sum_{Z<m_1\le \frac {Y}{m_2^4}}\psi\left((\frac {x}{
m_1m_2^4})^{\frac 15}\right)\ll \left(\frac
{x}{m_2^4}\right)^{\frac 18}\mathcal{L} +\frac
{Y}{m_2^4}\left(\frac {Y} {x}\right)^{\frac 15}\mathcal{L}.\end{aligned}$$
On taking $Z=x^{\frac 17}Y^{-\frac 17}$, it follows from (6.14) and (6.15) $$\begin{aligned}
&\sum_{m\le \frac {x}{y^5}}u_1(1;4;m)\psi\left((\frac x m)^{\frac
15}\right)\\
&\ll \sum_{m_1\le Z}\left(\frac {Y}{m_1}\right)^{\frac
14}+\sum_{m_2\le (\frac{Y}{Z})^{\frac 14}}\left(\left(\frac
{x}{m_2^4}\right)^{\frac 18} +\frac {Y}{m_2^4}\left(\frac {Y}
{x}\right)^{\frac 15}\right)\mathcal{L}\nonumber\\
&\ll Y^{\frac 14}Z^{\frac 34}+x^{\frac 18}Y^{\frac 18}Z^{-\frac
18}\mathcal{L}+Y^{\frac 65}x^{-\frac 15}\mathcal{L}\nonumber\\
&\ll x^{\frac {3}{28}}Y^{\frac 17}\mathcal{L}+Y^{\frac
65}x^{-\frac 15}\mathcal{L}\ll x^{\frac 14}y^{-\frac
57}\mathcal{L}+xy^{-6}\mathcal{L}\ll x^{\alpha
}\mathcal{L}^2.\nonumber\end{aligned}$$ This completes the proof of Theorem 9.
[s2]{}
R. C. Baker, The square-free divisor problem, Quart. J. Oxford, 45(2)(1994),269–277.
R. C. Baker, The square-free divisor problem II, Quart. J. Oxford, 47(2)(1996),133–146.
R. C. Baker and K. Powell, The distribution of $k$-free numbers, Acta Math. Hungar., 126(1-2)(2010),181–197.
R. Balasubramanian, K. Ramachandra and M. V. Subbarao , On the error function in the asymptotic formula for the counting function of k-full numbers , Acta Arith., 50(1988), 107–118.
Xiaodong Cao, The distribution of square-full integers, Period. Math. Hung., 28(1)(1994), 43–54; ibid., 34(3)(1997), 169–175.
Xiaodong Cao and Wenguang Zhai, Multiple exponential sums with monomials, Acta Arith., 92(3)(2000),195–213.
Xiaodong Cao and Wenguang Zhai, On certain arithmetical functions involving exponential divisors, Journal of Integer Sequences, Vol. 13 (2010),Article 10.3.7.
E. Cohen, Arithmetrial notes II, An estimate of Erdös and Szekers, Scripta Mathematica, 26(1963),353–356.
E. Cohen, On the distribution of certain sequences of integers, Amer. Math. Monthly, 70(1963),519–521.
E. Cohen, The average order of certain types of arithmetical functions: generalized $k$-free numbers and totient points, Monatsh. Math., 64(1960),251–262.
E. Cohen and K. J. Davis, Elementary estimates for certain types of integers, Acta Sci. Math., (Szeged)31(1970),363–371.
Hailiang Fen, The divisor problem over the set of $r$-free numbers, Master Degree Thesis, Shandong Normal University, 2006.
E. Fouvry and H. Iwaniec, Exponential sums with monomials, J. Number Theory, 33(3)(1989), 311–333 .
Jun Furuya and Wenguang Zhai, On the $k$-free divisor problem, Acta Arithmetica, Vol. 123, no. 3, 267–287.
Jun Furuya and Wenguang Zhai, On the $k$-free divisor problem(II), Acta Arithmetica, Vol. 132, no. 4, 351–358.
S. W. Graham and J. Pintz, The distribution of r-free numbers, Acta Math. Hungar,53(1-2)(1989),213–236.
P. Jr. Hagis, Some results concerning exponential divisors, Internat. J. Math. Math. Sci., 11(2) (1988), 343–349.
M. N. Huxley, Area, Lattice Points, and Exponential Sums, Oxford University Press, New York, 1996.
M. N. Huxley, Exponential sums and lattice points III, Proc. London Math. Soc., 87(3)(2003), 591–609.
A. Ivić, The Riemann Zeta-Function, John. Wiley & Sons, 1985.
A. Ivić, E. Krätzel, M. Kühleitner and W. G. Nowak, Lattice points in large regions and related arithmetic functions: recent developments in a very classic topic. (English summary) Elementare und analytische Zahlentheorie, 89¨C128, Schr.Wiss. Ges. Johann Wolfgang Goethe Univ.Frankfurt am Main, 20, Franz Steiner Verlag Stuttgart, Stuttgart, 2006.
Chaohua Jia, The distribution of square-free numbers, Sci. China Ser. A, 36(2)(1993),154–169.
I. Kátai and M. V. Subbarao, On the iterates of the sum of exponential divisors, Math. Pannon., 10(2) (1999), 153–158.
I. Kátai and M. V. Subbarao, On the distribution of exponential divisors, Ann. Univ. Sci. Budapest. Sect. Comput., 22 (2003), 161–180.
E. Kr" atzel, Lattice Points, Kluwer, Dordrecht-Boston-London, 1988.
M. Küleitner and W. G. Nowak, An omega theorem for a class of arithmetic functions, Math. Nachr., 165(1994), 79–98.
A. Kumchev, The k-free divisor problem, Monatsh. Math., 129(2000),321–327.
L. Lucht, On the sum of exponential divisors and its iterates, Arch. Math. (Basel), 27(4)(1976), 383–386.
A. Mercier and W. G. Nowak, Problème des diviseurs pour des valeurs polynômiales II, Ann. Sci. Math. Québec, 16(1),(1992),81–94.
N. Minculete and L. Tóth, Exponential unitary divisors,Ann. Univ. Sci. Budapest. Sect. Comput., 35(2011), 205–216.
H. L. Montgomery and R. C. Vaughan, On the distribution of square-free numbers, in Recent Progress in Analytic Number Theory, Vol.1, 247—256, Academic Press, London, 1981.
W. G. Nowak and M. Schmeier, Conditional asymptotic fomula for a class of arithmetic functions, Proc. Amer. Math. Soc., vol 103(3), 1988, 713–717.
W. G. Nowak, On the average order of arithmetic functions connected with the Riemann zeta-function, in Analysis, Geometry and Groups: A Riemann Legacy Volume(editted by H.M. Srivastava and Th.M. Rassias ), Pages 449–470, Hadronic Press Collect. Orig. Artic., Hadronic Press, Palm Harbor, Florida ,1993
Pan Chengdong and Pan Chengbiao, Elements of Analytic Number Theory(in Chinese), Science press, Beijing, China, 1991.
Y.-F.S. Pétermann and J. Wu, On the sum of exponential divisors of an integer, Acta Math. Hungar., 77(1-2) (1997), 159–175.
Winfried Recknagel, Über k-freie Zahlen als Summe von zwei Quadraten, Arch. Math., 52(1989), 233—236.
H.-E. Richert, Über die Anzahl Abelscher Gruppen gegebener Ordnung. I, Math. Zeit., 56(1952), 21–32.
O. Robert and P. Sargos, Three-dimensiaonal exponential sums with monomials, J. reine angew. Math., 591(2006), 1–20.
J. Sándor, On an exponential totient function, Sudia Univ. Babes-Bolyai, Math., 41(1996),91–94.
J. Sándor, A note on exponential divisors and related arithmetic functions, Sci. Magna, 1(1)(2005), 97–101.
P. Sargos and J. Wu, Multiple exponential sums with monomials and their applications in number theory, Acta Math. Hungar., 87(4),(2000),1107–1126.
A. Smati and J. Wu, On the exponential divisor function, Publ. Inst. Math. (Beograd) (N.S.), 61(75)(1997), 21–32.
E. G. Straus and M. V. Subbarao, On exponential divisors, Duke Math. J., 41(1974), 465–471.
M. V. Subbarao and V. C. Harris, A new generalization of Ramanujan’s sum, J. London Math. Soc., 41(1966),595–604.
M. V. Subbarao, On some arithmetic convolutions, in The theory of arithmetic functions, Lecture Notes in Mathematics, No. 251,247–271, Springer, 1972.
M. V. Subbarao and D. Suryanarayana, On the order of the error function of the $(k,r)$-integers, J. Number Theory, 6(1974),112–123.
M. V. Subbarao and D. Suryanarayana, On the order of the error function of the $(k,r)$-integers, Canad. Math. Bull., 20(3)(1977),397–399.
D. Suryanarayana, On the distribution of some generalized square-full integers, Pacific J. Mathematics, 72(2)(1977),547–555.
D. Suryanarayana, On the orders of the error functions of the $(k,
r)$ and $r$-free integers. Proceedings of the Conference on Number Theory (Mysore, 1979), pp. 157–204, Matscience Rep., 101, Aarhus Univ., Aarhus, 1980.
E. C. Titchmarsh, The Theory of the Riemann Zeta-Function, Oxford University Press, Oxford 1951.
L. Tóth, On certain arithmetic functions involving exponential divisors, Ann. Univ. Sci. Budapest. Sect. Comput., 24(2004), 285–294.
L. Tóth, On certain arithmetical functions involving exponential divisors II, Ann. Univ. Sci. Budapest. Sect. Comput., 27(2007), 155–166.
L. Tóth, An order result for the exponential divisor function, Publ. Math. Debrecen, 71(1-2)(2007), 165–171.
J. D. Vaaler, Some extremal problems in Fourier analysis, Bull. Amer. Math. Soc., 12(1985),183–216.
Jie Wu, Problème de diviseurs exponentiels et entiers exponentiellement sans facteur carré , J. Théor. Nombres Bordeaux, 7(1)(1995), 133–141.
Jie Wu, On the distribution of square-full integers, Arch. Math. (Basel), 77(3)(2001), 233—240.
Yan Kwang Feng, Some representation and distribution problems for generalized $r$-free integers, Ph.D. Thesis, University of Alberta, 1970. Wenguang Zhai and Xiaodong Cao, On the mean square of the error term for the asymmetric two-dimensional divisor problems (I), Monatsh. Math., 159(2010),185—209.
Authors’ addresses:
Xiaodong Cao\
Dept. of Mathematics and Physics,\
Beijing Institute of Petro-Chemical Technology,\
Beijing, 102617, P. R. China\
Email: [email protected]
Wenguang Zhai,\
Department of Mathematics,\
China University of Mining and Technology,\
Beijing 100083, P. R. China\
E-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider continuous–time Markov kinetics with a finite number of states and a positive equilibrium $P^*$. This class of systems is significantly wider than the systems with detailed balance. Nevertheless, we demonstrate that for an arbitrary probability distribution $P$ and a general system there exists a system with detailed balance and the same equilibrium that has the same velocity ${{\mathrm{d}}}P / {{\mathrm{d}}}t$ at point $P$. The results are extended to nonlinear systems with the generalized mass action law.'
address: 'Department of Mathematics, University of Leicester, Leicester, LE1 7RH, UK'
author:
- 'Alexander N. Gorban'
title: |
Local equivalence of reversible\
and general Markov kinetics
---
detailed balance ,Lyapunov function ,decomposition ,entropy ,uncertainty 02.50.Ga ,05.20.Dd
Introduction
============
Detailed balance and beyond \[Overview\]
----------------------------------------
The principle of detailed balance is one of the most celebrated results in kinetics. A kinetic system is represented as a mixture of independent elementary processes (collisions or elementary reactions, for example). Due to the principle of detailed balance, [*at equilibrium, each elementary process should be equilibrated by its reverse process.*]{} Kinetics is decomposed into pairs of mutually inverse processes and in many problems we can consider these pairs separately.
We study relations between the systems with and without detailed balance. In this Section, we briefly overview the main results of the work. Then, in Sec. \[History\] we review the history of the problem. We prove the local equivalence theorem for the Markov processes in Sec. \[Markov\] and give there the simple examples. The nonlinear generalizations are presented in Sec. \[Nonlin\].
In Sec. \[Markov\] we start from the first order kinetics without the detailed balance assumption. The general first order kinetic equation has the form: $$\label{MAsterEq0}
\frac{{{\mathrm{d}}}p_i}{{{\mathrm{d}}}t}= \sum_{j, \, j\neq i} (q_{ij}p_j-q_{ji}p_i) \, ,$$ where $q_{ij}$ ($i,j=1,\ldots, n$, $i\neq j$) are non-negative. This system of equations (master equations or Kolmogorov’s equations) describes dynamics of non-negative variables $p_i$ ($i=1,\ldots, n$). These variables may be considered as probabilities (then $\sum_i p_i =1$) or concentrations. For the corresponding states or components we use the notation $A_i$. In this notation, $q_{ij}$ is the [*rate constant*]{} for transitions $A_j \to A_i$.
Let us assume that system (\[MAsterEq0\]) has a positive equilibrium $P^*=(p_i^*)$, $p_i^*>0$: $$\label{MasterEquilibrium}
\sum_{j, \, j\neq i} q_{ij}p^*_j = \left(\sum_{j, \, j\neq i}
q_{ji}\right)p^*_i \;\mbox{ for all } i=1,\ldots n \, .$$ This is the so-called [*balance equation*]{}.
The [*detailed balance*]{} condition is much stronger. It assumes that the sums in the left and right hand sides of Eq. (\[MasterEquilibrium\]) are equal term by term: $$\label{detBal}
q_{ij}p^*_j=q_{ji}p^*_i \;\mbox{ for all } i,j=1,\ldots n, \, i \neq j \, .$$
For the number of states $n>2$, a simple comparison of dimensions demonstrates that there are much more general systems with the given positive equilibrium $P^*$ (\[MasterEquilibrium\]) (dimension is $(n-1)^2$) than the systems with detailed balance with equilibrium $P^*$ (\[detBal\]) (dimension is $\frac{n(n-1)}{2}$). Surprisingly, [*for every given distribution $P$, the set of possible velocities ${{\mathrm{d}}}P / {{\mathrm{d}}}t$ for general Markov kinetics with equilibrium $P^*$ is the same that for Markov kinetics with detailed balance and the same equilibrium.*]{} This is the central result of the paper (Theorem \[Theorem1\] in Sec. \[Markov\]).
We demonstrate this in two steps. First, we use the representation of a general Markov chain with a given positive equilibrium as a combination with non-negative coefficients of several simple cycles with the same equilibrium.
Secondly, we demonstrate this equivalence for a simple cycle of transitions with positive constants $$\label{cycle}
A_1 \to A_2 \to \ldots \to A_n \to A_1 \, .$$ For the equilibrium $P^*$ the constants of the cycle are $q_{i+1 i}=\kappa
/p_i^*$ (we use here the standard convention about the cyclic numeration, $n+1=1$).
Thus, if we observe the Markov kinetics at one point then we can not distinguish general systems from systems with detailed balance because the sets of possible velocities coincide. In particular, they have the same set of Lyapunov functions.
Our main results allow us to decompose any Markov kinetics (or generalized mass action law kinetics with semi-detailed balance) into pairs of mutually inverse elementary processes with the same equilibrium. If the system does not satisfy the principle of detailed balance then this decomposition depends on the state. Nevertheless, in some problems it is still convenient to consider these pairs separately.
In this paper, we give two examples of the application of Theorem \[Theorem1\]: the evaluation of logarithmic decrement for general Markov chains and a simple proof of the Morimoto $H$-theorem for all the Csiszár–Morimoto divergencies. We give also the nonlinear generalization of Theorem \[Theorem1\] for the systems which obey the generalized [*Mass Action Law*]{} (MAL).
Master equation is a source for many other kinetic equations. In particular, in Sec. \[Nonlin\] we consider complex reactions with intermediate compounds (Fig. \[Compounds\]) under two asymptotic conditions
- [The compounds $B_j$ are in fast equilibrium with the corresponding input or output reagents;]{}
- [They exist in very small concentrations compared to other components.]{}
For compounds transitions the first order kinetics is assumed because of small concentrations of compounds (only first order terms survive). These assumptions allow us to produce the reaction rates for the overall reaction from Fig. \[Compounds\] in the form of the generalized MAL: $$\label{generalizedMAL0}
r_{\rho}=\varphi_{\rho}\exp\left(\frac{\sum_i\alpha_{\rho i}
\mu_i}{RT}\right)\, , \; \varphi_{\rho} \geq 0 \, .$$ where $\mu_i$ is the chemical potential of the component $A_i$, $\rho$ is the reaction number, $\alpha_{\rho i}$ are the input stoichiometric coefficients (Fig. \[Compounds\]). Both $\alpha_{\rho i}$ and $\beta_{\rho i}$ are non-negative integers. We use notations $\boldsymbol{\alpha}_{\rho}$ and $\boldsymbol{\beta}_{\rho}$ for vectors wit coordinates $\alpha_{\rho i}$ and $\beta_{\rho i}$. The positive functions $\varphi_{\rho}$ are called the [*kinetic factors*]{} whereas $\exp\left({\sum_i\alpha_{\rho i} \mu_i}/RT\right)$ are the [*Boltzmann factors*]{}.
The balance condition of the first order kinetics of compounds (\[MasterEquilibrium\]) transforms in the semi-detailed balance condition (that is known also as the complex or the cyclic balance condition): $$\sum_{\rho, \,\boldsymbol{\alpha}_{\rho}=\boldsymbol{\nu}} \varphi_{\rho}\equiv
\sum_{\rho, \,\boldsymbol{\beta}_{\rho}=\boldsymbol{\nu}} \varphi_{\rho}$$ for any vector $\boldsymbol{\nu}$ from the set of all vectors $\{\boldsymbol{\alpha}_{\rho}, \boldsymbol{\beta}_{\rho}\}$. Kinetics with the generalized MAL and the semi-detailed balance conditions give the natural nonlinear generalizations of Markov processes. In particular, the entropy production for these systems at any nonequilibrium state is positive.
If we assume for the Markov kinetics of compounds that the positive equilibrium is the point of detailed balance (\[detBal\]) then the kinetic factors $\varphi_{\rho}$ satisfy the stronger condition of detailed balance: $$\varphi_{\rho}^+\equiv \varphi_{\rho}^- \; \mbox{ for all } \rho \, ,$$ where $\varphi_{\rho}^+$ is the kinetic factor for the direct reaction and $\varphi_{\rho}^-$ is the kinetic factor for the reverse reaction.
The class of systems with semi-detailed balance is much wider than the class of systems with detailed balance. Nevertheless, locally they coincide: [*the set of possible velocities for systems with semi-detailed balance coincide with the set of possible velocities for the systems with detailed balance for given thermodynamic functions and any given state*]{} (Sec. \[Nonlin\]).
The systems with generalized MAL and semi-detailed balance are the nonlinear analogs of the Markov processes and the local equivalence of the generalized MAL systems with detailed and semi-detailed balance is the analog and a consequence of Theorem \[Theorem1\].
A bit of history \[History\]
----------------------------
In 1872, Boltzmann introduced the principle of detailed balance for collisions and used it to prove his $H$-theorem [@Boltzmann1872]. Boltzmann’s proof of the positivity of entropy production for systems with detailed balance is very transparent because it is sufficient to prove this positivity just for a couple of mutually inverse elementary processes.
In 1887, Lorentz [@Lorentz1887] objected Boltzmann: he insisted that some collisions of polyatomic molecules do not have reverse collisions and cannot satisfy the principle of detailed balance. Immediately, Boltzmann realized that there exists a much weaker condition sufficient for the $H$-theorem [@Boltzmann1887]. In 1981, it was proven that the Lorentz objections are wrong and the principle of detailed balance is valid for polyatomic molecules [@CercignaniLampis1981]. This is not very surprising because the detailed balance follows from microreversibility (or $T$-invariance of the fundamental equations of mechanics or quantum mechanics). Nevertheless, the Boltzmann discovery is valuable by itself and is used for many kinetic equations.
This condition was rediscovered several times. It is known as the semi-detailed balance condition, the cyclic balance condition or the complex balance condition. In 1952, Stueckelberg proposed a proof of the semi-detailed balance condition for the Boltzmann equation [@Stueckelberg1952]. His proof is based on the Markov model of elementary events. Recently, the Stueckelberg approach was extended to prove the semi-detailed balance condition for the generalized MAL kinetics [@GorbanShahzad2011].
The complex balance condition for chemical kinetics was introduced by Horn and Jackson in 1972 [@HornJackson1972] independently of Boltzmann’s work. Now it is used for mathematical modeling in chemical kinetics and engineering [@SzederkHangos2011]. Boltzmann’s idea about cyclic balance developed in physical kinetics was independently rediscovered in the theory of Markov processes and it is proved that any recurrent Markov process can be decomposed into directed cycles [@Kalpazidou2006].
The principle of detailed balance was crucially important in the development of the Metropolis–Hastings and other Markov chain Monte Carlo algorithms from the very beginning [@Metropolis1953]. Technically, it is much easier to use the detailed balance conditions (\[detBal\]) than to follow more intricate balance conditions. Detailed balance was considered as a necessary condition for construction of Monte Carlo algorithms as a “systematic design principle" [@Katsoulakis2003]. It was demonstrated that it helps to reduce the uncertainty of some observables in stochastic numerics [@Noe2008].
Nevertheless, there are many examples of efficient Monte Carlo computations without detailed balance. Sometimes computational models without detailed balance are constructed because of the physical nature of the systems. For example, the models of inelastic processes in particle physics [@Grassberger1979] or in granular media [@Dean2003] may violate the principle of detailed balance. The general theories of stochastic cellular automata with Gibbsian equilibria but without compulsory detailed balance were developed [@Marroquin1991]. It is widely recognized that the balance equation (not the detailed balance) is necessary and sufficient condition for invariance (stationarity) of the desired equilibrium distribution. Under some more technical irreducibility conditions, the Monte Carlo simulations will converge to this equilibrium [@Manousiothakis1999; @Athenesa2007].
The interplay between reversible (with detailed balance) and irreversible Markov chains is non-trivial and important for many applications. Recently, it was demonstrated that the local deformation of the reversible Markov processes into irreversible ones helps to create efficient computational Monte Carlo algorithms [@Turitsyn2011].
Much efforts were applied to verification of microscopic reversibility and its consequences, the detailed balance conditions and the Onsager reciprocal relations, in many experimental systems [@Miller1960; @Thornton1971; @Driller1979; @Rettner1989; @YablonskyGor_atal2011]. To check experimentally the detailed balance conditions it is necessary to deal with a complex reaction that can be formally equilibrated without detailed balance. If not, then one tests not the detailed balance but just the equilibrium condition as it was mentioned in [@Henley1959].
The detailed balance conditions are very natural and appealing. They simplify many computations and proofs. Thus, they are used in many applications and models. For some physical and chemical systems, detailed balance has a solid background, the $T$-invariance of the fundamental equations, but it is also used beyond the proven microreversibility. When modelling with detailed balance meets some difficulty then the problem about relations between reversible and irreversible systems arises again and detailed balance is substituted by more general conditions. Nevertheless, the convenience, beauty and some intrinsic benefits of the phenomenon, have forced researchers to return to detailed balance if it is possible without a conflict with reality. There are many examples of this “pendulum" in scientific publications: accept detailed balance – criticize detailed balance – go to more general conditions – realize the benefits of detailed balance – return to detailed balance – ...
At the same time, the consequences of the principle of detailed balance are extended to the situations where it was not used before. Thus, recently the multiscale limit of the systems of reversible reactions was studied when some of the equilibrium concentrations tend to zero. The extended principle of detailed balance was proved for the systems with some irreversible reaction [@GorbanYablonsky2011; @GorbanMirYab2012].
In Theorem \[Theorem1\], we compare the sets of possible velocities, ${{\mathrm{d}}}P/{{\mathrm{d}}}t$, for two classes of systems: (i) general first order kinetics (\[MAsterEq0\]) with the given positive equilibrium $P^*$ and (ii) first order kinetics with detailed balance (\[detBal\]) and the same positive equilibrium.
Understanding the structure of the sets of possible velocities can provide additional information about attainable states of the system which is helpful in the modelling context. There are many other reasons too. It is known that different types of kinetics data bear different degrees of reliability. It would therefore be very attractive to study the consequences of the information of each level of reliability separately [@GorbanTree2012]. For example, uncertainty about equilibria in the system is usually significantly lower than that of the reaction rate constants. The value of the equilibrium gives us some information about dynamics: the set of possible velocities ${{\mathrm{d}}}P/{{\mathrm{d}}}t$ is not arbitrarily wide at a given state and for given equilibrium. For the systems with detailed balance, this set is a polyhedral cone which allows a simple description (Sec. \[Quasichem\]). Due to Theorem \[Theorem1\], however, this cone is also the set of possible velocities for the general master equation. Therefore, to distinguish the detailed balance systems from the general ones we have to involve data about ${{\mathrm{d}}}P/{{\mathrm{d}}}t$ for several significantly distant distributions $P$.
Another example is to employ the knowledge of the sets of possible velocities for estimating attainable regions for kinetics. The idea to use the equilibrium information to estimate the attainable sets in kinetics was proposed in 1964 by Horn [@Horn1964] and developed further in chemical kinetics and chemical engineering [@Gorban1979; @obh; @Glasser1987; @Hildebrandt1990; @GorbKagan2006] (for more detailed review see [@GorbanTree2012]). If the sets of possible velocities coincide then the estimated attainable regions coincide too. The knowledge of the sets of possible velocities is also important for the analysis of observability, identifiability and controllability of the systems.
Local equivalence of general Markov systems and systems with detailed balance \[Markov\]
========================================================================================
Global decomposition into cycles, local decomposition into steps, and the equivalence theorem
---------------------------------------------------------------------------------------------
Let us start from master equation (\[MAsterEq0\]). The coefficient $q_{ij}$ is the [*rate constant*]{} for transitions $A_j \to A_i$. Any set of non-negative coefficients $q_{ij}$ ($i\neq j$) corresponds to a master equation. Therefore, the set of all master equations (\[MAsterEq0\]) may be considered as the non-negative orthant in $\mathbb{R}_+^{n(n-1)} \subset \mathbb{R}^{n(n-1)}$. (The non-negative orthant is the set of all vectors with only non-negative components.)
We assume that system (\[MAsterEq0\]) has a positive equilibrium $P^*=(p_i^*)$, $p_i^*>0$ and the balance condition (\[MasterEquilibrium\]) holds. The sum of these $n$ [*balance conditions*]{} is a trivial identity. Let us delete any single equation from (\[MasterEquilibrium\]), for example, the last one (for $i=n$). Each of the remaining equations includes the variable $q_{in}$ which is not present in other equations ($i=1,\ldots ,n-1$). Therefore, for given positive $P^*$, there are $n-1$ independent conditions on $q_{ij}$ ($i,j=1,\ldots, n$, $i\neq j$) in (\[MasterEquilibrium\]).
Thus, the balance conditions (\[MasterEquilibrium\]) define for a given positive equilibrium a $(n-1)^2$-dimensional linear subspace $L^*$ in the $n(n-1)$ dimensional space of $q_{ij}$ ($i\neq j$). A vector of positive coefficients $q_{ij}^*=1/p_j^*$ satisfies (\[MasterEquilibrium\]) and belongs to $L^*$. This vector belongs to the interior of the non-negative orthant $\mathbb{R}_+^{n(n-1)}$. Therefore, the intersection $\mathbb{R}_+^{n(n-1)} \cap L^*$ includes a vicinity of $q_{ij}^*$ in $L^*$ and is a $(n-1)^2$-dimensional cone. Thus, the non-negative solutions of (\[MasterEquilibrium\]) form a $(n-1)^2$-dimensional closed cone in $\mathbb{R}_+^{n(n-1)}$.
The systems with detailed balance (\[detBal\]) for a given positive equilibrium form a smaller cone. Under these conditions, there are only $\frac{n(n-1)}{2}$ independent coefficients among $n(n-1)$ numbers $q_{ij}$. For example, we can arbitrarily select $q_{ij} \geq 0$ for $i>j$ and then take $q_{ij} = q_{ji}\frac{p_i^*}{p^*_j}$ for $i<j$. So, for given $P^*$, the cone of the detailed balance systems (\[detBal\]) can be considered as a non-negative orthant in $\mathbb{R}^{\frac{n(n-1)}{2}}$ embedded in $\mathbb{R}_+^{n(n-1)}$.
If the balance condition (\[MasterEquilibrium\]) holds then system (\[MAsterEq0\]) may be rewritten in a convenient equivalent form: $$\label{MAsterEq1}
\frac{{{\mathrm{d}}}p_i}{{{\mathrm{d}}}t}= \sum_{j, \, j\neq i}
q_{ij}p^*_j\left(\frac{p_j}{p_j^*}-\frac{p_i}{p_i^*}\right) \, .$$ With this form of master equation, it is straightforward to calculate the time derivative of the quadratic divergence, a weighted $l_2$ distance between $P$ and $P^*$, $H_2(P\|P^*)=\sum_i\frac{(p_i-p^*_i)^2}{p^*_i}$: $$\label{QuadLyap}
\frac{{{\mathrm{d}}}H_2(P\|P^*)}{{{\mathrm{d}}}t}=-\sum_{i,j, \, j\neq i}q_{ij}p^*_j\left(\frac{p_i}{p^*_i} - \frac{p_j}{p^*_j}\right)^2 \leq 0\, .$$ This time derivative is [*strictly negative*]{} if for a transition $A_j \to
A_i$ the rate constant is positive, $q_{ij}>0$, and $\frac{p_i}{p^*_i} \neq
\frac{p_j}{p^*_j}$. Hence, if the state $P$ is not an equilibrium (i.e., the right hand side in (\[MAsterEq1\]) is not zero) then $\frac{{{\mathrm{d}}}H_2(P\|P^*)}{{{\mathrm{d}}}t}<0$.
Let us introduce the following notation for a given number of states $n$:
- [$\mathcal{Q}^n_{\rm B}(P^*)$ is the cone of the vectors of non-negative coefficients $q_{ij}$ ($i\neq j$) which satisfy the balance conditions (\[MasterEquilibrium\]), that is, the set of all Markov processes with the equilibrium distribution $P^*$;]{}
- [$\mathcal{Q}^n_{\rm DB}(P^*)$ is the cone of the vectors of non-negative coefficients $q_{ij}$ ($i\neq j$) which satisfy the detailed balance conditions (\[detBal\]), that is, the set of all Markov processes with detailed balance and the equilibrium distribution $P^*$.]{}
If a system satisfies the detailed balance condition (\[detBal\]) then the balance condition (\[MasterEquilibrium\]) holds too: it holds term by term, even without summation. Therefore, $\mathcal{Q}^n_{\rm DB}(P^*) \subset \mathcal{Q}^n_{\rm B}(P^*)$. Moreover, comparing dimension we find that this inclusion is strong. Indeed, $\dim
\mathcal{Q}^n_{\rm B}(P^*)=(n-1)^2$, $\dim \mathcal{Q}^n_{\rm DB}(P^*)=\frac{n(n-1)}{2}$. If $n>2$ then $\frac{n(n-1)}{2} < (n-1)^2$, hence, $$\label{inclusion1}
\mathcal{Q}^n_{\rm DB}(P^*) \subsetneqq \mathcal{Q}^n_{\rm B}(P^*) \subsetneqq \mathbb{R}_+^{n(n-1)} \, .$$ The cone of the systems with detailed balance is, in some sense, much smaller than the cone of the systems with the balance condition: the difference between their dimensions is $\frac{(n-1)(n-2)}{2}$.
Now, let us consider the right hand side vector fields of the systems (\[MAsterEq0\]) at the point $P\neq P^*$. For each cone of the coefficients $q_{ij}$ the vectors of the possible velocities, ${{\mathrm{d}}}P/ {{\mathrm{d}}}t$, also form a cone. Let us introduce the following notation:
- [$\mathbf{Q}^n_{\rm B}(P,P^*)$ is the cone of the possible velocities, ${{\mathrm{d}}}P/ {{\mathrm{d}}}t$, at the point $P$ for $(q_{ij}) \in \mathcal{Q}^n_{\rm B}(P^*)$;]{}
- [$\mathbf{Q}^n_{\rm DB}(P,P^*)$ is the cone of of the possible velocities, ${{\mathrm{d}}}P/ {{\mathrm{d}}}t$, at point $P$ for $(q_{ij}) \in \mathcal{Q}^n_{\rm DB}(P^*)$.]{}
$\mathbf{Q}^n_{\rm B}(P,P^*)$ is the cone of all possible velocities for Markov kinetics at the point $P$ if the equilibrium is $P^*$. $\mathbf{Q}^n_{\rm DB}(P,P^*)$ is the cone of these velocities for Markov kinetics with detailed balance.
Surprisingly, [*for every given distribution $P$, the set of possible velocities ${{\mathrm{d}}}P
/ {{\mathrm{d}}}t$ for general Markov kinetics with equilibrium $P^*$ is the same that for Markov kinetics with detailed balance and the same equilibrium.*]{} The following theorem is the central result of this work.
\[Theorem1\]$\mathbf{Q}^n_{\rm B}(P,P^*)=\mathbf{Q}^n_{\rm DB}(P,P^*)$
This means that for every first order kinetic equation (\[MAsterEq0\]) with a given positive equilibrium $P^*$ and for every point $P\neq P^*$ there exists a first order kinetic equation with detailed balance and equilibrium $P^*$ that has the same velocity at $P$. At this point the right hand sides of the kinetic equations coincide.
Therefore, if we observe the Markov kinetics at one point then we can never distinguish general systems from systems with detailed balance. In particular, they have the same set of Lyapunov functions:
\[Corollary1\] If for a function $H(P,P^*)$, ${{\mathrm{d}}}H / {{\mathrm{d}}}t \leq 0$ for any system (\[MAsterEq0\]) with equilibrium $P^*$ and detailed balance then ${{\mathrm{d}}}H / {{\mathrm{d}}}t \leq 0$ for any system (\[MAsterEq0\]) with equilibrium $P^*$.
The system with detailed balance has $n(n-1)/2$ dimensions available to match a single $n$-dimensional velocity vector. Therefore, it is not surprising that the cones $\mathbf{Q}^n_{\rm DB}(P,P^*)$ has non-empty interior ([*solid cones*]{}). But this is not enough to cover any $n$-dimensional velocity vector using [*non-negative*]{} coefficients $q_{ij}$. This non-negativity condition defines the borders of the cones and we can a priori just state the inclusion $\mathbf{Q}^n_{\rm DB}(P,P^*) \subseteq
\mathbf{Q}^n_{\rm B}(P,P^*)$. The calculation of dimension does not give a hint about coincidence of these cones.
The proof of Theorem 1 is constructed in two steps. First, we prove that for every $P^*$ and $P$ the cone of possible velocities $\mathbf{Q}^n_{\rm
B}(P,P^*)$ is the convex hull of the velocities at point $P$ of the simple cyclic schemes, $A_{i_1} \to \ldots \to A_{i_k} \to A_{i_1}$ ($k\leq n$ and all the numbers $i_1, \ldots, i_k$ are different), with the same equilibrium $P^*$. Secondly, we prove that it is sufficient to take $k=2$.
We will characterize $\mathbf{Q}^n_{\rm B}(P,P^*)$ by its extreme rays. A ray with direction vector $x\neq 0$ is a set $\{\lambda x\}$ ($\lambda \geq 0$). $l$ is an extreme ray of a cone $\mathbf{Q}$ if for any $u \in l$ and any $x,y
\in \mathbf{Q}$, whenever $u = (x + y)/2$, we must have $x,y\in l$. If a closed convex cone does not include a whole straight line then it is the convex hull of its extreme rays [@Rockafellar1970].
\[Lemma:PointCone\] The cone $\mathbf{Q}^n_{\rm B}(P,P^*)$ does not include a whole straight line.
If $v\neq 0$ is a possible value of the right hand side of (\[MAsterEq1\]) then the derivative of $H_2(P\|P^*)$ in direction $v$ is strictly negative (\[QuadLyap\]). Therefore, it is impossible that both $v$ and $-v$ belong to $\mathbf{Q}^n_{\rm B}(P,P^*)$.
Let us consider a simple cyclic scheme, $A_{i_1} \to \ldots \to A_{i_k} \to
A_{i_1}$ ($k\leq n$ and all the numbers $i_1, \ldots, i_k$ are different). For a given positive equilibrium, $P^*$ the coefficients for this scheme belong to a ray: $$\label{CycleRates}
q_{i_{j+1} \, i_{j} }= \frac{\kappa}{p_{i_j}^*} \; (j=1, \ldots , k) \, ,$$ where $\kappa \geq 0$ is a constant and we use the standard convention that for a cycle $q_{i_{k+1} \, i_{k} }=q_{i_{1} \, i_{k} }$.
\[Lemma:CycleExistence\] If system (\[MAsterEq0\]) has a positive equilibrium $P^*$ then for every $A_i$ either all $q_{ji}=q_{ij}=0$ or the state $A_i$ belongs to a cycle with strictly positive rate constants.
Let $A_i$ not belong to a cycle with positive constants. We say that a state $A_j$ is reachable from a state $A_k$ if there exists a non-empty chain of transitions with non-zero coefficients which starts at $A_k$ and ends at $A_j$: $A_k \to \ldots \to A_j$. Let $\mathcal{A}_{i \downarrow}$ be the set of states reachable from $A_i$ and $\mathcal{A}_{i \uparrow}$ be the set of states $A_i$ is reachable from. $\mathcal{A}_{i \downarrow} \cap \mathcal{A}_{i \uparrow} =
\emptyset$ because $A_i$ does not belong to a cycle. If $\mathcal{A}_{i
\uparrow}$ is not empty then in equilibrium all the corresponding $p_j^*=0$ ($j
\in \mathcal{A}_{i \downarrow}$ because there is flow from $\mathcal{A}_{i
\uparrow}$ to $A_i$ and no flow back). If $\mathcal{A}_{i \downarrow}$ is not empty then in equilibrium $p^*_i=0$ because there is a flow from $A_i$ to $\mathcal{A}_{i \downarrow}$ and no flow back. Therefore, if the equilibrium is strictly positive and $A_i$ does not belong to a cycle then $\mathcal{A}_{i
\uparrow} = \mathcal{A}_{i \downarrow}=\emptyset$, hence, all $q_{ji}=q_{ij}=0$.
We will use the following simple general statement: Let $\mathcal{Q}$ be a cone in $\mathbb{R}^m$ without straight lines, $L$ be a linear map, $L:\mathbb{R}^m
\to \mathbb{R}^k$, and $\mathbf{Q}=L(\mathcal{Q})$ be a cone in $\mathbb{R}^k$ without straight lines. Then for every extreme ray $V \subset \mathbf{Q}$ there exists an extreme ray $W \subset \mathcal{Q}$ such that $L(W)=V$. (In other words, there always exists an extreme ray in the preimage of an extreme ray.) We will apply this statement to $\mathcal{Q}=\mathcal{Q}^n_{\rm B}(P^*)$ (the cone of all Markov processes with the given equilibrium $P^*$) and $\mathbf{Q}=\mathbf{Q}^n_B{(P,P^*)}$ (the cone of the possible velocities at point $P$ for all Markov processes with the given equilibrium). The map $L$ transforms the right hand side of the Kolmogorov equation (\[MAsterEq0\]) into its value at point $P$. This transformation “vector field $\mapsto$ its value at point $P$" is, obviously, a linear map.
\[Lemma:ExtremeCycles\]Any extreme ray of the cone $\mathcal{Q}^n_{\rm
B}(P^*)$ is a simple cycle with constants (\[CycleRates\]).
Let a non-zero Markov chain $Q$ with coefficients $q_{ij}$ belong to an extreme ray of $\mathcal{Q}^n_{\rm B}(P^*)$. Due to Lemma \[Lemma:CycleExistence\] this chain includes a simple cycle with non-zero coefficients, $A_{i_1} \to \ldots \to A_{i_k} \to A_{i_1}$ ($k\leq n$, all the numbers $i_1, \ldots, i_k$ are different, $q_{i_{j+1}\,i_{j}}>0$ for $j=1,\ldots, k$, and $i_{k+1}=i_1$). For sufficiently small $\kappa$ ($0<\kappa<\kappa_0$), $q_{i_{j+1}\,i_{j}}-\frac{\kappa}{p^*_{i_{j}}}>0$ ($j=1,\ldots, k$). Let $Q_{\kappa}$ be the same simple cycle with the coefficients (\[CycleRates\]). Then for $0<\kappa <\kappa_0$ vectors $Q\pm
Q_{\kappa}$ also represent Markov chains with the equilibrium $P^*$. Obviously, $Q=\frac{(Q+Q_{\kappa})+(Q-Q_{\kappa})}{2}$, hence, $Q$ should be proportional to $Q_{\kappa}$.
Due to this Lemma, every Markov chain with positive equilibrium is a convex combination of several simple cycles with the same equilibrium. This is the [*global decomposition*]{} of a Markov chain into simple cycles. “Global" here means that the same decomposition is valid for all distributions.
[*Now, we are in position to prove Theorem 1.*]{}
We will prove that any extreme ray of the cone $\mathbf{Q}^n_{\rm B}(P,P^*)$ corresponds to a simple cycle of length 2 (a [*step*]{}): $A_i \rightleftharpoons A_j$ with the rate constants (\[CycleRates\]) $q_{ij} = \frac{\kappa}{p^*_j}\, , \; q_{ji} =
\frac{\kappa}{p^*_ i}$.
According to Lemma \[Lemma:ExtremeCycles\], it is sufficient to prove that for any simple cycle with equilibrium $P^*$ and rate constants (\[CycleRates\]) and for any distribution $P$ the right hand side of the Kolmogorov equation (\[MAsterEq0\]) is a conic combination (a combination with non-negative real coefficients) of the right hand sides of this equation for simple cycles of length 2 at the same point $P$.
Let us prove this by induction on the cycle length $k$. For $k=2$ it is true (trivially). For a cycle of length $k>2$, $A_1 \to A_2 \to \ldots A_k \to A_1$, with the rate constants given by (\[CycleRates\]), the right hand side of equation (\[MAsterEq0\]) is the vector $\mathbf{v}_k$ with coordinates $$\label{cycleVelocity} (\mathbf{v}_k)_j=\frac{p_{j-1}}{p^*_{j-1}}
-\frac{p_{j}}{p^*_{j}}$$ Here, without loss of generality, we take $\kappa=1$, use index $j$ instead of $i_j$ and apply the standard convention regarding cyclic order. Other coordinates of $\mathbf{v}_k$ are zeros.
Let us decompose this $\mathbf{v}_k$ into a conic combination of a vector $\mathbf{v}_{k-1}$ for a cycle of length $k-1$ and a vector $\mathbf{v}_2$ for a cycle of length 2. The flux $A_{j}\to A_{j+1}$ is ${p_{j}}/{p^*_{j}}$. Let us find the minimum value of this flux and, for convenience, let us put this minimal flux in the first position by a cyclic permutation. The target cycle of length $k-1$ is $A_{2} \to \ldots A_{k} \to A_{2}$ with rate constants given by formula (\[CycleRates\]) ($\kappa=1$). We just delete the vertex with the smallest flux from the initial cycle of length $k$. The target cycle of length 2 is $A_1\rightleftharpoons A_2$ with the rate constants (\[CycleRates\]) $q_{21} = \frac{\kappa}{p^*_1}\, , \; q_{12} = \frac{\kappa}{p^*_ 2}$. We find the constant $\kappa$ from the conditions: $\mathbf{v}_k=\mathbf{v}_{k-1}+\mathbf{v}_2$ at the point $P$, hence, two following reaction schemes, (a) and (b), should have the same velocities, ${{\mathrm{d}}}P/{{\mathrm{d}}}t$:
$$\mbox{(a) }A_k { \overset{1/p_k^*}{\rightarrow}}A_1{ \overset{1/p_1^*}{\rightarrow}}A_2 \mbox{ and (b) }
A_k { \overset{1/p_k^*}{\rightarrow}}A_2 \, ; A_1
\underset{\kappa/p_2^*}{\overset{\kappa/p_1^*}{\rightleftharpoons}}A_2 \, .$$ From this condition, $$\kappa=\left(\frac{p_{k}}{p_{k}^*}-\frac{p_{1}}{p_{1}^*}\right)
\left(\frac{p_{2}}{p_{2}^*}-\frac{p_{1}}{p_{1}^*}\right)^{-1}$$ $\kappa \geq 0$ because ${p_{1}}/{p^*_{1}}$ is the minimal value of ${p_{j}}/{p^*_{j}}$. Finally, $\mathbf{v}_k=\mathbf{v}_{k-1}+\mathbf{v}_2$.
Further, we omit the index B or DB at the cone: $\mathbf{Q}^n_{\rm
B}(P,P^*)=\mathbf{Q}^n_{\rm DB}(P,P^*)=\mathbf{Q}^n (P,P^*)$.
It is necessary to stress that the decomposition of the right hand side of the Kolmogorov equation (\[MAsterEq0\]) into a conic combination of cycles of length 2 depends on the ordering of the ratios $p_i/p_i^*$ and cannot be performed for all values of $P$ simultaneously. Thus, this decomposition is [*local*]{}.
Quasichemical representation and the cones of possible velocities \[Quasichem\]
-------------------------------------------------------------------------------
For systems with detailed balance, the cone of possible velocities, $\mathbf{Q}^n_{\rm
DB}(P,P^*)$, is a polyhedral cone. For a given $P^*$, it is a piecewise constant function of $P$. The hyperplanes of the equilibria $A_i\leftrightarrow A_j$ divide the standard simplex of distributions into a finite number of polyhedra ([*compartments*]{}). In each compartment the dominant direction of every transition $A_i\leftrightarrow A_j$ is fixed and the cone of possible velocities is constant. Now we find that this construction provides the cones of possible velocities for general Markov kinetics and not only for systems with detailed balance. Let us describe these cones in detail.
The construction of cones of possible velocities was described in 1979 [@Gorban1979] for systems with detailed balance in the general setting for generalized MAL, for nonlinear chemical kinetics. These systems are represented by stoichiometric equations of the elementary reaction coupled with the reverse reactions: $$\label{Stoichiometric}
\alpha_{\rho 1}A_1+\ldots + \alpha_{\rho n}A_n \rightleftharpoons \beta_{\rho 1}A_1+\ldots
+ \beta_{\rho n}A_n\, ,$$ where $\alpha_{\rho i}, \, \beta_{\rho i} \geq 0$ are the stoichiometric coefficient, $\rho $ is the reaction number ($\rho =1, \ldots, m$). The stoichiometric vector of the $\rho $th reaction is an $n$ dimensional vector $\gamma_\rho $ with coordinates $\gamma_{\rho i} = \beta_{\rho i}- \alpha_{\rho
i}$.
The equilibria of the $\rho $th pair of reactions (\[Stoichiometric\]) form a surface in the space of concentrations. The intersection of these surfaces for all $\rho $ is the equilibrium (with detailed balance). These surfaces of the equilibria of the pairs of elementary reactions (\[Stoichiometric\]) divide the space of concentrations into several compartments. In each compartment the dominant direction of each reaction (\[Stoichiometric\]) is fixed and, hence, the cone of possible velocities is also constant. It is a piecewise constant function of concentrations (for a given temperature): $$\mathbf{Q}={\rm
cone}\{\gamma_{\rho } {\rm sign}(w_{\rho }) \, | \, \rho =1, \ldots , m\}\, .$$
For example, let us join the transitions $A_i \rightleftharpoons A_j$ in pairs (say, $i>j$) and introduce the [*stoichiometric vectors*]{} $\gamma^{ij}$ with coordinates: $$\gamma^{ij}_k=\left\{\begin{array}{ll}
-1 &\mbox{ if } k=i,\\
1 &\mbox{ if } k=j,\\
0 &\mbox{ otherwise}.
\end{array}
\right.$$ Let us rewrite the Kolmogorov equation for the Markov process with detailed balance (\[detBal\]) in the quasichemical form: $$\label{QuasiChemKol}
\frac{{{\mathrm{d}}}P}{{{\mathrm{d}}}t}=\sum_{i>j}w_{ij}^*\left(\frac{p_j}{p_j^*}-\frac{p_i}{p_i^*}\right) \gamma^{ji}\, .$$ Here, $w_{ij}^*=q_{ij}p^*_j=q_{ji}p^*_i$ is the equilibrium flux from $A_i$ to $A_j$ and reverse.
The cone of possible velocities for (\[QuasiChemKol\]) is $$\mathbf{Q}^n (P,P^*)={\rm cone}\left\{\gamma^{ji}
{\rm sign}\left(\frac{p_j}{p_j^*}-\frac{p_i}{p_i^*}\right) \, \Big| \, i>j \right\}\, .$$ Here, we use the three-valued sign function (with values $\pm 1$ and 0). In Fig. \[3stateCones\], the partition of the standard distribution simplex into compartments, and the cones (angles) of possible velocities are presented for the Markov chains with three states.
A set of distributions $U$ is [*positively invariant*]{} with respect to system (\[MAsterEq0\]) if for any initial distribution $P(0)\in U$, the solution of (\[MAsterEq0\]) $P(t)$ remains in $U$ for $t>0$. The bold broken lines in Fig. \[3stateCones\] follow along the extreme rays of the angles of possible velocities (clockwise or anticlockwise). They form the borders of a positively-invariant area for all the Markov chains with the given equilibrium $P^*$.
These borders give, for example, a simple estimate of the logarithmic decrement for Markov chains. For decaying oscillations, the logarithmic decrement is the natural logarithm of the ratio of any two successive amplitudes: $\delta
\triangleq \ln\frac{x_1}{x_2}$. For a complex eigenvalue $\lambda$, the period between two amplitudes $T=2\pi/|\Im \lambda|$ and $\delta = 2\pi \frac{|\Re
\lambda|}{|\Im \lambda|}$. For systems with detailed balance, eigenvalues are always real but for the general Markov chains they may be complex. For example, for the simple cycle $A_1\to A_2 \to A_3 \to A_1$ with the equilibrium equidistribution $p_{1,2,3}^*=1/3$, and the rate coefficients $\kappa$, the nonzero eigenvalues of the linear system (\[MAsterEq0\]) are $\lambda=\kappa
(-\frac{3}{2}\pm i\frac{\sqrt{3}}{2})$ and $\delta = 2\pi \sqrt{3}$.
Let us follow the clockwise border trajectory (Fig. \[3stateCones\]) starting from the state $A_1$ (the corresponding distribution is $P=(1,0,0)$). This state belongs to the line of equilibria of the transition $A_2
\rightleftharpoons A_3$. The first step is the equilibration of the transition $A_1 \rightleftharpoons A_2$ ($A_3$ does not change). After that, the equilibration of the transition $A_1 \rightleftharpoons A_3$ follows ($A_2$ does not change): $$\begin{split}
&(1,0,0) \mapsto
\left(\frac{p^*_1}{1-p^*_3},\frac{p^*_2}{1-p^*_3},0\right) \\
&\mapsto \left(\frac{(p^*_1)^2}{(1-p^*_3)(1-p^*_2)},\frac{p^*_2}{1-p^*_3},\frac{p^*_1
p^*_3 }{(1-p^*_3)(1-p^*_2)}\right)\, .
\end{split}$$ As the result of this sequence of equilibrations, when the clockwise border line again approaches the equilibrium line of the transition $A_2
\rightleftharpoons A_3$, the value of $p_1$ is $\frac{(p^*_1)^2}{(1-p^*_3)(1-p^*_2)}$. After this turn in angle $\pi$ every trajectory becomes closer to $P^*$. The contraction coefficient is $\frac{p^*_1p^*_2 p^*_3}{(1-p^*_1)(1-p^*_2)(1-p^*_3)}$ or less. The anticlockwise trajectory gives the same contraction. We estimated the logarithmic decrement from below: $$\label{DecremEval}
\delta \left(=2\pi \frac{|\Re \lambda|}{|\Im \lambda|}\right) \geq 2\ln
\left(\frac{(1-p^*_1)(1-p^*_2)(1-p^*_3)}{p^*_1p^*_2 p^*_3}\right) \, .$$
Two $H$-theorems \[H-theorems\]
-------------------------------
The most general form of the $H$-theorem for Markov processes was proposed by Morimoto [@Morimoto1963]. He used the following $H$-functions: for each convex function of the positive convex variable $h(x)$ the $h$-[*divergence*]{} between distributions $P$ and $P^*$ is $$\label{Morimoto}
\begin{split}
&H_h(P \| P^*)=\sum_i p^*_i h\left(\frac{p_i}{p_i^*}\right)\, .
\end{split}$$ At the same time these divergencies were studied by Csiszár [@Csiszar1963] and sometimes they are called the Csiszár–Morimoto divergences. These functions were introduced two years earlier by Rényi on the last page of his famous work [@Renyi1961] together with the hint about the $H$-theorem. For more details see [@GorbanGorbanJudge2010].
The time derivative of the Csiszár–Morimoto function $H_h(P \| P^*)$ (\[Morimoto\]) with respect to master equation (\[MAsterEq1\]) for a general Markov process is $$\label{ENtropyProd}
\begin{split}
&\frac{{{\mathrm{d}}}H_h(P \| P^*)}{{{\mathrm{d}}}t}=\sum_{i,j, \, j\neq i} q_{ij}p^*_j \\
&\times \left[h\left(\frac{p_i}{p_i^*}\right)-h\left(\frac{p_j}{p_j^*}\right) +
h'\left(\frac{p_i}{p_i^*}\right)\left(\frac{p_j}{p_j^*}-
\frac{p_i}{p_i^*}\right)\right] \leq 0
\end{split}$$ For a Markov process with detailed balance we use the quasichemical form of master equation (\[QuasiChemKol\]) and find immediately $$\label{ENtropyProdDB}
\begin{split}
&\frac{{{\mathrm{d}}}H_h(P \| P^*)}{{{\mathrm{d}}}t}=-\sum_{i,j, \, i>j} q_{ij}p^*_j \\
&\times\left(\frac{p_j}{p_j^*}-\frac{p_i}{p_i^*}\right)\left(h'\left(\frac{p_j}{p_j^*}\right)-h'\left(\frac{p_i}{p_i^*}\right)\right)
\leq 0 \, .
\end{split}$$ The inequality for the general Markov processes (\[ENtropyProd\]) follows from Jensen’s inequality in the differential form, $h'(x)(y-x)\leq h(y)-h(x)$. It is valid for left and right limits of $h'$ at any point $x >0$. The inequality for systems with detailed balance (\[ENtropyProdDB\]) follows from the monotonicity of $h'$. In full agreement with Corollary \[Corollary1\], the divergences $H_h(P \| P^*)$ (\[Morimoto\]) are Lyapunov functions for systems with detailed balance and for all the Markov processes as well. Theorem \[Theorem1\] has an even stronger corollary.
For every Markov process $Q$ with positive equilibrium $P^*$ and for a distribution $P\neq P^*$ there exists a Markov process $Q_{\rm DB}$ with the same equilibrium that obeys the detailed balance condition and has the following property: For every convex function $h$ the time derivative ${{{\mathrm{d}}}H_h(P \| P^*)}/{{{\mathrm{d}}}t}$ for $Q$ coincides at point $P$ with the time derivative of $ H_h(P \| P^*)$ at this point for $Q_{\rm DB}$.
Nonlinear kinetics: detailed balance versus semi-detailed balance \[Nonlin\]
============================================================================
The general equations of MAL without any restriction on the reaction rate constants demonstrate all types of non-trivial dynamic behavior, from multiple steady states to strange attractors [@Yablonskiiatal1991; @Ertl1990]. It is not a surprise because the MAL systems can approximate with arbitrary accuracy any smooth vector field which preserves the linear conservation laws and positivity of concentrations [@GorbanByYa1986; @Kowalski1993].
The systems with semi-detailed balance give the direct nonlinear generalization of the general Markov kinetics. They were introduced by Boltzmann for gas kinetics [@Boltzmann1887] and generalized later for MAL systems [@HornJackson1972; @GorbanShahzad2011; @SzederkHangos2011]. The systems with semi-detailed balance are the generalized MAL systems with additional relations between rate constants.
To produce these relations, let us follow the classical work [@Stueckelberg1952] and assume that behind the reaction mechanism (\[Stoichiometric\]) there is the reaction mechanism with intermediate [*compounds*]{} $B_{\rho}^{\pm}$ illustrated by Fig. \[Compounds\]. Each compound is associated with a formal input or output complex $\sum_i\alpha_{\rho i}A_i$ or $\sum_i \beta_{\rho i} A_i$. Such a complex may participate in several reactions. Let there be $k$ different vectors among $\{\boldsymbol{\alpha}_{\rho}, \boldsymbol{\beta}_{\rho}\}$ ($\boldsymbol{\alpha}_{\rho}=(\alpha_{\rho i})$ and $\boldsymbol{\beta}_{\rho}=(\beta_{\rho i})$). We denote these different vectors by $\boldsymbol{\nu}_j$ ($j=1, \ldots, k$). The correspondent complexes are $\Theta_j=\sum_i
\nu_{ji} A_i$. The reaction mechanism (\[Stoichiometric\]) takes the form of the list of transitions $\Theta_j \to \Theta_l$ and the extended reaction mechanism is the list of transitions $$\label{ExtendedStoi}
\Theta_j \rightleftharpoons B_j \to B_l \rightleftharpoons \Theta_l \, .$$
Stueckelberg introduced this representation for the collisions in Boltzmann’s equations and used two asymptotic assumptions:
- [The compounds $B_j$ are in fast equilibrium with the corresponding input or output reagents and the reactions $\Theta_j \rightleftharpoons B_j$ in (\[ExtendedStoi\]) are always close to equilibrium (this is the quasiequilibrium assumption, QE);]{}
- [They exist in very small concentrations compared to other components (this leads to the quasi steady state approximation, QSS).]{}
We call the intermediates $B_j$ compounds following the classical work of Michaelis and Menten [@MichaelisMenten1913]. In 1913, they introduced the same asymptotic assumptions and representation for an enzyme reaction and demonstrated that in this case the overall catalytic reaction obeys the MAL.
In more general settings, these two assumption, QE and QSS, allow us to produce the reaction rates for the rates of the overall reactions in the form of the generalized MAL. The rate of the reaction $$\sum_i\alpha_{\rho i}A_i \to \sum_i \beta_{\rho i} A_i$$ is the product of two factors, a standard Boltzmann factor $W_{\rho}$ and a kinetic factor $\varphi_{\rho} \geq 0$: $$\label{generalizedMAL}
r_{\rho}=\varphi_{\rho}W_{\rho}=\varphi_{\rho} \exp\left(\frac{\sum_i\alpha_{\rho i}
\mu_i}{RT}\right)\, ,$$ where $\mu_i$ is the chemical potential of the component $A_i$. The corresponding kinetic equation is $$\label{kinur}
\frac{{{\mathrm{d}}}N }{{{\mathrm{d}}}t}=V \sum_{\rho}r_{\rho} \boldsymbol{\gamma}_{\rho}\, , \;\;
(\gamma_{\rho i}=\beta_{\rho i} - \alpha_{\rho i})\, .$$ Here $N$ is the vector of composition ($N_i$ is the amount of $A_i$), and $V$ is the volume.
We use the notation $c_i$ for the concentration of $A_i$, $c$ is the vector of concentrations, $\varsigma_j$ is the concentration of $B_j$. The chemical potentials $\mu_i$ of the components $A_i$ are the partial derivatives of the free energy density, $\mu_i=\partial f(c,T)/\partial c_i$. The standard thermodynamic assumption about strong convexity of the function $f(c,T)$ for all $T$ is accepted.
Let us demonstrate how the generalized MAL follows from the QE and QSS approximations (for more details see Ref. [@GorbanShahzad2011]). The thermodynamic equilibria for the extended mixture are defined as the conditional minima of the free energy $F$. The free energy of a mixture of $A_i$ with small admixtures of the compounds $B_j$ is: $$\label{FreeEn1}
F=Vf(c,T)+VRT \sum_{j=1}^q \varsigma_j \left(\frac{u_j(c,T)}{RT}+\ln
\varsigma_j-1\right)\, .$$ The entropic terms $VRT \varsigma_j \ln \varsigma_j$ in this expression corresponds to the ideal gas equations $p_j=\varsigma_jRT$ for the partial pressure of the small admixtures of the compounds $B_j$. This ideal gas low may be valid not only in gases but for osmotic pressure of small admixtures in solutions (the Morse equation).
The thermodynamic equilibria of $B_j$ are: $\varsigma_j^{\rm
eq}=\varsigma_j^*/Z$, where $Z$ is a positive number and $\varsigma_j^*$ is the [*standard equilibrium*]{}: $$\label{StandEquili}
\varsigma_j^*(c,T)=\exp\left(-\frac{u_j(c,T)}{RT}\right)\, .$$ The thermodynamic equilibrium condition of the reactions $\Theta_j
\rightleftharpoons B_j$ under the condition of smallness of $\varsigma_j$ (QE+QSS) can be solved explicitly: $$\label{equilibrationEq}
\varsigma_j^{\rm qe}=\varsigma^*_j(c,T)\exp\left(\frac{\sum_i \nu_{ji}
\mu_i(c,T)}{RT}\right)\, .$$
The smallness of the concentration of the compounds implies that the rates of the reactions $B_i \to B_j$ in the extended mechanism (\[ExtendedStoi\]) are linear functions of their concentrations. Let the rate constants for this first order kinetics be $q_{ji}$.
In the selected approximations the extended reaction mechanism (\[ExtendedStoi\]) returns to the form $\Theta_j \to \Theta_l$. The reaction rate of the transition $\Theta_j \to \Theta_l$ in the quasiequilibrium approximation is $r_{lj}=q_{lj}
\varsigma_j^{\rm qe}$. This is exactly the generalized MAL (\[generalizedMAL\]) with $$\varphi_{lj}=q_{lj} \varsigma_j^*,\, \boldsymbol{\alpha}_{\rho}=\boldsymbol{\nu}_j, \,
\boldsymbol{\beta}_{\rho}=\boldsymbol{\nu}_l \, .$$
At the equilibrium $\varsigma^*/Z$, the first order kinetics of compounds should satisfy the general balance condition (\[MasterEquilibrium\]): $$\sum_j q_{lj} \varsigma_j^*= \sum_j q_{jl} \varsigma_l^* \, .$$ Therefore, the kinetic factors $\varphi_{\rho}$ satisfy the identity of semi-detailed balance: $$\label{semidetailed}
\sum_{\rho, \,\boldsymbol{\alpha}_{\rho}=\boldsymbol{\nu}} \varphi_{\rho}\equiv
\sum_{\rho, \,\boldsymbol{\beta}_{\rho}=\boldsymbol{\nu}} \varphi_{\rho}$$ for any vector $\boldsymbol{\nu}$ from the set of all vectors $\{\boldsymbol{\alpha}_{\rho},
\boldsymbol{\beta}_{\rho}\}$. This identity is exactly the Markov balance condition (\[MasterEquilibrium\]) for kinetics of compounds with equilibrium $\varsigma^*/Z$. It has a very transparent sense: the thermodynamic equilibrium is, at the same time, the equilibrium for the first order kinetics of compounds, i.e. the it satisfies the balance condition (\[MasterEquilibrium\]) for master equation.
Let us assume that the Markov kinetics of compounds satisfies the detailed balance condition (\[detBal\]) at the thermodynamic equilibrium: $$q_{lj}\varsigma_j^*=q_{jl}\varsigma^*_l \, .$$ Then the kinetic factors $\varphi_{\rho}$ satisfy the condition of detailed balance: $$\label{NlinDetailed}
\varphi_{\rho}^+\equiv \varphi_{\rho}^- \, ,$$ where $\varphi_{\rho}^+$ is the kinetic factor for the direct reaction and $\varphi_{\rho}^-$ is the kinetic factor for the reverse reaction.
This detailed balance condition assumes that the sums in the left and right hand sides of Eq. (\[semidetailed\]) are equal term by term. Therefore, it is stronger than the semi-detailed balance condition.
For linear systems, the semi-detailed balance condition turns into the standard balance condition (\[MasterEquilibrium\]) and the detailed balance condition (\[NlinDetailed\]) turns into (\[detBal\]). Of course, the class of systems with semi-detailed balance is much wider than the class of systems with detailed balance. Nevertheless, locally they coincide: [*for given thermodynamic functions (\[FreeEn1\]) and any given concentrations and temperature, the cone of possible velocities for systems (\[kinur\]) with semi-detailed balance coincides with the cone of the possible velocities for the systems with detailed balance.*]{}
Indeed, for the given values of concentrations we can perform the following three operations, (i) return from the generalized MAL to the first order kinetic equations of compounds, (ii) use Theorem \[Theorem1\] and find the system of compounds with detailed balance, which has the same velocity at the same point, and (iii) return back, to the generalized MAL. As a result, we get a kinetic system for the components $A_i$ with detailed balance, the same free energy $Vf(c,T)$ (\[FreeEn1\]), and the same velocity at the selected values of concentrations.
Conclusion
==========
The definition of detailed balance includes the rates of all transitions at equilibrium but observability of all these rates together is a very special situation. Typically, one can observe the overall system velocity, ${{\mathrm{d}}}P/{{\mathrm{d}}}t$, or just some components of this velocity but not the rates of individual transitions. According to our results, if we know the equilibrium distribution $P^*$ and observe the system velocity at one nonequilibrium point $P$ then we can never distinguish a general system from the systems with detailed balance. This is true for Markov kinetics as well as for the systems with the generalized MAL; detailed balance can never be distinguished from the semi-detailed balance if we know the equilibrium and observe the velocity at one nonequilibrium point. The difference between velocities of the general kinetic systems and the systems with detailed balance is hidden in the correlations between different nonequilibrium states (or, for example, in the continuous pieces of trajectories). The cone of possible velocities at a nonequilibrium state $P$ is a piece-wise constant function of $P$, which can be constructed explicitly for the systems with detailed balance (Fig. \[3stateCones\]), and the same construction is valid for the general kinetics. These results seem to be rather surprising.
For the nonlinear mass action systems, the systems with semi-detailed balance give the proper analogue of the general Markov kinetics. The conditions of semi-detailed balance were invented for the Boltzmann equation by Boltzmann [@Boltzmann1887], studied by Stueckelberg [@Stueckelberg1952] and rediscovered for the mass action kinetics by Horn and Jackson [@HornJackson1972]. Recently [@GorbanShahzad2011], it was proved that the generalized MAL with the semi-detailed balance condition always follows from the Markov kinetics of compounds in the Michaelis–Menten–Stueckelberg asymptotic. The class of the systems with semi-detailed balance is wider than the class of systems with detailed balance. Nevertheless, for a given equilibrium and for any given value of concentration these two classes have the same sets of possible velocities in the distribution space.
[00]{}
, [Lectures on gas theory]{}, [U. of California Press, Berkeley, CA]{}, [1964]{}.
H.-A. Lorentz, Über das Gleichgewicht der lebendigen Kraft unter Gasmolekülen. Sitzungsberichte der Kaiserlichen Akademie der Wissenschaften in Wien. 95 (2) (1887) 115–152.
, Neuer Beweis zweier Sätze über das Wärmegleichgewicht unter mehratomigen Gasmolekülen, [Sitzungsberichte der Kgl. Akademie der Wissenschaften in Wien]{} [95]{} (2) (1887) [153–164]{}.
C. Cercignani and M. Lampis, On the $H$-theorem for polyatomic gases, J. Stat. Phys. 26 (4) (1981) 795–801.
E.C.G. [Stueckelberg]{}, Théorème $H$ et unitarité de $S$, [Helv. Phys. Acta]{} [5]{} (1952) [577–580]{}.
A.N. Gorban and M. Shahzad, The Michaelis-Menten-Stueckelberg Theorem, [Entropy]{} [13]{} (2011) [966-1019]{}; arXiv:1008.3296
F. Horn, and R. Jackson, General mass action kinetics, [Arch. Ration. Mech. Anal.]{} [47]{} (1972) [81–116]{}.
G. Szederkényi and K.M. Hangos, Finding complex balanced and detailed balanced realizations of chemical reaction networks. [J. Math. Chem.]{} [49]{} (2011) [1163–1179]{}.
J. von Neumann, [Mathematical Foundations of Quantum Mechanics]{} [Princeton Univ. Press, Princeton, NJ]{}, [1996]{}.
S.L. Kalpazidou, Cycle Representations of Markov Processes, (Series: Applications of Mathematics, V. 28), Springer, New York, 2006.
N. Metropolis, A.W. Rosenbluth, M.N. Rosenbluth, A.H. Teller and E. Teller, Equations of State Calculations by Fast Computing Machines, J. Chem. Phys. 21 (6) (1953) 1087–1092.
M.A. Katsoulakis, A.J. Majda, and D.G. Vlachos, Coarse-grained stochastic processes and Monte Carlo simulations in lattice systems, J. Comput. Phys. 186 (1) (2003) 250–278
Frank Noé, Probability distributions of molecular observables computed from Markov models, J. Chem. Phys. 128 244103 (2008).
P. Grassberger and A. de la Torre, Reggeon field theory (Schlögl’s first model) on a lattice: MonteCarlo calculations of critical behaviour, Ann. Phys. 122 (2) (1979) 373–396.
D.S. Dean and A. Lefèvre, Possible Test of the Thermodynamic Approach to Granular Media, Phys. Rev. Lett. 90 198301 (2003)
J.L. Marroquin and A. Ramirez, Stochastic cellular automata with Gibbsian invariant measures, IEEE Transactions on Information Theory 37 (3) (1991) 541–551.
V.I. Manousiothakis and M.W. Deem, Strict detailed balance is unnecessary in Monte Carlo simulation, J. Chem. Phys. 110 (6) (1999) 2753–2756; arXiv:cond-mat/9809240.
M. Athènesa, Web ensemble averages for retrieving relevant information from rejected Monte Carlo moves, Eur. Phys. J. B 58 (2007) 83–95.
K.S. Turitsyn, M. Chertkov, and M. Vucelja, Irreversible Monte Carlo algorithms for efficient sampling, Physica D 240 (4–5) (2011) 410–414.
D.G. Miller, Thermodynamics of irreversible processes. The experimental verification of the Onsager reciprocal relations, Chem. Rev. 60 (1960), 15-37.
T. Thornton, C.M. Jones, J.K. Bair, M.D. Mancusi, and H.B. Willard, Test of Time-Reversal Invariance in the Reactions $^{16}\mbox{O}(d,a)^{14}\mbox{N}$ and $^{14}\mbox{N}(a,d)^{16}\mbox{O}$, Phys. Rev. C 3 (3) (1971) 1065–1086.
H. Driller, E. Blanke, H. Genz, A. Richter, G. Schrieder, and J.M. Pearson, Test of detailed balance at isolated resonances in the reactions $^{27}\mbox{Al+p}\rightleftharpoons ^{24}\mbox{Mg}+\alpha$ and time reversibility, Nuclear Physics A 317 (2–3) (1979) 300–312.
C. T. Rettner, E. K. Schweizer, and C. B. Mullins, Desorption and trapping of argon at a 2H–W(100) surface and a test of the applicability of detailed balance to a nonequilibrium system, J. Chem. Phys. 90, 3800 (1989)
G. S. Yablonsky, A.N. Gorban, D. Constales, V.V. Galvita, and G.B. Marin, Reciprocal relations between kinetic curves, EPL 93 20004 (2011).
E.M. Henley and B.A. Jacobsohn, Time Reversal in Nuclear Interactions, Phys. Rev. 113 (1) (1959), 225–233.
A.N. Gorban and G.S. Yablonsky, Extended detailed balance for systems with irreversible reactions, Chemical Engineering Science 66 (2011) 5388–5399; arXiv:1101.5280 \[cond-mat.stat-mech\].
A.N. Gorban, E.M. Mirkes and G.S. Yablonsky, Thermodynamics in the limit of irreversible reactions, Physica A, (2012) DOI:10.1016/j.physa.2012.10.009; arXiv:1207.2507 \[cond-mat.stat-mech\].
A.N. Gorban, Thermodynamic Tree: The Space of Admissible Paths, SIADS (in press); arXiv:1201.6315 \[cond-mat.stat-mech\].
, [Attainable regions in chemical reaction technique]{}, in [The Third European Symposium on Chemical Reaction Engineering]{}, Pergamon Press, London, UK, 1964, pp. 1–10.
A.N. [Gorban]{}, Invariant sets for kinetic equations, [ React. Kinet. Catal. Lett.]{} [10]{} (1979) [187–190]{}.
, [Equilibrium Encircling. Equations of Chemical Kinetics and their Thermodynamic Analysis]{}, Nauka Publ., Novosibirsk, 1984.
, [A geometric approach to steady flow reactors: the attainable region and optimisation in concentration space]{}, Ind. Eng. Chem. Res. 26 (1987) 1803–1810.
, [The attainable region and optimal reactor structures]{}, [ Chem. Eng. Sci.]{} 45 (1990) 2161–2168.
, [Thermodynamic Equilibria and Extrema: Analysis of Attainability Regions and Partial Equilibria]{}, Springer, New York, NY, 2006.
R.T. [Rockafellar]{}, [Convex Analysis]{}, [Princeton Univ. Press, Princeton, NJ]{}, [1997]{}.
Markov processes and the $H$-theorem, [J. Phys. Soc. Jap.]{} [12]{} (1963) [328–381]{}.
, Eine informationstheoretische Ungleichung und ihre Anwendung auf den Beweis der Ergodizität von Markoffschen Ketten, [Magyar. Tud. Akad. Mat. Kutat´o Int. Közl.]{} [8]{} (1963) [85–108]{}.
V. [1]{} [University of California Press: Berkeley, CA]{} (1961), pp. 547–561.
, Entropy: The Markov Ordering Approach, arXiv:1003.1377.
V.[32]{}, [Elsevier, Amsterdam, The Netherlands]{}, [1991]{}.
Nonlinear dynamics in the CO-oxidation on Pt single crystal surfaces, [Appl. Phys. A – Mater. Sci. Process]{} [51]{} (1990) [79–90]{}.
A.N. Gorban, V.I Bykov, and G.S. Yablonskii, Essays on chemical relaxation, [Nauka, Novosibirsk, ]{} 1986.
Universal formats for nonlinear dynamical systems, [Chem. Phys. Lett.]{} [209]{} (1993) [167–170]{}.
L. Michaelis and M. Menten, Die Kinetik der Intervintwirkung, [Biochem. Z.]{} [49]{} (1913) [333–369]{}.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The attenuation of very high energy [$\gamma$-rays]{} by pair production on the Galactic interstellar radiation field has long been thought of as negligible. However, a new calculation of the interstellar radiation field consistent with multi-wavelength observations by DIRBE and FIRAS indicates that the energy density of the Galactic interstellar radiation field is higher, particularly in the Galactic center, than previously thought. We have made a calculation of the attenuation of very high energy [$\gamma$-rays]{} in the Galaxy using this new interstellar radiation field which takes into account its nonuniform spatial and angular distributions. We find that the maximum attenuation occurs around 100 TeV at the level of about 25% for sources located at the Galactic center, and is important for both Galactic and extragalactic sources.'
author:
- 'Igor V. Moskalenko'
- 'Troy A. Porter'
- 'Andrew W. Strong'
title: Attenuation of VHE gamma rays by the Milky Way Interstellar Radiation Field
---
Introduction
============
The attenuation of very high energy (VHE) [$\gamma$-rays]{} by pair production on the Galactic interstellar radiation field (ISRF) has previously been considered to be negligible [@Nikishov62; @Protheroe1986]. The main contribution is thought to come from pair production on the cosmic microwave background (CMB) where the effective threshold for attenuation is $\sim$100 TeV and a maximum is reached at about 2000 TeV, currently accessible only via air-shower experiments. The Galactic ISRF photons are more energetic so that the effective threshold is lower ($\sim$100 GeV) and the attenuation increases slowly to a maximum around 100 TeV. This covers the energy range of present day Imaging Atmospheric Cherenkov Telescopes, such as the HESS instrument. A rough estimate of the attenuation of VHE [$\gamma$-rays]{} coming from the Galactic center (GC), which uses a new ISRF [@PS05], but assumes an isotropic angular distribution for the ISRF, shows that the effect is observable [@Z05]. Given the essential anisotropy of the Galactic ISRF with most of the photons going outwards from the inner Galaxy, the effect depends on the position of the source of VHE photons and its orientation relative to the observer in the Galactic plane. We calculate the attenuation of VHE [$\gamma$-rays]{} due to pair production with the Galactic photon field using the total ISRF over the entire Galaxy on a fine spatial grid which takes into account the nonuniform spatial and anisotropic angular distribution of background photons.
Interstellar Radiation Field
============================
The essential ingredients to calculate the Galactic ISRF are a model for the distribution of stars in the Galaxy, a model for the dust distribution and properties, and a treatment of scattering, absorption, and subsequent re-emission of the stellar light by the dust. We briefly describe our ISRF calculation, which is a further development of the work reported by [@PS05]; full details will be given in a forthcoming paper (Porter & Strong, in preparation).
Our stellar model assumes a type classification based on that used in the SKY model of [@Wainscoat1992]. It includes 87 stellar classes encompassing main sequence stars, AGB stars and exotics. For each stellar type there is a local star number density, scale height above the plane, fraction of local number density in each of several discrete spatial components, and spectrum in standard photometric filters. The stars are distributed in seven geometrical components: thin and thick disc, halo, bulge, bar, ring, and spiral arms. Spectra for normal stars are taken from the synthetic spectral library of [@Girardi2002]. Spectra for AGB stars and exotics are as given in the SKY model.
![Interstellar radiation field energy density. Local interstellar radiation field (upper): heavy solid line, total radiation field including CMB; heavy dashed line, contribution by stars; heavy dotted line, scattered light; heavy dot-dashed line, infra-red; thin solid line, local ISRF from [@Strong2000]. Data: squares, Apollo [@Henry1980]; triangles, DIRBE [@Arendt1998]; circles, FIRAS [@Finkbeiner1999]. Interstellar total radiation field radial variation (lower): solid line, $(R, z) = (0$ kpc, 0 kpc); dashed line, $(R, z) = (4$ kpc, 0 kpc); dotted line, $(R, z) = (12$ kpc, 0 kpc); dash-dotted line, $(R, z) = (16$ kpc, 0 kpc). Heavy lines are for our ISRF; thin lines for the ISRF of [@Strong2000]. \[fig1\]](f1a.eps){width="3.5in"}
![Interstellar radiation field energy density. Local interstellar radiation field (upper): heavy solid line, total radiation field including CMB; heavy dashed line, contribution by stars; heavy dotted line, scattered light; heavy dot-dashed line, infra-red; thin solid line, local ISRF from [@Strong2000]. Data: squares, Apollo [@Henry1980]; triangles, DIRBE [@Arendt1998]; circles, FIRAS [@Finkbeiner1999]. Interstellar total radiation field radial variation (lower): solid line, $(R, z) = (0$ kpc, 0 kpc); dashed line, $(R, z) = (4$ kpc, 0 kpc); dotted line, $(R, z) = (12$ kpc, 0 kpc); dash-dotted line, $(R, z) = (16$ kpc, 0 kpc). Heavy lines are for our ISRF; thin lines for the ISRF of [@Strong2000]. \[fig1\]](f1b.eps){width="3.5in"}
We assume a dust model including graphite, polycyclic aromatic hydrocarbons (PAHs), and silicate. Dust grains in the model are spherical and the absorption and scattering efficiencies for graphite, PAHs, and silicate grains are taken from [@Li2001]. The grain model abundance and size distribution are taken from [@Weingartner2001] (their best fit Milky Way model), and a purely neutral interstellar medium is assumed. We consider only coherent scattering, and a Henyey-Greenstein angular distribution function [@Henyey1941] is used in the scattering calculation. The stochastic heating of grains smaller than $\sim0.1$ $\mu$m is treated using the “thermal continuous” approach of [@Draine2001]; we calculate the equilibrium heating of larger dust grains by balancing absorption with re-emission as described by [@Li2001].
Dust is assumed to follow the Galactic gas distribution. We use the gas model for neutral and molecular hydrogen given by [@Moskalenko2002]. The radial variation in the Galactic metallicity gradient is taken to be 0.07 dex/kpc [@Strong2004b and references therein].
A cylindrical geometry is adopted for the radiation field calculation. Our calculations are simplified by assuming symmetry in azimuth and about the Galactic plane. The maximum radial extent is taken to be $R_{\rm max} = 20$ kpc. The maximum height above the Galactic plane is taken to be $z_{\rm max} = 5$ kpc. The Sun is located at $R_S = 8.5$ kpc from the GC. The Galaxy is divided into elements of equal volume and the total radiation field is calculated for each.
The radiation field from stellar and scattered light is obtained using a modified form of the so-called partial intensity method [@Baes2001]. Henceforth, we will refer to the total stellar and scattered light as the “optical” radiation field. The infra-red radiation field is obtained by using the optical radiation field to calculate the dust emissivity for stochastic and equilibrium heating. The dust emission is integrated to obtain the infra-red radiation field throughout the Galaxy.
Fig. \[fig1\] (upper panel) shows our calculated local ISRF including the CMB; also shown in the figure is the local ISRF calculated by [@Strong2000]. The agreement of our computed ISRF with the data is generally good. We note that our new ISRF provides an improved fit to the observations around 100 $\mu$m compared to the earlier model of [@Strong2000]. This is particularly important for our optical depth calculation, since the majority of the attenuation is on these more numerous less-energetic photons of the ISRF.
Fig. \[fig1\] (lower panel) shows the radial variation in the Galactic plane of our ISRF (thick lines) together with that of [@Strong2000] (thin lines). Toward the inner Galaxy, our ISRF predicts a significantly higher energy density than previously described, particularly for the infra-red component. This arises because of the coupling between the optical radiation field and the infra-red emission: in our calculation, the optical radiation field is used as direct input to the dust heating calculation, which was not done in previous work. The intense optical radiation field toward the inner Galaxy heats the dust to warmer temperatures, increasing the emission in the infra-red. Even though the optical emission does not cause a direct enhancement to the attenuation at GeV to TeV energies – the energy of these photons is typically a few eV, lowering the threshold, however their number density is too low to provide significant absorption – it is essential to calculate this component of the ISRF to obtain the correct emission and angular distribution for the infra-red. Further discussion of the new ISRF is deferred to a forthcoming paper (Porter & Strong, in preparation).
![Angles involved in the calculation: $(R,z,\alpha)$, galactocentric coordinates of the source of VHE photons; $(\theta_2,\phi_2)$, angles of the VHE photon; GC marks the Galactic center; and the observer’s position is marked by “sun”. \[fig2\]](f2.eps){width="3.5in"}
Calculations
============
The optical depth for VHE [$\gamma$-rays]{} is given by the general formula: $$\begin{aligned}
\tau_{\gamma\gamma}(E)&&= \label{tau}\\
&&\int_L dx \int d\varepsilon \int d\Omega_1 \,
\frac{dN(\varepsilon,\Omega_1,x)}{d\varepsilon d\Omega_1}\sigma_{\gamma\gamma}
(\varepsilon_c) (1-\cos\theta),\nonumber\end{aligned}$$ where $dN(\varepsilon,\Omega_1,x)/d\varepsilon d\Omega_1$ is the differential number density of background photons at the point $x$, $\varepsilon$ is the background photon energy, $d\Omega_1=d\cos\theta_1 d\phi_1$ is a solid angle, $\sigma_{\gamma\gamma}$ is the total cross section for the pair production process $\gamma\gamma\to e^+e^-$ [@JR80], $\varepsilon_c=[\frac12\varepsilon E (1-\cos\theta)]^{1/2}$ is the center-of-momentum system energy of a photon, and $\theta$ is the angle between the momenta of the two photons in the observer’s frame. The integral over $x$ should be taken along the path of the [$\gamma$-rays]{} from the source to the observer.
The ISRF is cylindrically symmetric so that the photon angular distribution depends on $R$ and $z$ only. In Fig. \[fig2\] we illustrate the galactocentric coordinate system $(R,z,\alpha)$. To calculate $\cos\theta$, the polar and azimuthal angles of the VHE photon, $\theta_2$ and $\phi_2$, are derived: $$\begin{aligned}
&& \rho^2 = R^2+R_s^2-2RR_s\cos\alpha,\nonumber\\
&& \left\{ \begin{array}{l}
\sin\phi_2 = -(R_s/\rho) \sin\alpha,\smallskip\\
\cos\phi_2 = -(\rho^2+R^2-R_s^2)/2R\rho,
\end{array} \right. \nonumber\\
&& \left\{ \begin{array}{l}
\sin\theta_2 = \rho\, (\rho^2+z^2)^{-1/2},\smallskip\\
\cos\theta_2 = -(1-\sin^2\theta_2)^{1/2},
\end{array} \right. \\
&& \cos\theta =\cos\theta_1\cos\theta_2+\sin\theta_1\sin\theta_2\cos(\phi_1-\phi_2),\nonumber
\label{angles}\end{aligned}$$ where $R_s$ is the galactocentric radius of the Sun. The integration of eq. (\[tau\]) is done numerically.
For the calculation of the optical depth in the CMB field we use the formula: $$\begin{aligned}
\tau_{\gamma\gamma}^{CMB}(E) = &&\\
\frac{-4kT}{(\hbar c)^3 \pi^2 E^2} &&
\int_L dx \int_{m_ec^2}^{\infty} d\varepsilon_c \varepsilon_c^3
\sigma_{\gamma\gamma}
(\varepsilon_c) \log\left( 1-e^{-\varepsilon_c^2/E kT}\right),\nonumber
\label{tau_cmb}\end{aligned}$$ where $kT$ is the CMB temperature, and $m_ec^2$ is the electron rest mass.
Results
=======
Fig. \[fig3\] shows the attenuation for selected positions $(R, z, \alpha)$ as a function of incident [$\gamma$-ray]{} energy. For sources located at the GC the attenuation is $\sim$12% at 30 TeV and $\sim$23% at 100 TeV. In Table \[TabTau\] we give our optical depth results for selected values of $R$ and $z$ for $\alpha = 0^\circ, 90^\circ$ and $180^\circ$ at 30 and 100 TeV without contribution by the CMB. The attenuation is strongest for VHE [$\gamma$-ray]{} sources located toward the inner Galaxy and on its farside.
To illustrate the effect of the anisotropic radiation field on the attenuation calculation, we show in Fig. \[fig4\] the ratio of the optical depths $\tau_{\gamma\gamma}/\tau_{\gamma\gamma}^{\rm iso}$, where $\tau_{\gamma\gamma}$ is calculated using the full angular distribution of the ISRF and $\tau_{\gamma\gamma} ^{\rm iso}$ is calculated assuming an isotropic distribution, for a source located at $z = 0$ kpc (upper panel) and $z = 5$ kpc (lower panel) emitting 100 TeV [$\gamma$-rays]{} as a function of position. The variation of $\tau_{\gamma\gamma}/\tau_{\gamma\gamma} ^{\rm iso}$ over the Galaxy as seen from Earth is non-trivial. For sources located in the Galactic plane between the GC and solar system the ratio is less than one since the majority of the ISRF photons are coming from the GC direction, $\cos\theta \rightarrow 1$ in eq. (\[tau\]) and the interactions are mainly following, leading to a lower pair production probability when the ISRF angular distribution is taken into account. The reverse situation, $\tau_{\gamma\gamma}/\tau_{\gamma\gamma} ^{\rm iso} > 1$, occurs for [$\gamma$-rays]{} interacting with the ISRF in the outer Galaxy, where the majority of interactions are now head-on, $\cos\theta \rightarrow -1$ in eq. (\[tau\]). For sources located beyond the GC, the ratio is $\sim 1$. The interpretation of this case is straightforward: [$\gamma$-rays]{} emitted on the farside of the Galaxy toward the solar system have mainly head-on absorption interactions until they reach the GC, whereupon the interactions become mainly following. The isotropic case averages the angular distribution, and therefore leads to $\tau_{\gamma\gamma}/\tau_{\gamma\gamma} ^{\rm iso} \sim 1$. We note that the optical depth ratio increases monotonically with $z$, and is a result of the progressively more head-on nature of the [$\gamma$-ray]{} absorption interactions for sources located at larger distances from the plane.
![Transmittance of VHE [$\gamma$-rays]{} as a function of [$\gamma$-ray]{} energy. Solid line: $(R, z, \alpha) = (0$ kpc, 0 kpc, $0^\circ)$ – $L = 8.5$ kpc; Dashed line: $(R, z, \alpha) = (20$ kpc, 0 kpc, $90^\circ)$ – $L = 21.8$ kpc; Dash-dotted line: $(R, z, \alpha) = (20$ kpc, 0 kpc, $180^\circ)$ – $L = 28.5$ kpc. Thick lines give the total transmittance curve including the ISRF and CMB. Left-most thin lines give the transmittance for the ISRF only; right-most thin lines for the CMB only. \[fig3\]](f3.eps){width="3.5in"}
[ccrrrrrrrr]{} 0 & 0 & & 0.12 & & & & 0.25 & &\
5 & 0 & & 0.01 & 0.15 & 0.22 & & 0.05 & 0.32 & 0.41\
10 & 0 & & 0.01 & 0.08 & 0.28 & & 0.02 & 0.21 & 0.54\
15 & 0 & & 0.02 & 0.07 & 0.31 & & 0.05 & 0.18 & 0.58\
20 & 0 & & 0.03 & 0.07 & 0.32 & & 0.06 & 0.16 & 0.60\
0 & 5 & & 0.03 & & & & 0.09 & &\
5 & 5 & & 0.01 & 0.04 & 0.06 & & 0.05 & 0.10 & 0.15\
10 & 5 & & 0.02 & 0.05 & 0.09 & & 0.04 & 0.12 & 0.21\
15 & 5 & & 0.03 & 0.05 & 0.12 & & 0.05 & 0.13 & 0.26\
20 & 5 & & 0.03 & 0.06 & 0.14 & & 0.06 & 0.13 & 0.29\
\[TabTau\]
{width="3.5in"} {width="3.5in"}
Our results show that the attenuation of VHE [$\gamma$-rays]{} by the Galactic radiation field may be marginally observable by the HESS instrument which has an effective sensitivity up to several tens of TeV [@Aharonian2005]. The attenuation of VHE [$\gamma$-rays]{} from the sources on the Galaxy’s farside will essentially steepen their spectra above $\sim$10 TeV. In any case correction of source spectra for absorption is required. Interestingly, observations by future high energy experiments of the steepening of the spectra of Galactic sources in the GC region and beyond may serve as a probe of the Galactic ISRF.
I. V. M. acknowledges partial support from NASA Astronomy and Physics Research and Analysis Program (APRA) grant. T. A. P. acknowledges partial support from the US Department of Energy.
Aharonian, F., et al., 2005, ApJ, accepted (astro-ph/0510397)
Arendt, R. G., et al., [1998, , 508, 74]{}
Baes, M. & Dejonghe, H., [2001, , 326, 722]{}
Draine, B. T. & Li, A., [2001, , 551, 807]{}
Finkbeiner, D., Davis, M. & Schlegel, D. J., [1999, , 524, 867]{}
Freudenreich, H. T., [1998, , 492, 495]{}
Girardi, L., et al., [2002, , 391, 195]{}
Henry, R. C., Anderson, R. C., & Fastie, W. G., [1980, , 239, 859]{}
Henyey, L. G. & Greenstein, J. L., [1941, , 93, 70]{}
Jauch, J. M., & Rohrlich, F., [1980, Theory of Photons and Electrons (New York: Springer-Verlag)]{}
Li, A. & Draine, B. T., [2001, , 554, 778]{}
Moskalenko, I. V., Strong, A. W., Ormes, J. F., & Potgieter, M. S., [2002, , 565, 280]{}
Nikishov, A. I., [1962, Soviet Physics JETP, 14, 393]{}
Porter, T. A. & Strong, A. W., 2005, in Proc. $29^{\rm th}$ Int. Cosmic Ray Conf. (Pune) (astro-ph/0507119)
Protheroe, R. J., [1986, , 221, 769]{}.
Strong, A. W., Moskalenko, I. V., & Reimer, O., [2000, , 537, 763]{}
Strong, A. W., Moskalenko, I. V., Reimer, O., Digel, S. & Diehl, R., [2004, , 422, L47]{}
Wainscoat, R. J., et al., [1992, , 83, 111]{}
Weingartner, J. C. & Draine, B. T., [2001, , 548, 296]{}
Zhang, J.-L., Bi, X.-J., & Hu, H.-B. 2005, preprint (astro-ph/0508236)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Neural decoders were shown to outperform classical message passing techniques for short BCH codes. In this work, we extend these results to much larger families of algebraic block codes, by performing message passing with graph neural networks. The parameters of the sub-network at each variable-node in the Tanner graph are obtained from a hypernetwork that receives the absolute values of the current message as input. To add stability, we employ a simplified version of the arctanh activation that is based on a high order Taylor approximation of this activation function. Our results show that for a large number of algebraic block codes, from diverse families of codes (BCH, LDPC, Polar), the decoding obtained with our method outperforms the vanilla belief propagation method as well as other learning techniques from the literature.'
author:
- |
Eliya Nachmani and Lior Wolf\
Facebook AI Research and Tel Aviv University
bibliography:
- 'codes.bib'
title: 'Hyper-Graph-Network Decoders for Block Codes'
---
Introduction
============
Decoding algebraic block codes is an open problem and learning techniques have recently been introduced to this field. While the first networks were fully connected (FC) networks, these were replaced with recurrent neural networks (RNNs), which follow the steps of the belief propagation (BP) algorithm. These RNN solutions weight the messages that are being passed as part of the BP method with fixed learnable weights.
In this work, we add compute to the message passing iterations, by turning the message graph into a graph neural network, in which one type of nodes, called variable nodes, processes the incoming messages with a FC network $g$. Since the space of possible messages is large and its underlying structure random, training such a network is challenging. Instead, we propose to make this network adaptive, by training a second network $f$ to predict the weights $\theta_g$ of network $g$.
This “hypernetwork” scheme, in which one network predicts the weights of another, allows us to control the capacity, e.g., we can have a different network per node or per group of nodes. Since the nodes in the decoding graph are naturally stratified and since a per-node capacity is too high for this problem, the second option is selected. Unfortunately, training such a hypernetwork still fails to produce the desired results, without applying two additional modifications. The first modification is to apply an absolute value to the input of network $f$, thus allowing it to focus on the confidence in each message rather than on the content of the messages. The second is to replace the $arctanh$ activation function that is employed by the check nodes with a high order Taylor approximation of this function, which avoids its asymptotes.
When applying learning solutions to algebraic block codes, the exponential size of the input space can be mitigated by ensuring that certain symmetry conditions are met. In this case, it is sufficient to train the network on a noisy version of the zero codeword. As we show, the architecture of the hypernetwork we employ is selected such that these conditions are met.
Applied to a wide variety of codes, our method outperforms the current learning based solutions, as well as the classical BP method, both for a finite number of iterations and at convergence of the message passing iterations.
Related Work
============
Over the past few years, deep learning techniques were applied to error correcting codes. This includes encoding, decoding, and even, as shown recently in [@kim2018deepcode], designing new feedback codes. The new feedback codes, which were designed by an RNN, outperform the well-known state of the art codes (Turbo, LDPC, Polar) for a Gaussian noise channel with feedback.
Fully connected neural networks were used for decoding polar codes [@gruber2017deep]. For short polar codes, e.g., $n=16$ bits, the obtained results are close to the optimal performance obtained with maximum a posteriori (MAP) decoding. Since the number of codewords is exponential in the number of information bits $k$, scaling the fully connected network to larger block codes is infeasible.
Several methods were introduced for decoding larger block codes ($n\geqslant100$). For example in [@nachmani2016learning] the belief propagation (BP) decoding method is unfolded into a neural network in which weights are assigned to each variable edge. The same neural decoding technique was then extended to the min-sum algorithm, which is more hardware friendly [@lugosch2017neural]. In both cases, an improvement is shown in comparison to the baseline BP method.
Another approach was presented for decoding Polar codes [@cammerer2017scaling]. The polar encoding graph is partitioned into sub-blocks, and the decoding is performed to each sub-block separately. In [@kim2018communication] an RNN decoding scheme is introduced for convolutional and Turbo codes, and shown to achieve close to the optimal performance, similar to the classical convolutional codes decoders Viterbi and BCJR.
Our work decodes block codes, such as LDPC, BCH, and Polar. The most relevant comparison is with [@nachmani2017learning], which improve upon [@nachmani2016learning]. A similar method was applied to Polar code in [@teng2019low], and another related work on Polar codes [@cammerer2017scaling] introduced a non-iterative and parallel decoder. Another contribution learns the nodes activations based on components from existing decoders (BF, GallagerB, MSA, SPA) [@vasic2018learning]. In contrast, our method learns the node activations from scratch.
The term [*hypernetworks*]{} is used to refer to a framework in which a network $f$ is trained to predict the weights $\theta_g$ of another network $g$. Earlier work in the field [@klein2015dynamic; @7410424] learned weights of specific layers in the context of tasks that required a dynamic behavior. Fuller networks were trained to predict video frames and stereo views [@jia2016dynamic]. The term itself was coined in [@ha2016hypernetworks], which employed such meta-functions in the context of sequence modeling. A Bayesian formulation was introduced in a subsequent work [@krueger2017bayes]. The application of hyper networks as meta-learners in the context of few-shot learning was introduced in [@bertinetto2016learning].
An application of hypernetworks for searching over the architecture space, where evaluation is done with predicted weights conditioned on the architecture, rather than performing gradient descent with that architecture was proposed in [@brock2018smash]. Recently, graph hypernetworks were introduced for searching over possible architectures [@zhang2018graph]. Given an architecture, a graph hypernetwork that is conditioned on the graph of the architecture and shares its structure, generates the weights of the network with the given architecture. In our work, a non-graph network generates the weights of a graph network. To separate between the two approaches, we call our method hyper-graph-network and not graph hypernetwork.
Background
==========
We consider codes with a block size of $n$ bits. It is defined by a binary generator matrix $G$ of size $k \times n$ and a binary parity check matrix $H$ of size $(n-k) \times n$.
The parity check matrix entails a Tanner graph, which has $n$ variable nodes and $(n-k)$ check nodes, see Fig. \[fig:tanner\_Trellis\](a). The edges of the graph correspond to the on-bits in each column of the matrix $H$. For notational convenience, we assume that the degree of each variable node in the Tanner graph, i.e., the sum of each column of $H$, has a fixed value $d_v$.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![(a) The Tanner graph for a linear block code with $n=5$, $k=2$ and $d_v=2$. (b) The corresponding Trellis graph, with two iteration.[]{data-label="fig:tanner_Trellis"}](tanner_1.png "fig:"){width=".18\textwidth"} ![(a) The Tanner graph for a linear block code with $n=5$, $k=2$ and $d_v=2$. (b) The corresponding Trellis graph, with two iteration.[]{data-label="fig:tanner_Trellis"}](trellis_1.png "fig:"){width=".38\textwidth"}
(a) (b)
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
The Tanner graph is unrolled into a Trellis graph. This graph starts with $n$ variable nodes and is then composed of two types of columns, variable columns and check columns. Variable columns consist of *variable processing units* and check columns consist of *check processing units*. [ $d_v$ variable processing units are]{} associate with each received bit, and the number of processing units in the variable column is, therefore, $E=d_v n$. The check processing units are also directly linked to the edges of the Tanner graph, where each parity check corresponds to a row of $H$. Therefore, the check columns also have $E$ processing units each. The Trellis graph ends with an output layer of $n$ variable nodes. See Fig. \[fig:tanner\_Trellis\](b).
Message passing algorithms operate on the Trellis graph. The messages propagate from [variable columns]{} to [check columns]{} and from [check columns]{} to [variable columns]{}, in an iterative manner. The leftmost layer corresponds to a vector of log likelihood ratios (LLR) $l \in \mathbb{R}^n$ of the input bits: $$l_v = \log\frac{\Pr\left(c_v=1 | y_v\right)}{\Pr\left(c_v=0 | y_v\right)},$$ where $v\in[n]$ is an index and $y_v$ is the channel output for the corresponding bit $c_v$, which we wish to recover.
Let $x^j$ be the vector of messages that a column in the Trellis graph propagates to the next column. At the first round of message passing $j=1$, and similarly to other cases where $j$ is odd, a variable node type of computation is performed, in which the messages are added:
$$x^{j}_e = x^{j}_{(c,v)} = l_v + \sum_{e'\in N(v)\setminus \{(c,v)\}} x^{j-1}_{e'},
\label{eq:base_var}$$
where each variable node is indexed the edge $e=(c,v)$ on the Tanner graph and $N(v)=\{(c,v) | H(c,v)=1\}$, i.e, the set of all edges in which $v$ participates. By definition $x^0=0$ and when $j=1$ the messages are directly determined by the vector $l$.
For even $j$, the check layer performs the following computations: $$x^{j}_e = x^j_{(c,v)} = 2arctanh \left( \prod_{e'\in N(c) \setminus \{(c,v)\}}{tanh \left ( \frac{x^{j-1}_{e'}}{2} \right ) }\right)
\label{eq:base_check}$$ where $N(c)=\{(c,v) | H(c,v)=1\}$ is the set of edges in the Tanner graph in which row $c$ of the parity check matrix $H$ participates.
A slightly different formulation is provided by [@nachmani2017learning]. In this formulation, the $tanh$ activation is moved to the variable node processing units. In addition, a set of learned weights $w_e$ are added. Note that the learned weights are shared across all iterations $j$ of the Trellis graph.
$$\label{eq:odd}
x^{j}_e = x^{j}_{(c,v)} = \tanh \left(\frac{1}{2}\left(l_v + \sum_{e'\in N(v)\setminus \{(c,v)\}} w_{e'}x^{j-1}_{e'}\right)\right), ~~~~~~~~~~~~~~~\text{if $j$ is odd}$$
$$\label{eq:even}
x^{j}_e = x^j_{(c,v)} = 2arctanh \left( \prod_{e'\in N(c) \setminus \{(c,v)\}}{x^{j-1}_{e'}}\right)~~~~~~~~~~~~~~~\text{if $j$ is even}$$
As mentioned, the computation graph alternates between variable columns and check columns, with $L$ layers of each type. The final layer marginalizes [ the messages from the last check layer]{} with the logistic (sigmoid) activation function $\sigma$, and output $n$ bits. The $v$th bit output at layer $2L+1$, in the weighted version, is given by: $$o_v = \sigma \left( l_v + \sum_{e'\in N(v)}\bar{w}_{e'} x^{2L}_{e'} \right),
\label{eq:base_final}$$ where $\bar{w}_{e'}$ is a second set of learnable weights.
Method
======
We suggest further adding learned components into the message passing algorithm. Specifically, we replace Eq. \[eq:odd\] (odd $j$) with the following equation: $$\label{eq:oddreplaced}
x^{j}_e = x^{j}_{(c,v)} = g(l_v,x^{j-1}_{N(v,{\hbox{\tikz{\draw[line width=0.2pt,line cap=round] (1.5pt,0) -- (0,3pt);}}}c)},\theta_g^j),$$ where $x^{j}_{N(v,{\hbox{\tikz{\draw[line width=0.2pt,line cap=round] (1.5pt,0) -- (0,3pt);}}}c)}$ is a vector of length $d_v-1$ that contains the elements of $x^{j}$ that correspond to the indices $N(v)\setminus \{(c,v)\}$, and $\theta_g^j$ has the weights of network $g$ at iteration $j$. In order to make $g$ adaptive to the current input messages at every variable node, we employ a hypernetwork scheme and use a network $f$ to determine its weights. $$\label{eq:f}
\theta_g^j = f(|x^{j-1}|,\theta_f)$$ where $\theta_f$ are the learned weights of network $f$. Note that $g$ is fixed to all variable nodes at the same column. We have also experimented with different weights per variable (further conditioning $g$ on the specific messages $x^{j-1}_{N(v,{\hbox{\tikz{\draw[line width=0.2pt,line cap=round] (1.5pt,0) -- (0,3pt);}}}c)}$ for the variable with index $e=(v,c)$). However, the added capacity seems detrimental.
The adaptive nature of the hypernetwork allows the variable computation, for example to neglect part of the inputs of $g$, in case the input message $l$ contains errors.
Note that the messages $x^{j-1}$ are passed to $f$ in absolute value (Eq. \[eq:f\]). The absolute value of the messages is sometimes seen as measure for the correctness, and the sign of the message as the value (zero or one) of the corresponding bit [@richardson2008modern]. Since we want the network $f$ to focus on the correctness of the message and not the information bits, we remove the signs.
The architecture of both $f$ and $g$ does not contain bias terms and employs the $tanh$ activations. The network $g$ has $p$ layers, i.e., $\theta_{g}=\left(W_1,...,W_p\right)$, for some weight matrices $W_i$. The network $f$ ends with $p$ linear projections, each corresponding to one of the layers of network $g$. As noted above, if a set of symmetry conditions are met, then it is sufficient to learn to correct the zero codeword. The link between the architectural choices of the networks and the symmetry conditions is studied in Sec. \[sym\_cond\].
Another modification is being done to the columns of the check variables in the Trellis graph. For even values of $j$, we employ the following computation, instead of Eq. \[eq:even\]. $$x^j_e=x^j_{(c,v)} = 2\sum_{m=0}^{q}\frac{1}{2m+1}\left ( \prod_{e'\in N(c) \setminus \{(c,v)\}}{x_{e'}^{j-1}} \right )^{2m+1}
\label{eq:our_check}$$ in which $arctanh$ is replaced with its Taylor approximation of degree $q$. The approximation is employed as a way to [ stabilize the training process]{}. The $arctanh$ activation, has asymptotes in $x=1,-1$, and training with it often explodes. Its Taylor approximation is a well-behaved polynomial, see Figure \[fig:taylor\].
![Taylor Approximation of the $arctanh$ activation function.[]{data-label="fig:taylor"}](Taylor_fig.png){width=".52\textwidth"}
Training
--------
In addition to observing the final output of the network, as given in Eq. \[eq:base\_final\], we consider the following marginalization for each iteration where $j$ is odd: $o^{j}_v = \sigma \left( l_v + \sum_{e'\in N(v)}\bar{w}_{e'} x^{j}_{e'} \right)$. Similarly to [@nachmani2017learning], we employ the cross entropy loss function, which considers the error after every check node iteration out of the $L$ iterations: $$\mathcal{L}=-\frac{1}{n}\sum_{h=0}^{L}\sum_{v=1}^{n} c_{v}\log(o^{2h+1}_{v})+(1-c_{v})\log(1-o^{2h+1}_{v})
\label{eq:loss}$$ where $c_{v}$ is the ground truth bit. This loss simplifies, when learning the zero codeword, to $-\frac{1}{n}\sum_{h=0}^{L}\sum_{v=1}^{n} \log(1-o^{2h+1}_{v})$.
The learning rate was $1e-4$ for all type of codes, and the Adam optimizer [@kingma2014adam] is used for training. The decoding network has ten layers which simulates $L=5$ iterations of a modified BP algorithm.
Symmetry conditions {#sym_cond}
===================
For block codes that maintain certain symmetry conditions, the decoding error is independent of the transmitted codeword [@richardson2008modern Lemma 4.92]. A direct implication is that we can train our network to decode only the zero codeword. Otherwise, training would need to be performed for all $2^k$ words. [ Note that training with the zero codeword should give the same results as training with all $2^k$ words.]{}
There are two symmetry conditions.
1. For a check node with index $(c,v)$ at iteration $j$ and for any vector $b\in \{0,1\}^{d_v-1}$ $$\Phi\left ( b^{\top}x^{j-1}_{N({\hbox{\tikz{\draw[line width=0.2pt,line cap=round] (1.5pt,0) -- (0,3pt);}}}v,c)} \right )= \left ( \prod_{1}^{K} b_{k}\right ) \Phi\left ( x^{j-1}_{N({\hbox{\tikz{\draw[line width=0.2pt,line cap=round] (1.5pt,0) -- (0,3pt);}}}v,c)} \right )
\label{eq:sym_check}$$ where $x^{j}_{N({\hbox{\tikz{\draw[line width=0.2pt,line cap=round] (1.5pt,0) -- (0,3pt);}}}v,c)}$ is a vector of length $d_v-1$ that contains the elements of $x^{j}$ that correspond to the indices $N(c)\setminus \{(c,v)\}$ and $\Phi$ is the activation function used, e.g., $arctanh$ or the truncated version of it.
2. For a variable node with index $(c,v)$ at iteration $j$, which performs computation $\Psi$ $$\Psi\left(-l_v,-x^{j-1}_{N(v,{\hbox{\tikz{\draw[line width=0.2pt,line cap=round] (1.5pt,0) -- (0,3pt);}}}c)}\right )= -\Psi\left (l_v,x^{j-1}_{N(v,{\hbox{\tikz{\draw[line width=0.2pt,line cap=round] (1.5pt,0) -- (0,3pt);}}}c)}\right)
\label{eq:sym_var_0}$$ In the proposed architecture, $\Psi$ is a FC neural network ($g$) with $tanh$ activations and no bias terms.
Our method, by design, maintains the symmetry condition on both the variable and the check nodes. This is verified in the following lemmas.
Assuming that the check node calculation is given by Eq. then the proposed architecture satisfies the first symmetry condition.
In our case the activation function $\Phi$ is Taylor approximation of $arctanh$. Let the input message at $j$ be $x^{j}_{N({\hbox{\tikz{\draw[line width=0.2pt,line cap=round] (1.5pt,0) -- (0,3pt);}}}v,c)} = \left(x^{j}_{1}, \dots, x^{j}_{K} \right)$ for $K=d_v-1$. We can verify that:
$$\begin{aligned}
x^{j}(b_{1}x^{j-1}_{1}, ..., b_{K}x^{j-1}_{K})
&= 2\sum_{m=0}^{q}\frac{1}{2m+1} ( \prod_{k=1}^{K}{b_{k}x^{j-1}_{k}} )^{2m+1} = 2( \prod_{k=1}^{K}{b_{k}} ) \sum_{m=0}^{q}\frac{1}{2m+1} ( \prod_{k=1}^{K}{x^{j-1}_{k}} )^{2m+1}\\
&= ( \prod_{k=1}^{K}{b_{k}} )x_{j}(x^{j-1}_{1}, ..., x^{j-1}_{K})\qedhere
\label{eq:sc_check_proof}
\end{aligned}$$
where the second equality holds since $2m+1$ is odd.
Assuming that the variable node calculation is given by Eq. and Eq. , $g$ does not contain bias terms and employs the $tanh$ activation, then the proposed architecture satisfies the variable symmetry condition.
Let $K=d_v-1$ and $x^{j}_{N(v,{\hbox{\tikz{\draw[line width=0.2pt,line cap=round] (1.5pt,0) -- (0,3pt);}}}c)} = \left(x^{j}_{1}, \dots, x^{j}_{K} \right)$. In the proposed architecture for any odd $j\geqslant0$, $\Psi$ is given as $$g\left(l_v,x^{j-1}_{1},\dots, x^{j-1}_{K}, \theta^{j}_g\right) = tanh\left ( W_{p}^\top \enspace ... \enspace tanh\left (W_{2}^\top tanh\left ( W_{1}^\top \left(l_v,x^{j-1}_{1},\dots, x^{j-1}_{K}\right) \right )\right )\right )$$
where $p$ is the number of layers and the weights $W_{1},...,W_{p}$ constitute $\theta^{j}_g=f(|x^{j-1}|,\theta_f)$.
For real valued weights $\theta^{lhs}_g$ and $\theta^{rhs}_g$, since $tanh(x)$ is an odd function for any real value input, if $\theta^{lhs}_g = \theta^{rhs}_g$ then $g\left(l_v,x^{j-1}_{1},\dots, x^{j-1}_{K},\theta^{lhs}_g \right) = - g\left(-l_v,-x^{j-1}_{1},\dots, -x^{j-1}_{K},\theta^{rhs}_g \right)$. In our case, $\theta^{lhs}_g = f(|x^{j-1}|,\theta_f) = f(|-x^{j-1}|,\theta_f) = \theta^{rhs}_g$.
Experiments
===========
In order to evaluate our method, we train the proposed architecture with three classes of linear block codes: Low Density Parity Check (LDPC) codes [@gallager1962low], Polar codes [@arikan2008channel] and Bose–Chaudhuri–Hocquenghem (BCH) codes [@bose1960class]. All generator matrices and parity check matrices are taken from [@channelcodes].
Training examples are generated as a zero codeword transmitted over an additive white Gaussian noise. For validation, we use the generator matrix $G$, in order to simulate valid codewords. Each training batch contains examples with different Signal-To-Noise (SNR) values.
The hyperparameters for each family of codes are determined by practical considerations. For Polar codes, which are denser than LDPC codes, we use a batch size of $90$ examples. We train with SNR values of $1dB,2dB,..,6dB$, where from each SNR we present $15$ examples per single batch. For BCH and LDPC codes, we train for SNR ranges of $1-8dB$ (120 samples per batch). In our results we report, the test error up to an SNR of $6dB$, since evaluating the statistics for higher SNRs in a reliable way requires the evaluation of a large number of test samples (recall that in train, we only need to train on a noisy version of a single codeword). However, for BCH codes, which are the focus of the current literature, we extend the tests to $8dB$ in some cases.
In our experiments, the order of the Taylor series of $arctanh$ is set to $q=1005$. The network $f$ has four layers with $32$ neurons at each layer. The network $g$ has two layer with $16$ neurons at each layer. For BCH codes, we also tested a deeper configuration in which the network $f$ has four layers with $128$ neurons at each layer.
The results are reported as bit error rates (BER) for different SNR values (dB). Fig. \[fig:ber\_snr\] shows the results for sample codes, and Tab. \[tab:ber\_snr\] lists results for more codes. As can be seen in the figure for Polar(128,96) code with five iteration of BP we get an improvement of $0.48dB$ over [@nachmani2017learning]. For LDPC MacKay(96,48) code, we get an improvement of $0.15dB$. For the BCH(63,51) with large $f$ we get an improvement of $0.45dB$ and with small $f$ we get a similar improvement of $0.43dB$. Furthermore, for every number of iterations, our method obtains better results then [@nachmani2017learning]. We can also observe that our method with $5$ iteration achieve the same results as [@nachmani2017learning] with $50$ iteration, for BCH(63,51) and Polar(128,96) codes. Similar improvements were also observe for other BCH and Polar codes. [ Fig. \[fig:ber\_snr\](e) provides experiments for large and non-regular LDPC codes - WARN$(384,256)$ and TU-KL$(96,48)$. As can be seen, our method improves the results, even in non-regular codes where the degree varies. Note that we learned just one hypernetwork $g$, which corresponds to the maximal degree and we discard irrelevant outputs for nodes with lower degrees.]{} In Tab. \[tab:ber\_snr\] we present the negative natural logarithm of the BER. For the 15 block codes tested, our method get better results then the BP and [@nachmani2017learning] algorithms. This results stay true for the convergence point of the algorithms, i.e. when we run the algorithms with $50$ iteration.
[lc@[ ]{}c@[ ]{}cc@[ ]{}c@[ ]{}cc@[ ]{}c@[ ]{}cc@[ ]{}c@[ ]{}c]{}
Method & & & &\
(lr)[2-4]{} (lr)[5-7]{} (lr)[8-10]{} (lr)[11-13]{} & 4 & 5 & 6 & 4 & 5 & 6 & 4 & 5 & 6 & 4 & 5 & 6\
\
Polar (63,32) & 3.52 & 4.04 & 4.48 & 4.14 & 5.32 & 6.67 & 4.25 & 5.49 & 7.02 & — & — & —\
Polar (64,48) & 4.15 & 4.68 & 5.31 & 4.77 & 6.12 & 7.84 & 4.91 & 6.48 & 8.41 & — & — & —\
Polar (128,64)& 3.38 & 3.80 & 4.15 & 3.73 & 4.78 & 5.87 & 3.89 & 5.18 & 6.94 & — & — & —\
Polar (128,86) & 3.80 & 4.19 & 4.62 & 4.37 & 5.71 & 7.19 & 4.57 & 6.18 & 8.27 & — & — & —\
Polar (128,96) & 3.99 & 4.41 & 4.78 & 4.56 & 5.98 & 7.53 & 4.73 & 6.39 & 8.57 & — & — & —\
LDPC (49,24) & 5.30 & 7.28 & 9.88 & 5.49 & 7.44 & 10.47 & 5.76 & 7.90 & 11.17 & — & — & —\
LDPC (121,60) & 4.82 & 7.21 & 10.87 & 5.12 & 7.97 & 12.22 & 5.22 & 8.29 & 13.00 & — & — & —\
LDPC (121,70) & 5.88 & 8.76 & 13.04 & 6.27 & 9.44 & 13.47 & 6.39 & 9.81 & 14.04 & — & — & —\
LDPC (121,80) & 6.66 & 9.82 & 13.98 & 6.97 & 10.47 & 14.86 & 6.95 & 10.68 & 15.80 & — & — & —\
MacKay (96,48) & 6.84 & 9.40 & 12.57 & 7.04 & 9.67 & 12.75 & 7.19 & 10.02 & 13.16 & — & — & —\
CCSDS (128,64) & 6.55 & 9.65 & 13.78 & 6.82 & 10.15 & 13.96 & 6.99 & 10.57 & 15.27 & — & — & —\
BCH (31,16) & 4.63 & 5.88 & 7.60 & 4.74 & 6.25 & 8.00 & 5.05 & 6.64 & 8.80 & 4.96 & 6.63 & 8.80\
BCH (63,36) & 3.72 & 4.65 & 5.66 & 3.94 & 5.27 & 6.97 & 3.96 & 5.35 & 7.20 & 4.00 & 5.42 & 7.34\
BCH (63,45) & 4.08 & 4.96 & 6.07 & 4.37 & 5.78 & 7.67 & 4.48 & 6.07 & 8.45 & 4.41 & 5.91 & 7.91\
BCH (63,51) & 4.34 & 5.29 & 6.35 & 4.54 & 5.98 & 7.73 & 4.64 & 6.08 & 8.16 & 4.67 & 6.19 & 8.22\
\
Polar (63,32) & 4.26 & 5.38 & 6.50 & 4.22 & 5.59 & 7.30 & 4.59 & 6.10 & 7.69 & — & — & —\
Polar (64,48) & 4.74 & 5.94 & 7.42 & 4.70 & 5.93 & 7.55 & 4.92 & 6.44 & 8.39 & — & — & —\
Polar (128,64) & 4.10 & 5.11 & 6.15 & 4.19 & 5.79 & 7.88 & 4.52 & 6.12 & 8.25 & — & — & —\
Polar (128,86) & 4.49 & 5.65 & 6.97 & 4.58 & 6.31 & 8.65 & 4.95 & 6.84 & 9.28 & — & — & —\
Polar (128,96) & 4.61 & 5.79 & 7.08 & 4.63 & 6.31 & 8.54 & 4.94 & 6.76 & 9.09 & — & — & —\
LDPC (49,24)& 6.23 & 8.19 & 11.72 & 6.05 & 8.34 & 11.80 & 6.23 & 8.54 & 11.95 & — & — & —\
MacKay (96,48) & 8.15 & 11.29 & 14.29 & 8.66 & 11.52 & 14.32 & 8.90 & 11.97 & 14.94 & — & — & —\
BCH (63,36) & 4.03 & 5.42 & 7.26 & 4.15 & 5.73 & 7.88 & — & — & — & 4.29 & 5.91 & 8.01\
BCH (63,45) & 4.36 & 5.55 & 7.26 & 4.49 & 6.01 & 8.20 & — & — & — & 4.64 & 6.27 & 8.51\
BCH (63,51) & 4.58 & 5.82 & 7.42 & 4.64 & 6.21 & 8.21 & — & — & — & 4.80 & 6.44 & 8.58\
To evaluate the contribution of the various components of our method, we ran an ablation analysis. We compare (i) our complete method, (ii) a method in which the parameters of $g$ are fixed and $g$ receives and additional input of $|x^{j-1}|$, (iii) a similar method where the number of hidden units in $g$ was increased to have the same amount of parameters of $f$ and $g$ combined, (iv) a method in which $f$ receives the $x^{j-1}$ instead of the absolute value of it, (v) a variant of our method in which $arctanh$ replaces its Taylor approximation, and (vi) a similar method to the previous one, in which gradient clipping is used to prevent explosion. The results, reported in Tab. \[tab:ablation\] demonstrate the advantage of our complete method. We can observe that without hypernetwork and without the absolute value in Eq. \[eq:f\], the results degrade below those of [@nachmani2017learning]. We can also observe that for (ii), (iii) and (iv) the method reaches the same low quality performance. For (v) and (vi), the training process explodes and the performance is equal to a random guess. [ In (vi), we train our method while clipping the arctanh at multiple threshold values ($\textrm{TH}=0.5, 1, 2, 4, 5$, applied to both the positive and negative sides, multiple block codes BCH(31,16), BCH(63,45), BCH(63,51), LDPC (49,24), LDPC (121,80), POLAR(64,32), POLAR(128,96), $L=5$ iterations). In all cases, the training exploded, similar to the no-threshold vanilla arctanh (v). In order to understand this, we observe the values when arctanh is applied at initialization for our method and for [@nachmani2016learning; @nachmani2017learning]. In [@nachmani2016learning; @nachmani2017learning], which are initialized to mimic the vanilla BP, the activations are such that the maximal arctanh value at initialization is 3.45. However in our case, in many of the units, the value explodes at infinity. Clipping does not help, since for any threshold value, the number of units that are above the threshold (and receive no gradient) is large. Since we employ hypernetworks, the weights $\theta^j_g$ of the network $g$ are dynamically determined by the network $f$ and vary between samples, making it challenging to control the activations $g$ produces. This highlights the critical importance of the Taylor approximation for the usage of hypernetworks in our setting.]{} The table also shows that for most cases, the method of [@nachmani2017learning] slightly benefits from the usage of approximated $arctanh$.
--------------------------------------------------- ------ ------ ------ ------ ------ ------
Code
(lr)[2-3]{} (lr)[4-5]{} (lr)[6-7]{} Variant/SNR 4 6 4 6 4 6
\(i) Complete method 4.96 8.80 4.41 7.91 4.67 8.22
\(ii) No hypernetwork 2.94 3.85 3.54 4.76 3.83 5.18
\(iii) No hypernetwork, higher capacity 2.94 3.85 3.54 4.76 3.83 5.18
\(iv) No abs in Eq. \[eq:f\] 2.86 3.99 3.55 4.77 3.84 5.20
\(v) Not truncating $arctanh$ 0.69 0.69 0.69 0.69 0.69 0.69
\(vi) Gradient clipping 0.69 0.69 0.69 0.69 0.69 0.69
[@nachmani2017learning] 4.74 8.00 3.97 7.10 4.54 7.73
[@nachmani2017learning] with truncated $arctanh$ 4.78 8.24 4.34 7.34 4.53 7.84
--------------------------------------------------- ------ ------ ------ ------ ------ ------
: Ablation analysis. The negative natural logarithm of BER results of our complete method are compared with alternative methods. Higher is better.[]{data-label="tab:ablation"}
Conclusions
===========
We presents graph networks in which the weights are a function of the node’s input, and demonstrate that this architecture provides the adaptive computation that is required in the case of decoding block codes. Training networks in this domain can be challenging and we present a method to avoid gradient explosion that seems more effective, in this case, than gradient clipping. By carefully designing our networks, important symmetry conditions are met and we can train efficiently. Our results go far beyond the current literature on learning block codes and we present results for a large number of codes from multiple code families.
### Acknowledgments {#acknowledgments .unnumbered}
[ We thank Sebastian Cammerer and Chieh-Fang Teng for the helpful discussion and providing code for deep polar decoder. The contribution of Eliya Nachmani is part of a Ph.D. thesis research conducted at Tel Aviv University.]{}
$
\begin{array}{c@{~}c}
\includegraphics[width=.51\textwidth]{Polar_128_96_rebattle.png}&
\includegraphics[width=.51\textwidth]{MacKay_96_48.png} \\
(a) & (b)\\
\end{array}$
$
\begin{array}{c@{~}c}
\includegraphics[width=.51\textwidth]{BCH_63_51_small_f_32_neurons_rebattle.png}&
\includegraphics[width=.51\textwidth]{BCH_63_51_large_f_128_neurons.png} \\
(c) & (d)\\
\end{array}$
$
\begin{array}{c@{~}c}
\includegraphics[width=.51\textwidth]{Non_Regular_LDPC_-_WRAN_384_256_TU-KL_96_48_rebattle.png} \\
(e)\\
\end{array}$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'It is well known that the three parameters that characterize the Kerr black hole (mass, angular momentum and horizon area) satisfy several important inequalities. Remarkably, some of these inequalities remain valid also for dynamical black holes. This kind of inequalities play an important role in the characterization of the gravitational collapse. They are closed related with the cosmic censorship conjecture. In this article recent results in this subject are reviewed.'
author:
- |
Sergio Dain\
Facultad de Matemática, Astronomía y Física, FaMAF,\
Universidad Nacional de Córdoba,\
Instituto de Física Enrique Gaviola, IFEG, CONICET,\
Ciudad Universitaria (5000) Córdoba, Argentina.
title: Geometric inequalities for black holes
---
Geometric inequalities in General Relativity {#sec:geom-ineq}
============================================
A classical example of a geometric inequality is the isoperimetric inequality for closed plane curves given by $$\label{eq:54}
L^2 \geq 4\pi A\quad (=\text{ circle}),$$ where $A$ is the area enclosed by a curve $C$ of length $L$. In (\[eq:54\]) equality holds if and only if $C$ is a circle, see figure \[fig:1\]. For a review on this subject see [@Osserman78].
{width="3cm"} {width="3cm"}
\[fig:1\]
The inequality (\[eq:54\]) applies to complicated geometric objects (i.e. arbitrary closed planar curves). The equality in (\[eq:54\]) is achieved only for an object of “optimal shape” (i.e. the circle) which is described by few parameters (in this case only one: the radius). Moreover, this object has a variational characterization: the circle is uniquely characterized by the property that among all simple closed plane curves of given length $L$, the circle of circumference $L$ encloses the maximum area.
General Relativity is a geometric theory, hence it is not surprising that geometric inequalities appear naturally in it. Many of these inequalities are similar in spirit as the isoperimetric inequality (\[eq:54\]). In particular, all the geometric inequalities discussed in this article will have the same structure as (\[eq:54\]): the inequality applies for a rich class of objects and the equality only applies for an object of “optimal shape” (always indicated in parenthesis as in (\[eq:54\])). This object, like the circle, can be described by few parameters and it has also a variational characterization.
However, General Relativity is also a physical theory. It is often the case that the quantities involved have a clear physical interpretation and the expected behavior of the gravitational and matter fields often suggests geometric inequalities which can be highly non-trivial from the mathematical point of view. The interplay between physics and geometry gives to geometric inequalities in General Relativity their distinguished character. These inequalities relate quantities that have both a physical interpretation and a geometrical definition.
The plan of this article follows this interplay between physics and mathematics. In section \[sec:physical-picture\] we present the physical motivations for the black holes geometric inequalities. In section \[sec:theorems\] we summarize some theorems where these inequalities have been recently proved. Finally, in section \[sec:open-problems-recent\] we list relevant open problems and we also describe recent results on geometric inequalities for bodies.
Physical picture {#sec:physical-picture}
================
An important example of a geometric inequality is the positive mass theorem. Let $m$ be the total ADM mass on an asymptotically flat complete initial data such that the dominant energy condition is satisfied. Then we have $$\label{eq:1}
0\leq m\quad (=\text{ Minkowski}).$$ The mass $m$ is a pure geometrical quantity [@Arnowitt62][@Bartnik86][@chrusciel86]. However, from the geometrical mass definition, without the physical picture, it would be very hard even to conjecture the inequality (\[eq:1\]). In fact the proof of the positive mass theorem turns out to be very subtle [@Schoen79b][@Schoen81][@witten81].
A key assumption in the positive mass theorem is that the matter fields should satisfy an energy condition. This condition is expected to hold for all physically realistic matter. This kind of general properties which do not depend very much on the details of the model are not easy to find for a macroscopic object. And hence it is difficult to obtain simple and general geometric inequalities among the parameters that characterize ordinary macroscopic objects. Black holes represent a unique class of very simple macroscopic objects and hence they are natural candidates for geometrical inequalities. Nevertheless, in section \[sec:open-problems-recent\] we will present also a geometric inequality valid for ordinary bodies.
The black hole uniqueness theorem ensures that stationary black holes in vacuum are characterized by the Kerr exact solution of Einstein equations [^1]. For simplicity we will not consider the electromagnetic field in this article, however most of the results presented here can be generalized to include that case.
It is somehow remarkable that the same family of solutions of Einstein equations that describe the unique stationary black hole (i.e. the Kerr metric) also describe naked singularities. In effect, the Kerr metric depends on two parameters: the mass $m$ and the angular momentum $J$. This metric is a solution of Einstein vacuum equations for any choice of the parameters $m$ and $J$. However, it represents a black hole if and only if the following remarkably inequality holds $$\label{eq:2}
\sqrt{|J|}\leq m.$$ Otherwise the spacetime contains a naked singularity. Figure \[fig:2\] shows the parameter space of the Kerr solution. Extreme black holes are defined by the equality in (\[eq:2\]). These black holes lie at the boundary between naked singularities and black holes. For most of the inequalities discussed in this article, extreme black holes play the role of the circle in the isoperimetric inequality (\[eq:54\]): they reach the equality and they represent objects of “optimal shape”.
![A point in this graph is a Kerr solution with parameters $m$ and $J$. The horizontal axis where $m=0$ is Minkowski space. The Schwarzschild solution is given by the vertical axis where $J=0$. In the gray region the parameters satisfy the inequality (\[eq:2\]) and hence the Kerr solution describe a black hole. The boundary of this region is given by the equality in (\[eq:2\]), these solutions are called extreme black holes. In the white region, excluding the horizon axis, the Kerr solution contains a naked singularity. That includes also the negative mass region. []{data-label="fig:2"}](kerr-parametros-p.pdf){width="70.00000%"}
The area of the horizon of the Kerr black hole is given by the simple but very important formula $$\label{eq:3}
A=8\pi \left(m^2+ \sqrt{m^4-J^2} \right).$$ From equation (\[eq:3\]) we deduce that the following three geometric inequalities hold for a Kerr black hole $$\begin{aligned}
\sqrt{\frac{A}{16\pi}} &\leq m &(=\text{Schwarzschild}),\label{eq:pen}\\
\sqrt{|J|} &\leq m &(= \text{Extreme Kerr}),\label{eq:mj}\\
8\pi |J| &\leq A &(= \text{Extreme Kerr}).\label{eq:JA}\end{aligned}$$ As expected from the discussion above, the inequality (\[eq:mj\]) is needed to define the black hole horizon area in (\[eq:3\]): if (\[eq:mj\]) does not hold, then the expression (\[eq:3\]) is not a real number. We have listed this inequality again here to emphasize its connection with the other two in the following discussion. Inequalities (\[eq:pen\]) and (\[eq:JA\]) follow from (\[eq:mj\]) and (\[eq:3\]). Note that these inequalities relate the three relevant parameters of the Kerr black hole $(m,J,A)$.
Let us discuss the physical meaning of the inequalities (\[eq:pen\]), (\[eq:mj\]) and (\[eq:JA\]). In the inequality (\[eq:pen\]), the difference $$\label{eq:4b}
m-\sqrt{\frac{A}{16\pi}},$$ represents the rotational energy of the Kerr black hole. This is the maximum amount of energy that can be extracted from the black hole by the Penrose process [@Christodoulou70]. When the difference (\[eq:4b\]) is zero, the black hole has no angular momentum and hence it is the Schwarzschild black hole.
From Newtonian considerations, we can interpret the inequality (\[eq:mj\]) as follows [@Wald71]. In a collapse the gravitational attraction ($\approx
m^2/r^2$) at the horizon ($r \approx m $) dominates over the centrifugal repulsive forces ($\approx J^2/mr^3$).
Finally, concerning the inequality (\[eq:JA\]), the black hole temperature is given by the following formula $$\label{eq:5c}
\kappa= \frac{1}{4 m} \left(1-\frac{(8\pi J)^2 }{A^2} \right).$$ The temperature is positive if and only if the inequality (\[eq:JA\]) holds. Moreover the temperature is zero if and only if the equality in (\[eq:JA\]) holds and hence the black hole is extreme.
There exists another relevant geometrical inequality which can be deduced from the formula (\[eq:3\]) $$\label{eq:10}
8\pi \left( m^2-\sqrt{m^4-J^2} \right) \leq A \quad (= \text{Extreme Kerr}).$$ Remarkably, as it was pointed out in [@Khuri:2013wha] for the case of the electric charge and in [@Dain:2013qia] for the present case of angular momentum, the inequality (\[eq:10\]) can be deduced purely from the inequalities (\[eq:mj\]) and (\[eq:JA\]) (i.e. without using the equality (\[eq:3\])) by simple algebra. Namely $$\begin{aligned}
\label{eq:11}
m^2 &= \sqrt{m^4-J^2+J^2},\\
& \leq |J| +\sqrt{m^4-J^2}, \label{eq:11b} \\
& \leq \frac{A}{8\pi}+ \sqrt{m^4-J^2},\label{eq:11c}\end{aligned}$$ where in the line (\[eq:11b\]) we have used (\[eq:mj\]) and in line (\[eq:11c\]) we have used (\[eq:JA\]). In that sense, the inequalities (\[eq:pen\]), (\[eq:mj\]) and (\[eq:JA\]) are more fundamental than (\[eq:10\]). However, the inequality (\[eq:10\]) is important by itself since it related with the Penrose inequality with angular momentum, see [@Khuri:2013wha] [@Dain:2013qia].
We have seen that for stationary black holes the inequalities (\[eq:pen\]), (\[eq:mj\]) and (\[eq:JA\]) are straightforward consequences of the area formula (\[eq:3\]).
![Schematic representation of an initial data for a non-stationary black hole. The black ring represents a trapped surface. Outside and inside the trapped surface the gravitational field is highly dynamical.[]{data-label="fig:3"}](dynamical-bh-p.pdf){width="6cm"}
However, black holes are in general non stationary, see figure \[fig:3\]. Astrophysical phenomena like the formation of a black hole by gravitational collapse or a binary black hole collision are highly dynamical. For such systems, the black hole can not be characterized by few parameters as in the stationary case. In fact, even stationary but non-vacuum black holes have a complicated structure (for example black holes surrounded by a rotating ring of matter, see the numerical studies in [@Ansorg05]). Remarkably, inequalities (\[eq:pen\]), (\[eq:mj\]) and (\[eq:JA\]) extend (under appropriate assumptions) to the fully dynamical regime. Moreover, these inequalities are deeply connected with properties of the global evolution of Einstein equations, in particular with the cosmic censorship conjecture.
To discuss the physical arguments that support these inequalities in the dynamical regime it is convenient to start with the inequality . For a dynamical black hole, the physical quantities that are well defined are the total ADM mass $m$ of the spacetime and the area $A$ of the black hole horizon. The total mass $m$ of the spacetime measures the sum of the black hole mass and the mass of the gravitational waves surrounding it. In the stationary case, the mass of the black hole is equal to the total mass of the spacetime, but this is no longer true for a dynamical black hole. The mass $m$ is a global quantity, it carries information on the whole spacetime. In contrast, the area of the horizon $A$ is a quasi-local quantity, it carries information on a bounded region of the spacetime.
It is well known that the energy of the gravitational field cannot be represented by a local quantity (i.e. a scalar field). The best one can hope is to obtain a quasi-local expression. The same applies to the angular momentum. In general, it is difficult to find physically relevant quasi-local quantities like mass and angular momentum (see the review article [@Szabados04]). However, in axial symmetry, there is a well defined notion of quasi-local angular momentum: the Komar integral of the axial Killing vector. Moreover, the angular momentum is conserved in vacuum. That is, axially symmetric gravitational waves do not carry angular momentum.
Then, for axially symmetric dynamical black holes we have two well defined quasi-local quantities: the area of the horizon $A$ and the angular momentum $J$. Note that the inequality relates only quasi-local quantities.
Using $A$ and $J$ we can define the quasi-local mass for a dynamical black hole by the Kerr formula (\[eq:3\]), that is $$\label{eq:masa}
{m_{bh}}= \sqrt{\frac{A}{16\pi}+\frac{4\pi J^2}{A}}.$$ This is, in principle, just a definition. Since ${m_{bh}}$ is given by the Kerr formula (\[eq:3\]) it automatically satisfies the inequalities (\[eq:mj\]) and (\[eq:pen\]). However, the relevant question is: does ${m_{bh}}$ describes the quasi-local mass of a non-stationary black hole? This question is closed related to the validity of the inequality (\[eq:JA\]) in the dynamical regime. In order to answer it let us analyze the evolution of ${m_{bh}}$.
For a dynamical black hole, by the area theorem, we know that the horizon area $A$ increase with time, see figure \[fig:4\].
![The area theorem. The horizon area of a dynamical black hole increase with time.[]{data-label="fig:4"}](area-th-p.pdf){width="6cm"}
In general, the quasi-local mass of the black hole is not expected to be a monotonically increasing quantity. Energy can be extracted from a rotating black hole by the Penrose process. However, if we assume axial symmetry then the angular momentum will be conserved at the quasi-local level. On physical grounds, one would expect that in this situation the quasi-local mass of the black hole should increase with the area, since there is no mechanism at the classical level to extract mass from the black hole. In effect, the Penrose process involves an interchange of angular momentum between the black hole and the exterior. But the angular momentum transfer is forbidden in axial symmetry. Then, both the area $A$ and the quasi- local mass ${m_{bh}}$ should monotonically increase with time in axial symmetry.
Let us take a time derivative of ${m_{bh}}$. To analyze this, it is illustrative to write down the complete differential, namely the first law of thermodynamics $$\label{eq:mq}
\delta {m_{bh}}= \frac{\kappa}{8 \pi} \delta A + \Omega_H \delta J,$$ where $$\label{eq:7}
\kappa= \frac{1}{4{m_{bh}}} \left(1-\frac{(8\pi J)^2 }{A^2} \right),\quad
\Omega_H=\frac{4\pi J}{A \,{m_{bh}}}.$$ In equation (\[eq:mq\]) we have followed the standard notation for the formulation of the first law; we emphasize, however, that in our context this equation is a trivial consequence of . In axial symmetry $\delta
J=0$ and hence we obtain $$\delta {m_{bh}}= \frac{\kappa}{8 \pi} \delta A.$$ By the area theorem we have $$\delta A \geq 0.$$ Then $\delta {m_{bh}}\geq 0$ if and only if $\kappa \geq 0$, that is $\delta {m_{bh}}\geq
0$ if and only if the inequality (\[eq:JA\]) holds. Then, it is natural to conjecture that this inequality should be satisfied for any axially symmetric black hole. If the horizon violates (\[eq:JA\]), then in the evolution the area will increase but the mass ${m_{bh}}$ will decrease. This will indicate that the quantity ${m_{bh}}$ does not have the desired physical meaning. Also, a rigidity statement is expected. Namely, the equality in (\[eq:JA\]) is reached only by the extreme Kerr black hole where $\kappa=0$.
This inequality provides a remarkable quasi-local measure of how far a dynamical black hole is from the extreme case, namely an ‘extremality criteria’ in the spirit of [@Booth:2007wu], although restricted only to axial symmetry. In the article [@Dain:2007pk] it has been conjectured that, within axially symmetry, to prove the stability of a nearly extreme black hole is perhaps simpler than a Schwarzschild black hole. It is possible that this quasi-local extremality criteria will have relevant applications in this context. Note also that the inequality allows to define, at least formally, the positive temperature of a dynamical black hole $\kappa$ by the formula (\[eq:7\]) (see Refs. [@Ashtekar03] [@Ashtekar02] for a related discussion of the first law in dynamical horizons). If inequality holds, then ${m_{bh}}$ defines a non-trivial quantity that increase monotonically with time, like the black hole area $A$.
It is important to emphasize that the physical arguments presented above in support of are certainly weaker in comparison with the ones behind the Penrose inequalities that support the inequalities (\[eq:pen\]) and (\[eq:mj\]) that we will discuss bellow. A counter example of any of these inequality will prove that the standard picture of the gravitational collapse is wrong. On the other hand, a counter example of will just prove that the quasi-local mass is not appropriate to describe the evolution of a non-stationary black hole. One can imagine other expressions for quasi-local mass, may be more involved, in axial symmetry. On the contrary, reversing the argument, a proof of will certainly suggest that the mass has physical meaning for non-stationary black holes as a natural quasi-local mass (at least in axial symmetry). Also, the inequality provide a non trivial control of the size of a black hole valid at any time.
In a seminal article Penrose [@Penrose73] proposed a remarkably physical argument that connects global properties of the gravitational collapse with geometric inequalities on the initial conditions. That argument lead to the well known Penrose inequality (\[eq:pen\]) for dynamical black holes (without any symmetry assumption). In the following we review this argument imposing axial symmetry, where angular momentum is conserved. And, more important, we include a relevant new ingredient: we assume that the inequality (\[eq:JA\]) holds.
We will assume that the following statements hold in a gravitational collapse:
- Gravitational collapse results in a black hole (weak cosmic censorship).
- The spacetime settles down to a stationary final state. We will further assume that at some finite time all the matter have fallen into the black hole and hence the exterior region is vacuum.
Conjectures (i) and (ii) constitute the standard picture of the gravitational collapse. Relevant examples where this picture is confirmed (and where the role of angular momentum is analyzed) are the collapse of neutron stars studied numerically in [@Baiotti:2004wn] [@Giacomazzo:2011cv].
The black hole uniqueness theorem implies that the final stationary state postulated in (ii) is given by the Kerr black hole. Let us denote by $m_0,
J_0, A_0$, respectively, the mass, angular momentum and horizon area of the remainder Kerr black hole. Penrose argument runs as follows. Take a Cauchy surface $S$ in the spacetime such that the collapse has already occurred. This is shown in figure \[fig:5\].
![The Penrose diagram of a gravitational collapse. The initial Cauchy surface is denoted by $S$. The area $A$ increase along the event horizon. The mass $m$ decrease along null infinity. We have assumed axial symmetry and hence the angular momentum remains constant along null infinity $J=J_0$.[]{data-label="fig:5"}](penrose-ineq-js-p.pdf){width="50.00000%"}
Let ${\Sigma}$ denotes the intersection of the event horizon with the Cauchy surface $S$ and let $A$ be its area. Let $(m, J)$ be the total mass and angular momentum at spacelike infinity. These quantities can be computed from the initial surface $S$. By the black hole area theorem we have that the area of the black hole increase with time and hence $$\label{eq:15}
A_0\geq A.$$ Since gravitational waves carry positive energy, the total mass of the spacetime should be bigger than the final mass of the remainder Kerr black hole $$\label{eq:4}
m\geq m_0.$$ The difference $m-m_0$ is the total amount of gravitational radiation emitted by the system.
To related the initial angular momentum $J$ with the final angular momentum $J_0$ is much more complicated. Angular momentum is in general non-conserved. There exists no simple relation between the total angular momentum $J$ of the initial conditions and the angular momentum $J_0$ of the final black hole. For example, a system can have $J=0$ initially, but collapse to a black hole with final angular momentum $J_0\neq 0$. We can imagine that on the initial conditions there are two parts with opposite angular momentum, one of them falls in to the black hole and the other escape to infinity. Axially symmetric vacuum spacetimes constitute a remarkable exception because the angular momentum is conserved. In that case we have $$\label{eq:59}
J=J_0.$$ For a discussion of this conservation law in detail see [@dain12] and reference therein.
We have assumed that the inequality holds, then by the discussion above we have that the quasi-local mass ${m_{bh}}$ increase with time, that is $$\label{eq:8b}
{m_{bh}}\leq m_0.$$ We emphasize that this inequality is highly non-trivial. The quantity ${m_{bh}}$ is computed on the initial surface $S$, in contrast to compute $m_0$ we need to known the whole spacetime. Using (\[eq:8b\]) and (\[eq:4\]) we finally obtain $$\label{eq:6}
\sqrt{\frac{A}{16\pi}+\frac{4\pi J^2}{A}}= {m_{bh}}\leq m.$$ This inequality has the natural interpretation that the mass of the black hole ${m_{bh}}$ should always be smaller than the total mass of the spacetime $m$. The inequality (\[eq:6\]) represents a generalization of the Penrose inequality with angular momentum. This inequality implies $$\label{eq:mjd}
\sqrt{|J|}\leq m.$$ In fact, the inequality can be deduced directly by the same heuristic argument without using the area theorem. It depends only on the following assumptions
- Gravitational waves carry positive energy.
- Angular momentum is conserved in axial symmetry.
- In a gravitational collapse the spacetime settles down to a final Kerr black hole.
Let us summarize the discussion of this section. For an axially symmetric, dynamical black hole, the following two geometrical inequalities are expected $$\begin{aligned}
8\pi |J| & \leq A \quad (=\text{Extreme Kerr horizon}), \label{eq:JAd}\\
\sqrt{\frac{A}{16\pi}+\frac{4\pi J^2}{A}} & \leq m \quad (=\text{Kerr black
hole}). \label{eq:pendj}
\end{aligned}$$ The inequality is quasi-local and the inequality is global. The global inequality implies the following two inequalities $$\begin{aligned}
\sqrt{\frac{A}{16\pi}} &\leq m \quad (=\text{Schwarzschild}),\label{eq:penddd}\\
\sqrt{|J|} &\leq m. \quad (=\text{extreme Kerr black hole}).\label{eq:mjdd}\end{aligned}$$ That is:
> *The three geometrical inequalities (\[eq:pen\]), (\[eq:mj\]) and (\[eq:JA\]) valid for the Kerr black holes are expected to hold also for axially symmetric, dynamical black holes.*
The Penrose inequality is valid also without the axial symmetry assumption. It is important to emphasize that all the quantities involved in the geometrical inequalities above can be calculated on the initial surface. For simplicity, we have avoided the distinction between event horizon and apparent horizons (defined in terms of trapped surfaces) to calculate the area $A$. This point is important for the Penrose inequality (see the discussion in [@Mars:2009cj]) but not for the other inequalities which are the main subject of this review. In particular the horizon area $A$ in (\[eq:JAd\]) is the area of an appropriated defined trapped surface.
A counter example of the global inequality (\[eq:pendj\]) will imply that cosmic censorship is not true. Conversely a proof of it gives indirect evidence of the validity of censorship, since it is very hard to understand why this highly nontrivial inequality should hold unless censorship can be thought of as providing the underlying physical reason behind it.
The inequalities (\[eq:pen\]), (\[eq:mj\]) and (\[eq:JA\]) can be divided into two groups:
1. $\sqrt{\frac{A}{16\pi}}\leq m$: the area appears as lower bound.
2. $\sqrt{|J|} \leq m$ and $ 8\pi |J| \leq A $: the angular momentum appears as lower bound and the area appears as upper bound.
The mathematical methods used to study these two groups are, up to now, very different. This review is mainly concerned with the second group.
Finally, we mention that for the Kerr black hole there exists a remarkable equality of the form $(8\pi J)^2 =A^+A^-$, where $A^+$ and $A^-$ denote the areas of event and Cauchy horizon (see figure \[fig:kerr-diag\]). This equality has been proved for general stationary spacetimes in the following series of articles [@Ansorg:2009yi] [@Hennig:2009aa] [@Ansorg:2008bv]. It has recently received considerable attention in the string community (see [@Cvetic:2010mn] and [@Visser:2012wu] and references therein). The key property used in these studies is that the product of horizon areas is independent of the mass of the black hole. It is interesting to note that there exists, up to now, no generalization of this kind of equality (or a related inequality) to the dynamical regime.
Theorems {#sec:theorems}
========
The Penrose inequality $$\label{eq:9}
\sqrt{\frac{A}{16\pi}} \leq m \quad (=\text{Schwarzschild}),$$ has been intensively studied. It is a very relevant geometric inequality for black holes since it is valid without any symmetry assumption. For a comprehensive review on this subject see [@Mars:2009cj]. The most important results concerning this inequality are the proofs of Huisken-Ilmanen [@Huisken01] and Bray [@Bray01] for the Riemannian case. The general case remains open. Also, there is up to now no result concerning the Penrose inequality with angular momentum (\[eq:pendj\]) discussed in the previous section.
In the following we present a sample of the main results concerning inequalities (\[eq:mjdd\]) and (\[eq:JAd\]) that have been recently proved.
For the global inequality (\[eq:mjdd\]) we have the following theorem.
\[t:1\] Consider an axially symmetric, vacuum, asymptotically flat and maximal initial data set with two asymptotics ends. Let $m$ and $J$ denote the total mass and angular momentum at one of the ends. Then, the following inequality holds $$\label{eq:60}
\sqrt{|J|} \leq m \quad (= \text{Extreme Kerr}).$$
For the precise definitions, fall off conditions an assumptions on the initial data we refer to original articles cited bellow.
The first proof of the global inequality (\[eq:60\]) was provided in a series of articles [@Dain05c], [@Dain05d], [@Dain05e] which end up in the global proof given in [@Dain06c]. The proof is based on a variational characterization of the extreme Kerr initial data. In [@Chrusciel:2007dd] and [@Chrusciel:2007ak] the result was generalized and the proof simplified. In [@Chrusciel:2009ki] [@Costa:2009hn] the charge was included. In [@Schoen:2012nh] relevant improvements on the rigidity statements were made. In particular in that article it was proved the first rigidity result including charge and a measure of the distance to extreme Kerr black hole was introduced. In [@zhou12] the result was proved with the maximal condition replaced by a small trace assumption for the second fundamental form of the initial data. Related results concerning the force between black holes were proved in [@Clement:2012np]. Finally, the mass formula and the variational techniques involved in the proof of the inequality (\[eq:60\]) were very recently used to study the linear stability of the extreme Kerr black hole [@Dain:2014iba].
Under the hypothesis of theorem \[t:1\] (namely, vacuum and axial symmetry) the angular momentum is defined as conserved quasi-local integral. In particular, if the topology of the manifold is trivial (i.e. ${\mathbb{R}^3}$), then the angular momentum is zero and hence theorem \[t:1\] reduces to the positive mass theorem. In order to have non-zero angular momentum we need to allow non-trivial topologies, for example manifolds with two asymptotic ends as it is the case in theorem \[t:1\]. An important initial data set that satisfies the hypothesis of the theorem is provided by an slice $t=constant$ in the Kerr black hole in the standard Boyer-Lindquist coordinates, see figures \[fig:kerr-diag\] and \[fig:kerr-extrem-diag\]. The non-extreme initial data have a different geometry as the extreme initial data. The former are asymptotically flat at both ends. In contrast, extreme initial data, which reach the equality in (\[eq:60\]), have one asymptotically flat end and one cylindrical end, see figure \[fig:non-extreme-id\]. That geometry represents the “optimal shape” with respect to the inequality (\[eq:60\]). Figure \[fig:non-extreme-id\] is the analog of figure \[fig:1\] for the geometrical inequality (\[eq:60\]).
![Conformal diagram of the non-extreme Kerr black hole. The points $i_0$ represent spacelike infinity. The surface $S$ have two identical asymptotically flat ends $i_0$.[]{data-label="fig:kerr-diag"}](kerr-color-p.pdf){width="6cm"}
![Conformal diagram of the extreme Kerr black hole. The point $i_0$ represents spacelike infinity, the point $i_c$ represent the cylindrical end. The surface $S$ has one asymptotically flat end $i_0$ and one cylindrical end $i_c$.[]{data-label="fig:kerr-extrem-diag"}](kerr-extremo-color-p.pdf){width="5cm"}
![On the left, an the initial data with two asymptotically flat ends, like the non-extreme Kerr black holes. For these data the strict inequality holds. On the right, the data of extreme Kerr black hole, with one asymptotically flat and one cylindrical end. For this data the equality holds.[]{data-label="fig:non-extreme-id"}](kerr-initial-data-iso-p.pdf "fig:"){width="30.00000%"} ![On the left, an the initial data with two asymptotically flat ends, like the non-extreme Kerr black holes. For these data the strict inequality holds. On the right, the data of extreme Kerr black hole, with one asymptotically flat and one cylindrical end. For this data the equality holds.[]{data-label="fig:non-extreme-id"}](kerr-ex-initial-data-iso-p.pdf "fig:"){width="35.00000%"}
Regarding the quasi-local inequality (\[eq:JAd\]) we have the following result.
\[t:2\] Given an axisymmetric closed marginally trapped and stable surface ${\Sigma}$, in a spacetime with non-negative cosmological constant and fulfilling the dominant energy condition, it holds the inequality $$\label{eq:JAt}
8\pi |J| \leq A \quad (= \text{Extreme Kerr throat}),$$ where $A$ and $J$ are the area and angular momentum of ${\Sigma}$.
This is a pure spacetime and local result. That is, there is no mention of a three-dimensional initial hypersurface where the two-dimension surface ${\Sigma}$ is embedded. Axisymmetry is only imposed on ${\Sigma}$. Moreover, this theorem does not assume vacuum. The matter fields can have also angular momentum and it can be transferred to the black hole, however the inequality (\[eq:JAt\]) remains true even for that case. It is important to note that the angular momentum that appears in (\[eq:JAt\]) is the gravitational one (i.e. the Komar integral). In fact this inequality is non-trivial even for the Kerr-Newman black hole, see the discussion in [@dain12].
Theorem \[t:2\] has the following history. The quasi-local inequality (\[eq:JAt\]) was first conjectured to hold in stationary spacetimes surrounded by matter in [@Ansorg:2007fh]. In that article the extreme limit of this inequality was analyzed and also numerical evidences for the validity in the stationary case was presented (using the numerical method and code developed in [@Ansorg05]). In a series of articles [@hennig08] [@Hennig:2008zy] the inequality (\[eq:JAt\]) (including also the electromagnetic charge) was proved for that class of stationary black holes. See also the review article [@Ansorg:2010ru].
In the dynamical regime, the inequality (\[eq:JAt\]) was conjectured to hold in [@dain10d] based on the heuristic argument mentioned in section \[sec:physical-picture\]. In that article also the main relevant techniques for its proof were introduced, namely the mass functional on the surface and its connections with the area. A proof (but with technical restrictions) was obtained in [@Acena:2010ws] [@Clement:2011kz]. The first general and pure quasi-local result was proven in [@Dain:2011pi], where the relevant role of the stability condition for minimal surfaces was pointed out. The generalization to trapped surfaces and non-vacuum has been proved in [@Jaramillo:2011pg]. The electromagnetic charge was included in [@Clement:2011np] and [@Clement:2012vb]. This inequality has been extended to higher dimensions in [@Hollands:2011sy] and [@Paetz:2013rka]. In [@Yazadjiev:2012bx] [@Yazadjiev:2013hk] and [@Fajman:2013ffa] it has been also extended to Einstein-Maxwell dilaton gravity. In [@Reiris:2013jaa] related inequalities that involve the shape of the black hole were proved.
![Axially symmetric two-surface. The axial Killing vector $\eta$ is tangent to the surface. The null vectors $\ell^a$ and $k^a$ are normal to ${\cal S}$[]{data-label="fig:axial-2s"}](axial-two-surface.pdf){width="20.00000%"}
To describe the concept of stable trapped surface (this condition was first introduced in [@andersson08]) used in theorem \[t:2\] let us consider an axially symmetric closed two-surface ${\Sigma}$ with the topology of a two-sphere. The surface ${\Sigma}$ is embedded in the spacetime. Let $\ell^a$ and $k^a$ be null vectors spanning the normal plane to ${\Sigma}$ and normalized as $\ell^a k_a =
-1$, see figure \[fig:axial-2s\]. The expansion is defined by $\theta^{(\ell)}= \nabla_a\ell^a$, where $\nabla$ is the spacetime connection. The surface ${\Sigma}$ is marginally trapped if $\theta^{(\ell)}=0$. Given a closed marginally trapped surface ${\Sigma}$ we will refer to it as spacetime stably outermost if there exists an outgoing ($-k^a$-oriented) vector $X^a= \gamma \ell^a - \psi k^a$, with $\gamma\geq0$ and $\psi>0$, such that the variation of $\theta^{(\ell)}$ with respect to $X^a$ fulfills the condition $$\label{e:stability_condition}
\delta_X \theta^{(\ell)} \geq 0.$$ Here $\delta$ denotes a variation operator associated with a deformation of the surface ${\Sigma}$ (c.f. for example [@Booth:2006bn] [@andersson08])). For maximal initial data the stability condition (\[e:stability\_condition\]) is closed related with the stability condition for minimal surfaces (see [@Dain:2011kb], [@Jaramillo:2011pg]). The stability of a minimal surface is the requirement that the area is a local minimum.
The extreme throat geometry, with angular momentum $J$, was defined in [@dain10d] (see also [@Acena:2010ws] and [@Dain:2011pi]). This concept captures the local geometry near the horizon of an extreme Kerr black hole. The extreme throat is the asymptotic limit in the cylindrical end of an extreme Kerr black hole, see figure \[fig:throat-cd\] and \[fig:throat-id\]. Both the intrinsic and extrinsic geometry of this surface are fixed. For example, it has an intrinsic metric given by $$\label{eq:gamma0}
|J| \left( (1+\cos^2\theta) d\theta^2+ \frac{4\sin^2\theta}{(1+\cos^2\theta)} d\phi^2 \right).$$ It is an oblate sphere with respect to the axis of rotation (see figure \[fig:arb-st\], on the right).
![Location of the extreme Kerr throat surface ${\Sigma}$ in the spacetime. []{data-label="fig:throat-cd"}](kerr-extremo-s-p.pdf){width="50.00000%"}
![Location of the extreme Kerr throat surface ${\Sigma}$ on the initial data. []{data-label="fig:throat-id"}](kerr-ex-initial-data-s-p.pdf){width="40.00000%"}
![On the left, an arbitrary axially symmetric stable two surface. For this kind of surface the strict inequality holds. On the right, the extreme throat sphere, where the equality holds.[]{data-label="fig:arb-st"}](axial-two-surface-iso-p.pdf "fig:"){width="30.00000%"} ![On the left, an arbitrary axially symmetric stable two surface. For this kind of surface the strict inequality holds. On the right, the extreme throat sphere, where the equality holds.[]{data-label="fig:arb-st"}](extreme-throat-sphere-p.pdf "fig:"){width="30.00000%"}
The extreme Kerr throat achieve the equality in (\[eq:JAt\]), this surface has the “optimal shape” with respect this inequality. It has also a variational characterization. Figure \[fig:arb-st\] is the analog of figures \[fig:1\] and \[fig:non-extreme-id\] for inequality (\[eq:JAt\]).
The results in theorem \[t:2\] has been used in a recent non-existence proof of stationary black holes binaries [@Neugebauer:2013ee] [@Neugebauer:2011qb] [@Chrusciel:2011iv].
The rigidity statement in theorem \[t:2\] (namely that the equality in implies that the surface is an extreme Kerr throat) has been proved in a different context: for extreme isolated horizon and near-horizon geometries of extremal black holes in [@Hajicek:1974oua], [@Lewandowski:2002ua] and [@Kunduri:2008rs], see also the review article [@lrr-2013-8] and reference therein.
Open problems and recent results on bodies {#sec:open-problems-recent}
==========================================
In this final section I would like to present the main open problems regarding the black holes geometrical inequalities discussed in the previous sections. My aim is to present open problems which are relevant (and probably involve the discovery of new techniques) and at the same time they appear feasible to solve. For more details see the review article [@dain12]. The open problem mentioned there regarding the inclusion of the electric charge in the quasi-local inequality (\[eq:JAt\]) have been solved [@Clement:2011np] [@Clement:2012vb].
For the global inequality (\[eq:60\]) there are two main open problems, which involve generalizations of the assumptions in theorem \[t:1\]:
- Remove the maximal condition.
- Generalization for asymptotic flat manifolds with multiple ends.
Concerning the maximal condition, as we mention above, in a recent article [@zhou12] this assumption have been replaced by a small trace condition. See also the discussion in [@dain12]. The most relevant open problem is the second one. The physical heuristic argument presented in section \[sec:physical-picture\] applies to that case and hence there little doubt that the inequality holds. This problem is related with the uniqueness of the Kerr black hole with degenerate and disconnected horizons. It is probably a hard problem. There are very interesting partial results in [@Chrusciel:2007ak] and also numerical evidences in [@Dain:2009qb].
Probably the most important open problem for geometrical inequalities for axially symmetric black holes is the following:
- Prove the Penrose inequality with angular momentum (\[eq:6\]).
We mention in section \[sec:physical-picture\] that there is a clear physical connection between the global inequality (\[eq:60\]) and the Penrose inequality with angular momentum in axial symmetry (\[eq:6\]). However, the techniques used to prove the inequality (\[eq:60\]) are very different than the one used to prove the classical Penrose inequality (\[eq:9\]) (see the discussion in [@dain12]).
For the quasi-local inequality (\[eq:JAt\]) the two main problems are the following:
- A generalization of the inequality (\[eq:JAt\]) without axial symmetry.
- A generalization of the inequality (\[eq:JAt\]) for ordinary bodies.
The problem of finding versions of inequality (\[eq:JAt\]) without any symmetry assumption, in contrast with the other open problems presented above, is not a well-defined mathematical problem since there is no unique notion of quasi-local angular momentum in the general case. However, exploring the scope of the inequality in regions close to axial symmetry (in some appropriate sense) can perhaps provide such a notion. From the physical point of view, we do not see any reason why this inequality should only hold in axial symmetry. Note that the global inequality (\[eq:60\]) only holds in axial symmetry. This is clear from the physical point of view (see the discussion in [@dain12]) and in [@huang11] highly non-trivial counter examples have been constructed.
Finally, concerning the second problem there have been recently some results in [@Dain:2013gma]. Consider a rotating body ${U}$ with angular momentum $J({U})$, see figure \[fig:body\]. Let ${\mathcal{R}}({U})$ be a measure (with units of length) of the size of the body.
![Axially symmetric rotating body.[]{data-label="fig:body"}](body3.pdf){width="2.6cm"}
In [@Dain:2013gma], the following universal inequality for all bodies is conjectured $$\label{eq:22}
{\mathcal{R}}^2({U}) \apprge \frac{G}{c^3} |J({U})|,$$ where $G$ is the gravitational constant and $c$ the speed of light. The symbol $\apprge$ is intended as an order of magnitude, the precise universal (i.e. independent of the body) constant will depend on the definition of ${\mathcal{R}}$. We have reintroduced in (\[eq:22\]) the fundamental constants in order to make more transparent the discussion bellow.
The arguments in support of the inequality (\[eq:22\]) are based in the following three physical principles:
- The speed of light $c$ is the maximum speed.
- For bodies which are not contained in a black hole the following inequality holds $$\label{eq:2b}
{\mathcal{R}}({U}) \apprge\frac{G}{c^2} m({U}),$$ where $m({U})$ is the mass of the body.
- The inequality (\[eq:22\]) holds for black holes.
Let us discuss these assumptions. Item (i) is clear. Item (ii) is called the *trapped surface conjecture* [@Seifert79]. Essentially, it says that if the reverse inequality as in (\[eq:2b\]) holds then a trapped surface should enclose ${U}$. That is: if matter is enclosed in a sufficiently small region, then the system should collapse to a black hole. This is related with the *hoop conjecture* [@thorne72] (see also [@Wald99] [@PhysRevD.44.2409] [@Malec:1992ap] ). The trapped surface conjecture has been proved in spherical symmetry [@Bizon:1989xm] [@Bizon:1988vv] [@Khuri:2009dt] and also for a relevant class of non-spherical initial data [@Malec:1991nf]. The general case remains open but it is expected that some version of this conjecture should hold.
Concerning item (iii), the area $A$ is a measure of the size of a trapped surface, hence the inequality (\[eq:JAt\]) represents a version of for axially symmetric black holes. If we include the physical constants, this inequality has the form $$\label{eq:5}
A\geq8\pi\frac{G}{c^3} |J|.$$ In fact the inequality (\[eq:5\]) was the inspiration for the inequality (\[eq:22\]). A possible generalization of (\[eq:5\]) for bodies is to take the area $A(\partial {U})$ of the boundary $\partial {U}$ of the body ${U}$ as measure of size. But unfortunately the area of the boundary is not a good measure of the size of a body in the presence of curvature. In particular, an inequality of the form $A(\partial {U}) \apprge G c^{-3} |J({U})| $ does not holds for bodies. The counter example is essentially given by a rotating torus in the weak field limit, with large major radius and small minor radius. The details of this calculation will be presented in [@Anglada13].
Using the three physical principles (i), (ii) and (iii) in [@Dain:2013gma] it is argued that the inequality (\[eq:22\]) should hold. One of the main difficulties in the study of inequalities of the form is the very definition of the measure of size. In fact, despite the intensive research on the subject, there is no know universal measure of size such that the trapped surface conjecture (or, more general, the hoop conjecture) holds (see the interesting discussions in [@Malec:1992ap] [@Gibbons:2012ac] [@Senovilla:2007dw] [@Reiris:2013jaa]). However, the remarkable point is that in order to find an appropriate measure of size ${\mathcal{R}}$ such that holds it is not necessary to prove first , and hence we do not need to find the relevant measure of mass $m({U})$ for the trapped surface conjecture. In [@Dain:2013gma] a size measure is proposed and for that measure the following version of the inequality has been proved for constant density bodies. This theorem is a consequence of the Schoen-Yau theorem [@schoen83d].
\[t:3\] Consider a maximal, axially symmetric, initial data set that satisfy the dominant energy condition. Let ${U}$ be an open set on the data. Assume that the energy density is constant on ${U}$. Then the following inequality holds $$\label{eq:7d}
{\mathcal{R}}^2({U}) \geq \frac{24}{\pi^3}\frac{G}{c^3} |J({U})|.$$
The definition of the radius ${\mathcal{R}}$ in (\[eq:7d\]) is as follow. Let ${\mathcal{R}_{SY}}({U})$ be the Schoen-Yau radius defined in [@schoen83d]. This radius is expressed in terms of the largest torus that can be embedded in ${U}$. See figure \[fig:sy-torus\].
Consider a region ${U}$ with a Killing vector $\eta^i$ with norm $\lambda$, we define the radius ${\mathcal{R}}$ by $$\label{eq:8}
{\mathcal{R}}({U}) = \frac{2}{\pi} \frac{\left(\int_{U}\lambda \right)^{1/2}}{{\mathcal{R}_{SY}}({U})}.$$ The definition of the radius (\[eq:8\]) is, no doubt, very involved. It is not expected to be the optimal size measure for a body. It should be considered, together with theorem \[t:3\], as an example where the conjecture (\[eq:22\]) can be proved with the current available mathematical techniques. For examples and further discussion on this radius we refer to [@Dain:2013gma].
![On the left, the Schoen-Yau ${\mathcal{R}_{SY}}$ radius for a body is defined in terms of the biggest embedded torus. On the right, the same torus is showed on the plane orthogonal to the axial Killing vector. On that plane the torus is a circle, and the radius ${\mathcal{R}_{SY}}$ is related to the radius of the biggest embedded circle.[]{data-label="fig:sy-torus"}](toro-interno-p.pdf "fig:"){width="3.2cm"} ![On the left, the Schoen-Yau ${\mathcal{R}_{SY}}$ radius for a body is defined in terms of the biggest embedded torus. On the right, the same torus is showed on the plane orthogonal to the axial Killing vector. On that plane the torus is a circle, and the radius ${\mathcal{R}_{SY}}$ is related to the radius of the biggest embedded circle.[]{data-label="fig:sy-torus"}](sy-radius-as-2-p.pdf "fig:"){width="1.8cm"}
This article is based on the longer review article [@dain12], we refer to that article for more details. The two main differences with respect to [@dain12] are the following. First, several new results appeared after the publication of [@dain12]. These results have been included here. Second, the physical arguments in section \[sec:physical-picture\] have been significantly improved and clarified, based on the discussion in [@Dain:2013qia].
Acknowledgements {#acknowledgements .unnumbered}
================
This work was supported by grant PICT-2010-1387 of CONICET (Argentina) and grant Secyt-UNC (Argentina).
[10]{}
A. Aceña, S. Dain, and M. E. Gabach Clément. Horizon area – angular momentum inequality for a class of axially symmetric black holes. , 28(10):105014, 2011, 1012.2413.
L. Andersson, M. Mars, and W. Simon. Stability of marginally outer trapped surfaces and existence of marginally outer trapped tubes. , 12(4):853–888, 2008.
P. Anglada, S. Dain, and O. Ortiz. In preparation.
M. Ansorg and J. Hennig. . , 25:222001, 2008, 0810.3998.
M. Ansorg and J. Hennig. . , 102:221102, 2009, 0903.5405.
M. Ansorg, J. Hennig, and C. Cederbaum. . , 43:1205–1210, 2011, 1005.3128.
M. Ansorg and D. Petroff. Black holes surrounded by uniformly rotating rings. , 72:024019, 2005, gr-qc/0505060.
M. Ansorg and H. Pfister. . , 25:035009, 2008, 0708.4196.
R. Arnowitt, S. Deser, and C. W. Misner. The dynamics of general relativity. In L. Witten, editor, [*Gravitation: An Introduction to Current Research*]{}, pages 227–265. Wiley, New York, 1962, gr-qc/0405109.
A. Ashtekar and B. Krishnan. Dynamical horizons: Energy, angular momentum, fluxes and balance laws. , 89:261101, 2002, gr-qc/0207080.
A. Ashtekar and B. Krishnan. Dynamical horizons and their properties. , 68:104030, 2003, gr-qc/0308033.
L. Baiotti et al. . , D71:024035, 2005, gr-qc/0403029.
R. Bartnik. The mass of an asymptotically flat manifold. , 39(5):661–693, 1986.
P. Bizon, E. Malec, and N. O’Murchadha. . , 61:1147–1450, 1988.
P. Bizon, E. Malec, and N. O’Murchadha. . , 6:961–976, 1989.
I. Booth and S. Fairhurst. . , D75:084019, 2007, gr-qc/0610032.
I. Booth and S. Fairhurst. . , D77:084005, 2008, 0708.2209.
H. L. Bray. Proof of the riemannian penrose conjecture using the positive mass theorem. , 59:177–267, 2001, math.DG/9911173.
D. Christodoulou. Reversible and irreversible transforations in black-hole physics. , 25:1596–1597, 1970.
P. Chru[ś]{}ciel. Boundary conditions at spatial infinity from a [H]{}amiltonian point of view. In [*Topological properties and global structure of space-time ([E]{}rice, 1985)*]{}, volume 138 of [*NATO Adv. Sci. Inst. Ser. B Phys.*]{}, pages 49–59. Plenum, New York, 1986.
P. T. Chrusciel. . , 323:2566–2590, 2008, 0710.3680.
P. T. Chrusciel, M. Eckstein, L. Nguyen, and S. J. Szybka. . , 28:245017, 2011, 1111.1448.
P. T. Chru[ś]{}ciel, Y. Li, and G. Weinstein. Mass and angular-momentum inequalities for axi-symmetric initial data sets. [II]{}. [A]{}ngular-momentum. , 323(10):2591–2613, 2008, arXiv:0712.4064.
P. T. Chrusciel and J. Lopes Costa. . , 26:235013, 2009, 0909.5625.
P. T. Chruściel, J. L. Costa, and M. Heusler. Stationary black holes: Uniqueness and beyond. , 15(7), 2012.
J. L. Costa. Proof of a [D]{}ain inequality with charge. , 43(28):285202, 2010, 0912.0838.
M. Cvetic, G. Gibbons, and C. Pope. . , 106:121301, 2011, 1011.0008.
S. Dain. Angular momemtum-mass inequality for axisymmetric black holes. , 96:101101, 2006, gr-qc/0511101.
S. Dain. Proof of the (local) angular momemtum-mass inequality for axisymmetric black holes. , 23:6845–6855, 2006, gr-qc/0511087.
S. Dain. A variational principle for stationary, axisymmetric solutions of einstein’s equations. , 23:6857–6871, 2006, gr-qc/0508061.
S. Dain. The inequality between mass and angular momentum for axially symmetric black holes. , 17(3-4):519–523, 2008, arXiv:0707.3118 \[gr-qc\].
S. Dain. Proof of the angular momentum-mass inequality for axisymmetric black holes. , 79(1):33–67, 2008, gr-qc/0606105.
S. Dain. Extreme throat initial data set and horizon area-angular momentum inequality for axisymmetric black holes. , 82(10):104010, Nov 2010, 1008.0019.
S. Dain. Geometric inequalities for axially symmetric black holes. , 29(7):073001, 2012, 1111.3615.
S. Dain. Inequality between size and angular momentum for bodies. , 112:041101, Jan 2014, 1305.6645.
S. Dain and I. Gentile de Austria. , 2014, 1402.2848.
S. Dain, J. L. Jaramillo, and M. Reiris. . , 29(3):035013, 2012, 1109.5602.
S. Dain, M. Khuri, G. Weinstein, and S. Yamada. . , D88:024048, 2013, 1306.4739.
S. Dain and O. E. Ortiz. . , D80:024045, 2009, 0905.0708.
S. Dain and M. Reiris. Area—angular-momentum inequality for axisymmetric black holes. , 107(5):051101, Jul 2011, 1102.5215.
D. Fajman and W. Simon. , 2013, 1308.3659.
E. Flanagan. Hoop conjecture for black-hole horizon formation. , 44:2409–2420, Oct 1991.
M. E. Gabach Clément. , 2011, 1102.3834.
M. E. Gabach Clément. . , 29:165008, 2012, 1201.4099.
M. E. Gabach Clément and J. L. Jaramillo. . , D86:064021, 2012, 1111.6248.
M. E. Gabach Clément, J. L. Jaramillo, and M. Reiris. . , 30:065017, 2013, 1207.6761.
M. E. Gabach Clément and M. Reiris. . , D88:044031, 2013, 1306.1019.
B. Giacomazzo, L. Rezzolla, and N. Stergioulas. . , D84:024022, 2011, 1105.0122.
G. Gibbons. , 1460:90–100, 2012, 1201.2340.
J. Hennig and M. Ansorg. . , 10:1075–1095, 2009, 0904.2071.
J. Hennig, M. Ansorg, and C. Cederbaum. A universal inequality between the angular momentum and horizon area for axisymmetric and stationary black holes with surrounding matter. , 25(16):162002, 2008.
J. Hennig, C. Cederbaum, and M. Ansorg. . , 293:449–467, 2010, 0812.2811.
S. Hollands. . , 29:065006, 2012, 1110.5814.
L.-H. Huang, R. Schoen, and M.-T. Wang. Specifying angular momentum and center of mass for vacuum initial data sets. , 306:785–803, 2011, 1008.4996. 10.1007/s00220-011-1295-9.
G. Huisken and T. Ilmanen. The inverse mean curvature flow and the [R]{}iemannian [P]{}enrose inequality. , 59:352–437, 2001.
P. Hájiček. . , 36(4):305–320, 1974.
J. L. Jaramillo, M. Reiris, and S. Dain. . , D84:121503, 2011, 1106.3743.
M. A. Khuri. . , D80:124025, 2009, 0912.3533.
M. A. Khuri, S. Yamada, and G. Weinstein. , 2013, 1306.0206.
H. K. Kunduri and J. Lucietti. . , 50:082502, 2009, 0806.2051.
H. K. Kunduri and J. Lucietti. Classification of near-horizon geometries of extremal black holes. , 16(8), 2013.
J. Lewandowski and T. Pawlowski. . , 20:587–606, 2003, gr-qc/0208032.
E. Malec. . , 67:949–952, 1991.
E. Malec. . , B22:829, 1992.
M. Mars. . , 26:193001, 2009, 0906.5566.
G. Neugebauer and J. Hennig. . , 62:613–630, 2012, 1105.5830.
G. Neugebauer and J. Hennig. , 2013, 1302.0573.
R. Osserman. The isoperimetric inequality. , 84(6):1182–1238, 1978.
T.-T. Paetz and W. Simon. , 2013, 1302.3052.
R. Penrose. Naked singularities. , 224:125–134, 1973.
R. Schoen and S. T. Yau. On the proof of the positive mass conjecture in general relativity. , 65(1):45–76, 1979.
R. Schoen and S. T. Yau. The energy and the linear momentum of space-times in general relativity. , 79(1):47–51, 1981.
R. Schoen and S. T. Yau. The existence of a black hole due to condensation of matter. , 90(4):575–579, 1983.
R. Schoen and X. Zhou. Convexity of reduced energy and mass angular momentum inequalities. , 14(7):1747–1773, 2013.
H. Seifert. Naked singularities and cosmic censorship: Comment on the current situation. , 10(12):1065–1067, 1979.
J. M. Senovilla. . , 81:20004, 2008, 0709.0695.
L. B. Szabados. Quasi-local energy-momentum and angular momentum in [GR]{}: A review article. , 7(4), 2004. cited on 8 August 2005.
K. Thorne. Nonspherical gravitational collapse: A short review. In J. Klauder, editor, [*Magic Without Magic: John Archibald Wheeler. A Collection of Essays in Honor of his Sixtieth Birthday*]{}, pages 231–258. W.H. Freeman, San Francisco, 1972.
M. Visser. . , D88:044014, 2013, 1205.6814.
R. Wald. Final states of gravitational collapse. , 26(26):1653–1655, 1971.
R. Wald. Gravitational collapse and cosmic censorship. In B. R. Iyer and B. Bhawal, editors, [*Black Holes, Gravitational Radiation and the Universe*]{}, volume 100 of [*Fundamental Theories of Physics*]{}, pages 69–85. Kluwer Academic, Dorddrecht, 1999, gr-qc/9710068.
E. Witten. A new proof of the positive energy theorem. , 80:381–402, 1981. 10.1007/BF01208277.
S. Yazadjiev. . , 30:115010, 2013, 1301.1548.
S. S. Yazadjiev. . , D87:024016, 2013, 1210.4684.
X. [Zhou]{}. . , Sept. 2012, 1209.1605.
[^1]: It is worth mention that important aspects of the black hole uniqueness problem remain still open, see recent review article [@lrr-2012-7] and reference therein.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The renormalised quark Dyson-Schwinger equation is studied in the limit of the renormalised current heavy quark mass $m_R\rightarrow \infty$. We are particularly interested in the analytic pole structure of the heavy quark propagator in the complex momentum plane. Approximations in which the quark-gluon vertex is modelled by either the bare vertex or the Ball-Chiu Ansatz, and the Landau gauge gluon propagator takes either a gaussian form or a gaussian form with an ultraviolet asymptotic tail are used.'
address: ' Department of Theoretical Physics, Research School of Physical Sciences and Engineering, Australian National University, Canberra, ACT 0200, Australia\'
author:
- 'C. J. Burden'
title: The analytic structure of heavy quark propagators
---
Introduction
============
The solution of approximate Dyson-Schwinger equations (DSE) has proved to be an effective means for modelling quark propagators in hadronic physics [@RW94]. Recent calculations within the genre of models which we shall refer to as the DSE technique include those of the light hadron spectrum [@BSpapers; @JM93] and of electromagnetic form factors of the pion and kaon [@R96]. Although no rigorous proof exists, it is the philosophy of the DSE technique that one possible signal of confinement in QCD should be the absence of timelike poles in the quark propagator [@RWK92]. It has furthermore been conjectured that the propagator $S(p)$ could be an entire function in the complex $p^2$ plane [@M86; @BRW92]. Such a scenario would, for instance, avoid certain unpleasant consequences which can result when modelling mesons via the Bethe-Salpeter equation which samples the quark propagator over a region of the complex plane.
Determining the analytic structure of fermion propagators in QCD [@MH92; @SC92] or other confining theories [@M94; @AB96] by the direct solution of model DSEs is not easy. It appears that the pole or branch cut structure obtained in any particular model is heavily dependent on the approximations employed. In general, two aspects of the quark DSE must be approximated: the quark-gluon vertex, and the gluon propagator. In this paper, we shall look at both these aspects within the heavy quark sector. In existing numerical studies in the light quark sector which produce propagators with conjugate singularities [@MH92; @SC92], the quark gluon vertex has usually been approximated by the bare vertex (the so called rainbow approximation). In ref. [@BRW92], however, it was shown that an entire function propagator can be obtained if the vertex function is modelled by a more sophisticated form respecting the Ward-Takahashi identity. This suggests that it is worthwhile exploring the importance of accurately modelling the quark-gluon vertex, as well as the gluon propagator, when studying the analytic structure of the quark propagator.
In a recent development, the DSE technique has been extended to the realm of heavy quarks [@BL97; @DL97] in a way inspired by heavy quark effective theory (HQET) [@N94]. The purpose of this exercise was twofold. Firstly, if one acknowledges the success of the DSE technique in the light quark sector, it is clear that the dynamics of confined particles is driven by non-perturbative dynamical self dressing. In HQET, non-perturbative self dressing and the detailed analytic structure of the heavy quark propagator are largely ignored. It is important to know whether this is justified, or whether the successes of HQET are purely fortuitous. Secondly, one has the hope that an accurate determination of the heavy quark propagator will eventually prove useful for building phenomenological models of heavy quark hadrons.
In ref. [@BL97] a preliminary attempt is made to calculate the spectrum of heavy quark–light antiquark mesons by using the combination of rainbow DSE and ladder Bethe-Salpeter equation (BSE). It is found that, within the limitations of the model, the pole structure of the heavy quark propagator prevents solution of the meson BSE. This is clearly a shortcoming of the approximations involved. In ref. [@DL97] the heavy quark DSE is examined from the point of view of the gauge technique. This is essentially an improvement on the rainbow approximation to the quark-gluon vertex which is designed to respect the Ward-Takahashi identity. An alternative approach, and one which we follow in this paper, is to replace the bare vertex Ansatz with the Ball-Chiu vertex Ansatz [@BC80]. We shall see that in the heavy fermion limit the gauge technique and the Ball-Chiu vertex are equivalent.
Regarding the gluon propagator, our treatment differs from ref. [@DL97] in that it is principally numerical, enabling us to concentrate on a more realistic class of model gluon propagators. Specifically we study the simple gaussian Ansatz employed in ref. [@BL97], designed to model the infrared enhanced behaviour of the gluon propagator, and a model proposed by Frank and Roberts (FR) [@FR96] which includes both an enhanced infrared behaviour and the known asymptotic ultraviolet behaviour. In order to deal with the FR propagator, it has been necessary to formulate a properly renormalised version of the heavy quark DSE of ref. [@BL97].
In summary, our main finding is as follows: Improving the vertex Ansatz does little to improve the analytic structure of the quark propagator. However, improving the Ansatz employed for the gluon propagator, particularly by including a realistic asymptotic ultraviolet tail, moves poles in the heavy quark propagator to a less intrusive part of the complex momentum plane. This portends well for future application of the heavy quark DSE technique.
The layout of the paper is as follows. In Section II we summarise the renormalised quark DSE and the approximations we shall be employing for the quark-gluon vertex and gluon propagator. In Section III we summarise the heavy quark formalism and derive a renormalised DSE for the heavy quark propagator. Numerical solutions to these equations are discussed in Section IV. Conclusions are drawn and suggestions for the direction of future work are given in Section V.
The quark Dyson-Schwinger equation
==================================
Our starting point is the renormalised quark DSE [@RW94] $$\Sigma'(p,\Lambda) = Z_1(\mu^2,\Lambda^2)\frac{4g^2}{3}
\int^\Lambda \frac{d^4q}{(2\pi)^4}\,
D_{\mu\nu}(p - q) \gamma_\mu S(q) \Gamma_\nu(q,p), \label{DSEqn}$$ where we have used a Euclidean metric in which timelike vectors satisfy $p^2 = -p_{\rm Minkowski}^2 < 0$, and for which $\{\gamma_\mu,\gamma_\nu\} = 2\delta_{\mu \nu}$. Our aim is to solve the DSE for the renormalised quark propagator $S(p,\mu)$, which we write in the form $$\begin{aligned}
S(p,\mu) & = & \frac{1}{i\gamma\cdot p A(p^2,\mu^2) + B(p^2,\mu^2)}
\nonumber \\
& = & \frac{1}{Z_2(\mu^2,\Lambda^2)[i\gamma\cdot p + m_0(\Lambda)]
+ \Sigma^{'}(p,\Lambda)} . \label{genprop}\end{aligned}$$ The unrenormalised self energy is written $$\Sigma'(p,\Lambda) = i\gamma\cdot p \left[A'(p^2,\Lambda^2) - 1\right]
+ B'(p^2,\Lambda^2). \label{Sigpr}$$
If the renormalisation scale is set such that $$\left.S(p)\right|_{p^2 = \mu^2}=\frac{1}{i\gamma\cdot p + m_R(\mu^2)},
\label{mufix}$$ it follows from Eqs. (\[genprop\]) and (\[Sigpr\]) that renormalised and bare quantities are related by $$Z_2(\mu^2,\Lambda^2) = 2 - A'(\mu^2,\Lambda^2), \label{Zfix}$$ $$A(p^2,\mu^2) = 1 + A'(p^2,\Lambda^2) - A'(\mu^2,\Lambda^2), \label{AApr}$$ $$B(p^2,\mu^2) = m_R(\mu^2) + B'(p^2,\Lambda^2) - B'(\mu^2,\Lambda^2).
\label{BBpr}$$
The set of equations (\[DSEqn\]) and (\[Zfix\]) to (\[BBpr\]) together with the ‘abelian approximation’ $Z_1 = Z_2$ can be solved numerically for the propagator functions $A$ and $B$ once the renormalised quark-gluon vertex function $\Gamma_\mu$, the renormalised gluon propagator $D_{\mu \nu}$, the renormalisation point $\left(\mu,m_R(\mu)\right)$ and cutoff $\Lambda$ are specified. The precise forms of the quark-gluon vertex and gluon propagator are unknown and must be modelled by appropriate Ansätze. We next summarise the Ansätze employed in this paper.
Quark-gluon vertex
------------------
The most general form of the quark-gluon vertex consistent with Lorentz and CPT invariance, satisfying the Ward identity $i\Gamma_\mu(p,p) = \partial^p_\mu S^{-1}(p)$ and Ward Takahashi identity[^1] $i(p - q)_\mu \Gamma_\mu(p,q) = S^{-1}(p) - S^{-1}(q)$, and free of kinematic singularities has been given by Ball and Chiu [@BC80]. It takes the form $$\Gamma_\mu(p,q) = \Gamma_\mu^{\rm BC}(p,q) + \Gamma_\mu^{\rm T}(p,q) ,
\label{genver}$$ where $$\begin{aligned}
\Gamma_\mu^{\rm BC}(p,q) & = & \frac{1}{2} \left[A(p^2) + A(q^2)\right]
\gamma_\mu \nonumber \\
& + & \frac{(p + q)_\mu}{(p^2 - q^2)}\left\{ \left[A(p^2) - A(q^2)\right]
\frac{\gamma\cdot p + \gamma\cdot q}{2} -
i \left[B(p^2) - B(q^2)\right] \right\}, \label{BCver}\end{aligned}$$ and $\Gamma_\mu^{\rm T}(p,q)$ is an otherwise unconstrained piece satisfying the conditions $(p - q)_\mu \Gamma_\mu^{\rm T}(p,q) = 0$ and $\Gamma_\mu^{\rm T}(p,p) = 0$.
We mention two well studied vertex Ansätze falling within this class. The first of these, introduced by Curtis and Pennington [@CP90] to ensure multiplicative renormalisability in quantum electrodynamics, is defined by setting the transverse piece $\Gamma_\mu^{\rm T}$ equal to $$\Gamma_\mu^{{\rm T}_{\rm CP}}(p,q) = \frac{A(p^2) - A(q^2)}{2d(p,q)}
\left[\gamma_\mu(p^2 - q^2) - (p + q)_\mu
(\gamma\cdot p - \gamma\cdot q ) \right], \label{CPTr}$$ with $$d(p,q) = \frac{(p^2 - q^2)^2 + \left[ M^2(p^2) + M^2(q^2)^2 \right]^2}
{p^2 + q^2} ,$$ where $M = B/A$.
The second of these, proposed by Haeri [@H91], takes the form $$\begin{aligned}
\Gamma_\mu^{\rm H}(p,q) & = & \frac{p^2 A(p^2) - q^2 A(q^2)}{p^2 - q^2}
\gamma_\mu + \frac{A(p^2) - A(q^2)}{p^2 - q^2}
\gamma\cdot p \gamma_\mu \gamma\cdot q \nonumber \\
& & -i \frac{B(p^2) - B(q^2)}{p^2 - q^2}
\left(\gamma\cdot p \gamma_\mu + \gamma_\mu \gamma\cdot q\right).
\label{Hver}\end{aligned}$$ It satisfies the above criteria and therefore must be of the form Eq. (\[genver\]). Munczek [@Mpr] has shown that the Haeri vertex is identical to the spectral representation of the vertex used in the gauge technique, and this has in turn been employed in ref. [@DL97].
Gluon propagator
----------------
In a general covariant gauge, the gluon propagator takes the form $$g^2 D_{\mu \nu}(k) =
\left(\delta_{\mu \nu} - \frac{k_\mu k_\nu}{k^2}\right) \Delta(k^2)
+ g^2 \xi \frac{k_\mu k_\nu}{k^4}, \label{gprop}$$ where $\xi$ is the gauge fixing parameter. Perhaps the simplest Ansatz which has proved useful for modelling QCD is the ‘infrared dominant’ model [@M86; @BRW92] $$\Delta_{\rm IR}(k^2) = \frac{3}{16}(2\pi)^4 \mu^2 \delta^4(k). \label{delpr}$$ In applications to hadronic physics, $\mu$ is usually taken to be of the order of 1 GeV [@MN83], which is the typical scale of QCD. An obvious disadvantage of this model is that it neglects completely any intermediate or ultraviolet behaviour. One can go some way towards incorporating some intermediate energy structure by using the computationally convenient gaussian model $$\Delta_{\rm G}(k^2) = \frac{3}{16}(2\pi)^4 \frac{\mu^2}{\alpha^2 \pi^2}
e^{-k^2/\alpha}. \label{gauss}$$ We note that the infrared dominant gluon model $\Delta_{\rm IR}$, together with the minimal Ball-Chiu vertex $\Gamma_{\mu \nu}(p,q) = \Gamma_{\mu \nu}^{\rm BC}(p,q)$ defines precisely the light quark model considered in ref. [@BRW92], whereas the gaussian model $\Delta_{\rm G}$ has been used in our earlier rainbow approximation studies of heavy quarks [@BL97]. Below we shall explore the effect on the analytic structure of the heavy quark propagator of combining either $\Delta_{\rm IR}$ or $\Delta_{\rm G}$ with the Ball-Chiu vertex.
A more realistic model gluon propagator which goes some way toward modelling the asymptotically free ultraviolet behaviour of QCD (neglecting logarithmic corrections) has been proposed by Frank and Roberts [@FR96]. It takes the form $$\Delta_{\rm FR}(k^2) = 4\pi^2 d \left[4\pi^2m_t^2\delta^4(k) +
\frac{1 - e^{k^2/(4m_t^2)}}{k^2} \right], \label{FRprop}$$ where $d = 12/(33 - 2N_f)$, $N_f = 3$ is the number of light quark flavours, and $m_t \approx 0.69$ GeV is a parameter fitted to a range of calculated pion observables. In our numerical calculations, we find it more convenient to consider a gaussian smeared version of the FR propagator given by $$\Delta_{\rm GFR}(k^2) = (2\pi)^4 \frac{m_t^2 d}{\alpha^2 \pi^2}
e^{-k^2/\alpha} + 4\pi^2 d \frac{1 - e^{k^2/(4m_t^2)}}{k^2}.
\label{GFRprop}$$ This will enable comparisons to be made with the pure gaussian model $\Delta_G$.
In the next section we consider the heavy quark limit of the quark Dyson-Schwinger equation. Our earlier analysis of this limit [@BL97] employed the heavily damped gluon propagator $\Delta_{\rm G}$, and so was free from ultraviolet divergences. In order to deal with a more realistic model, such as Eq. (\[GFRprop\]), it is necessary to develop a properly renormalised version of the formalism.
The heavy quark limit
=====================
Heavy quark propagator
----------------------
In the absolute limit of heavy renormalised quark masses, $m_R(\mu^2) \rightarrow \infty$, the dressed quark propagator Eq. (\[genprop\]) is dominated by the bare form $S^{-1}_{\rm bare} = i\gamma\cdot p + m_R$. However it is important to isolate from the full inverse propagator finite order self energy corrections to the bare inverse propagator which drive confining and remnant chiral symmetry breaking effects. To this end we write the momentum variables occurring in the DSE Eq. (\[DSEqn\]) as $$p_\mu = im_R(\mu^2) v_\mu + k_\mu, \hspace{5 mm}
q_\mu = im_R(\mu^2) v_\mu + k'_\mu, \label{kmudef}$$ where, for convenience, we take $v = ({\bf 0},1)$, so $k\cdot v = k_4$. We then write the renormalised quark propagator functions as $$A(p^2,\mu^2) = 1 + \frac{\Sigma_A(K,\kappa)}{m_R(\mu^2)}, \label{sigAdef}$$ $$B(p^2,\mu^2) = m_R(\mu^2) + \Sigma_B(K,\kappa), \label{sigBdef}$$ where we have defined the independent momentum variable $$K = \frac{p^2 + m_R^2}{2im_R} = k_4 + \frac{k^2}{2im_R}, \label{Kdef}$$ and renormalisation point $$\kappa = \frac{\mu^2 + m_R^2}{2im_R},$$ in the complex $K$ plane. The change of dependent variable $p^2 \rightarrow K$ induced by the transformation Eq. (\[Kdef\]) is illustrated in Fig. \[fig1\]. In general, when working to zeroth order in $1/m_R$, one can use the approximation $K = k_4$ (an approximation which was used in ref. [@BL97]). An exception to this rule, relevant to models such as Eq. (\[GFRprop\]) for which the gluon propagator is not heavily ultraviolet damped, is in the denominator of the heavy quark propagator in the integrand in the quark DSE. This point will become clearer at the end of Section IIIB.
From Eqs. (\[genprop\]), (\[sigAdef\]) and (\[sigBdef\]) we have $$S(p,\mu) = \frac{1 + \gamma_4}{2} \frac{1}{iK + \Sigma(K,\kappa)}
+ O\left(\frac{1}{m_R}\right) , \label{hprop}$$ where we have defined the heavy quark self energy $$\Sigma(K,\kappa) = \Sigma_B(K,\kappa) - \Sigma_A(K,\kappa).\label{Sigdef}$$ We find in general that the DSE leads to a single integral equation for the complex valued function $\Sigma(K,\kappa)$. The form given by Eq.(\[hprop\]) represents the heavy quark propagator in the dominant region near the the bare propagator mass pole $p^2 = -m_R^2$. Obtaining an integral equation for $\Sigma(K)$ involves the change of integration $\int d^4q \rightarrow \int d^4k' =
\int_{-\infty}^\infty dk'_4 \int_0^\infty d\left|{\bf k}\right|
\left|{\bf k}\right|^2$ induced by the change of variable Eq. (\[kmudef\]). For this change of integration to be valid, the propagator, and hence the functions $A(q^2)/[q^2A(q^2)^2 + B(q^2)^2]$ and $B(q^2)/[q^2A(q^2)^2 + B(q^2)^2]$, must be analytic over the shaded region in Fig. \[fig1\], $${\rm Re}\,(q^2) > -m_R^2 + \frac{\left({\rm Im}\,(q^2)\right)^2}{4m_R^2}.$$ Equivalently, the function defined by $$\sigma_{\rm Q}(K,\kappa) = \frac{1}{iK + \Sigma(K,\kappa)}, \label{sgdef}$$ must be analytic over the shaded region ${\rm Im}\, K < 0$.
The confinement criterion that $S(p)$ should be free from timelike poles on the negative real $p^2$ axis translates in the heavy quark case to a requirement that $\sigma_{\rm Q}$ should be free from poles on the imaginary $K$ axis. The stronger conjecture [@BRW92], that the quark propagator should be an entire function of $p^2$ translates in the heavy quark formalism to a conjecture that $\sigma_{\rm Q}$ should be an entire function of $K$.
From Eq. (\[hprop\]), we have that the renormalisation condition Eq. (\[mufix\]) is equivalent to $$\left.S(p,\mu)\right|_{p^2 = \mu^2} =
\left. \frac{1 + \gamma_4}{2}\frac{1}{iK} \right|_{K = \kappa}\, ,$$ to zeroth order in $1/m_R$. Typically we choose $\kappa$ to be on the negative imaginary $K$ axis, as the heavy quark propagator asymptotes to the bare propagator as $K \rightarrow -i\infty$, as can be seen from Fig. \[fig1\].
Heavy quark DSE
---------------
To illustrate the derivation of the renormalised heavy quark DSE, we choose Landau gauge ($\xi = 0$ in Eq. (\[gprop\])) and, for the time being, work with the rainbow or bare vertex approximation $$\Gamma_\mu(p,q) = \gamma_\mu.$$ Using Dirac trace identities to project out from Eqs. (\[DSEqn\]) to (\[Sigpr\]) a pair of coupled integral equations gives $$\begin{aligned}
\lefteqn{A'(p^2,\Lambda^2) = }\nonumber \\
& & \!\!\!\!\!\! 1 + \frac{4Z_1}{3p^2}
\int^\Lambda \frac{d^4q}{(2\pi)^4}\, \left[p\cdot q +
2\frac{p\cdot(p - q) q\cdot(p - q)}{(p - q)^2}\right]
\Delta\left[(p - q)^2\right] \frac{A(q^2)}{q^2 A^2 + B^2}, \label{ASDE}\end{aligned}$$ $$B'(p^2,\Lambda^2) = 4Z_1 \int^\Lambda \frac{d^4q}{(2\pi)^4}\,
\Delta\left[(p - q)^2\right] \frac{B(q^2)}{q^2 A^2 + B^2}. \label{BSDE}$$ Substituting $$\frac{A(q^2)}{q^2 A(q^2)^2 + B(q^2)^2} =
\frac{1}{2m_R}\,\frac{1}{iK' + \Sigma(K')} + O\left(\frac{1}{m_R^2}\right) ,$$ $$\frac{B(q^2)}{q^2 A(q^2)^2 + B(q^2)^2} =
\frac{1}{2}\,\frac{1}{iK' + \Sigma(K')} + O\left(\frac{1}{m_R}\right) ,$$ into Eqs. (\[ASDE\]) and (\[BSDE\]) gives $$\begin{aligned}
m_R\left[A'(p^2,\Lambda^2) - 1\right] & = &
\frac{2Z_1}{3} \int^\Lambda
\frac{d^4k'}{(2\pi)^4}\,\left[1 + 2\frac{(k_4 - k'_4)^2}{(k - k')^2}
\right] \nonumber \\
& & \times \Delta\left[(k - k')^2\right] \frac{1}{iK'+\Sigma(K',\kappa)}
+ O\left(\frac{1}{m_R}\right) , \label{AprSDE}\end{aligned}$$ and $$B'(p^2,\Lambda^2) = 2Z_1 \int^\Lambda \frac{d^4k'}{(2\pi)^4}\,
\Delta\left[(k - k')^2\right] \frac{1}{iK'+\Sigma(K',\kappa)}
+ O\left(\frac{1}{m_R}\right) . \label{BprSDE}$$ From Eqs. (\[AApr\]), (\[BBpr\]), (\[sigAdef\]), (\[sigBdef\]) and (\[Sigdef\]) we have $$\Sigma(K,\kappa) = \left[B'(p^2,\Lambda^2) - m_R A'(p^2,\Lambda^2)\right]
-[p^2\rightarrow\mu^2],$$ which gives $$\Sigma(K,\kappa) = \frac{4}{3}
\int^\Lambda \frac{d^4k'}{(2\pi)^4}\,
\left[ \frac{\left|{\bf k} - {\bf k}'\right|^2}{(k - k')^2}
\Delta\left[(k - k')^2\right] \frac{1}{iK'+\Sigma(K',\kappa)}
- (K\rightarrow\kappa) \right] , \label{hsde1}$$ where $K$ on the left hand side is related to $k$ under the integrand via Eq. (\[Kdef\]), and a similar definition exists relating $K'$ and $k'$. Here we have assumed that Eqs. (\[Zfix\]) and (\[AprSDE\]) imply $Z_1 = 1 + O(1/m_R)$.
As noted above, $K$ and $k_4$ are interchangeable to leading order, except where they occur in the denominator of the heavy fermion propagator $1/[iK' + \Sigma(K')]$ in the integrand of Eq. (\[hsde1\]). This is because, for any gluon propagator $\Delta$ with a realistic asymptotic UV behaviour, all powers of $k'$ must be retained in the denominator to maintain the same degree of divergence in the heavy fermion DSE as in the original equation (\[DSEqn\]). With this observation, and choosing $k_\mu = ({\bf 0},K)$ and the renormalisation point $({\bf 0},\kappa)$, we arrive at the leading order heavy quark DSE $$\begin{aligned}
\Sigma(K,\kappa) & = & \frac{4}{3} \int^\Lambda \frac{d^4k'}{(2\pi)^4}\,
\frac{1}{ik_4' + k'^2/(2m_R) + \Sigma(k'_4,\kappa)} \nonumber \\
& & \times \left\{ \frac{\left|{\bf k}'\right|^2
\Delta\left[(K - k'_4)^2 + \left|{\bf k}'\right|^2 \right]}
{(K - k'_4)^2 + \left|{\bf k}'\right|^2} - (K \rightarrow\kappa) \right\}.
\label{hsde2}\end{aligned}$$ We show in the Appendix that, with the smeared FR gluon propagator Ansatz Eq. (\[GFRprop\]), and assuming a hierarchy of scales $$m_t,\;\left|K\right|,\;\left|\kappa\right| \ll m_R \ll \Lambda, \label{hier}$$ the integral in Eq. (\[hsde2\]) is independent of the ultraviolet cutoff $\Lambda$, and the heavy quark self energy behaves like $$\Sigma(K,\kappa) \sim 2id(\kappa - K) \ln\left(\frac{m_R}{m_t}\right),$$ as $m_R\rightarrow\infty$. With the more severely truncated gaussian gluon propagator Eq. (\[gauss\]), the renormalisation point $\kappa$ can be taken to $-i\infty$ and the ${k'}^2/(2m_R)$ term in the denominator of the integrand ignored with impunity.
Choice of renormalisation point
-------------------------------
At the end of the day, physical quantities must be insensitive to the choice of renormalisation point $\kappa$. In this section we note that the freedom to choose the renormalisation point is equivalent to the notion of a ‘residual mass’ in HQET [@FNL92; @N94], that is, the notion that to zeroth order in $1/m_R$, physical quantities computed in HQET do not depend on the choice of $m_R$.
After formally carrying out the spatial momentum integration in Eq. (\[hsde2\]), one obtains an equation generically of the form $$\Sigma(K,\kappa) = \frac{1}{\sigma_Q(K,\kappa)} - iK
= \int_{-\infty}^\infty dK' \, \left[T(K - K') - T(\kappa - K')\right]
\sigma_Q(K',\kappa), \label{frmly}$$ for some kernel $T$, and with $\sigma_Q$ defined by Eq (\[sgdef\]). It is possible to show from this generic form that the effect of making a change of renormalisation point $\kappa_{\rm old} \rightarrow \kappa_{\rm new}$ is equivalent to a shift of the quark propagator solution along the imaginary $K$ axis: $$\sigma_Q(K,\kappa_{\rm new}) = \sigma_Q(K - i\delta m,\kappa_{\rm old}),
\label{shift}$$ where $\delta m$ is the solution to $$\frac{1}{i \kappa_{\rm new}} =
\sigma_Q(\kappa_{\rm new} - i\delta m,\kappa_{\rm old}), \label{delm}$$ that is, the shift is that required to ensure that the new heavy quark propagator pass through the renormalisation point $\sigma_Q(\kappa_{\rm new},\kappa_{\rm new}) = 1/i\kappa_{\rm new}$. Referring to Fig. \[fig1\] we see that this is equivalent to a shift in the position of the origin of the $K$ plane along the $p^2$ axis corresponding to changing $m_R$ by an amount $\delta m$.
One can also demonstrate that the mass differences between any two heavy quark–light antiquark meson states calculated from the Bethe-Salpeter formalism set out in ref. [@BL97] is independent of the renormalisation point.
Results
=======
Our main concern in this paper is to compare how the analytic structure of the heavy quark propagator solutions is affected by the approximations employed both for the quark-gluon vertex and the gluon propagator. In an earlier work [@BL97] an attempt was made to study the heavy quark–light antiquark meson spectrum using a combination of rainbow DSE and ladder BSE. It was found that, if a simple gaussian Ansatz is used for the gluon propagator, complex conjugate poles occur in the heavy quark propagator which prevent solution of the meson BSE. Below we systematically explore the movement of the poles as the bare vertex of the rainbow approximation is replaced by a Ball-Chiu vertex, and as the gaussian gluon propagator is replaced by the more realistic Frank and Roberts propagator.
Gaussian gluon propagator
-------------------------
### Rainbow approximation
As noted in the previous section, if the gluon propagator Eq. (\[gauss\]) is used, we may set the renormalisation point $\kappa = -i\infty$ and ignore the ${k'}^2/(2m_R)$ term in Eq. (\[hsde2\]). Furthermore, there is no need to distinguish between the independent momentum variables $K$ and $k_4$. With these simplifications the Landau gauge, rainbow heavy quark DSE becomes $$\Sigma(k_4) = \frac{4}{3} \int \frac{d^4k'}{(2\pi)^4} \,
\frac{\left|{\bf k} - {\bf k}'\right|^2}{(k - k')^2}
\frac{\Delta\left((k - k')^2\right)}
{ik_4' + \Sigma(k_4')}. \label{baresde}$$ Choosing the infrared dominant gluon propagator Eq. (\[delpr\]), the DSE reduces to an algebraic equation with solution $$\Sigma(k_4) =
\left \{ \begin{array}{lrl}
\frac{1}{2} \left(-ik_4 + \sqrt{\frac{3\mu^2}{4} - k_4^2} \right)
& \mbox{if } & 0\le k_4 < \frac{\sqrt{3}\mu}{2} , \\
& & \\
\frac{i}{2} \left(- k_4 + \sqrt{k_4^2 - \frac{3\mu^2}{4}} \right)
& \mbox{if } & k_4\ge \frac{\sqrt{3}\mu}{2} ,
\end{array} \right. \label{baresoln}$$ or, using the definition (\[sgdef\]), $$\sigma_{\rm Q}(k_4) =
\left \{ \begin{array}{lrl}
\frac{8}{3\mu^2} \left(-ik_4 + \sqrt{\frac{3\mu^2}{4} - k_4^2} \right)
& \mbox{if } & 0\le k_4 < \frac{\sqrt{3}\mu}{2} , \\
& & \\
\frac{8i}{3\mu^2} \left(- k_4 + \sqrt{k_4^2 - \frac{3\mu^2}{4}} \right)
& \mbox{if } & k_4\ge \frac{\sqrt{3}\mu}{2} ,
\end{array} \right. \label{delsig}$$
Alternatively, choosing the gaussian gluon propagator Eq. (\[gauss\]) and carrying out the $d^3{\bf k}$ integration, we obtain the integral equation [@GR80] $$\begin{aligned}
\lefteqn{\Sigma(k_4) = \frac{\mu^2}{2\alpha^2\sqrt{\pi}}
\int_{-\infty}^\infty dk_4 \, \frac{1}{ik_4' + \Sigma(k_4')} \times}
\nonumber \\
& & \left\{\sqrt{\alpha} \left[\frac{\alpha}{2} - (k_4 - k_4')^2 \right]
e^{-(k_4 - k_4')^2/\alpha} + \sqrt{\pi} \left|k_4 - k_4'\right|^3
{\rm erfc}\left(\frac{\left|k_4 - k_4'\right|}
{\sqrt{\alpha}}\right) \right\},
\label{bareie}\end{aligned}$$ where ${\rm erfc}\,z = 1 - {\rm erf}\,z$ is the complementary error function. This equation can be solved numerically.
### Ball-Chiu vertex
If any of the minimal Ball-Chiu Ansatz Eq. (\[BCver\]), the Curtis-Pennington Ansatz Eqs.(\[genver\]) and (\[CPTr\]), or the Haeri Ansatz Eq. (\[Hver\]) is used in place of the bare vertex, together with the Landau gauge gluon propagator, we obtain in place of Eq. (\[baresde\]) the equation $$\Sigma(k_4) = \frac{4}{3} \int \frac{d^4k'}{(2\pi)^4} \,
\frac{\left|{\bf k} - {\bf k}'\right|^2}{(k - k')^2}
\frac{\Delta\left((k - k')^2\right)}{ik_4' + \Sigma(k_4')}
\left[1 + \frac{\Sigma(k_4) - \Sigma(k'_4)}{i(k_4 - k'_4)}\right].
\label{bcsde}$$ It is interesting to note that, within the set of vertex Ansätze we have considered, the heavy quark propagator is insensitive to the transverse part of the vertex. This is not difficult to understand for the Curtis-Pennington vertex, in which the transverse part is heavily damped by the presence of the factor $M^4 \sim m^4$ in the denominator $d(p,q)$. However, in the case of the Haeri vertex there is no such obvious mechanism, and one is led to question whether the heavy quark propagator may be insensitive to a broad class of Ansätze satisfying the criteria specified above Eq. (\[genver\]).
Taking the gluon propagator to be the infrared dominant form Eq. (\[delpr\]), gives the differential equation $$\Sigma(k_4) = \frac{3\mu^2}{16i} \frac{d}{dk_4}\ln[ik_4 + \Sigma(k_4)],$$ which, together with the boundary condition $\sigma_{\rm Q}(k_4)
\rightarrow 0$ as $k_4 \rightarrow -i\infty$, admits the solution $$\sigma_{\rm Q}(k_4) = \beta\left[\sqrt{\pi} e^{-\beta^2k_4^2}
+ 2iF(-\beta k_4) \right], \label{bcsoln}$$ where $\sigma_Q$ is defined by Eq. (\[sgdef\]), $\beta = 2\sqrt{2}/\mu\sqrt{3}$ and $$F(z) = e^{-z^2}\int_0^z e^{t^2} dt = \frac{i\sqrt{\pi}}{2} e^{-z^2}
{\rm erf}(-iz),$$ is Dawson’s integral. We note that this solution is an entire function of $k_4$, which, as pointed out earlier, is a desirable feature of a quark propagator. This comes as no surprise, as it simply the heavy quark limit of the model considered in ref. [@BRW92], in which it was demonstrated that the combination of Ball-Chiu vertex and infrared dominant gluon propagator leads to an entire function propagator for all values of the bare current quark mass.
It is of interest to determine to what extent this analytic structure is a feature of the Ball-Chiu vertex, and to what extent it is a feature of the infrared dominant gluon propagator. If the infrared dominant propagator is replaced by the gaussian smeared form Eq. (\[gauss\]), we obtain the integral equation $$\begin{aligned}
\lefteqn{\Sigma(k_4) = \frac{\mu^2}{2\alpha^2\sqrt{\pi}}
\int_{-\infty}^\infty dk_4 \, \frac{1}{ik_4' + \Sigma(k_4')} \times}
\nonumber \\
& & \left\{\sqrt{\alpha} \left[\frac{\alpha}{2} - (k_4 - k_4')^2 \right]
e^{-(k_4 - k_4')^2/\alpha} + \sqrt{\pi} \left|k_4 - k_4'\right|^3
{\rm erfc}\left(\frac{\left|k_4 - k_4'\right|}
{\sqrt{\alpha}}\right) \right\}
\nonumber \\
& & \times
\left[1 + \frac{\Sigma(k_4) - \Sigma(k'_4)}{i(k_4 - k'_4)}\right],
\label{bcie}\end{aligned}$$ which can be solved numerically.
### Numerical results: gaussian gluon propagator
For the purpose of determining the analytic structure of the heavy fermion propagator obtained from the DSE with the gaussian gluon propagator Eq. (\[gauss\]), it is sufficient to look at the one parameter family of models obtained by scaling either $\mu$ or $\alpha$ to unity. We choose to scale $\mu$ to unity, which amounts to working with a set of dimensionless quantities $$\hat{k}_4 = k_4/\mu, \hspace{5 mm} \hat{\alpha} = \alpha/\mu^2,
\hspace{5 mm} \hat{\sigma}_{\rm Q} = \mu\sigma_{\rm Q}.$$ This choice enables us to recover the infrared dominant model in the limit $\alpha \rightarrow 0$.
In Figs. \[fig2\] and \[fig3\] we plot the heavy quark self energy $\Sigma(k_4)$ as a function of real $k_4$ obtained from the bare vertex DSE Eq. (\[bareie\]) and the Ball-Chiu vertex DSE Eq. (\[bcie\]) for $\hat{\alpha} = 1$, 2 and 3. These results are obtained by iterating from an initial guess and using a Simpson’s rule quadrature. We find that the derivative-like terms in Eq. (\[bcie\]) prevent a numerical solution for values of $\hat{\alpha}$ less than 1, as numerical noise in the function values becomes unstable with respect to iteration at small values of $k_4$. This problem is a general feature of numerical treatments of DSEs with Ball-Chiu-like vertices. Also plotted are the $\hat{\alpha} = 0$ analytic results Eqs. (\[baresoln\]) and (\[bcsoln\]). In all cases the self energy is characterised by a real part which peaks at zero and an imaginary part which peaks near the typical scale of the model $k_4 \sim \mu$. The self energy for negative real $k_4$ can be obtained from these results using the reflection property $\Sigma(-k_4^*) =
\Sigma(k_4)^*$.
To solve for the heavy quark propagator away from the real $k_4$ axis we shift the contour of integration into the complex plane parallel to the real $k_4$ axis and again solve iteratively. We note that, to determine $\Sigma(k_4)$ for complex arguments, it is necessary to move the contour of integration to pass through the point $k_4$. This is because the radial part of the $d^3{\bf k}'$ integration, carried out in going from Eq. (\[baresde\]) to Eq. (\[bareie\]) or from Eq. (\[bcsde\]) to Eq. (\[bcie\]), creates a pinch singularity at $k'_4 = k_4$ in the error function term in Eq. (\[bareie\]) or Eq. (\[bcie\]).
We have carried out a search for poles in the propagator function $\sigma_Q(k_4)$ for a range of values of $\alpha$ for both the bare and Ball-Chiu vertex. Our results are listed in Table \[tab\]. In all cases we find that the only observed poles occur for ${\rm Im}\, k_4 > 0$, and that $\sigma_Q$ dies away to small values and is free from singularities over that part of the shaded region in Fig. \[fig1\] accessible to our computer program. Of course we are unable to pass the contour of numerical integration through the pole itself, and these results are attained by extrapolation from results of contours which we gradually moved deeper into the complex plane. We were unable to obtain a reasonable extrapolation for the Ball-Chiu vertex at $\alpha = 1$, again because of the iterative instability problem associated with the derivative-like term in Eq. (\[bcie\]).
We also list in Table \[tab\] the results of using the model gluon propagator $$g^2 D_{\mu \nu}(k^2) =
\delta_{\mu \nu} \Delta_{\rm G}(k^2) , \label{flg}$$ where $\Delta_{\rm G}$ is given by Eq. (\[gauss\]). Forms such as Eq. (\[flg\]) are frequently used in phenomenological modelling (see for instance ref. [@BSpapers]) and are sometimes referred to as propagators in a ‘Feynman-like gauge’, though of course they are generally not of the form of Eq. (\[gprop\]). The sole advantage of the Feynman-like gauge is that it leads to considerably simplified calculations. In our case it is possible to locate poles more accurately because there is no pinch singularity requiring the contour of integration to pass through the point in question. Once the propagator has been solved on the real $k_4$ axis, the value of the propagator can be calculated at any point in the complex plane by integrating once along the real axis. Nevertheless, we have also repeated our pole calculations by shifting the contour and extrapolating as in the Landau gauge case as a check on the the consistency of the two methods and find that they agree to within the accuracy given in Table \[tab\] of the corresponding Landau gauge results. A Feynman-like gauge propagator was also used in ref. [@BL97] dealing with the ladder Bethe-Salpeter equation for the heavy quark–light antiquark system. There it was demonstrated that the model with bare quark gluon vertex and gaussian Feynman-like gauge gluon propagator had no solutions because of poles in the heavy and light quark propagators. Ideally, one would like improvements in the DSE approximations to move the poles further from the real $k_4$ axis to avoid the region of the $k_4$ plane sampled by a Bethe-Salpeter calculation.
From Table \[tab\] we conclude that simply replacing the bare vertex by the Ball-Chiu vertex in itself does nothing to improve the pole structure of the heavy quark propagator, either for the Landau gauge gluon propagator or the Feynman-like gluon propagator. In particular we find that, as the gaussian width $\alpha$ increases, a mass pole pole moves in along the imaginary $k_4$ axis. A pole on the imaginary axis indicates that the fermion can propagate as a free particle, and the position on the positive imaginary axis gives the contribution to the quark mass from the dynamical self dressing. For the bare vertex, the pole splits into conjugate pairs either side of the imaginary axis as $\alpha$ decreases. In this instance the quark becomes a confined particle. Numerical difficulties described above prevented us from confirming that the same situation occurs in the case of the Ball-Chiu vertex. As $\alpha \rightarrow 0$ we must recover the solution Eq. (\[bcsoln\]), which is an entire function with an essential singularity at infinity.
We see from Table \[tab\] that, when poles occur, their position remains almost unchanged in going from Landau to Feynman-like gauge if the Ball-Chiu vertex is used, but not if the bare vertex is used. In a properly formulated gauge covariant calculation, the position of any propagator mass pole should be independent of the gauge fixing procedure [@AF79]. While we certainly do not claim that that our treatment is gauge covariant, it is amusing to note that replacing the bare vertex by the Ball-Chiu vertex (and Feynman gauge by the computationally convenient Feynman-like gauge) appears to go some way towards satisfying this requirement.
Frank and Roberts gluon propagator
----------------------------------
We now return to the renormalised DSE Eq. (\[hsde2\]) with the gluon propagator $\Delta$ set equal to the smeared FR Ansatz Eq. (\[GFRprop\]). For numerical simplicity we shall restrict ourselves to the rainbow approximation $\Gamma_\mu(p,q) = \gamma_\mu$. We set $$K = X + iY, \hspace{10 mm} \kappa = i\eta.$$ Assuming the contour of integration can be deformed to pass through $k'_4 = K$ for the integral of the first term in chain brackets in Eq.(\[hsde2\]), and through $k'_4 = \kappa$ for the second term, we further set $$k'_4 = x + iY \hspace{3 mm}\mbox{ and } \hspace{3 mm} k'_4 = x + i\eta,$$ respectively in each of these two terms. We also make the replacement $k'^2/(2m_R) \rightarrow (x^2 + \left|{\bf k}'\right|^2)/(2m_R)$ without affecting $\Sigma$ to leading order. This gives $$\begin{aligned}
\lefteqn{ \Sigma(X + iY, i\eta) = \frac{4}{3}
\int_{-\infty}^\infty \frac{dx}{2\pi}
\int\frac{d^3{\bf k}'}{(2\pi)^3}} \nonumber \\
& &
\left\{ \frac{1}{ix - Y + (x^2 + \left|{\bf k}'\right|^2)/(2m_R)
+ \Sigma(x + iY)}
\frac{\left|{\bf k}'\right|^2}{(x - X)^2 + \left|{\bf k}'\right|^2}
\Delta\left[(x - X)^2 + \left|{\bf k}'\right|^2\right] \right. \nonumber \\
& & - \left. \frac{1}{ix - \eta + (x^2 + \left|{\bf k}'\right|^2)/(2m_R)
+ \Sigma(x + i\eta)}
\, \frac{\left|{\bf k}'\right|^2}{x^2 + \left|{\bf k}'\right|^2}
\Delta\left[x^2 + \left|{\bf k}'\right|^2\right]\right\} .
\label{sg2} \end{aligned}$$
For the purpose of carrying out the numerics, it is convenient to change to the polar coordinates $x = r \cos\phi$, $\left|{\bf k}'\right| = r \sin\phi$, giving finally $$\begin{aligned}
\lefteqn{ \Sigma(X + iY, i\eta) = \frac{1}{3\pi^3}
\int_0^\infty dr \int_0^\pi d\phi\, r^3\sin^2\phi} \nonumber \\
& &
\left\{ \frac{1}{ir\cos\phi - Y + r^2/(2m_R)
+ \Sigma(r\cos\phi + iY)} \frac{r^2\sin^2\phi}{r^2 - 2Xr\cos\phi +X^2}
\Delta(r^2 - 2Xr\cos\phi +X^2) \right. \nonumber \\
& & - \left. \frac{1}{ir\cos\phi - \eta + r^2/(2m_R)
+ \Sigma(r\cos\phi + i\eta)}\sin^2\phi \Delta(r^2)\right\}.
\label{hsde3}\end{aligned}$$ This equation is first solved numerically along the line ${\rm Im}\,K= \eta$, (i.e. $Y = \eta$), and the function $\Sigma$ along this line is stored for subsequent calculations at arbitrary $Y$.
### Numerical results: Frank and Roberts propagator
We have numerically solved Eq. (\[hsde3\]) over a region of the complex $K$ plane with the smeared FR gluon propagator $\Delta_{\rm GFR}$. Our parameter choices are $m_t = 0.69$ GeV, in agreement with ref. [@FR96], and $\alpha
= 16m_t^2 \hat{\alpha}d/3 = 0.5643$ (GeV)$^2$ corresponding to $\hat{\alpha} = 0.5$. The choice of $\alpha$ is designed so that a comparison can be made between the full FR propagator and the gaussian propagator Eq. (\[gauss\]) obtained by keeping only the first term in Eq. (\[GFRprop\]). From Table \[tab\] we know that gaussian propagator $\Delta_{\rm G}$ with the parameter choice $\hat{\alpha} = 0.5$ results in a pair of conjugate poles in the heavy quark propagator. The heavy quark mass is set to $m_R = 5.0$ GeV.
In Fig. \[fig4\] we plot the modulus $\left|\sigma_Q(K)\right|$ of the heavy quark propagator for the full gaussian FR propagator $\Delta_{\rm GFR}$ and in Fig. \[fig5\] plot the same quantity using only the gaussian part of $\Delta_{\rm G}$ with the parameters otherwise unchanged. Both calculations have been done using the renormalisation point $\kappa = -1.0i$ GeV, and to clarify the comparison the same region of the $K$ plane is displayed in both plots. The calculation of $\sigma_Q(K)$ involves a shift of integration path to a contour parallel to the real axis passing through the point $K$. Deforming the contour to include points behind the pole is a numerically tedious exercise which is unlikely to enhance our understanding, so no results are given for the part of the $K$ plane in Fig. \[fig5\].
In the process of carrying out our computations, we have observed that the shift property resulting from changes of renormalisation point, namely Eq. (\[shift\]), is indeed respected by our numerical solutions. In fact, the full plot in Fig. \[fig4\] was pieced together by altering the renormalisation point to obtain solutions in strips of the complex plane parallel to the real axis, and using Eq. (\[delm\]) to match solutions where strips overlapped.
We note a clear movement of the propagator pole further away from the real $K$ axis when the asymptotic ultraviolet tail is included in the gluon propagator. Since it is the proximity of the pole to the real axis which prevented a solution to BSE in our earlier studies, this suggests that this movement of the pole portends well for future possible studies of heavy quark mesons if careful attention is paid to the asymptotic ultraviolet behaviour of the gluon propagator.
However, a note of caution is in order. A different choice of renormalisation point would result in the plots in Figs. \[fig4\] and \[fig5\] shifting by different amounts respectively along the imaginary axis, ensuring that both plots pass through the same point $\sigma_Q(\kappa) = 1/i\kappa$. Consequently, the actual amount by which the pole moves away from the real $K$ axis as a result of adding an asymptotic tail to the gluon propagator is an artefact of the choice of renormalisation point, though the movement will always be [*away*]{} from the real axis. Of course only by carrying though the BSE calculation completely can one say for certain whether bound state meson solutions can be obtained.
Conclusions and outlook
=======================
We have explored the analytic structure of heavy quark propagators following a recently proposed formalism which borrows ideas both from the DSE technique and HQET. It is our belief that, if the successes of HQET are to be properly understood, we must first understand how the non-perturbative dynamics of QCD affect the heavy quark propagator. Within the light quark sector the analytic structure of the quark propagator is perhaps best understood in terms of model Dyson-Schwinger equations. It is therefore a worthwhile exercise to extend the DSE technique to the heavy quark limit.
The initial attempt in this direction [@BL97] failed essentially because the approximations used led to spurious propagator poles which prevented solution of the bound state Bethe-Salpeter equations. Two approximations were involved: modelling of the quark-gluon vertex and of the the gluon propagator. We have focused on each of these aspects in turn in this paper. In order to deal with an improved Ansatz for the gluon propagator with a realistic asymptotic ultraviolet behaviour, it has been necessary to formulate a properly renormalised version of the heavy quark DSE technique proposed in ref. [@BL97]. As an interesting corollary to our formalism we observe that the freedom to choose the renormalisation point is tantamount to the freedom in zeroth order HQET to choose the heavy quark mass up to a residual mass.
We have first examined the effect of replacing the bare vertex with Ansätze based on the Ball-Chiu vertex [@BC80], which is primarily designed to satisfy the Ward-Takahashi identity. Specifically, we have considered the minimal Ball-Chiu vertex and two variants: that proposed by Curtis and Pennington [@CP90] and that proposed by Haeri [@H91]. The two variants differ from the minimal vertex by the inclusion of extra transverse components. We find that, to zeroth order in the inverse of the heavy quark mass, the heavy quark propagator is insensitive to which of the three above Ansätze is used. One is led to question to what extent transverse additions to the minimal Ball-Chiu vertex can be ignored in determining the leading order heavy quark propagator. If they can be be ignored in general there are immediate benefits in using the heavy fermion limit as a test-case for studies of confining field theories.
In our numerical calculations we began with the Landau gauge form of the model gaussian gluon propagator $\Delta_{G}$, Eq. (\[gauss\]), which was employed (together with the bare vertex) in previous studies [@BL97]. Unfortunately, we find no improvement in the propagator pole structure in going from the bare vertex to the Ball-Chiu vertex while maintaining a gaussian gluon propagator. That is to say, timelike mass poles indicating non-confinement, or conjugate poles which are likely to interfere with the successful solution to bound state problems, are not removed simply by improving the quark gluon vertex Ansatz alone. However, in the limit in which the width of the gaussian gluon propagator is taken to zero (the ‘infrared dominant model’), we do obtain an entire function heavy quark propagator, free from singularities except an essential singularity at infinity. This is consistent with the equivalent finite quark mass calculation [@BRW92], and may provide a useful propagator for phenomenological modelling purposes.
We conclude then that it is most likely the remaining approximation, namely the gaussian model gluon propagator, which is responsible for the poor analytic structure previously obtained for the quark propagator. To explore this possibility, we have replaced the simple gaussian gluon propagator Ansatz of ref. [@BL97] by a gaussian smeared version of the more sophisticated Frank and Roberts Ansatz $\Delta_{\rm FR}$ given by Eq. (\[FRprop\]). In this case, convergence of the integral in the DSE which is lost by naively retaining only the lowest order of the $1/m_R$ expansion of the quark propagator must be restored by judiciously including at least the spatially dependent $O(1/m_R)$ part: $$S(p) = \frac{1 + \gamma_4}{2} \frac{1}
{ik_4 + \left|{\bf k}\right|^2/2m_R + \Sigma(k_4)} +
O\left(m_R^{-1}\right).$$ The renormalised current quark mass $m_R$ then becomes an ultraviolet regulator, and in the case of the FR propagator the lowest order contribution to the mass expansion of the heavy quark self energy behaves as $\ln(m_R/m_t)$, where $m_t$ delineates the scale at which the asymptotic ultraviolet behaviour of the FR propagator sets in.
Our numerical solutions of the heavy quark DSE show a clear movement of the offensive propagator poles away from that part of the complex momentum plane likely to be sampled by a Bethe-Salpeter calculation of heavy quark meson states. However we caution that the amount by which the poles shift is, strictly speaking, dependent on the choice of renormalisation point. Without carrying through the the Bethe-Salpeter analysis one cannot say for sure that the problem is solved. A further study of the BSE for heavy mesons is expected to be the focus of future work.
Our work has also thrown up a couple of other interesting questions worthy of attention. Firstly, it should be possible to check directly to what extent the propagator pole structure is invariant with respect to the choice of gauge fixing parameter $\xi$. As the postions of propagator poles should be gauge independent [@AF79], this provides a straightforward measure of the ability of a particular vertex Ansatz to respect gauge covariance of the model.
Secondly, one is led to question the meaning of propagator poles obtained from a Euclidean DSE formalism. Whether propagator poles obtained in this way are an artefact of the approximations used or whether they are a genuine property of quark propagators has been an open question for some time [@RW94; @M94][^2]. It is possible that the heavy fermion limit may help to shed some light on this problem by developing a Bethe-Salpeter formalism for heavy quarkonium states. It is well known that the non-relativistic limit of the Bethe-Salpeter equation for a heavy fermion–heavy antifermion bound state can be written in the form of a Schrödinger equation [@FS82]. The derivation typically assumes physical mass poles in the fermion propagators whose residues contribute to the resulting Schrödinger equation. An analogous derivation for the case of propagator poles which have moved off the timelike momentum axis as a result of a confining gluon propagator may help both with interpretation of quark propagator poles and with understanding the success of heavy quark potential models.
Acknowledgement {#acknowledgement .unnumbered}
===============
The author is grateful to C. D. Roberts and P. Maris for helpful discussions.
Appendix: convergence of the heavy quark DSE {#appendix-convergence-of-the-heavy-quark-dse .unnumbered}
============================================
Consider the heavy quark DSE Eq. (\[hsde2\]), and suppose we assume for the gluon propagator Ansatz the asymptotic ultraviolet behaviour $$\Delta(k^2) \sim \frac{4\pi^2 d}{k^2} \hspace{5 mm}\mbox{ for $k^2 > m_t^2$},$$ where $m_t$ is a scale parameter typically of the order of 1 GeV. This is the behaviour exhibited by the smeared FR Ansatz Eq. (\[GFRprop\]). We demonstrate here that the right hand side of Eq. (\[hsde2\]) is the finite difference of two logarithmically divergent integrals.
We begin with the change of variables $$k'_4 = r \cos\phi, \hspace{5 mm} \left|{\bf k}'\right| = r \sin\phi,$$ and hence $$\int^\Lambda d^4k' = 4\pi\int dk'_4 \int d\!\left|{\bf k}'\right|
\,\left|{\bf k}'\right|^2
= 4\pi\int_0^\Lambda dr \int_0^\pi d\phi \,r^3 \sin^2\phi.$$ This gives $$\begin{aligned}
\Sigma(K,\kappa) & = & \frac{1}{3\pi^3} \int_0^\Lambda dr \int_0^\pi d\phi\,
\frac{r^5\sin^4\phi }{ir\cos \phi + r^2/(2m_R) + \Sigma(r\cos\phi)}
\nonumber \\
& & \times \left[ \frac{\Delta(K^2 - 2Kr\cos\phi + r^2)}
{K^2 - 2Kr\cos\phi + r^2} - (K\rightarrow\kappa) \right] . \label{pol}\end{aligned}$$ Implicit in this equation is a hierarchy of scales given by Eq. (\[hier\]). Simply counting powers of $r$ in the integrand, we see that each of the two terms diverges as $$m_R\int^\Lambda \frac{dr}{r} \sim m_R \ln \Lambda.$$
On the other hand consider the difference of the two terms. For $r > m_t$ the part in square brackets is $$\begin{aligned}
\lefteqn{ \left[ \frac{4\pi^2 d}
{(K^2 - 2Kr\cos\phi + r^2)^2} - (K\rightarrow\kappa) \right] }
\nonumber \\
& = & 4\pi^2d\left[\frac{4(K - \kappa)\cos\phi}{r^5} +
\frac{2(K^2 - \kappa^2)(6\cos^2\phi - 1)}{r^6}\right] +
O\left(\frac{1}{r^7}\right). \end{aligned}$$ Neglecting $\Sigma(r\cos\phi)$ in Eq. (\[pol\]) for large $r$, we can approximate the propagator contribution to the integrand by $$\frac{1}{ir \cos\phi + r^2/(2m_R)} =
- \frac{i\cos\phi - r/2m_R}{r(\cos^2\phi + r^2/4m_R^2)}.$$ Then, taking into account the hierarchy (\[hier\]), the contribution to the integrand for $r>m_t$ is approximately $$\frac{16id(\kappa - K)}{3\pi}\int_{m_t}^\infty \frac{dr}{r}
\int_0^\pi d\phi\,\frac{\sin^4\phi \cos^2\phi}{\cos^2\phi + r^2/4m_R^2}$$ giving $$\Sigma(K,\kappa) \sim 2id(\kappa - K) \ln\left(\frac{m_R}{m_t}\right)
\hspace{5 mm}\mbox{as $m_R\rightarrow\infty$}.$$ The last step can be achieved using the crude approximation $$\frac{\cos^2\phi}{\cos^2\phi + r^2/4m_R^2} \approx \left\{
\begin{array}{lr} 1 & \mbox{ if $r<4m_R$} \\
(4m_R^2/r^2)\cos^2\phi & \mbox{ if $r>4m_R$}
\end{array} \right. .$$
C. D. Roberts and A. G. Williams, Prog. Part. and Nucl. Phys. [**33**]{}, 475 (1994). C. J. Burden, et al., Phys. Rev. C [**55**]{}, 2649 (1997); R. T. Cahill and S. M. Gunner, Phys. Lett. B [**359**]{}, 281 (1995). P. Jain and H. J. Munczek, Phys. Rev. D [**48**]{}, 5403 (1993), and references therein, P. Maris and C. D. Roberts, [*$\pi$- and $K$-meson Bethe-Salpeter amplitudes*]{}, nucl-th/9708029. C. D. Roberts, Nucl. Phys. A [**605**]{}, 475 (1996); C. J. Burden, et al., Phys. Lett. B [**371**]{}, 163 (1996). C. D. Roberts, A. G. Williams and G. Krein, Int. J. Mod. Phys. A [**7**]{}, 5607 (1992). H. Munczek, Phys. Lett. B [**175**]{}, 215 (1986). C. J. Burden, C. D. Roberts and A. G. Williams, Phys. Lett. B [**285**]{}, 347 (1992). P. Maris and H. A. Holties, Int. J. Mod. Phys. A [**7**]{}, 5369 (1992). S. J. Stainsby, and. R. T. Cahill, Int. J. Mod. Phys. A [**7**]{} 7541 (1992). P. Maris, Phys. Rev. D [**50**]{} 4189 (1994), and references therein. T. W. Allen, and. C. J. Burden, Phys. Rev. D [**53**]{} 5842 (1996). C. J. Burden and D.-S. Liu, Phys. Rev. D [**55**]{}, 367 (1997). R. Delbourgo and D.-S. Liu, Phys. Rev. D [**56**]{}, 27 (1997). M. Neubert, Phys. Rep. [**245**]{}, 259 (1994). J. S. Ball and T.-W. Chiu, Phys. Rev. D [**22**]{}, 2542 (1980). M. R. Frank and C. D. Roberts, Phys. Rev. C [**53**]{}, 390 (1996). D. C. Curtis and M. R. Pennington, Phys. Rev. D [**42**]{}, 4165 (1990); Phys. Rev. D [**46**]{}, 2663 (1992). B. Haeri, Phys. Rev. D [**43**]{}, 2701 (1991). H. Munczek (private communication). For a summary of the proof of equivalence between the Haeri vertex and the spectral representation, see C. J. Burden and C. D. Roberts, Phys. Rev. D [**47**]{}, 5581 (1993). H. J. Munczek and A. M. Nemirovsky, Phys. Rev. D [**28**]{}, 181 (1983); R. T. Cahill and C. D. Roberts, Phys. Rev. D [**32**]{}, 2419 (1985). A. F. Falk, M. Neubert and M. Luke, Nucl. Phys. B [**388**]{}, 363 (1992). I. S. Gradshteyn and I. M. Ryzhik, [*Table of integrals, series and products*]{}, Academic, New York, 1980, Eq. (3.446.2). D. Atkinson and M. P. Fry, Nucl. Phys. B [**156**]{}, 301 (1979). D. McKay and H. Munczek, Phys. Rev. D [**55**]{}, 2455 (1997). D. Flamm and F. Schöberl, [*Introduction to the quark model of elementary particles*]{}, (Gordon and Breach, New York, 1982).
------------------------------- --------------------- ----------- ----------------------- -----------
$\hat{\alpha} = \alpha/\mu^2$ bare vertex BC vertex bare vertex BC vertex
0.5 $\pm0.510 + 0.415i$ ? $\pm0.5466 + 0.5109i$ ?
1 $\pm0.378 + 0.506i$ ? $\pm0.3383 + 0.6758i$ $0.4429i$
2 $0.466i$ $0.31i$ $0.3844i$ $0.3130i$
3 $0.307i$ $0.252i$ $0.2880i$ $0.2554i$
4 $0.251i$ $0.219i$ $0.2408i$ $0.2210i$
------------------------------- --------------------- ----------- ----------------------- -----------
: Position $k_4/\mu$ of poles in the heavy quark propagator closest to the real $k_4$ axis, using the gaussian gluon propagator Ansatz $\Delta_{\rm G}$. Numerical instabilities prevent an accurate location of the pole in the case indicated by a question mark.[]{data-label="tab"}
[^1]: Strictly speaking, it is the Slavnov-Taylor identities, which include ghost contributions, and not the Ward-Takahashi identities which are relevant to QCD. By using the Ball-Chiu vertex we are effectively ignoring the ghost self energy and ghost-quark scattering kernel. This is a commonly used approximation [@RW94].
[^2]: In an interesting recent development, McKay and Munczek [@MM97] have examined the analytic structure of quark propagators when an extra constraint that solutions of the DSE should be Fourier transformable is imposed.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Eric Werner [^1]\
University of Oxford\
Department of Physiology, Anatomy and Genetics,\
and Department of Computer Science,\
Le Gros Clark Building, South Parks Road, Oxford OX1 3QX\
email: [email protected]\
bibliography:
- 'StemCells.bib'
date:
nocite:
- '[@Nakada2011]'
- '[@Werner2003b; @Werner2005; @Werner2007a; @Werner2009; @Werner2010]'
- '[@Tyson1991; @Tyson2003]'
- '[@Dick2010; @Dick2009; @Dick2008]'
- '[@Shackleton2010; @Shackleton2010a; @Shackleton2010b; @Shackleton2010c]'
- '[@Gardner1999; @Gardner2001; @Gardner2006; @Gardner2007; @Tesar2007]'
- '[@Sonnenschein1999]'
- '[@Bizzarri2008]'
title: Stem Cell Networks
---
=1
**Abstract**
> *We present a general computational theory of stem cell networks and their developmental dynamics. Stem cell networks are special cases of developmental control networks. Our theory generates a natural classification of all possible stem cell networks based on their network architecture. Each stem cell network has a unique topology and semantics and developmental dynamics that result in distinct phenotypes. We show that the ideal growth dynamics of multicellular systems generated by stem cell networks have mathematical properties related to the coefficients of Pascal’s Triangle. The relationship to cancer stem cells and their control networks is indicated. The theory lays the foundation for a new research paradigm for understanding and investigating stem cells. The theory of stem cell networks implies that new methods for generating and controlling stem cells will become possible.*
[**Key words**]{}: [stem cells, stem cell networks, cene, cenome, developmental control networks, cancer stem cells, stochastic stem cell networks, stochastic cancer stem cell networks, metastatic hierarchy, linear networks, geometric cancer networks, systems biology, computational biology, multiagent systems, muticellular modeling, simulation, cancer modeling, cancer simulation]{}
Introduction
============
Stem cells are crucial in embryogenesis as well as maintaining our bodies. They are the source of tissue regeneration, generating skin, hair, and replacing damaged cells in wound healing. They are also the source of many cancers. Understanding the nature of stem cells and what controls them is, therefore, of great importance. The aim of this essay is to understand the essence of what makes stem cells run. To achieve this we present a general theory of developmental networks that lie hidden in stem cells and control them. The theory is abstract and powerful describing the universal principles and architectures that underly all stem cell networks. This allows us to model and simulate all stem cell types. And that enables us to do computational experiments by running different networks in virtual stem cells and observing their behavior in a virtual space-time context of other cells.
\[sec:G3OctalTetrahedron\]
The result is a deep understanding of the properties of real stem cells and their controlling networks. It also gives us a profound insight into the nature of the role of cancer stem cells in metastases. The theory predicts that there exists a metastatic hierarchy that corresponds directly to the stem cell network hierarchy. We will describe the stem cell network hierarchy and show that this network hierarchy has direct causal links to this previously unknown hierarchy of metastases. This has important implications for the classification, diagnosis and treatment of metastases in cancer. The theory also implies that it is in principle possible to construct methodologies that can transform cancer stem cell networks into a harmless form giving us the potential to ultimately cure and even eliminate such cancers.
Developmental control networks
------------------------------
We postulate that multicellular development is controlled by networks. We call these networks [*developmental control networks*]{} or [*cenes*]{} for control genes. The global developmental control network is the entire network that controls the development of an organism. This global cene is sometimes called the [*cenome*]{} [@Werner2011a].
[*Stem cell networks*]{} form a subclass of developmental networks. They have a well defined network architectures with a range of topologies. We will show that stem cell networks form a hierarchy based on their network topology. The properties of stem cells are determined by the properties of their network architecture. It is the network architecture that gives stem cells their primary defining property, namely, the capacity to self-renew while generating an unlimited set of progeny of various cell types. The network locality determines the stem cell maturity and the capacity of its progeny to differentiate to various cell types. Thus, the potency of a stem cell’s daughter cell is determined by the locality of that stem cell network in the global developmental network.
Definitions: Two kinds of stem cells
------------------------------------
We distinguish two independent properties of stem cells: Self-renewal (iteropotency) and the capacity to transform into various cell types (transpotency). These two properties are often confused in the scientific and popular science literature.
### Iteropotent stem cells
We define a cell to be [*iteropotent*]{} if it has the capacity of self-renewal. When an [*iteropotent stem cell*]{} divides it generates two daughter cells, one that has the same control state as the parent and one that is in a new developmental control state. The [*developmental control state*]{} of a cell is defined by the set nodes in the developmental network that are active in that cell. Examples of self-renewing, iteropotent stem cells are skin (epithelial) and hair stem cells, bone marrow stem cells, and cancer stem cells. [*Aniteropotent cells*]{} are by definition non-self-renewing cells.
### Transpotent stem cells
We define a cell to be [*transpotent*]{} if it has the capacity to differentiate into other cell types. The [*degree of transpotency*]{} is the variety and number of cell types into which a transpotent cell can differentiate. Transpotent stem cells are usually, roughly classified into embryonic, totipotent, pluripotent, and oligopotent stem cells. Transpotency may be conditional and context dependent. [*Natural transpotency*]{} is a natural, inherent property of the cell. It is the natural property of a cell to transform itself based on its context and developmental network state. [*Artificial transpotency*]{} is an artificial, experimentally induced property of cells that would not naturally occur.
### Relations between transpotency and iteropotency of stem cells
Because self-renewal and transpotency are orthogonal properties of cells they can occur in any combination: A transpotent stem cell need not have the capacity of self-renewal even though it may generate stem cells that do have that capacity. So too, the two properties of stem cells can coexist in the same cell. For example, a self-renewing stem cell may also be transpotent, but if it differentiates it may or may not retain its property of self-renewal. Finally, a self-renewing stem cell may may not be transpotent and yet generate transpotent cells. All these different cases can be more precisely defined and modeled by different developmental stem cell networks.
The graphical network formalism
-------------------------------
We will use an abstract graphical network language to describe developmental networks. Differential equations are often used to describe the quantitative output of cell proliferation. While our graphical networks can in part be reductively formalized by differential equations, such reductions loose many of the architectural properties and advantages of graphical developmental networks. A good deal of theoretical and practical insight is lost in reductions to differential equations. Since many of the biological parameters and molecular implementation of stem cell networks is still unknown and perhaps never will be fully known, graphical representation languages of developmental networks help us to understand developmental processes and may even enable us to control developmental processes without necessarily knowing or understanding their full molecular implementation. This is not to say that modeling attempts with differential equations is not worthwhile. Indeed, we use differential equations to model the physics of cell interactions. One goal is to link up lower level differential equation formalisms of molecular networks with our higher level abstract graphical representations of developmental control networks. Any cooperations to this end are welcomed by the author.
Plan
----
The main focus of this essay will be on self-renewing, iteropotent stem cell networks. We first look at deterministic stem cell networks. We describe meta-stem cell networks and their mathematical properties. We discover a stem cell hierarchy of networks and show that their proliferative properties are related to the geometric numbers and the coefficients of Pascal’s Triangle. Next, we describe the close relationship between normal stem cell networks and cancer stem cell networks. We apply the theory of the hierarchy of stem cell networks to metastases. We show that there is a direct correspondence between the stem cell network hierarchy and a metastatic hierarchy. The theory of a hierarchy of metastases makes testable predictions and has significant implications for cancer diagnosis and treatment. We then add stochasticity to stem cell networks and investigate the range of stochastic stem cell networks. It turns out that adding stochasticity makes stem cell networks extremely flexible. Next, we add cell communication to stem cell networks. Finally, we combine deterministic, stochastic and communication stem cell networks.
A hierarchy of stem cell networks {#sec:NG}
=================================
Traditionally, stem cells have been viewed as cells that endlessly self-regenerate while producing terminal daughter cells of some type regulated by some terminal developmental network. However, there may exist stem cells produce other stem cells. We refer to stem cells that produce stem cells as [*meta-stem cells*]{} or as [*higher order stem cells*]{}. The order of a stem cell depends on the number of linear loops in its developmental control network. A [*3rd-order stem cell network*]{} contains three loops and produces 2nd-order and 1st-order stem cells. In turn, a [*2nd-order stem cell network*]{} contains two loops and produces only 1st-order stem cells. [*1st-order stem cells*]{} are controlled by single loop and produce terminal cells or cells with limited proliferative potential. Cells with limited proliferative potential that can engage in a small number of further divisions are called [*progenitor cells*]{}. We will see that stem cell networks have proliferation or growth potential related to the geometric numbers and the coefficients of Pascal’s Triangle. Without preconditions in an idealized setting without physical constraints, a [*$k-$th order stem cell network*]{} would have the proliferations properties of an $k$-th order geometric network (see \[sec:Gk\]). We will thus sometimes refer to stem cell networks as [*geometric networks*]{} when emphasizing their geometric growth properties and to distinguish them from exponential networks (Werner [@Werner2011b]).
First order linear stem cells {#sec:NSC1}
-----------------------------
A first order stem cell produces no additional stem cells. Instead a first order stem cell $A$ produces cells $B$ that is in a different control state than its parent cell $A$. The stem cell thus produces one daughter cell that inherits the control state of its parent which is $A$ itself and one daughter cell $B$ that is in a new control state. Thus stem cells have a self reflexive control system. Note, that the cell $B$ may still be multi-potent in that it may be controlled by a network that generates a whole multicellular system.
First order stem cells have conditional control networks. This means that their activation depends on not just being linked into another network that activates them but that they can only be active if the conditions $\Phi$ are satisfied.
\(m) \[matrix of math nodes, row sep=3em, column sep=3em, text height=1.5ex, text depth=0.25ex\] [ & & & &\
\_[1]{} A\_[1]{} && B && &&\
& & & &\
]{}; (m-2-1) edge \[pot2, blue\](m-3-2); (m-3-2) edge \[pot2, blue\] (m-2-3); (m-2-1) edge \[selfloop1, cross line\] node\[nodedescr\] [$\alpha_{1}$]{} (m-2-1);
(m-2-1) edge (m-2-3); (m-2-3) edge (m-2-5);
In terms of network topology, first order stem cells are similar to first order geometric cancer networks in that both contain one loop Werner [@Werner2011b] They contain one loop and generate cells linearly, one at a time. The newly generated cells $B$ are terminal cells that do not proliferate or they may be progenitor cells that are controlled by a terminal developmental network (see Werner [@Werner2011b; @Werner2011a] for more on developmental control networks, stem cell and cancer networks. The cells $B$ may also be transpotent cells that can be induced naturally or artificially to differentiate into various cell types.
Second order geometric meta-stem cells {#sec:DSC2}
--------------------------------------
Basic first order stem cells can only produce terminal cells. A first order stem cell cannot generate any further stem cells and a regular developmental network that links to a single stem cell network will only generate one stem cell. Thus, to produce multiple stem cells we need another type of developmental network. One way multiple stem cells can be produced is by networks of the type NI$_k$ () that could produce $2^k$ identical first order stem cells. But this is still limited by number of loops $k$ in those networks. Another method is to have meta-stem cell networks that can endlessly produce 1st order stem cells. It is possible that organisms use both strategies.
Thus, there are at least two methods of producing stem cells. A developmental network may link to a stem cell network to generate one or more stem cells. Or there are meta-stem cells that can produce an unlimited supply of stem cells. A third method uses stochastic dedifferentiation of terminal or progenitor cells to their 1st order stem cell parent network. ([@Werner2011b]).
Meta-stem cells, like first order stem cells, can also have conditional activation in that being linked into another network is not sufficient for their potential activation. Their preconditions $\Phi_{2}$ must be satisfied at each pass of the loop to execute the next loop. In the case of cancer stem cells the preconditions may be permanently switched on or they may be null.
\(m) \[matrix of math nodes, row sep=3em, column sep=3em, text height=1.5ex, text depth=0.25ex\] [ & & & &\
\_[2]{} A\_[2]{} && \_[1]{} A\_[1]{} && B && C\
& \_[1]{} & & &\
]{}; (m-2-1) edge \[selfloop1, cross line\] node\[nodedescr\] [$ \alpha_{2} $]{} (m-2-1); (m-2-1) edge \[pot2,blue\](m-3-2); (m-3-2) edge \[pot2,blue\] (m-2-3); (m-2-3) edge \[selfloop1, cross line\] node\[nodedescr\] [$ \alpha_{1} $]{} (m-2-3); (m-2-3) edge \[pot2,blue\](m-3-4); (m-3-4) edge \[pot2,blue\] (m-2-5);
(m-2-1) edge (m-2-3); (m-2-3) edge (m-2-5); (m-2-5) edge (m-2-7);
3rd-Order stem cell networks NG$_3$ with $3$ loops {#sec:G3}
--------------------------------------------------
Adding another loop to a meta-stem cell network results in a meta-meta-stem cell network or third order stem cell network.
\(m) \[matrix of math nodes, row sep=3em, column sep=3em, text height=1.5ex, text depth=0.25ex\] [ & & & &\
A\_3 && A\_2 && A\_1 && D\
& & & & &\
]{}; (m-2-1) edge \[selfloop1, cross line\] node\[nodedescr\] [$ \alpha_3 $]{} (m-2-1); (m-2-1) edge \[pot2\](m-3-2); (m-3-2) edge \[pot2\] (m-2-3); (m-2-3) edge \[selfloop1, cross line\] node\[nodedescr\] [$ \alpha_2 $]{} (m-2-3); (m-2-3) edge \[pot2\](m-3-4); (m-3-4) edge \[pot2\] (m-2-5); (m-2-5) edge \[selfloop1, cross line\] node\[nodedescr\] [$ \alpha_1 $]{} (m-2-5); (m-2-5) edge \[pot2\](m-3-6); (m-3-6) edge \[pot2\] (m-2-7);
(m-2-1) edge (m-2-3); (m-2-3) edge (m-2-5);s (m-2-5) edge (m-2-7);
Meta-stem cells
---------------
Geometric networks of the above type suggest a new category of stem cells, namely, meta-stem cells (see ). A [*linear*]{} or [*first order stem cell*]{} is a cell that produces other cells (that are not stem cells) by means of a linear developmental network. [*Meta-stem cells*]{} are stem cells that produce stem cells. Thus a 2nd-order geometric network with two loops contains a meta-stem cell $A_{2}$ that produces linear stem cells of type $A_{1}$. A 3rd-order three loop geometric network contains a [*meta-meta-stem*]{} cell $A_{3}$ that produces meta-stem cells $A_{2}$ that produce stem cells $A_{1}$ that produce cells of type D. And, so on.
Whether a geometric meta-stem cell network is a cancerours network will depend on the properties of the cells and how they function in the organism. Any stem cell or meta-stem cell network can by mutations become an exponential cancer (see below). A transformation of linear stem cell network into a meta-stem cell network will make the stem cell proliferate according to below and be potentially harmful. Indeed, even the metastatic behavior of cancer stem cells will be seen to be directly related to the properties of their geometric, meta-stem cell networks ().
Geometric stem cell networks NG$_k$ with $k$ loops {#sec:Gk}
--------------------------------------------------
\(m) \[matrix of math nodes, row sep=3em, column sep=3em, text height=1.5ex, text depth=0.25ex\] [ & & & &\
A\_k && A\_[k-1]{} && A\_1 && D\
& & & & &\
]{}; (m-2-1) edge \[selfloop1, cross line\] node\[nodedescr\] [$ \alpha_k $]{} (m-2-1); (m-2-1) edge \[pot2\](m-3-2); (m-3-2) edge \[pot2\] (m-2-3); (m-2-3) edge \[selfloop1, cross line\] node\[nodedescr\] [$ \alpha_{k-1} $]{} (m-2-3); (m-2-5) edge \[pot2\](m-3-6); (m-3-6) edge \[pot2\] (m-2-7);
(m-2-5) edge \[selfloop1, cross line\] node\[nodedescr\] [$ \alpha_1 $]{} (m-2-5);
(m-2-1) edge (m-2-3); (m-2-3) edge (m-2-5);s (m-2-5) edge (m-2-7);
Mathematical properties of geometric stem cell networks {#sec:Pascal}
-------------------------------------------------------
A stem cell network of the above type with one or more loops in a linear connected sequence leads has interesting mathematical properties. It turns out that such networks are directly related to Pascal’s Triangle, binomial coefficients and geometric numbers. Under the appropriate interpretation, one loop simple produces a line. Two loops produce a triangle. Three loops produce an equilateral pyramid (tetrahedron), four loops a pentalope (a four dimensional pyramid), etc. In general, an network of $k$ linearly connected loops will produce an $k$-dimensional pyramid. Or viewed numerically, after $n$ synchronous rounds of ideal division, a cancer network with $k$ linearly connected single loops produce the sum of the first $k$ binomial coefficients at level $n$ of Pascal’s triangle.
Recall that by definition: $${n \choose k} = \frac{n!}{k!(n - k)!}$$
Then the number of cells produced by a single cell that is controlled by an stem cell network with $k$ linearly connected loops, after undergoing $n$ rounds of synchronous cell division division is given by the following formula (for $n > 0$): $${\mbox Cells}(n, k) = \sum_{i = 0}^k {n \choose i} = \sum_{i = 0}^k \frac{n!}{i!(n - i)!}
\label{eq:Gnk}%was GGformula$$
The above formula shows the direct relationship between such stem cell networks and the binomial coefficients of Pascal’s triangle, i.e., the coefficients the binomial theorem.
Given the following standard definitions:
1. Linear number: $$Lin(n) = 1 + n = 1 + {n \choose 1} \mbox{ when } n \geq 1$$
2. Triangular number: $$Tri(n) = \frac{n^2 + n}{2} = \frac{n(n + 1)}{2} = {n + 1 \choose 2} %\left( \begin{array}{c} n \\ 2 \end{array} \right)
\label{eq:TriangularNo}$$
3. Tetrahedral number: $$Tet(n) = \frac{n(n + 1)(n+2)}{6} = {n + 2 \choose 3} %\left( \begin{array}{c} n \\ 3 \end{array} \right)
\label{eq:TetrahedralNo}$$
4. Pentalope number: $$Pen(n) = \frac{n(n + 1)(n+2)(n + 3)}{24} = {n +3 \choose 4} %\left( \begin{array}{c} n \\ 4 \end{array} \right)$$
Then the number of cells, Cells$(n,k)$, that develop after $n$ rounds of synchronous division obey the sum of the coefficients of Pascal’s Triangle where $n$ is the height of the triangle and $k$ is horizontal coordinate corresponding to the number of single connected loops in the regulatory network. These sums correspond to the volumes of the corresponding $k$-dimensional geometric form in an ideal $k$-dimensional discrete space:
1. For one loop we get a $1$-dimensional structure where its length gives the number of cells: $$Cells(n, 1) = Lin(n) = 1 + n
\label{eq:G1}$$
2. Two loops add the triangular number to give the area of a $2$-dimensional triangle: $$\begin{aligned}
Cells(n, 2) & = & Lin(n) + \textcolor{red}{Tri(n-1)} \\ \STRUT
& = & 1 + n + \textcolor{red}{\frac{n(n - 1) }{2}} \\ %\\ \STRUT
& = & \sum_{i = 0}^2 {n \choose i} = {n \choose 0} + {n \choose 1} + \textcolor{red}{{n \choose 2} }
\label{eq:G2}\end{aligned}$$
3. For three loops add the tetrahedral number to give the volume of $3$-dimensional pyramid: $$\begin{aligned}
Cells(n, 3) & = & Lin(n) + Tri(n-1) + \textcolor{red}{Tet(n-2)} \\ \STRUT
& = & 1 + n + \frac{n(n - 1)}{2} + \textcolor{red}{ \frac{n(n-1)(n-2)}{6}} \\ \STRUT
& = & \sum_{i = 0}^2 {n \choose i} = {n \choose 0} + {n \choose 1} + {n \choose 2} + \textcolor{red}{{n \choose 3} }
\label{eq:G3}\end{aligned}$$
4. For four loops add the pentalope number to give the volume of a $4$-dimensional pyramid: $$\begin{aligned}
Cells(n, 4) & = & Lin(n) + Tri(n-1)+ Tet(n-2) + \textcolor{red}{Pen(n-3)} \\ \STRUT
& = & 1 + n + \frac{n(n - 1)}{2} + \frac{n(n-1)(n-2)}{6} + \textcolor{red}{\frac{n(n-1)(n-2)(n-3)}{24} }\\ \STRUT
& = & {n \choose 0} + {n \choose 1} + {n \choose 2} + {n \choose 3} + \textcolor{red}{{n \choose 4} }
\label{eq:G4}
\end{aligned}$$
5. For $k$ loops sum the sequence of numbers through ${\frac{n!}{k!(n - k)!}} $ to give the volume of a $k$-dimensional pyramid. Given $n > 0$ : $$\begin{aligned}
Cells(n, k) & = & Lin(n) + Tri(n-1)+ Tet(n-2) + \ldots + \textcolor{red}{{n \choose k}}\\ \STRUT
& = & 1 + n + \frac{n(n - 1)}{2} + \frac{n(n-1)(n-2)}{6} + \dots
+ \textcolor{red}{\frac{n!}{k!(n - k)!}} \\ \STRUT
& = & {n \choose 0} + {n \choose 1} + {n \choose 2} + \dots + \textcolor{red}{{n \choose k} }\\ \STRUT
& = & \sum_{i = 0}^k {n \choose i} = \sum_{i = 0}^k \frac{n!}{i!(n - i)!}
\label{eq:longGnk}
\end{aligned}$$
6. So, in general we have: $$\begin{aligned}
Cells(n, k) = \left\{
\begin{array}{ll}
1 & \mbox{if $n = 0$} \\
\sum_{i = 0}^k \frac{n!}{i!(n - i)!}
= \sum_{i = 0}^k {n \choose i} %\left( \begin{array}{c} n \\ i \end{array} \right)
& \mbox{otherwise}
\end{array}
\right.\end{aligned}$$ \[eq:generalGnk\]
This view shows the direct relationship between this type of cancer and stem cell network and the geometric numbers. What is fascinating is that these ideal networks have a generative competence with such interesting numerical properties. Their mathematical history spans back from the binomial theorem, to Pascal’s triangle, to the Greek’s discovery of geometric numbers.
Stem cell production in development
===================================
Stem cells are created in a developmental context. This is done by linking in a stem cell network into the normal developmental network. The locality of origin of the link in the developmental network to the stem cell network determines the properties of the cells that a stem cell network generates. For example, heart stem cells are different from liver stem cells even if one can coaxed into assuming the properties of the other.
Multi-identical linear stem cell network
----------------------------------------
A linear $k$-identical cell stem cell network $NI_{k}L = NI_{k}G_{1}$ links an identical cell network NI$_k$ that generates $k$ identical cells with a downstream linear stem cell network $NL = G_{1}$. The simplest case is where the number of identical linear stem cells is $k=2$:
\(m) \[matrix of math nodes, row sep=3em, column sep=3em, text height=1.5ex, text depth=0.25ex\] [ & & & &\
A && B && C && D\
& & & c &\
]{}; (m-2-1) edge \[inPot1, DarkGreen\]node\[nodedescr\] [$ b $]{}( m-2-3); (m-2-1) edge \[inPot2, DarkBlue\]node\[nodedescr\] [$ b $]{}( m-2-3); (m-2-3) edge \[pot2, blue\](m-3-4); (m-3-4) edge \[pot2, blue\] (m-2-5); (m-2-3) edge \[selfloop1, cross line\] node\[nodedescr\] [$ b $]{} (m-2-3);
(m-2-1) edge (m-2-3); (m-2-3) edge (m-2-5); (m-2-5) edge (m-2-7);
If $2$ is the number of identical cells controlled by the network and $n$ is the number of loops synchronously executed by those cells then in this case the number of cells after $n$ cell divisions is: $$\Cells(n) = 2 + (2 \times n)$$
The **phenotype** of such a network would have identical linearly growing stem cells in several places at once. They would have started simultaneously (up to some error range).
More generally, let NI$_k$ be a network of $k$ identical divisions that generates $2^k$ cells. Then we can generate a stem cell network NLI$_k$ if we link an upstream NI$_k$ to a simple linear stem cell network of type G1 (see ), for example as follows:
\(m) \[matrix of math nodes, row sep=3em, column sep=2em, text height=1.5ex, text depth=0.25ex\] [ & & & & & &\
A\_0 && A\_1 & & A\_2 && A\_[k-1]{} && A\_k = B && C\
& & & & & &\
]{};
(m-2-1) edge \[inPot1,DarkGreen, cross line\] node\[nodedescr\] [$\alpha_1$]{} (m-2-3) (m-2-1) edge \[inPot2, DarkBlue, cross line\] node\[nodedescr\] [$\alpha_1$]{} (m-2-3) (m-2-3) edge \[inPot1, DarkGreen , cross line\] node\[nodedescr\] [$\alpha_2$]{} (m-2-5) (m-2-3) edge \[inPot2, DarkBlue, cross line\] node\[nodedescr\] [$\alpha_2$]{} (m-2-5) (m-2-7) edge \[in100Pot1, DarkGreen, cross line\] node\[nodedescr\] [$\alpha_{k}$]{} (m-2-9) (m-2-7) edge \[in100Pot2, DarkBlue, cross line\] node\[nodedescr\] [$\alpha_{k}$]{} (m-2-9) (m-2-9) edge \[selfloop1, red, cross line\] node\[nodedescr\] [$\alpha_{k}$]{} (m-2-9) (m-2-9) edge \[inPot2, blue, cross line\] node\[nodedescr\] [$c$]{} (m-2-11); (m-2-1) edge (m-2-3); (m-2-3) edge (m-2-5); (m-2-5) edge (m-2-7); (m-2-7) edge (m-2-9); (m-2-9) edge (m-2-11);
The below is function describing the ideal rate of growth after $n$ synchronous divisions where $k$ is the number of identical daughter cell divisions nodes in the network in .
$$\begin{aligned}
\mbox{Cells}(n,k) = \left\{
\begin{array}{ll}
2^n & \mbox{if $n <= k$} \\
2^k + 2^{k}\times(n - k + 1) = 2^k\times(n -k +2) & \mbox{for $n > k$}
\end{array}
\right.\end{aligned}$$
Stem cells that generate identical cells
----------------------------------------
\(m) \[matrix of math nodes, row sep=3em, column sep=2em, text height=1.5ex, text depth=0.25ex\] [ & & & & & &\
A\_[L]{} && A\_0 & & A\_1 && A\_[2]{} && A\_[k-1]{} && A\_[k]{}=B\
& & & & & &\
]{};
(m-2-3) edge \[inPot1, DarkGreen , cross line\] node\[nodedescr\] [$\alpha_1$]{} (m-2-5) (m-2-3) edge \[inPot2, DarkBlue, cross line\] node\[nodedescr\] [$\alpha_1$]{} (m-2-5) (m-2-5) edge \[inPot1,DarkGreen, cross line\] node\[nodedescr\] [$\alpha_2$]{} (m-2-7) (m-2-5) edge \[inPot2, DarkBlue, cross line\] node\[nodedescr\] [$\alpha_2$]{} (m-2-7) (m-2-9) edge \[in100Pot1, DarkGreen, cross line\] node\[nodedescr\] [$\alpha_{k}$]{} (m-2-11) (m-2-9) edge \[in100Pot2, DarkBlue, cross line\] node\[nodedescr\] [$\alpha_{k}$]{} (m-2-11) (m-2-1) edge \[selfloop1, red, cross line\] node\[nodedescr\] [$\alpha_{L}$]{} (m-2-1) (m-2-1) edge \[inPot2, blue, cross line\] node\[nodedescr\] [$\alpha_{0}$]{} (m-2-3); (m-2-1) edge (m-2-3); (m-2-3) edge (m-2-5); (m-2-5) edge (m-2-7); (m-2-7) edge (m-2-9); (m-2-9) edge (m-2-11);
Stem cell networks commute with identical cell networks
-------------------------------------------------------
It turns out that in terms of ultimate cell production the network in is equivalent to the network in . This is because the simple network NG$_{1}$ commutes with NI networks. More generally, given NI$_{k}$ then any element NG$_{1}$ in NG$_{1}$ commute with the elements of NI$_{k}$ to generate the same ultimate cell proliferation even while the multicellular form may vary.
Stem cell networks and cancer stem cells {#sec:NSC}
========================================
Normal networks that come closest to cancer networks are stem cell networks. What distinguishes normal from cancer stem cell networks is not primarily their network architecture but rather when and where a stem cell network is linked and activated, and what further networks the stem cell network activates. A cancer stem cell network is either formed by mutation in the wrong place in the global developmental network, it is activated inappropriately in the multicellular system, or it activates inappropriate, nonfunctional networks with abnormal and possibly pathological cellular or multicellular phenotypes relative to the overall multicellular system.
From our theory it follows that there are at least two main kinds of stem cell networks, linear and geometric. However, linear stem cell networks are just a special case of geometric stem cell networks, namely, 1st order geometric stem cell networks. These are also fundamental cancer stem cell networks. A normal linear stem cell produces a cell of a particular type that is either terminal itself or activates a terminal network. A second order geometric stem cell network is a meta-stem cell network that produces linear stem cells. Thus, linear stem cells produce no further stem cells. Meta stem cells do produce stem cells.
Since a linear stem cell cannot produce further stem cells there must exist meta-stem cells that produce more than one stem cell. These in turn are produced by yet further upstream embryonic networks. The original embryonic fertilized egg is controlled by the global embryonic network that is mostly terminal with the exception of its stem cells. It consists of many subnetworks that may be multiply employed.
A linear cancer stem cell produces no additional cancer cells. Interestingly, we will see that the order of a cancer meta-stem cell network controlling a cancer cell determines the possible metastases producible by that cancer cell (see ).
A further possibility is that there exist normal [*exponential stem cell networks*]{} that are activated by cell signaling, in effect a communication network linked with a cytogenic control network to produce cells quickly on demand. However, such networks are be dangerous, leading to a proliferative explosion if something goes wrong with the communication network.
Normal stem cells and cancer stem cells
---------------------------------------
Stem cells have the function of continually producing new cells in tissue that is meant to regenerate. Stem cells have functional developmental control networks with preexisting loops. This makes the progression to cancer a shorter route, because the linear, proliferating network already exists. We just need to mutate in one more loop to makes it exponential and then with additional mutations it becomes invasive. Any slow growing cancers that suddenly become very fast growing may be an instance of this process.
If a stem cell proliferates conditionally, i.e., based on some contextual cellular or environmental condition, then a stem cell can transform into a cancer stem cell by a situation or process that constantly activates the antecedent condition necessary for the stem cell to divide. This can happen if the activating signal is constant or the receptor system for that signal is mutated to be constantly on. It may be the case that a stem cell may already be potentially, linearly, geometrically or exponentially, proliferative, but its antecedent conditions may not be satisfied. Hence, if the conditional system remains functional the potential stem cell proliferation whether it be linear or exponential will not show itself. However, when the antecedent conditions for proliferation are fulfilled or the antecedent testing system is mutated to a constant on-state then the stem cell cancer will flower. An example of this would be a stem cell whose proliferation depends on some signal such as a hormone that, together with a signal transduction pathway, activates the stem cell proliferative loop. Then if the signal is constantly turned on or the signal transduction mechanism is in a constant “signal received” state, then the stem cell will proliferate either linearly, geometrically or exponentially depending on the structure of its controlling network.
A further danger with some types of stem cells is that they may already be functionally, tissue invasive in order to move to the appropriate site, as, for example, in wound healing. This invasive property would then make a stem cell cancer even more dangerous.
Stem cells by themselves are not cancerous, but they have the properties of conditional, possibly invasive, linear, geometric or exponential cancers. The boundary between a cancer and normal cell may not always be clear since in the wrong context a normal stem cell may be cancerous.
Distinguishing linear cancer cells from stem cells based purely on the network architecture is difficult if not impossible. Stem cells generate normal cell that are functional in the context of the organism. Any stem cell that generates such “normal” cells inappropriately is a cancer cell. Thus, the dividing line between linear cancers and stem cells may be indistinct. Thus, a normal stem cell network becomes cancerous if it is activated inappropriately with respect to the overall functioning of a biological system in the organism, or if the cell differentiation networks the stem network activates generate abnormal cells (e.g., invasive cells, or having abnormal phenotype) or cells that generate structures inappropriately in the overall system.
Cancer susceptibility of stem cells
-----------------------------------
We see why stem cells are especially susceptible to develop into cancer networks. First, all that has to happen is that the first order or second order conditional antecedents are mutated to be always true, then the stem cell networks, once activated produces cells indefinitely independent of outside signaling. Second, any of the links from the other non-proliferative daughter cell differentiation site could loop back to a first or second order loop activation point to produce an exponential cancer.
Having two or more proliferative loops does not necessarily result in an exponential cancer. However, they can transform into exponential cancers. For example, using the WHO-grading scheme, astrocytoma grad I, which is slow growing and benign, appears to be a linear cancer of type NG, astrocytoma grade II -> grade III. Grade III is relatively fast growing compared to grade II, but can be present for several years and then suddenly change to grade IV Glioblastoma Multiforme (GBM). GBM are fast growing and spread quickly. .
Hierarchy of metastases from geometric cancer networks
======================================================
One of the most important consequences of the hierarchical nature of stem cell networks is that they imply the existence of a corresponding metastatic hierarchy. Geometric cancer stem cell networks form a hierarchy of interlinked stem cell networks such that an k-th order stem cell network generates a cell controlled by an k-1 order stem cell network. Thus, 3rd order or meta-meta-stem cell networks produce cells controlled by 2nd order, meta-stem cells. A 2nd order or meta-stem cell generates a cell controlled by a 1st order, linear stem cell network. At the base of the hierarchy is a terminal network controlling a cell generated by a 1st order, linear stem cell.
This hierarchy of control in higher order stem cell networks has direct consequences for the types of metastases that can, in principle, be generated by such higher order geometric networks. In other words. the metastatic potential of a cancer stem cell is determined by the order of its geometric control network. As we will see, this the dynamic phenotypic properties of the metastases are a result of their place in the metastatic hierarchy. This can be used for diagnosis and reverse inference as to the type of stem cell network controlling the metastatic tumor.
Note, however, that there may also be functional uses of higher order stem cell networks for the production and distribution of stem cells in healthy multicellular systems. Meta stem cells and their stem cell network hierarchy could be used by organisms to generate and deliver lower level stem cells to various parts of the body.
\[sec:GMetastases\]
A 2nd order geometric cancer cell generates 1st order linear cancer cells (see ). If these 1st order linear cancer cells metastasize then we have a phenotype of multiple linear cancer growths in different regions of the body. For example, one might have multiple slow growing cysts in different regions of the body. In this case, these cysts are metastases that were generated from a single meta-stem cell that generates progenitor cells that generate cysts. Other examples, might include slowly progressing cancers that move in the lymphatic system. Here again in may, but need not, be the case that a single meta-cancer stem cell is generating sub-cancers that have the overt phenotype of a uniform cancer progressing in some direction across a region or through the whole body.
Secondary and tertiary metastases generated by 3rd order cancer networks
------------------------------------------------------------------------
In general, k-th order geometric cancer networks $G_{k}$ (see ) generated cells controlled by $k-1$ order geometric cancer networks $G_{k-1}$. These $G_{k-1}$ networks, in turn, generate cells controlled by $k-2$ order cancer networks $G_{k-2}$, etc. Therefore, third-order geometric cancer networks $G3$ () , such as the above simulated network in , generate cells controlled by second-order cancer networks $G2$, (). In turn, $G2$-networks generate cells controlled by 1st-order geometric networks $G1$, (). $G1$-networks are linear cancer networks which generate terminal cells $G0$ that do not proliferate.
Relating metastatic phenotype with geometric cancer networks
------------------------------------------------------------
Since any of the cells generated by a cancer network have the potential to metastasize by moving to other regions of the organism, each primary tumor controlled by a cancer network will have a distinct metastatic phenotype. Therefore, we can use the phenotype of the metastases to draw inferences about the nature and architecture of the network generating that phenotype. Moreover, if we find the metastases that together form related patterns as described for the geometric metastases above, it warrants looking for the primary tumor generating the secondary and tertiary metastases.
Treatment options for metastases generated by geometric cancer networks
-----------------------------------------------------------------------
Treatment would focus on eliminating or transforming cells with the higher level control networks generating the less dangerous lower level networks. Treatment can be by cell death or by network transformations that block or modify the cancer network. Different transformations are necessary for different cancer networks. Once metastases have been created by a higher level network, it is no longer sufficient to block or transform the higher level initiating subnetwork. Instead, each of the cells with lower level cancer subnetworks also have to be transformed or destroyed. If they are 1st order linear networks, they may be relatively harmless since they only produce cells that no longer proliferate. However, this will depend on their location and effect on the other multicellular contexts in which they reside. Because of cell interactions such as cell signaling, stochastic reactivation of higher level cancer networks such cells, depending on their differentiation state and the cellular context, may still be dangerous.
Stochastic reactivation of terminal cells in geometric networks
---------------------------------------------------------------
If terminal cells can reactivate a proliferative control state in a network, then the more numerous the number of terminal cells the more likely it is that one or more of them will become an active cancer cell. Hence, given the possibility of stochastic activation of ancestral cancer networks, formerly passive metastatic cells can become active tumors.
Stochastic differentiation in stem cell networks {#sec:NSSC}
================================================
It has been argued that the traditional theory of linear stem cells is wrong [@Jones2007]. On the traditional theory of stem cells, stem cells are immortal and constant in number. However, experimental data appears to show that for some stem cells appear to generate further stem cells. On a deterministic theory stem cells that produce stem cells are what we have termed meta-stem cells. However, another model would have stem cells divide and differentiate stochastically either exponentially into two stem cells, linearly into a stem cell and a terminal cell, or into two terminal cells. The evidence, suggests that epidermal cells divide stochastically according to the distribution 8% double stem cells, 84% one stem cell and one terminal cell, and 8% two terminal cells. The biological mechanism is not known.
Historical background: A network that generates Till’s stochastic stem cell model
---------------------------------------------------------------------------------
One of the first models of stem cells was the stochastic model of hemopoietic cell proliferation \[Till et.al. [@Till1964]\]. Till distinguished what he called “colony forming cells” that have the capacity to form colonies from differentiated cells without that capacity. When a colony develops from a single cell only a small number of these cells have “colony-forming capacity”. The rest are differentiated cells without this capacity. This he considered a “birth-and-death” process, the generation of colony cells being a birth process and the generation of differentiation a death process. According to the model, a given cell with colony-forming capacity can either divide into two colony forming cells (with probability $p_{2}$) or differentiate into a terminal cell (with probability $p_{0} = 1 - p_{2}$) having no colony forming capacity. The case of generating mixtures of one colony-forming cells and one differentiated cell was considered a birth process followed by death and was handled by adjusting the probabilities.
Production numbers of blood cells is relatively constant under normal conditions. Under stress or increased demand there is a rapid increase in production of blood cells. The properties of hemopoiesis (blood cell creation) implies that the production of differentiated cells is precisely controlled. Hence, Till argues, there must be control mechanisms. Till then considers if cell proliferation (the numbers of cells produced) is also under precise control or lax control. He argues for lax control and that the data suggests cell proliferation is stochastic. While a stem cell network was not given by Till, the behavior of the Till model of colony forming cells can be generated by following stochastic stem cell network:
\(m) \[matrix of math nodes, row sep=3em, column sep=3em, text height=1.5ex, text depth=0.25ex\] [ && &&\
S && J1 && T &&\
&& && &&\
]{}; (m-2-1) edge \[inPot1, red, cross line\] node\[nodedescr\] [$j_1$]{} (m-2-3) (m-2-1) edge \[inPot2, in=-90, blue, cross line\] node\[nodedescr\] [$j_{2}$]{} (m-2-3); (m-2-3) edge \[inPot1, out=90, in=90,distance=5cm, green, cross line\] node\[stochasticNodestyle\] [$p_{2}$]{} (m-2-1) (m-2-3) edge \[in2Pot1,in=120,distance=3cm,green, cross line\] node\[stochasticNodestyle\] [$p_{0} = 1 -p_{2}$]{} (m-2-5); (m-2-1) edge (m-2-3); (m-2-3) edge (m-2-5); (m-2-5) edge (m-2-7);
Under this network () a stem cell divides into identical daughter cells J1. Each J1 cell can with probability $p_{2}$ dedifferentiate into a stem cell S or differentiate with probability $p_{0} = 1-p_{2}$ into a terminal cell T. Hence, unlike Till’s original model, this network explicitly handles all the possible outcomes of stem cell division. Upon stem cell division, the network allows the generation of two stem cells (S,S), as well as the mixed case (S,T) of one stem cell and one differentiated cell, and the case (T,T) of two differentiated cells.
Till’s model has been criticized because it predicts that stem cells eventually differentiate into terminal cells and, thus, no longer self renew. However, this depends on the probability distribution. If $p$ is high then exponential proliferation outruns terminal cell differentiation.
A central problem with this network is that the key to its behavior is dependent on the value of the probability $p_{2}$. On the one hand, if the probability $p_{2}$ of stem cell self-renewal is high, it tends toward exponential growth of stem cells. This leads to too many stem cells versus progenitor and terminal, specialized cells. On the other hand, if the self-renewal probability $p_{2}$ is low, it eventually leads to the elimination of all stem cells. In the latter case, this network does not represent a stem cell with unlimited proliferative potential. In this network, no dedifferentiation is possible for terminal cells. Hence, the greater the probability $p_{0} = 1-p_{2}$ that the terminal network T is activated, the more likely is the permanent quiescence of the parent stem cell.
If the probability $p_{2}=1$ we have exponential growth of stem cells. If $p_{2}=0$ we have differentiation to the terminal cells T. Hence, this network cannot model the developmental dynamics of deterministic linear networks of the type fig:G1 . However, with the right choice of the probability distribution, it can approximate a relatively constant production of stem cells and terminal cells.
In their Monte Carlo simulation the probability used was $p_{2} = 0.6$ and $p_{0}=0.4$. In our Monte Carlo simulation, using their probability distribution, the network proliferation dynamics tends toward a 21% proportion of stem cells after 52 generations. This appears to be much higher than the empirical data for hemopoietic stem cells. However, these percentages depend on how many cells are generated by the terminal cell T. If T is not a terminal cell and instead a progenitor cell controlled by a terminal network T\*, and is instead a terminal cell state that does not divide further (as was the case in Till’s original model), then the stem cell percentage is lower being 12% after 20 generations of cell division. We add progenitor cell capacity to T in the network in () below.
Note, in this sort of modeling approach, precise differentiation networks (architectures) are coupled with stochastic activation networks. The topology of the network architecture sets the boundaries for what stochastic paths are possible at all.
The Till stochastic network extended to generate progenitor cells
-----------------------------------------------------------------
It is very simple to extend Tills original model of hemopoietic (blood) stem cells (S) \[Till et.al. [@Till1964]\] to include progenitor networks that control the proliferative capacity of stem cell progeny. We simply link the terminal cell state to a bounded network T\* that in the case below is simply a network that results in one further cell division to produce two distinct daughter cells T1 and T2. But, T\* could be any bounded network. Thus, in principle, given the right network T\*, such progenitor cells could develop into arbitrarily complex structures.
\(m) \[matrix of math nodes, row sep=3em, column sep=3em, text height=1.5ex, text depth=0.25ex\] [ && && &t\_[1]{} &\
S && J1 && T\* && T1 && T2 &&\
&& && && t\_[2]{} &&\
]{}; (m-2-1) edge \[inPot1, red, cross line\] node\[nodedescr\] [$j_1$]{} (m-2-3) (m-2-1) edge \[inPot2, in=-90, blue, cross line\] node\[nodedescr\] [$j_{2}$]{} (m-2-3); (m-2-3) edge \[inPot1, out=90, in=90,distance=5cm, green, cross line\] node\[stochasticNodestyle\] [$p_{2}$]{} (m-2-1) (m-2-3) edge \[in2Pot1,in=120,distance=3cm,green, cross line\] node\[stochasticNodestyle\] [$p_{0} = 1 -p_{2}$]{} (m-2-5); (m-2-5) edge \[pot2,blue\](m-3-7); (m-3-7) edge \[pot2,blue\] (m-2-9); (m-2-5) edge \[pot2,red\](m-1-6); (m-1-6) edge \[pot2,red\] (m-2-7);
(m-2-1) edge (m-2-3); (m-2-3) edge (m-2-5); (m-2-5) edge (m-2-7); (m-2-7) edge (m-2-9);
As with the previous network , the proliferative potential of stem cells under this network is dependent on the probability $p_{2}$. Since no dedifferentiation is possible in this network for terminal cells T1 and T2. Furthermore, it is not an inherent property of the network topology that stem cells self renew. However, if the dedifferentiation probability $p_{2}$ of self renewal is high enough then stem cells do self renew and may even dominate the colonies they generate.
If the probability $p_{2}=1$ we have exponential growth of stem cells, with no progenitor cells T\*. If $p_{2}=0$ we have differentiation to to the terminal progenitor network T\* which in this case ends in just two terminal cells T1 and T2. Hence, this network cannot model the developmental dynamics of deterministic linear networks of the type G1 . However, with the right choice of the probability distribution, it can approximate that a relatively constant ratio of stem cells, progenitor and terminal cells.
It follows that under the Till model of stem cells we do not need dedifferentiation of terminal cells to maintain stem cell capacity. All we need is for the 1st order stem cells to be physically loose, mixing in with the terminal cells in the tumor as it grows. In this case, we would see proliferating stem cells within the tumor in the context of what appear to be only terminal cells. It may appear as if terminal cells have spontaneously dedifferentiated into stem cells, but this need not be the case since it may be the result of stem cell mixing.
What could also be happening is that we have a 2nd order geometric stem cell network that is generating 1st order stem cells which due to developmental control in conjunction with physics leads to a mixture of 1st order stem cells with terminal cells. Since the control state of a cell my be hidden by the overt differentiation state of the cell, the 1st order stem cells may be indistinguishable relative to a set of markers from terminal cells. Stochastically or under particular conditions these 1st order stem cells may then begin to proliferate generating terminal or progenitor cells.
A flexible stochastic network architecture with exponential and linear potential
--------------------------------------------------------------------------------
\(m) \[matrix of math nodes, row sep=2em, column sep=3em, text height=1.5ex, text depth=0.25ex\] [ && && && & & &\
S && J1 && J2 && T && &&\
&& && && && & &\
]{}; (m-2-1) edge \[inPot1, red, cross line\] node\[nodedescr\] [$j_1$]{} (m-2-3) (m-2-1) edge \[inPot2, in=-90, blue, cross line\] node\[nodedescr\] [$j_2$]{} (m-2-5); (m-2-3) edge \[inPot1, out=90, in=90,distance=4cm, green, cross line\] node\[stochasticNodestyle\] [$p$]{} (m-2-1) (m-2-3) edge \[in2Pot1,green, cross line\] node\[stochasticNodestyle\] [$1-p$]{} (m-2-7) (m-2-5) edge \[inPot2, out=-125, in= -90, distance=6cm, green, cross line\] node\[stochasticNodestyle\] [$q$]{} (m-2-1) (m-2-5) edge \[inPot2, out=-60, in= -90, distance=3cm, green, cross line\] node\[stochasticNodestyle\] [$1-q$]{} (m-2-7); (m-2-1) edge (m-2-3); (m-2-3) edge (m-2-5); (m-2-5) edge (m-2-7); (m-2-7) edge (m-2-9);
In the network in is very flexible. Depending on the probability distribution, the network can range between being exponential, linear, or terminal, as well as every mixture in between. Furthermore, the network can be deterministic, stochastic or mixture of both. This flexibility is partly the result of separating out the probability distributions for the behaviors of the two daughter cells. It shows that the architecture of the network imposes constraints on what kinds of developmental dynamics are in principle possible. The probability distribution presupposes a network architecture of possible developmental paths.
The cell S is only a stem cell stochastically and not intrinsically when $p < 1$ and $q < 1$. When $p = q = 1$ the network is deterministic exponential. When $p = 1, q = 0$ or $p = 0, q = 1$ the network is deterministic linear, i.e., a deterministic 1st order geometric stem cell network. If $p = q = 0$ the network is terminal. When $p = 1$ and $q < 1$, or when $q=1$ and $p<1$, then the network is mixed deterministic linear with stochastic exponential tendencies.
If probabilities $p = q$ then as $p$ and $q$ approach $1$ the network approaches the behavior of a deterministic exponential network. However, if $p$ and $q$ are different then this network can simulate a linear stochastic network as well when, for example, $p$ approaches $1$ and $q$ approaches $0$, or vice versa. As the probabilities $p$ and $q$ decrease, the more frequently the cancer stem cell results in a terminal tumor that does not develop further because it consists only of cells of terminal type T. This shows that stochastic cancer stem cell networks can in some cases go into spontaneous remission. While this network can also exhibit exponential growth even in a stochastically linear probability distribution, because of the two backward loops, there is a diminishing probability that it remains exponential. Thus, whether this network results in linear or exponential proliferation depends on the probability distribution.
For the network XLSSC, if the probabilities $p = 1- q$ then the higher the probability of $p$ the more the network approximates a deterministic linear developmental network. Since, in this case the distribution is anti-symmetric, the cell population partition of cell types consists of an equal number of exponential stem cells and terminal cells, with the majority of cells being linear stem cells. This corresponds to the observed distribution in epidermal basal stem cells. If, on the other hand we have a symmetric distribution where $p = q$ then the higher the probability of $p$ the more the network approximates a deterministic exponential network. Thus the type of cancer network we have depends on the probability distributions over the connecting stochastic links.
The network XLSSC () has some similarity to the Till model (TillSSC ), when $p=q$ (and value of probability $p=p_{2}$ as in the original model by Till [@Till1964] model) and T. The Till model forces that daughter cells of S are the same. In contrast, this model is more flexible in that, depending on the probability distribution ($p$, $q$), it can model both exponential and linear dynamics of multicellular development.
Linear stochastic stem cell network LSSC
----------------------------------------
\(m) \[matrix of math nodes, row sep=3em, column sep=3em, text height=1.5ex, text depth=0.25ex\] [ && && &&\
L && J1 && T && && && && &&\
]{}; (m-2-1) edge \[inPot1, red, cross line\] node\[nodedescr\] [$j_1$]{} (m-2-3) (m-2-1) edge \[inPot2, in=-90, blue, cross line\] node\[nodedescr\] [$b$]{} (m-2-5); (m-2-3) edge \[inPot1, out=90, in=90,distance=4cm, green, cross line\] node\[stochasticNodestyle\] [$p$]{} (m-2-1) (m-2-3) edge \[in2Pot1,in=120,distance=3cm,green, cross line\] node\[stochasticNodestyle\] [$1-p$]{} (m-2-5); (m-2-1) edge (m-2-3); (m-2-3) edge (m-2-5); (m-2-5) edge (m-2-7);
\(m) \[matrix of math nodes, row sep=3em, column sep=3em, text height=1.5ex, text depth=0.25ex\] [ && && &&\
L && J1 && T && && && && &&\
]{}; (m-2-1) edge \[inPot1, in=140, red, cross line\] node\[nodedescr\] [$j_1$]{} (m-2-3) (m-2-1) edge \[inPot2, in=-90, blue, cross line\] node\[nodedescr\] [$t$]{} (m-2-5); (m-2-3) edge \[inPot2, out=-90, in=-90,distance=4.5cm, green, cross line\] node\[stochasticNodestyle\] [$p$]{} (m-2-1) (m-2-3) edge \[in2Pot1,in=110,out=40,distance=4.5cm,green, cross line\] node\[stochasticNodestyle\] [$1-p$]{} (m-2-3); (m-2-1) edge (m-2-3); (m-2-3) edge (m-2-5); (m-2-5) edge (m-2-7);
1st-Order Geometric Cancer Networks with open stochastic dedifferentiation {#sec:G1SD}
--------------------------------------------------------------------------
\(m) \[matrix of math nodes, row sep=3em, column sep=3em, text height=1.5ex, text depth=0.25ex\] [ & & & &\
S && J && T &&\
& & & &\
]{}; (m-2-1) edge \[selfloop1, cross line\] node\[nodedescr\] [$ a $]{} (m-2-1) (m-2-1) edge \[inPot2, in=-110, blue, cross line\] node\[nodedescr\] [$j$]{} (m-2-3); (m-2-3) edge \[inPot2, out=-80, in=-90,distance=4cm, green, cross line\] node\[stochasticNodestyle\] [$p$]{} (m-2-1) (m-2-3) edge \[in2Pot1,in=90,out=80,distance=4cm,green, cross line\] node\[stochasticNodestyle\] [$1-p$]{} (m-2-5); (m-2-1) edge (m-2-3) (m-2-3) edge (m-2-5);
1st-Order Geometric Cancer Networks with closed stochastic dedifferentiation
-----------------------------------------------------------------------------
In the above network dedifferentiation can occur once prior to permanent differentiation in a terminal cell state. The effect of the stochasticity can only be seen in the tumor as a whole that is by the cancer stem cell. In the following network dedifferentiation is a constant possibility for the life of the cell.
\[sec:G1SDc\]
\(m) \[matrix of math nodes, row sep=3em, column sep=3em, text height=1.5ex, text depth=0.25ex\] [ & & & &\
S && J && T &&\
& & & &\
]{}; (m-2-1) edge \[selfloop1, cross line\] node\[nodedescr\] [$ a $]{} (m-2-1) (m-2-1) edge \[inPot2, in=-110, blue, cross line\] node\[nodedescr\] [$j$]{} (m-2-3); (m-2-3) edge \[inPot2, out=-80, in=-90,distance=4cm, green, cross line\] node\[stochasticNodestyle\] [$p$]{} (m-2-1) (m-2-3) edge \[in2Pot1,in=110,out=40,distance=4.5cm,green, cross line\] node\[stochasticNodestyle\] [$1-p$]{} (m-2-3); (m-2-1) edge (m-2-3) (m-2-3) edge (m-2-5);
In this linear network , a cell of type $S$ is a 1st-order stem cell. It produces progenitor cells of type J that have a continual stochastic dedifferentiation potential. A cell J can either dedifferentiate to its parent state S or maintains its present state J. In this closed stochastic network, a cell J continues to have a dedifferentiation potential for the life of that cell. There is no path to permanent terminal differentiation. This contrasts with the network where the cell only has a limited time frame to dedifferentiate prior to its permanent differentiation into a terminal cell T. Thus, all in a population of cells generated by S have the potential to dedifferentiate. In contrast, in the network if the probability $p$ is low, most cells in the population generated by S will be terminal cells with no dedifferentiation potential.
Closed exponential stem cell network with stochastic delays
-----------------------------------------------------------
Depending on the probability distribution, the next stochastic network proliferative properties can vary from deterministic terminal, linear, or exponential to stochastic terminal, linear or exponential.
\(m) \[matrix of math nodes, row sep=3em, column sep=3em, text height=1.5ex, text depth=0.25ex\] [ && && && & & &\
S && J1 && J2 && T && &&\
&& && && && & &\
]{}; (m-2-1) edge \[inPot1, red, cross line\] node\[nodedescr\] [$j_1$]{} (m-2-3) (m-2-1) edge \[inPot2, in=-125, blue, cross line\] node\[nodedescr\] [$j_2$]{} (m-2-5);
(m-2-3) edge \[inPot1, out=90, in=90,distance=4cm, green, cross line\] node\[stochasticNodestyle\] [$p$]{} (m-2-1) (m-2-3) edge \[in2Pot1,in=130,out=60,distance=3cm,green, cross line\] node\[stochasticNodestyle\] [$1-p$]{} (m-2-5) (m-2-5) edge \[inPot2, out=-90, in= -90, distance=6cm, red, cross line\] node\[stochasticNodestyle\] [$q$]{} (m-2-1) (m-2-5) edge \[inPot1, out=45, in= 110, distance=5cm, red, cross line\] node\[stochasticNodestyle\] [$1-q$]{} (m-2-5); (m-2-1) edge (m-2-3); (m-2-3) edge (m-2-5); (m-2-5) edge (m-2-7); (m-2-7) edge (m-2-9);
Unlike LSSC , if the probability $q > 0$ in cell state J2 then there are no true, probability independent, terminal cells in DXSSC . True terminal cells have no possible stochastically available paths that lead to a reversal of differentiation to a more dedifferentiated cell type. Instead, all cell states are stochastic with possible developmental paths that lead to the original founder stem cell S. The cell in state J2 has a probability $q$ of dedifferentiating to S. The higher the probabilities $q$ and $p$ the more this network mirrors the behavior of a deterministic exponential network. If either $p$ or $q$ is low while the other high it is more similar to a linear developmental network. Thus the probability distribution determines the similarity to linear or exponential networks. Even if the probability of $q$ is very low, as the number of J2 cell increase it becomes more and more likely that one of them will dedifferentiate to S. Thus, the cancer pulls away from linearity with an increasing cell population size.
If probabilities $q = p = 0$ then the network is deterministic terminal network leading to only one cell division. If $q = 0$ and $p = 1$ we have a deterministic linear stem cell network. However, if the probability $q = 0$ and $p > 0$, then the network is a stochastic linear network where the lower the probability $p$ the greater the chance that proliferation terminates. If either $p = 1$ or $p = 0$, and if $q = 1$ then the network is a deterministic exponential stem cell network. If $q > 0$ then no matter how small the nonzero probability $q$ is, proliferation increases with time as more and more cells are created.
[**Remark:**]{} There is another way to escape the maze of exponential proliferation. Cell proliferation can be influenced by the number of times the self-loop at J2 is allowed to repeat. If it is it is only allowed to repeat only a limited number of times relative then become quiescent, then, depending on the probability $q$, it could result in a terminal differentiation state after some stochastic tries. In other words, if, in parallel with the repetitions of the self-entry loop at J2, there exists a separate simultaneous, parallel process of terminal differentiation by means of some counting mechanism (where the cell at J2 becomes quiescent after some finite number of counting steps), then this network could lead to the ultimate termination of cell proliferation or at most linear proliferation. The counting mechanism could be dependent on the number of loops executed at J2 or it could be a function of some other temporal variable.
A 1st order geometric/exponential stochastic stem cell network with stochastic progenitor cell dedifferentiation
----------------------------------------------------------------------------------------------------------------
\(m) \[matrix of math nodes, row sep=3em, column sep=, text height=1.5ex, text depth=0.25ex\] [ && && && & & &\
G && J1 && J2 && B && J3 && J4 && T\^[\*]{} &&\
&& && && && && &\
]{}; (m-2-3) edge \[inPot1, out=90, in=90,distance=4cm, green, cross line\] node\[stochasticNodestyle\] [$p$]{} (m-2-1) (m-2-3) edge \[in2Pot1,out=60,in=110,distance=3.5cm,green, cross line\] node\[stochasticNodestyle\] [$1-p$]{} (m-2-7) (m-2-5) edge \[in2Pot1,out=-60,in=-110,distance=2.5cm,black, cross line\] node\[stochasticNodestyle\] [$q$]{} (m-2-7) (m-2-5) edge \[in2Pot2,in=-90,out=-90, distance=3.5cm,black, cross line\] node\[stochasticNodestyle\] [$1-q$]{} (m-2-1) (m-2-9) edge \[inPot1, out=60, in= 110, distance=3cm, green, cross line\] node\[stochasticNodestyle\] [$s$]{} (m-2-11) (m-2-9) edge \[inPot1, out=90, in= 80, distance=4cm, green, cross line\] node\[stochasticNodestyle\] [$1-s$]{} (m-2-13) (m-2-11) edge \[inPot2, out=-90, in= -120, distance=6cm, red\] node\[stochasticNodestyle\] [$1-r$]{} (m-2-1) (m-2-11) edge \[inPot2, out=-60, in= -70, distance=3cm, red, cross line\] node\[stochasticNodestyle\] [$r$]{} (m-2-13); (m-2-1) edge \[inPot1, red, cross line\] node\[nodedescr\] [$j_1$]{} (m-2-3) (m-2-1) edge \[inPot2, out=-60, in=-120,distance=2.5cm, blue\] node\[nodedescr\] [$j_{2}$]{} (m-2-5)(m-2-7) edge \[inPot1, red, cross line\] node\[nodedescr\] [$j_{3}$]{} (m-2-9) (m-2-7) edge \[inPot2, in=-120, blue\] node\[nodedescr\] [$j_{4}$]{} (m-2-11); (m-2-1) edge (m-2-3); (m-2-3) edge (m-2-5); (m-2-5) edge (m-2-7); (m-2-7) edge (m-2-9); (m-2-9) edge (m-2-11); (m-2-11) edge (m-2-13);
This stochastic stem cell network () generates progenitor cells B that generate cells J2 that can dedifferentiate to stem cells G or differentiate to a terminal network $T^{*}$. Depending on the probability distribution, this network can exhibit terminal, linear-geometric, and exponential developmental dynamics. If the probability distribution is such that $p=q=r=s=1$ then this is equivalent to a deterministic 1st order geometric stem cell network. A 1st-order stem cell (G) stochastically divides to produce either two progenitor cell B, or or two stem cells G or one progenitor cell B and one stem cell G. and cell J1. The cells in states J1 or J2 stochastically loop back to activate the stem cell G or to activate the progenitor cell $B$. The progenitor cell $B$ divides into a semi-terminal cell $T^{*}$ by way of J3 or in divides into two cells of type J4. J4 either terminates with $T^{*}$ or it dedifferentiates into a stem cell G.
Its behavior approaches a deterministic 1st-order geometric stem cell network as the probabilities $p$ and $q$ approach $1$. However, the subnetwork activated by this 1st-order geometric network can still stochastically dedifferentiate into one or two new stem cells G. Cancers controlled by such a network can go into spontaneous remission because of the fact that for all points in all possible paths in the network there is the possibility of reaching a terminal cell state. However, because there is the possibility at J1, J2 and J4 to dedifferentiate to earlier, upstream stem cell states, the relative numbers of stem cells to terminal cells will be higher than in networks that have fewer dedifferentiation pathways. Note, too that the dedifferentiation pathways introduce several possibilities for exponential growth. Even if dedifferentiation probabilities are very small, any increase in these probabilities could have a significant destabilizing influence on the resulting tumor.
A broad spectrum Linear or 2nd order geometric or exponential stochastic stem cell network
------------------------------------------------------------------------------------------
\(m) \[matrix of math nodes, row sep=3em, column sep=, text height=1.5ex, text depth=0.25ex\] [ && && && & & &\
G && J1 && J2 && B && J3 && J4 && T &&\
&& && && && && &\
]{}; (m-2-3) edge \[inPot1, out=90, in=90,distance=5cm, green, cross line\] node\[stochasticNodestyle\] [$p$]{} (m-2-1) (m-2-3) edge \[in2Pot1,out=60,in=110,distance=3.5cm,green, cross line\] node\[stochasticNodestyle\] [$1-p$]{} (m-2-7) (m-2-5) edge \[in2Pot1,out=-60,in=-110,distance=3.5cm,black, cross line\] node\[stochasticNodestyle\] [$q$]{} (m-2-7) (m-2-5) edge \[in2Pot2,in=-90,out=-90, distance=4cm,black, cross line\] node\[stochasticNodestyle\] [$1-q$]{} (m-2-1) (m-2-9) edge \[inPot1, out=90, in= 75, distance=5cm, green, cross line\] node\[stochasticNodestyle\] [$s$]{} (m-2-7) (m-2-9) edge \[inPot1, out=60, in= 120, distance=3cm, green, cross line\] node\[stochasticNodestyle\] [$1-s$]{} (m-2-13) (m-2-11) edge \[inPot2, out=-90, in= -90, distance=4cm, red\] node\[stochasticNodestyle\] [$1-r$]{} (m-2-7) (m-2-11) edge \[inPot2, out=-60, in= -90, distance=3cm, red, cross line\] node\[stochasticNodestyle\] [$r$]{} (m-2-13); (m-2-1) edge \[inPot1, red, cross line\] node\[nodedescr\] [$j_1$]{} (m-2-3) (m-2-1) edge \[inPot2, out=-60, in=-120,distance=2.5cm, blue\] node\[nodedescr\] [$j_{2}$]{} (m-2-5)(m-2-7) edge \[inPot1, red, cross line\] node\[nodedescr\] [$j_{3}$]{} (m-2-9) (m-2-7) edge \[inPot2, in=-120, blue\] node\[nodedescr\] [$j_{4}$]{} (m-2-11); (m-2-1) edge (m-2-3); (m-2-3) edge (m-2-5); (m-2-5) edge (m-2-7); (m-2-7) edge (m-2-9); (m-2-9) edge (m-2-11); (m-2-11) edge (m-2-13);
In , a 2nd-order meta-stem cell (G) divides to produce a 1st-order stem cell $B$ and cell J1. The cell J1 stochastically loops back to activate the meta-stem cell G or it activates the first order linear stem cell $B$. The 1st-order stem cell $B$ divides into a semi-terminal cell C and a cell J2. The cell J2 stochastically activates either its parent stem cell $B$, the semi-terminal cell C or dedifferentiates to G. The cell C also can stochastically dedifferentiate to any of the previous stem cell fates (G, B) or differentiate into the final terminal cell T.
The network consists of two linked 1st-order linear stochastic networks. Its behavior approaches a deterministic 2nd-order geometric stem cell network as the probabilities $p$ and $q_1$ approach $1$. Cancers controlled by such a network can go into spontaneous remission because of the fact that for all points in all possible paths in the network there is the possibility of reaching a terminal cell state.
However, because there is the possibility at J2 and C to dedifferentiate to earlier, upstream network states, the relative numbers of stem cells to terminal cells will be higher than in networks that have fewer dedifferentiation pathways. Note, too that the dedifferentiation pathways introduce several possibilities for exponential growth. Even if their probabilities ($q_3$, $r_2$, $r_3$) are very small, any increase in these probabilities could have a significant destabilizing influence on the resulting tumor.
The deterministic possible behaviors include: $(p=0, q= 1, s=0, r=1)$ is terminal. $(p=1, q= 0)$ is exponential. $(p=1, q= 1, s=0, r=1)$ is linear. $(p=1, q= 1, s=1, r=1)$ is 2nd order geometric. $(p=1, q= 1, s=1, r=0)$ is a linear network linked to an exponential network.
A Master stochastic geometric and exponential network
-----------------------------------------------------
We now present a generalized, broad spectrum, master network where each probability distribution results in a particular geometric or exponential network. One way to make the developmental network control theory more acceptable to traditional mathematicians and physicists is to formulate the networks in terms of differential equations. Given that stochastic networks in the cancer paper can be made so general as to describe any network based on the probability distribution, perhaps one can use stochasticity in differential equations to attain one Master Equation that describes all the major types of networks. Particular probability distributions would then give particular network behavior. The Master Network and Master Equation can be made fully general by allowing arbitrarily many loops.
\(m) \[matrix of math nodes, row sep=3em, column sep=, text height=1.5ex, text depth=0.25ex\] [ && && && & & &&\
N\_[1]{} && J\_[1]{} && J\_[2]{} && N\_[2]{} && J\_[3]{} && J\_[4]{} && N\_[3]{} & N\_[i]{} & N\_[k]{}\
&& && && && && &&\
]{}; (m-2-3) edge \[inPot1, out=90, in=90,distance=5cm, green, cross line\] node\[stochasticNodestyle\] [$p_{1}$]{} (m-2-1) (m-2-3) edge \[in2Pot1,out=60,in=110,distance=3.5cm,green, cross line\] node\[stochasticNodestyle\] [$1-p_{1}$]{} (m-2-7) (m-2-5) edge \[in2Pot1,out=-60,in=-110,distance=3.5cm,black, cross line\] node\[stochasticNodestyle\] [$q_{1}$]{} (m-2-7) (m-2-5) edge \[in2Pot2,in=-90,out=-90, distance=4cm,black, cross line\] node\[stochasticNodestyle\] [$1-q_{1}$]{} (m-2-1) (m-2-9) edge \[inPot1, out=90, in= 75, distance=5cm, green, cross line\] node\[stochasticNodestyle\] [$p_{2}$]{} (m-2-7) (m-2-9) edge \[inPot1, out=60, in= 120, distance=3cm, green, cross line\] node\[stochasticNodestyle\] [$1-p_{2}$]{} (m-2-13) (m-2-11) edge \[inPot2, out=-90, in= -90, distance=4cm, red\] node\[stochasticNodestyle\] [$1-q_{2}$]{} (m-2-7) (m-2-11) edge \[inPot2, out=-60, in= -90, distance=3cm, red, cross line\] node\[stochasticNodestyle\] [$q_{2}$]{} (m-2-13); (m-2-1) edge \[inPot1, red, cross line\] node\[nodedescr\] [$j_1$]{} (m-2-3) (m-2-1) edge \[inPot2, out=-60, in=-120,distance=2.5cm, blue\] node\[nodedescr\] [$j_{2}$]{} (m-2-5)(m-2-7) edge \[inPot1, red, cross line\] node\[nodedescr\] [$j_{3}$]{} (m-2-9) (m-2-7) edge \[inPot2, in=-120, blue\] node\[nodedescr\] [$j_{4}$]{} (m-2-11); (m-2-1) edge (m-2-3); (m-2-3) edge (m-2-5); (m-2-5) edge (m-2-7); (m-2-7) edge (m-2-9); (m-2-9) edge (m-2-11); (m-2-11) edge (m-2-13); (m-2-13) edge (m-2-14); (m-2-14) edge (m-2-15);
2nd Order Geometric stochastic stem cell network
------------------------------------------------
\(m) \[matrix of math nodes, row sep=3em, column sep=3em, text height=1.5ex, text depth=0.25ex\] [ && && && & & &\
G && J1 && B && J2 && C &&\
&& && && && & &\
]{}; (m-2-1) edge \[inPot1, red, cross line\] node\[nodedescr\] [$j_1$]{} (m-2-3) (m-2-1) edge \[inPot2, in=-90, blue, cross line\] node\[nodedescr\] [$b$]{} (m-2-5) (m-2-5) edge \[inPot1, red, cross line\] node\[nodedescr\] [$j_2$]{} (m-2-7) (m-2-5) edge \[inPot2, in=-90, blue, cross line\] node\[nodedescr\] [$c$]{} (m-2-9); (m-2-3) edge \[inPot1, out=90, in=90,distance=5cm, green, cross line\] node\[stochasticNodestyle\] [$p$]{} (m-2-1) (m-2-3) edge \[in2Pot1,in=120,distance=3cm,green, cross line\] node\[stochasticNodestyle\] [$1-p$]{} (m-2-5) (m-2-7) edge \[inPot1, out=90, in= 75, distance=5cm, green, cross line\] node\[stochasticNodestyle\] [$q$]{} (m-2-5) (m-2-7) edge \[inPot1, out=60, in= 90, distance=3cm, green, cross line\] node\[stochasticNodestyle\] [$1-q$]{} (m-2-9); (m-2-1) edge (m-2-3); (m-2-3) edge (m-2-5); (m-2-5) edge (m-2-7); (m-2-7) edge (m-2-9);
In The cell J1 stochastically loops back to activate the meta-stem cell G or it activates the first order linear stem cell $B$. The 1st-order stem cell $B$ divides into a terminal cell C and a cell J2. The cell J2 stochastically activates either the stem cell $B$ or the terminal cell $C$. The network consists of two linked 1st-order linear stochastic networks. Its behavior approaches a deterministic 2nd-order geometric stem cell network as the probabilities $p$ and $q$ approach $1$. Cancers controlled by such a network can go into spontaneous remission because of the fact that for all points in all possible paths in the network there is the possibility of reaching a terminal cell state.
Geometric stochastic stem cell network with dedifferentiation
-------------------------------------------------------------
\(m) \[matrix of math nodes, row sep=2.5em, column sep=, text height=1.5ex, text depth=0.25ex\] [ && && && & & &\
G && J1 && B && J2 && C && T &&\
&& && && && && &\
]{}; (m-2-1) edge \[inPot1, red, cross line\] node\[nodedescr\] [$j_1$]{} (m-2-3) (m-2-1) edge \[inPot2, in=-120, blue, cross line\] node\[nodedescr\] [$b$]{} (m-2-5) (m-2-5) edge \[inPot1, red, cross line\] node\[nodedescr\] [$j_2$]{} (m-2-7) (m-2-5) edge \[inPot2, in=-120, blue, cross line\] node\[nodedescr\] [$c$]{} (m-2-9); (m-2-3) edge \[inPot1, out=90, in=90,distance=5cm, green, cross line\] node\[stochasticNodestyle\] [$p$]{} (m-2-1) (m-2-3) edge \[in2Pot1,in=120,distance=3cm,green, cross line\] node\[stochasticNodestyle\] [$1-p$]{} (m-2-5) (m-2-7) edge \[inPot1, out=90, in= 75, distance=5cm, green, cross line\] node\[stochasticNodestyle\] [$q_1$]{} (m-2-5) (m-2-7) edge \[inPot1, out=60, in= 120, distance=3cm, green, cross line\] node\[stochasticNodestyle\] [$q_2$]{} (m-2-9) (m-2-7) edge \[inPot2, out=-120, in= -90, distance=4cm, green, cross line\] node\[stochasticNodestyle\] [$q_3$]{} (m-2-1) (m-2-9) edge \[inPot1, out=70, in= 110, distance=3cm, black, cross line\] node\[stochasticNodestyle\] [$r_1$]{} (m-2-11) (m-2-9) edge \[inPot2, out=-90, in= -90, distance=4cm, black, cross line\] node\[stochasticNodestyle\] [$r_2$]{} (m-2-5) (m-2-9) edge \[inPot2, out=-60, in= -120, distance=7cm, black, cross line\] node\[stochasticNodestyle\] [$r_3$]{} (m-2-1); (m-2-1) edge (m-2-3); (m-2-3) edge (m-2-5); (m-2-5) edge (m-2-7); (m-2-7) edge (m-2-9);(m-2-9) edge (m-2-11);
In let probabilities $q_1 + q_2 +q_3 = 1$ and $r_1+r_2+r_3 = 1$. A 2nd-order meta-stem cell (G) divides to produce a 1st-order stem cell $B$ and cell J1. The cell J1 stochastically loops back to activate the meta-stem cell G or it activates the first order linear stem cell $B$. The 1st-order stem cell $B$ divides into a semi-terminal cell C and a cell J2. The cell J2 stochastically activates either its parent stem cell $B$, the semi-terminal cell C or dedifferentiates to G. The cell C also can stochastically dedifferentiate to any of the previous stem cell fates (G, B) or differentiate into the final terminal cell T. The network consists of two linked 1st-order linear stochastic networks. Its behavior approaches a deterministic 2nd-order geometric stem cell network as the probabilities $p$ and $q_1$ approach $1$. Cancers controlled by such a network can go into spontaneous remission because of the fact that for all points in all possible paths in the network there is the possibility of reaching a terminal cell state. However, because there is the possibility at J2 and C to dedifferentiate to earlier, upstream network states, the relative numbers of stem cells to terminal cells will be higher than in networks that have fewer dedifferentiation pathways. Note, too that the dedifferentiation pathways introduce several possibilities for exponential growth. Even if their probabilities ($q_3$, $r_2$, $r_3$) are very small, any increase in these probabilities could have a significant destabilizing influence on the resulting tumor.
Modeling deterministic stem cell networks with stochastic networks
------------------------------------------------------------------
For stochastic stem cell networks the probability distribution over a network topology determines their dynamic properties. If the stochastic network allows all possible paths in a set of linear, 2nd-order geometric and exponential networks, then the dynamic behavior of each is approximated as the corresponding embedded links are assigned probabilities that match existence or nonexistence of links in the particular network type, For example, if a link exists in a network embedded in the stochastic network, then it assigned a probability close to 1 or 1 itself if we want an exact match. If the link does not exist in the embedded network, it is assigned a probability close to 0 or 0 itself for an exact match. Hence, the probability distribution distinguishes different network types that are embeddable in the topology or architecture of a given stochastic network.
Transformations of probability distributions change stochastic stem cell behavior
---------------------------------------------------------------------------------
If we allow the probability distribution over a stochastic network to change with time because of external factors (such as ultra violet radiation) then the same stochastic stem cell network may exhibit various proliferative phenotypes, appearing alternatively as being in remission, linear, geometric or exponential. Thus, meta-probability functions are at work here that conditionally change the probability distribution.
Communication in stem cell networks
===================================
We distinguish one-way, reactive communication protocols in stem cell networks and two-way, interactive communication protocols in stem cell networks. Reactive communication stem cell networks only react to external signals but do not send signals themselves. Interactive communicating stem cell networks interact with the networks of other cells via signaling protocols by both sending signals and receiving signals to and from other cells.
Reactive communication stem cell networks
-----------------------------------------
This class of stem cell networks reacts to external and internal cell signals and, thereby, activating the stem cell network if the cell receives the appropriate signal. Unlike more complex cases where there are intercellular signaling protocols involved, here the communication and the reaction are one-sided. A cell sends a signal to which the stem cell reacts by activating its stem cell network.
A reactive signal based geometric network architecture with 1st and 2nd order geometric potential
-------------------------------------------------------------------------------------------------
\(m) \[matrix of math nodes, row sep=2em, column sep=3em, text height=1.5ex, text depth=0.25ex\] [ && && && & & &\
S && M && L && T && &&\
&& && && && & &\
]{}; (m-2-1) edge \[inPot1, red, receivestyle, cross line\] node\[receivesigstyle\] [$\sigma_{1}$]{} (m-2-3) (m-2-1) edge \[inPot2, in=-110, blue, receivestyle, cross line\] node\[receivesigstyle\] [$\sigma_{2}$]{} (m-2-5); (m-2-3) edge \[selfloop1, out=80, distance=4cm, green, cross line\] node\[nodedescr\] [s]{} (m-2-1) (m-2-3) edge \[in2Pot1,green, cross line\] node\[nodedescr\] [$l$]{} (m-2-5); (m-2-5) edge \[selfloop1, out=60, in= 125, purple,cross line\] node\[nodedescr\] [$l$]{} (m-2-5) (m-2-5) edge \[inPot2, out=-60, in= -90, distance=3cm, purple,cross line\] node\[nodedescr\] [$t$]{} (m-2-7); (m-2-1) edge (m-2-3); (m-2-3) edge (m-2-5); (m-2-5) edge (m-2-7); (m-2-7) edge (m-2-9);
A pure reactive signal based geometric network architecture with 1st and 2nd order geometric potential
------------------------------------------------------------------------------------------------------
\(m) \[matrix of math nodes, row sep=2em, column sep=3em, text height=1.5ex, text depth=0.25ex\] [ && && && & & &\
S\_[M]{} && M && S\_[L]{} && L && T &&\
&& && && && & &\
]{}; (m-2-1) edge \[inPot1, red, receivestyle, cross line\] node\[receivesigstyle\] [$\sigma_{1}$]{} (m-2-3) (m-2-1) edge \[inPot2, in=-90, distance=4cm, blue, receivestyle, cross line\] node\[receivesigstyle\] [$\sigma_{2}$]{} (m-2-5) (m-2-5) edge \[inPot1, red, receivestyle, cross line\] node\[receivesigstyle\] [$\sigma_{3}$]{} (m-2-7) (m-2-5) edge \[inPot2, in=-80, distance=4cm, blue, receivestyle, cross line\] node\[receivesigstyle\] [$\sigma_{4}$]{} (m-2-9); (m-2-3) edge \[selfloop1, out=80, distance=4cm, green, cross line\] node\[nodedescr\] [$s_{M}$]{} (m-2-1) (m-2-3) edge \[inPot2, out=-60, in=-120, green, cross line\] node\[nodedescr\] [$s_{L}$]{} (m-2-5); (m-2-7) edge \[selfloop1, out=80, in= 90, purple,cross line\] node\[nodedescr\] [$s_{L}$]{} (m-2-5) (m-2-7) edge \[inPot2, out=-60, in= -120, purple,cross line\] node\[nodedescr\] [$t$]{} (m-2-9); (m-2-1) edge (m-2-3); (m-2-3) edge (m-2-5); (m-2-5) edge (m-2-7); (m-2-7) edge (m-2-9);
Stem cell networks with interactive communication {#sec:SigC}
=================================================
A type of conditional cancer are [*social cancers*]{} that depend on cell signaling to be active. This variety of cancer occurs when the genomic network interacts with receptors, signals and signal transduction pathways. In that case we can have what might be called social cancers. For example, a cell $A$ will signal $\alpha$ to cell $B$ and differentiates to A1, $B$ on receiving the signal divides into B1 and B1. B1 sends a signal $\beta$ back to $A$ and then B1 dedifferentiates to $B$ (its network loops back). When A1 receives $\beta$ its network loops back into state $A$. Now the process repeats with the $A$ cells sending signals $\alpha$ to $B$ cells. One can see that this process is potentially exponential as long as we have sufficient cells of type $A$ with sufficient signal capacity to continue to activate all the developing $B$ cells.
With conditional cancers that depend on social communication, the growth rate of the cancer will depend on the accessibility of the signal. If the receiver of the signal requires direct contact with the sender then even if the cytogenic cell contains a conditional exponential network, since the signal may not be received by those daughter cells the exponential potential may not be realized. Furthermore, if it is a linear network then if the passive daughter cell is interposed between the sender and the receiving cytogenic cell then growth will stop after several divisions when the signal no loner can reach the cytogenic cells. If later, because of physical pressure or other conditions, the cytogenic cell is again in close enough proximity to the sender then the cancer can start again.
Interactive signaling mono generative networks {#sec:SigC1}
----------------------------------------------
In the figure below we emphasize the communication links. The network has one cell only signaling and the other cell signaling and conditionally dividing, however it leaves open whether the cytogenic subnetwork is linear or exponential. The more detailed communication networks and their properties will described in the sections below.
\[x=1.00mm, y=1.00mm, inner xsep=0pt, inner ysep=0pt, outer xsep=0pt, outer ysep=0pt\] (51.00,42.86) rectangle +(107.34,54.79); (128.68,70.26) circle (25.39mm); (80.06,70.68) circle (23.96mm); (64.12,65.99) – (61.98,68.12) (66.94,65.99) – (64.81,68.12) (69.77,65.99) – (67.64,68.12) (72.60,65.99) – (70.47,68.12) (75.43,65.99) – (73.30,68.12) (78.26,65.99) – (76.13,68.12) (81.09,65.99) – (78.95,68.12) (83.92,65.99) – (81.78,68.12) (86.74,65.99) – (84.61,68.12) (89.57,65.99) – (87.44,68.12) (92.40,65.99) – (90.27,68.12) (95.23,65.99) – (93.10,68.12) (96.27,67.78) – (95.93,68.12); (61.73,67.38) – (62.47,68.12) (63.16,65.99) – (65.30,68.12) (65.99,65.99) – (68.12,68.12) (68.82,65.99) – (70.95,68.12) (71.65,65.99) – (73.78,68.12) (74.48,65.99) – (76.61,68.12) (77.31,65.99) – (79.44,68.12) (80.13,65.99) – (82.27,68.12) (82.96,65.99) – (85.09,68.12) (85.79,65.99) – (87.92,68.12) (88.62,65.99) – (90.75,68.12) (91.45,65.99) – (93.58,68.12) (94.28,65.99) – (96.27,67.98); (61.73,65.99) rectangle +(34.54,2.13); (113.48,64.71) – (113.11,65.08) (116.31,64.71) – (114.39,66.63) (119.14,64.71) – (117.22,66.63) (121.96,64.71) – (120.05,66.63) (124.79,64.71) – (122.87,66.63) (127.62,64.71) – (125.70,66.63) (130.45,64.71) – (128.53,66.63) (133.28,64.71) – (131.36,66.63) (136.11,64.71) – (134.19,66.63) (138.93,64.71) – (137.02,66.63) (141.76,64.71) – (139.84,66.63) (144.59,64.71) – (142.67,66.63) (147.42,64.71) – (145.50,66.63) (149.57,65.39) – (148.33,66.63); (113.11,65.03) – (114.71,66.63) (115.62,64.71) – (117.54,66.63) (118.45,64.71) – (120.37,66.63) (121.28,64.71) – (123.20,66.63) (124.11,64.71) – (126.03,66.63) (126.94,64.71) – (128.86,66.63) (129.77,64.71) – (131.68,66.63) (132.59,64.71) – (134.51,66.63) (135.42,64.71) – (137.34,66.63) (138.25,64.71) – (140.17,66.63) (141.08,64.71) – (143.00,66.63) (143.91,64.71) – (145.83,66.63) (146.74,64.71) – (148.66,66.63) (149.57,64.71) – (149.57,64.72); (113.11,64.71) rectangle +(36.46,1.92); (103.52,75.16) rectangle +(2.35,1.92); (105.22,76.23) .. controls (108.27,76.71) and (111.33,77.14) .. (114.39,77.51) .. controls (117.44,77.87) and (120.71,78.10) .. (123.13,76.23) .. controls (124.60,75.10) and (125.46,73.37) .. (125.69,71.54) .. controls (125.80,70.69) and (125.78,69.83) .. (125.69,68.98) .. controls (125.64,68.48) and (125.57,67.98) .. (125.48,67.48); (125.48,67.48) – (126.40,70.22) – (125.65,69.58) – (125.01,70.33) – (125.48,67.48) – cycle; (100.75,59.81) rectangle +(2.13,2.56); (100.53,61.30) .. controls (95.60,55.27) and (87.00,53.73) .. (80.28,57.68) .. controls (77.79,59.14) and (75.81,61.28) .. (73.88,63.43) .. controls (73.31,64.07) and (72.74,64.71) .. (72.17,65.35); (72.17,65.35) – (73.52,62.80) – (73.57,63.78) – (74.56,63.73) – (72.17,65.35) – cycle; (131.66,64.93) – (122.07,56.61) – (103.94,61.30); (104.62,61.13) circle (0.70mm); (77.08,68.55) – (85.82,80.49) – (101.81,76.44); (101.81,76.44) circle (0.70mm); (83.90,81.56) node\[anchor=base west\][$\alpha$]{}; (122.07,53.20) node\[anchor=base west\][$\beta$]{}; (53.00,51.00) node\[anchor=base west\][$A$]{}; (148.00,49.00) node\[anchor=base west\][$B$]{}; (73.00,68.00) .. controls (75.71,71.40) and (77.13,75.65) .. (77.00,80.00) .. controls (76.89,83.52) and (75.18,86.97) .. (72.00,87.00) .. controls (68.49,87.03) and (66.73,83.02) .. (67.00,79.00) .. controls (67.26,75.03) and (68.65,71.21) .. (71.00,68.00); (71.00,68.00) – (70.15,70.76) – (69.91,69.80) – (68.95,70.03) – (71.00,68.00) – cycle;
An [*in vivo*]{} instance of this network type is seen in bone cancer where a signal from the adjacent tissue is required for the other to become cancerous, for example, see (Logothetis [@Logothetis2005]).
Interactive signaling mono-linear network architecture
------------------------------------------------------
The following network controls two cells types A and B that communicate by cell signaling. The cell A divides only when it receives the signal $\beta$ from cell B. The cell B only sends its signal after it has received the signal $\alpha$ from cell A.
\[sec:SigL1\]
\(m) \[matrix of math nodes, row sep=3em, column sep=, text height=1.5ex, text depth=0.25ex\] [ && && && &&\
A && A\_[1]{} && A\_[2]{} &T & B && B\_[1]{} &\
&& && && && & &\
]{}; (m-2-1) edge \[inPot1, green, receivestyle, cross line\] node\[receivesigstyle\] [$\beta$]{} (m-2-3); (m-2-7) edge \[inPot1, out=90, green, receivestyle, cross line\] node\[receivesigstyle\] [$\alpha$]{} (m-2-9);
(m-2-5) edge \[selfloop2, out=-80, distance=4cm, red, cross line\] node\[nodedescr\] [a]{} (m-2-1) (m-2-5) edge \[in2Pot1,red, cross line\] node\[nodedescr\] [$t$]{} (m-2-6); (m-2-3) edge \[selfloop2, out=-90, in=-120, distance=2cm, black, cross line\] node\[nodedescr\] [$\alpha_{2}$]{} (m-2-5) (m-2-9) edge \[selfloop2, out=-60, distance=3cm, black, cross line\] node\[nodedescr\] [b]{} (m-2-7); (m-2-3) edge \[ in2Pot1, out= 60, in=125, blue,snakesendstyle,cross line\] node\[sendsigstyle\] [$\alpha$]{} (m-2-7) (m-2-9) edge \[ in2Pot1, distance=5cm, blue,snakesendstyle,cross line\] node\[sendsigstyle\] [$\beta$]{} (m-2-1); (m-2-1) edge (m-2-3); (m-2-3) edge (m-2-5); (m-2-5) edge (m-2-6); (m-2-6) edge (m-2-7); (m-2-7) edge (m-2-9); (m-2-9) edge (m-2-10);
Dual cancer signaling networks {#sec:SigC2}
------------------------------
A the interacting between two cooperative cell types, called partners, is driven by a network where each partner has a separate role driven by a different subnetwork with signaling driving their actions. In the figure below we emphasize the communication links. The network has both cells signaling and conditionally dividing, however it leaves open whether the network is linear or exponential. The more detailed communication networks and their properties will described in the following sections.
\[x=1.00mm, y=1.00mm, inner xsep=0pt, inner ysep=0pt, outer xsep=0pt, outer ysep=0pt\] (51.00,42.86) rectangle +(107.34,54.79); (128.68,70.26) circle (25.39mm); (80.06,70.68) circle (23.96mm); (64.12,65.99) – (61.98,68.12) (66.94,65.99) – (64.81,68.12) (69.77,65.99) – (67.64,68.12) (72.60,65.99) – (70.47,68.12) (75.43,65.99) – (73.30,68.12) (78.26,65.99) – (76.13,68.12) (81.09,65.99) – (78.95,68.12) (83.92,65.99) – (81.78,68.12) (86.74,65.99) – (84.61,68.12) (89.57,65.99) – (87.44,68.12) (92.40,65.99) – (90.27,68.12) (95.23,65.99) – (93.10,68.12) (96.27,67.78) – (95.93,68.12); (61.73,67.38) – (62.47,68.12) (63.16,65.99) – (65.30,68.12) (65.99,65.99) – (68.12,68.12) (68.82,65.99) – (70.95,68.12) (71.65,65.99) – (73.78,68.12) (74.48,65.99) – (76.61,68.12) (77.31,65.99) – (79.44,68.12) (80.13,65.99) – (82.27,68.12) (82.96,65.99) – (85.09,68.12) (85.79,65.99) – (87.92,68.12) (88.62,65.99) – (90.75,68.12) (91.45,65.99) – (93.58,68.12) (94.28,65.99) – (96.27,67.98); (61.73,65.99) rectangle +(34.54,2.13); (113.48,64.71) – (113.11,65.08) (116.31,64.71) – (114.39,66.63) (119.14,64.71) – (117.22,66.63) (121.96,64.71) – (120.05,66.63) (124.79,64.71) – (122.87,66.63) (127.62,64.71) – (125.70,66.63) (130.45,64.71) – (128.53,66.63) (133.28,64.71) – (131.36,66.63) (136.11,64.71) – (134.19,66.63) (138.93,64.71) – (137.02,66.63) (141.76,64.71) – (139.84,66.63) (144.59,64.71) – (142.67,66.63) (147.42,64.71) – (145.50,66.63) (149.57,65.39) – (148.33,66.63); (113.11,65.03) – (114.71,66.63) (115.62,64.71) – (117.54,66.63) (118.45,64.71) – (120.37,66.63) (121.28,64.71) – (123.20,66.63) (124.11,64.71) – (126.03,66.63) (126.94,64.71) – (128.86,66.63) (129.77,64.71) – (131.68,66.63) (132.59,64.71) – (134.51,66.63) (135.42,64.71) – (137.34,66.63) (138.25,64.71) – (140.17,66.63) (141.08,64.71) – (143.00,66.63) (143.91,64.71) – (145.83,66.63) (146.74,64.71) – (148.66,66.63) (149.57,64.71) – (149.57,64.72); (113.11,64.71) rectangle +(36.46,1.92); (103.52,75.16) rectangle +(2.35,1.92); (105.22,76.23) .. controls (108.27,76.71) and (111.33,77.14) .. (114.39,77.51) .. controls (117.44,77.87) and (120.71,78.10) .. (123.13,76.23) .. controls (124.60,75.10) and (125.46,73.37) .. (125.69,71.54) .. controls (125.80,70.69) and (125.78,69.83) .. (125.69,68.98) .. controls (125.64,68.48) and (125.57,67.98) .. (125.48,67.48); (125.48,67.48) – (126.40,70.22) – (125.65,69.58) – (125.01,70.33) – (125.48,67.48) – cycle; (100.75,59.81) rectangle +(2.13,2.56); (100.53,61.30) .. controls (95.60,55.27) and (87.00,53.73) .. (80.28,57.68) .. controls (77.79,59.14) and (75.81,61.28) .. (73.88,63.43) .. controls (73.31,64.07) and (72.74,64.71) .. (72.17,65.35); (72.17,65.35) – (73.52,62.80) – (73.57,63.78) – (74.56,63.73) – (72.17,65.35) – cycle; (131.66,64.93) – (122.07,56.61) – (103.94,61.30); (104.62,61.13) circle (0.70mm); (77.08,68.55) – (85.82,80.49) – (101.81,76.44); (101.81,76.44) circle (0.70mm); (83.90,81.56) node\[anchor=base west\][$\alpha$]{}; (122.07,53.20) node\[anchor=base west\][$\beta$]{}; (53.00,51.00) node\[anchor=base west\][$A$]{}; (148.00,49.00) node\[anchor=base west\][$B$]{}; (128.00,67.00) .. controls (127.86,70.35) and (127.53,73.69) .. (127.00,77.00) .. controls (126.34,81.15) and (126.38,85.81) .. (130.00,87.00) .. controls (134.48,88.47) and (138.37,83.08) .. (137.00,77.00) .. controls (136.07,72.88) and (133.55,69.29) .. (130.00,67.00); (130.00,67.00) – (132.61,68.23) – (131.62,68.33) – (131.72,69.32) – (130.00,67.00) – cycle; (73.00,68.00) .. controls (75.71,71.40) and (77.13,75.65) .. (77.00,80.00) .. controls (76.89,83.52) and (75.18,86.97) .. (72.00,87.00) .. controls (68.49,87.03) and (66.73,83.02) .. (67.00,79.00) .. controls (67.26,75.03) and (68.65,71.21) .. (71.00,68.00); (71.00,68.00) – (70.15,70.76) – (69.91,69.80) – (68.95,70.03) – (71.00,68.00) – cycle;
What we have here is a signal loop that is essential for the whole cancer network. The dual cancer signaling network in consists of two subnetworks, and A-subnetwork for cell type A and a B-subnetwork for cell type B. Once activated, the A-subnetwork (e.g., ) causes a signal $\alpha$ to be sent to cell $B$ followed by activation of subnetwork A$_{2}$ whereupon the cell divides into two daughter cells at least one of which enters the signal receptor state A which enables the cell to receive signals of type $\beta$. The signaling partner subnetwork of cell type B is activated by a signal transduction cascade initiated by the receipt of signal $\alpha$. The B-subnetwork in state B$_{1}$ directs the sending of a signal $\beta$ to cell A. After sending the signal, the B-subnetwork enters state $B_{2}$ that directs $B$ to divide into two daughter cells, one or both of which loop back, to enter the receptor state B.
Clinically, we would observe depends on nature of the cytogenic subnetworks, as well as the developing morphology of the tumor in the tissue. The cells A and B only proliferate if have a neighbor that is a signaling partner. Thus, the growth of B cells depends on being adjacent to A cells and vice versa. We would observe the stopping of cell proliferation if either the A or the B cells are removed or some other agent interferes with the signaling loop. The latter can be effected either by inhibiting the receptor for $\alpha$ or $\beta$, the signal $\alpha$ or $\beta$, or by interfering with either the$\alpha$or $\beta$ signal transduction pathway. Alternatively, one could interfere with the genome by inhibiting the areas responsible for initiation of signaling or the inhibiting the areas responsible for cellular division. One can see there are many potential areas where one can break the cancer loop. In actual in vivo systems one would choose that with minimal risk of side effects.
We now investigate some of the particular types of dual cancer signaling networks.
A interactive signaling dual linear network architecture {#sec:SigL2}
--------------------------------------------------------
\(m) \[matrix of math nodes, row sep=3em, column sep=, text height=1.5ex, text depth=0.25ex\] [ && && && && &&&&&\
A && A\_[1]{} && A\_[2]{} &T\_[1]{}& B && B\_[1]{} & & B\_[2]{} &T\_[2]{}\
&& && && && & &\
]{}; (m-2-1) edge \[inPot1, green, receivestyle, cross line\] node\[receivesigstyle\] [$\beta$]{} (m-2-3); (m-2-7) edge \[inPot1, out=90, green, receivestyle, cross line\] node\[receivesigstyle\] [$\alpha$]{} (m-2-9);
(m-2-5) edge \[selfloop2, out=-80, distance=4cm, red, cross line\] node\[nodedescr\] [a]{} (m-2-1) (m-2-5) edge \[in2Pot1,out=80, red, cross line\] node\[nodedescr\] [$t_{1}$]{} (m-2-6); (m-2-11) edge \[in2Pot1, red, in=90, cross line\] node\[nodedescr\] [$t_{2}$]{} (m-2-12) (m-2-11) edge \[selfloop2, out=-60, distance=4cm, red, cross line\] node\[nodedescr\] [b]{} (m-2-7); (m-2-3) edge \[selfloop2, out=-90, in=-120, distance=2cm, black, cross line\] node\[nodedescr\] [$\alpha_{2}$]{} (m-2-5) (m-2-9) edge \[selfloop2, out=-90, in=-120,distance=2cm, black, cross line\] node\[nodedescr\] [$\beta_{2}$]{} (m-2-11); (m-2-3) edge \[ in2Pot1, out= 60, in=125, distance=4cm, blue, snakesendstyle,cross line\] node\[sendsigstyle\] [$\alpha$]{} (m-2-7) (m-2-9) edge \[ in2Pot1, distance=6cm, blue,snakesendstyle,cross line\] node\[sendsigstyle\] [$\beta$]{} (m-2-1); (m-2-1) edge (m-2-3); (m-2-3) edge (m-2-5); (m-2-5) edge (m-2-6); (m-2-6) edge (m-2-7); (m-2-7) edge (m-2-9); (m-2-9) edge (m-2-11); (m-2-9) edge (m-2-11); (m-2-11) edge (m-2-12);
Hybrid communicative and stochastic stem cell networks
======================================================
It is possible in principle for a stem cell network to become stochastic as a result of communication. In other words, it is conditionally stochastic in response to cell signaling.
So too a stochastic process can lead to cell signaling which can lead to the, possibly stochastic, activation of another developmental network.
Discussion: Stem cell concepts
==============================
Formal and informal definitions of transpotency
-----------------------------------------------
This section can be skipped or read later since the graphical formalism with which we will represent stem cell networks may be more intuitive.
Given a cell type A, let $\theta_{1}\ldots\theta_{k}$ be the minimal number of external operations $\theta_{i}$ required to transform a cell type A into a cell type B. Then $B = \theta_{1}\ldots\theta_{k}A$. Then we say that cell type A is [*k-distant*]{} from cell type B and B is a [*k-transform*]{} of A. Note, if B is a k-transform of A that does not necessarily imply that A is a k-transform of B since the operations may be irreversible. B is a [*k-transform*]{} of A if there exist a number of operations $\theta_{1}\ldots\theta_{k}$ such that $B = \theta_{1}\ldots\theta_{k}A$. B is a [*transform*]{} of A if there exists some $k$ operators such that B is a k-transform of A. The transpotency of a cell type A is then definable as the set of all ordered pairs $<B_{i}, k_{i}>$ where $B_{i}$ is a $k_{i}$-transform of A. The [*operational distance*]{} from A to B is number of operators $\theta_{i}, i = 1 \ldots k$ required to transform A to some type B.
Another aspect of transpotency has to do with stochasticity. It can take many attempts to transform a cell type A into B. The operator sequence has to be applied to many cells of type A before one of them changes into type B. Thus, the sequence of operators $\theta_{1}\ldots\theta_{k}$ applied to A only results in B with some probability $p$. Thus the transformative distance of A from B is not just a one dimensional function of the operational distance, but also should include the probability that an application of the operators to A will result in B. The lower the transform probability $p$ the farther A is to B in transformational space. And, if the probability $p = 0$ then B is [*inaccessible*]{} from A. If the probability $p=1$ then there is a deterministic relationship that insures $\theta_{1}\ldots\theta_{k}A \rightarrow B$, where $\rightarrow$ means “yields” or “transforms into”. If $0 \leq p \leq 1$ then $\theta_{1}\ldots\theta_{k}A \stackrel{p}{\rightarrow} B$, where $\stackrel{p}{\rightarrow}$ means “The operations $\theta_{1}\ldots\theta_{k}$ on A [*transform A into B with probability*]{} $p$”.
Problems with definitions of transpotency
-----------------------------------------
Clearly as experimental procedures evolve the operations required to transform A into B will change and the operational distance $k$ may decrease. Still this formalization makes it clear that the larger operational distance the more resistant A is to becoming B. It is also evident that artificial transpotency need not be correlated with natural transpotency.
Moreover, the existence of operations via transcription factors that can dedifferentiate a cell to a transpotent totipotent or pluripotent cell and then induce that cell to differentiate become a specific cell type break down the difference between transpotent cells and normal cells making the definition of transpotent stem cell all inclusive and, thereby, useless.
In normal development, the [*natural transpotency*]{} of a cell depends on the natural conditions and multicellular context of in which the cell finds itself, e.g., cell signaling in a multicellular context. [*Artificial transpotency*]{} is experimentally induced transpotency resulting from viral, chemical or physical extraneous, experimental impingements.
The degree of transpotency of a cell is not precisely defined because it is ultimately an operational definition that depends on what operations are performed on cells to make them differentiate into other cells. And since a few transcription factors can make some cells dedifferentiate into stem cells the class of artificial transpotent cells has no clear boundaries. Hence, if sufficiently powerful experimental operations are allowed that can change any cell differentiation state to almost any other state, then this calls into question the whole concept of artificial transpotent, non self-renewing stem cells, since their transpotency can no longer operationally distinguished from normal cells . The only distinguishing feature would be the number and type of operators that are required to convert one cell type into another. The greater the number of steps and the more difficult the operations are the more resistant a cell is to transformation.
Conceptual ambiguities
----------------------
There is a conceptual ambiguity as to what stem cells are [^2]. One the one hand stem cells are viewed as having the capacity of endless self-renewal, and, other the other hand, stem cells are viewed as having the capacity to differentiate into multiple cell types. These two views of stem cells are distinct and refer to independent properties of cells. They are different conceptions of what stem cells are and what they can do. The difference is in the capacity of a cell to generate multiple progeny while maintaining its original, unchanged cell type, versus the capacity of a cell to change its differentiation state by transforming into different possible cell types.
The two attributes of self-renewal and transformative potential can coexist in the same cell. However, if a stem cell has both capacities and differentiates into some other cell type it may loose its capacity to self-renew.
Let us call stem cells that can endlessly generate other cells while maintaining their own cell type, [*iteropotent stem cells*]{}. While stem cells that have the capacity to differentiate or transform into various cell types [*transpotent stem cells*]{}. Yet, a third conception of stem cells combines the self-renewal with a transformative capacity by distributing the capacities over parent and daughter cells. It restricts the capacity of self-renewal to the parent cell and the transformative potential to daughter cells. In this version, stem cells have the capacity of endless self-renewal while they generate daughter cells that may, but need not, have the capacity to differentiate into multiple cell types.
In summary, stem cells may be purely iteropotent having the capacity of self-renewal but not have transformative potency (totipotent, pluripotent, multipotent or oligopotent), or they may be both self-renewing and have transformative potency but they may loose their self-renewing capacity once they differentiate, or they may be strictly self-renewing but their daughter cells may have transformative potency. A cell with transformative potency may generate a stem cell. A stem cell may generate another stem cell. As we will see, all these conceptions of stem cells are realizable by different developmental stem cell networks.
Iteropotent versus transpotent stem cells
-----------------------------------------
For example, a transpotent stem cell such as a totipotent or pluripotent stem cell may differentiate into a skin cell or heart cell or neuron. Or it may transform into a progenitor cell that generates a limited number of new cells of various types. If, however, the stem cell is an iteropotent stem cell then it may, given the right conditions, continually generate skin cells, or heart cells or neurons. While the same stem cell may be both a transpotent as well as an iteropotent stem cell, it need not be. Some stem cells may be transpotent but not iteropotent while others may be iteropotent but not transpotent. Some stem cells may generate transpotent stem cells that are not iteropotent. And, as we will see iteropotent stem cells can generate other stem cells. In the latter case, we will call them [*meta-stem cells*]{}.
It can be even more intertwined. A daughter cell of a stem cell may stochastically or in response to an external signal, dedifferentiate into its parent stem cell control state. [*Differentiation*]{} includes a change in the cell’s control state. [*Dedifferentiation*]{} means that the cell control state jumps from its given control state to an earlier upstream control state in its global developmental network. The [*locality*]{} of the stem cell network in the global network determines the particular phenotypic properties of the stem cell and that stem cell’s progeny.
Thus as we change the theoretical framework from a gene-centered view of development to a control network view of development a new conceptual framework becomes available that allows more precise definitions of stem cells and their properties.
The more immature a daughter cell is, the more cell types it can differentiate into. The cell type it becomes may depend on the cell types of the cellular context into which it is born. The more mature the parent stem cell is the more constrained is the cell type of its daughters[^3] .
The transpotent stem cells can be totipotent cells, pluripotent, multipotent or oligopotent. A totipotent cell can generate an entire organism. However, it is not necessarily an iteropotent stem cell since it does not self-renew, only one embryo is formed from a given cell under normal development. Stem cell that produced totipotent cells is a theoretical possibility. Then each daughter cell of such a stem cell could produce an embryo. A pluripotent cell can differentiate into multiple cell types. Multipotent cells are even more restricted in the cell types they can differentiate into. These gradations are vague and difficult to define precisely but they are all based on the cell’s capacity to differentiate into a set of cell types that then may or may not be able to generate entire organisms, organs, multicellular tissue or terminate in one or more cells of some type. If we look at the control network guiding such cells then a totipotent cell is controlled by the global developmental control network or cenome of the organism. Transformative pluripotent cells are controlled by some subnetwork of the global network. Such transpotent networks need not have the architecture of iteropotent stem cells. Generative stem cells are defined by their developmental control network architecture or topology. It is this architecture that is the basis of their capacity to self-renew.
This hierarchy is based on a hierarchical network architecture. This architecture is also the basis of metastatic hierarchies in cancer. A group of totipotent stem cells interact with each other to prevent multiple embryos from forming, one potential embryo from each totipotent cell.
The molecular implementation of stem cell networks involves transcription factors and possibly as yet undiscovered RNA based regulators. The basis of all regulation is an addressing system that maps addresses to potential control areas of the genome and the cell.
Dedifferentiation of daughter cells into stem cells
---------------------------------------------------
Dedifferentiation potential in the daughter cell introduces yet another layer of complexity. A stem cell may produce a daughter cell that has the capacity to dedifferentiate into its parent stem cell state. The dedifferentiation may be stochastic based on some probability or it may be induced by cell signaling or other cellular contextual information. In that case the original stem cell is conditionally a meta-stem cell since it has parented a stem cell but it is also conditionally exponential since it has generated two cells of its own cell type. However, even exponential networks that are communication dependent need not proliferate exponentially (Werner [@Werner2011b]).
Stem cell locality
------------------
The global developmental control network [^4] is the control network responsible for the embryogenesis and development of multicellular organisms. Stem cells can be of various types that depend on the position that their own local control network has in the global developmental network. The ability of daughter cells of stem cells to assume various cell types depends on the developmental context in which they find themselves. The earlier the stem cell network is linked into and activated by the global developmental network the more likely that the stem cell daughter cells are immature and pluripotent, pluripotent meaning that they can differentiate into multiple cell types. If the stem cell network is linked later into a developmental path within the global control network then the more likely the stem cell will produce daughter cells that are less pluripotent, being relatively fixed in their phenotype.
Distinct stem cell concepts
---------------------------
Stem cell properties are the result of the properties of the developmental cell network that controls their behavior. Thus, the attributes of stem cells are directly linked to the properties of their controlling developmental network. Hence, using network properties we can distinguish different attributes of stem cells:
Developmental networks by definition are proliferative, meaning that they lead to one or more cell divisions. A network has potentially infinite proliferative capacity if it contains loops that lead to endless repetition of a developmental network.
1. [**Stem cell potency**]{} is the capacity of a cell to differentiate into various cell types such as being totipotent, pluripotent and oligopotent. This capacity is the result of an underlying developmental subnetwork. This developmental subnetwork may or may not be proliferative, but need not have self-renewing loops.
2. [**Stem cell self-generating capacity**]{} is determined by the stem cell network architecture which includes the hierarchical order of the self-renewing stem cell network, e.g. 1st-order, 2nd-order, kth-order geometric network.
3. [**Stem cell proliferative potential**]{} is the capacity of a cell to generate a set of cells when the network it links to is run. [^5] [^6]
4. [**Downstream potential of a net state**]{} Downstream accessible nodes of a given net state. A net state is a set of net nodes. The execution of a net state partitions the developmental space into a set of subspaces of multicellular systems. Each subspace is the result of different developmental paths due to various possible conditions.
5. [**The execution potential**]{} is the set of network paths that are accessible from a given cell network control state. The downstream network from a cell state is the subnetwork that is accessible from that cell state. It includes every accessible cell history from the given cell state.
A cell history is a path in the developmental network. Given a cell state $\sigma$ let $\Omega(\sigma) = \sigma^{*} = \{H \in \Omega \mid \exists t \in \Psi \wedge n \in \sigma, \text{ such that } H_{t}= n \}$ [^7]
6. [**Stem cell locality**]{} describes the position, by way of the linkage relationships, of the local stem cell network within the global developmental network of the organism. Stem cell locality may influence stem cell differentiation potency.
Note, that these distinct stem cell attributes are described in terms of the properties of the underlying stem cell network. Through these network properties we can distinguish the overt dynamic stem cell phenotype: [*Stem cell potency*]{} to differentiate into various cells types (totipotent, pluripotent, oligopotent) versus [*stem cell proliferative capacity*]{} (the order of the meta-stem cell network) versus [*stem cell network locality*]{} (where the stem cell network is linked into the global developmental network).
A so-called stem cell that has the capacity to differentiate into a subset of cell types may not even have stem cell proliferative capacity. For example, a totipotent embryonic stem cell need not be self-renewing even though it can generate stem cells that are self-renewing. Hence, the term stem cell is being used equivocally for different network properties. Differentiation potency is an independent dimension from stem cell proliferative capacity as governed by a stem cell network. To clearly define stem cells we need to define them through their underlying developmental networks.
Stem cell classification and methods that induce cell differentiation
---------------------------------------------------------------------
Furthermore, when an experimentalist manipulates a cell by various chemical or physical means to differentiate into some specific cell type, it does not mean that that cell has a natural potential to differentiate into that cell type. The [*natural differentiation potential*]{} is given by the cells local active developmental network. External experimental manipulations may force an [*artificial differentiation potential*]{} by forcing the activation of totally different developmental control network that then may give the cell a new proliferative dynamic and capacity to differentiate to a new, uncharacteristic phenotype.
While forced cell differentiation may have bioengineering and medical applications, it should not be used to define the capacity of the unforced, natural stem cell. That being said, placing the cell in a new multicellular context or the external introduction of cell signals such as hormones, while maintaining the cell’s integrity, can help decipher the conditional network activation potential of the cell. Hence, the boundary between forced and unforced cell differentiation may be difficult to delineate.
[2]{}
[^1]: Balliol Graduate Centre, Oxford Advanced Research Foundation (http://oarf.org), Cellnomica, Inc. (http://cellnomica.com). We gratefully acknowledge the use of Cellnomica’s Software Suite to construct the cancer and stem cell networks used to model and simulate all the multicellular processes that generated the [*in silico*]{} cancers described and illustrated in this paper. ©Werner 2016. All rightsreserved.
[^2]: I am not referring to the controversy as to whether stem cells are deterministic or have a stochastic dedifferentiation potential (Dick [@Dick2008]).
[^3]: We need to distinguish stem cell network hierarchy from the stem cell maturity which has to do with network locality.
[^4]: The global developmental network is called the cenome in (Werner [@Werner2011a; @Werner2011b]).
[^5]: The ambiguity: A cell may divide into two daughter cells. These in turn can divide further. So we need to distinguish the whole [**downstream potential**]{} of a cell from its immediate daughter generating potential. e.g., a stem cell because it is self-renewing has infinite potential over time. The problem is giving labels to different properties of developmental networks and their dynamic behavior in a developing system of cells.
[^6]: Distinguish network accessibility from the resulting execution of that accessible downstream network. Note, in the case of a first order stem cell that network can be very small, and yet has potentially infinite generative/proliferative potential. Because of the infinite loop, the downstream proliferative potential is of a 1st order stem cell is infinite.
[^7]: The $H$ in $\Omega$ can be infinite. The problem is that a state is a set of nodes which may be in different parts of the network. Hence, the nodes need not be on the same history or path. While the cell state history is linear, the control state is nonlinear since it can be on several paths simultaneously. Hence, the history of a “path” of a cell control state is a set of histories or linear paths. It can branch out or shrink in the future. So at any point it is a set of nodes, and over time it is a set of paths in the network. It is like parallel programs. There can be interpretation conflicts. Any control pointer can change as a result of cell signaling or a new one can be added. But the receptor and its interpretation has to be thrown first, unless it is part of the innate non-genomic or pre-genomic or epigenomic IES.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'It is shown how the exchange interaction, the dipole-dipole interaction, and the Dzyaloshinsky-Moriya interaction between electronic spin-density fluctuations emerge naturally from a field-theoretic framework that couples electrons to the fluctuating electromagnetic potential. Semi-quantitative estimates are given to determine when the dipole-dipole interaction, which is often neglected, needs to be considered, and various applications are discussed, with an emphasis on weak ferromagnets and on helimagnets.'
author:
- 'D. Belitz'
- 'T.R. Kirkpatrick'
title: 'Quantum Electrodynamics and the Origins of the Exchange, Dipole-Dipole, and Dzyaloshinsky-Moriya Interactions in Itinerant Fermion Systems'
---
Introduction {#sec:I}
============
Understanding the origin of ferromagnetism was one of the success stories of applying quantum mechanics to solid-state systems. Classically, magnetic moments interact via the dipole-dipole interaction, which is much too weak to explain magnetic order at as high a temperature as is observed in, e.g., iron or nickel.[@Ashcroft_Mermin_1976] The explanation of this conundrum was found to be the exchange interaction mechanism, which leads to a spin-spin interaction that is governed by the Coulomb interaction via the Pauli principle. This was first understood in the context of atomic and molecular physics in the 1920s, and applied to solid-state physics in the 1950s.[@Ashcroft_Mermin_1976] Somewhat ironically, a straightforward application of the exchange interaction concept leads to a spin-spin interaction that is too strong, as the relevant energy scale is the atomic scale, or roughly $100,000\,{\text K}$. Many-body and band-structure effects renormalize this scale and bring it down to the observed ferromagnetic scale of rougly $1,000\,{\text{K}}$ or lower.[@Lonzarich_Taillefer_1985] This is still much larger than the dipole-dipole scale, and the latter is often neglected in the discussion of ferromagnets. When it is considered, e.g., for its influence on the critical behavior,[@Aharony_Fisher_1973; @Frey_Schwabl_1994; @Ma_1976] it is usually added phenomenologically to models that describe the exchange interaction. Another spin-spin interaction that has been of interest lately is the Dzyaloshinsky-Moriya (DM) interaction.[@Dzyaloshinsky_1958; @Moriya_1960] It results (in systems with suitable lattice structures) from the spin-orbit interaction, has been derived from microscopic models, and is believed to be responsible for the helical magnetic order observed in MnSi and FeGe.[@Bak_Jensen_1980] Rough estimates show that the DM interaction and the dipole-dipole interaction are of about the same strength, and should thus be considered together.[@Maleyev_2006] Furthermore, in weak ferromagnets, which order only at low temperatures, all three interactions can be comparable in strength, which can make the dipole-dipole and DM interactions crucial.
In this paper we provide a comprehensive derivation of all of these effects within one unified framework, namely, a field-theoretic description of electrons and photons. Starting with finite-temperature quantum electrodynamics (QED) coupled to a field-theoretic description of finite-density quasi-relativistic electrons we show that the exchange, dipole-dipole, and DM interactions all appear naturally upon integrating out the photons. The exchange and DM interactions arise from integrating out the scalar part of the electromagnetic potential; the dipole-dipole interaction, from integrating out the vector potential. Furthermore, the DM and dipole-dipole interactions are indeed of the same order in the relativistic corrections to the Schr[ö]{}dinger equation (i.e., of second order in $\vF/c$, with $\vF$ the Fermi velocity and $c$ the speed of light, or of second order in the fine structure constant $\alpha$).
Integrating out the fermions then leads to an effective theory for quantum magnets that generalizes and replaces the Hertz-Millis theory[@Hertz_1976; @Millis_1993] and its generalizations.[@Belitz_Kirkpatrick_Vojta_2005] More generally, the theory provides a derivation of spin-spin interactions in itinerant Fermi systems in general, whether or not they are in a parameter regime where they develop long-range magnetic order. Our results are therefore relevant, for instance, for fermionic atoms in optical traps or on optical lattices.[@Fregoso_Fradkin_2010]
This paper is organized as follows. In Sec. \[sec:II\] we consider, as a warm-up and to introduce various concepts, classical magnets, and show how the vector potential coupling to the magnetization gives rise to the dipole-dipole interaction. In Sec. \[sec:III\] we develop the technical machinery for dealing with quantum magnets and provide the derivations mentioned above. In Sec. \[sec:IV\] we discuss our results and provide a summary and conclusion. Some technical details are relegated to various appendices.
Effective Action for Classical Ferromagnets and Helimagnets {#sec:II}
===========================================================
We now proceed to derive an effective action for magnets that includes the effects of the fluctuating electromagnetic potential. We first consider the classical case as a warm-up; we will generalize to the quantum case in Sec.\[sec:III\].
Dipole-dipole interaction {#subsubsec:II.A}
-------------------------
Consider a classical model for a ferromagnet with a three-component order parameter ${\bm M}$. In addition to the field ${\bm M}({\bm x})$ we need to consider the electromagnetic vector potential ${\bm A}({\bm x})$, and the partition function $Z$ is given by \[eqs:2.1\] Z &=& D\[[M]{},[A]{}\] e\^[S\[[M]{},[A]{}\]]{} \[eq:2.1a\]\
&& D\[[M]{}\] e\^[-[F]{}\[[M]{}\]/T]{}. \[eq:2.1b\] The model is defined by specifying the action $S$, and in Eq. (\[eq:2.1b\]) we have anticipated integrating out the vector potential to obtain an effective action ${\cal F}$ in terms of the order parameter only. $T$ denotes the temperature, so ${\cal F}$ is the free energy in mean-field approximation. Throughout this paper we will use units such that Boltzmann’s constant and Planck’s constant are equal to unity, $k_{\text{B}} = \hbar = 1$.
For the order-parameter part of $S$, we consider an $O(3)$-symmetric $\phi^4$-theory, \[eqs:2.2\] S\_M = \_V d[x]{} . \[eq:2.2a\] $S_M$ represents a Landau-Ginzburg-Wilson (LGW) theory of an isotropic ferromagnet with volume $V \to \infty$. The parameter $t$ contains the exchange interaction that leads to a magnetic ordering transition. In mean-field approximation this transition occurs at $t=0$, with $t>0$ describing the paramagnetic phase, and $t<0$ the ferromagnetically ordered one. $a>0$ and $u>0$ are two additional model parameters, and $(\nabla{\bm M})^2 =
\partial_i\,M_j\,\partial^i\,M^j$. Here, and throughout the paper, summation over repeated vector, tensor, and spinor indices is implied unless otherwise noted. Note that $S_M$ is separately invariant under rotations in ${\bm M}$ (spin) space and real space, respectively.
The magnetization[@magnetization_footnote] ${\bm M}$ couples linearly to the curl of the magnetic vector potential ${\bm A}$: S\_ = \_V d[x]{} [M]{}([x]{})(([x]{})), \[eq:2.2b\] with $\muB = e/2\me c$ the Bohr magneton in terms of the electron charge $e$, the electron mass $\me$, and the speed of light $c$. ${\bm A}$ and ${\bm\nabla}\times{\bm A}$ transform as vectors in real space, and therefore $S_{\text{c}}$ is invariant only under co-rotations of spin space and real space. It is this coupling of the magnetization to the fluctuating vector potential that allows one to consider the magnetization as having a particular direction in real space. The vector potential is governed by S\_A = \_V d[x]{} , \[eq:2.2c\] with $\rho$ any real number. The first term in Eq. (\[eq:2.2c\]) is the magnetic energy, and the second term with coupling constant $1/\rho$ is a gauge fixing term. One popular choice is $\rho = 0$, which enforces a Coulomb gauge, ${\bm\nabla}\cdot{\bm A} = 0$; another one is the Feynman gauge, $\rho=1$.[@Ryder_1985; @gauge_fixing_footnote] Either choice ensures a finite ${\bm A}$-propagator. In Coulomb gauge, it is A\_i([k]{}) A\_j(-[k]{})= 4T . \[eq:2.3\] The vector potential can now be integrated out exactly, which leads to an effective action in terms of ${\bm M}$ only. Alternatively, we can consider the magnetic induction ${\bm B} = {\bm\nabla}\times{\bm A}$ the fundamental field to be integrated out. In that case, the gauge fixing condition needs to be replaced by a constraint that enforces the Maxwell equation ${\bm\nabla}\cdot{\bm B} = 0$. That is, the Eqs. (\[eq:2.2b\]), (\[eq:2.2c\]), and (\[eq:2.3\]) are replaced by[@monopoles_footnote] S\_ = \_V d[x]{} [M]{}([x]{})([x]{}), \[eq:2.2b’\] S\_A = \_V d[x]{} \_[0]{}, \[eq:2.2c’\] B\_i([k]{}) B\_j(-[k]{})= 4T(\_[ij]{} - [k]{}\_i [k]{}\_j). \[eq:2.3’\] Either way we find \[eqs:2.4\] &=& \_V d[x]{} + 2\^2 \_[k]{} d\_[ij]{}([k]{})M\_i([k]{})M\_j(-[k]{}), \[eq:2.4a\] with d\_[ij]{}([k]{}) = [k]{}\_i [k]{}\_j . \[eq:2.4b\] The terms generated by integrating out the vector potential we recognize as the leading contribution to the dipole-dipole interaction[@Aharony_Fisher_1973; @Ma_1976] plus a shift of the Landau parameter $t$ by $4\pi\muB^2$. The scalar potential $\varphi({\bm x})$, whose gradient is the electric field, does not lead to any magnetic interactions in a classical theory. This changes once the system is treated quantum mechanically, see Sec. \[sec:III\] below.
Renormalization, and higher order terms {#subsec:II.B}
---------------------------------------
The dipole-dipole operator $d_{ij}$, Eq. (\[eq:2.4b\]), transforms as a rank-two tensor in momentum (or real) space, and ${\bm M}$ transforms as a vector in spin space. Consequently, the dipole-dipole interaction is invariant under co-rotations in real space and spin space. This raises the question of other terms in the action that have the same symmetry properties. For instance, $({\bm\nabla}\cdot{\bm M})^2$ is allowed by symmetry. This term, and terms of higher order in the gradient, are generated by a renormalization of the action ${\cal F}$, as we now proceed to show.
A renormalization of the action ${\cal F}$ generates additional terms by virtue of the anisotropic $M$-propagator, which now reads M\_i([k]{})M\_j(-[k]{})= + . \[eq:2.5\] For instance, two-loop diagrams of the structure shown in Fig. \[fig:2.1\]
-0mm ![A diagram that generates a $({\bm\nabla}\cdot{\bm M})^2$ term.[]{data-label="fig:2.1"}](fig_2.1.eps "fig:"){width="3.0cm"}
both renormalize $d_{ij}({\bm k})$ and lead to a new vertex \_[k]{} k\_ik\_jM\_i([k]{})M\_j(-[k]{}), \[eq:2.6\] as well as to higher order anisotropic gradient terms. Equation (\[eq:2.6\]) represents the $({\bm\nabla}\cdot{\bm M})^2$ term that was mentioned above. Conversely, if one starts with a theory that contains a $({\bm\nabla}\cdot{\bm
M})^2$ term, which is allowed by symmetry and hence should be included in any Landau theory, then the leading dipole-dipole term will be generated in perturbation theory even if it was not included in the bare action. The complete LGW action for a classical, isotropic Heisenberg ferromagnet, up to terms quadratic in gradients and quartic in the order parameter, thus reads \[eqs:2.7\] &=& \_V d[x]{} \
&& - \_[V]{} d[x]{}d[y]{} M\_i([x]{}) d\_[ij]{}([x]{}-[y]{}) M\_j([y]{})\
&& + \_V d[x]{} (([x]{}))\^2. \[eq:2.7a\] Here $d_{ij}({\bm x}-{\bm y})$ is the Fourier transform of $d_{ij}({\bm k})$ in Eq. (\[eq:2.4b\]), namely, d\_[ij]{}([x]{}-[y]{}) = . \[eq:2.7b\] $r$ is the bare distance from the critical point in the effective LGW theory that takes into account the effect of the vector potential, and $a$, $u$, $d_0$, and $d_2$ are the remaining Landau parameters. As is clear from the above discussion, one expects the bare values of $d_0$ and $d_2$ to be small of order $1/c^2$ compared to the other parameters in natural units. These two parameters are usually set equal to zero in elementary treatments of classical Heisenberg ferromagnets.
Dzyaloshinsky-Moriya interaction, and helimagnets {#subsubsec:II.C}
-------------------------------------------------
The terms in the action so far are all even in the gradient operator, and hence invariant under spatial inversion. The spin-orbit interaction can eliminate this requirement by coupling the electron spins to the underlying lattice, provided the crystal structure is not inversion invariant. Dzyaloshinsky[@Dzyaloshinsky_1958] and Moriya[@Moriya_1960] (DM) showed that to linear order in the spin-orbit interaction $\gso$ the relevant term is \_ = d[x]{} [M]{}([x]{})( ([x]{})), \[eq:2.8\] with $c_{\,\text{DM}} \propto \gso$. At a classical level, this term is purely phenomenological. DM showed how to derive it in the context of quantum mechanics, and in Sec. \[sec:III\] we will see how it arises in the context of field theory.
Quantum Systems {#sec:III}
===============
We now turn to a quantum mechanical description of itinerant fermion systems in general, and certain types of magnetism in particular. We will show how the exchange interaction, the dipole-dipole interaction, and the DM interaction naturally arise in the context of a field-theoretic description of itinerant electrons. The former two lead to ferromagnetism, and the latter, if it is present, to helimagnetism.
Action {#subsec:III.A}
------
In Appendix \[app:A\] we list the complete action for free quasi-relativistic electrons, to order $1/c^2$, coupled to the electromagnetic field. Several terms in the complete action are not relevant for our present purposes. The only effect of the Darwin term, Eq. (\[eq:A.4g\]), is to modify the Coulomb interaction on length scales given by the electronic Compton wave length $\lambdae = 1/\me c$. The relativistic mass enhancement, the last term in Eq.(\[eq:A.4b\]), is a higher-order gradient term that is small compared to terms of the same form that are generated by renormalizing the final effective action. Finally, the Landau diamagnetic terms, Eq. (\[eq:A.4f\]), give rise to diamagnetism and, in the presence of an external magnetic field, Landau levels. These effects are physically very different from ferromagnetism or helimagnetism, and we do not consider them here. Finally, in quantum electrodynamics the Fadeev-Popov ghost field does not couple to any other fields. Its only effect is to subtract the contribution of the unphysical longitudinal photon polarization to the free energy, and one has to keep it only if one is interested in the absolute value of the latter. Neglecting all of these terms, we thus consider the following action: S\[[|]{},;A\_\] &=& dx [|]{}\_(x) \_(x)\
&&-60pt + dxA\_(x)A\_(x)\
&&-60pt - ie dx (x)n(x) + dx [B]{}(x)\_(x)\
&&-60pt + dx\_[\_1]{}(x)\_[\_1 \_2]{}((x)) \_[\_2]{}(x).\
\[eq:3.1\] Here $x \equiv ({\bm x},\tau)$ comprises real space position ${\bm x}$ and imaginary time $\tau$, and $\int dx \equiv \int_V d{\bm x} \int_0^{1/T} d\tau$. $\bar\psi_{\sigma}$ and $\psi_{\sigma}$ are Grassmann-valued fields for electrons with spin projection $\sigma$, and the first term in Eq.(\[eq:3.1\]) describes free electrons with chemical potential $\mu$.[@Negele_Orland_1988] $A_{\mu} \equiv (\phi,-A_i)$ ($\mu = 0,1,2,3$; $i=1,2,3$) denotes the electromagnetic potential, with $\phi$ the scalar potential and ${\bm A} = (A_1,A_2,A_3)$ the vector potential, and the second term in in Eq. (\[eq:3.1\]) describes the free electromagnetic field in Feynman gauge (i.e., $\rho=1$ in Eq. (\[eq:A.4c\])).[@Kaputsa_1989] Note that both 4-vector potential fields in the second term carry covariant indices; that is, the $A$-action is euclidian.[@Euclidian_footnote] $n(x) =
{\bar\psi}_{\sigma}(x)\,\psi_{\sigma}(x)$ and ${\bm n}_{\text{s}}(x) =
{\bar\psi}_{\sigma_1}(x)\,{\bm\sigma}_{\sigma_1 \sigma_2}\,\psi_{\sigma_2}(x)$ are the electronic number and spin density,[@magnetization_footnote] respectively, with ${\bm\sigma} = (\sigma^x,\sigma^y,\sigma^z)$ the Pauli matrices, and the third term in Eq. (\[eq:3.1\]) describes the coupling of the electrons to the electromagnetic field, with ${\bm B} =
{\bm\nabla}\times{\bm A}$ the magnetic induction.[@ionic_density_footnote] Finally, the last term in Eq. (\[eq:3.1\]) describes the spin-orbit interaction. Note that both terms coupling the scalar potential $\varphi$ to the fermions carry an extra factor of $i$ compared to what one might expect from the first quantized Hamiltonian. This has the same origin as the Euclidian metric mentioned above.[@Euclidian_footnote]
Equation (\[eq:3.1\]) describes a continuum model. Some of the effects we are interested in are present only in the presence of certain types of lattices, and we will comment later on the modifications that occur if the electrons are put on a lattice.
Integrating out the photons {#subsec:III.B}
---------------------------
The action, Eq. (\[eq:3.1\]), depends only bilinearly on the electromagnetic potential. The latter can therefore be integrated out exactly, albeit at the expense of creating four-fermion terms. The latter represent electron-electron interactions that are mediated by the exchange of virtual photons. Technically, we need the photon propagator, which we can read off the second term in Eq.(\[eq:3.1\]): \[eqs:3.2\] A\_(x)A\_(y)= \_[D]{}(x-y), \[eq:3.2a\] with (x-y) = -4(\_\^2/c\^2 + \^2)\^[-1]{}(x-y), \[eq:3.2b\] or, in Fourier space, (k) = 4/(\_n\^2/c\^2 + [k]{}\^2). \[eq:3.2c\] Here $k \equiv (i\Omega_n,{\bm k})$ comprises a bosonic Matsubara frequency $\Omega_n = 2\pi T n$ and a wave vector ${\bm k}$.
The result of integrating out the photons exactly is very complicated and involves interacting electronic modes in both the spin-singlet and spin-triplet channels, the particle-particle and particle-hole channels, and all angular-momentum channels. We will restrict ourselves to those terms that are most relevant for magnetism, i.e. interactions between spin-density fluctuations, or modes in the $s$-wave particle-hole spin-triplet channel.
### $O(1/c^0)$: Exchange interaction {#subsubsec:III.B.1}
We organize the various contributions to the effective electron-electron interaction in powers of $1/c$. To zeroth order only the term coupling $\varphi$ to the number density $n$ in Eq. (\[eq:3.1\]) contributes. Integrating out $\varphi$ leads to a Coulomb interaction \[eqs:3.3\] S\_ = - dxdy n(x)v\_([x]{}-[y]{}) (\_x - \_y) n(y), \[eq:3.3a\] with v\_([x]{}) = e\^2/. \[eq:3.3b\] Here we have neglected the dynamical nature of the photon propagator ${\cal D}$ and have replaced it by its value at $\Omega_n=0$. The reason for this approximation is that Fermi-liquid effects lead to a dynamical screening of the Coulomb interaction that is a much larger effect than the relativistic dynamics inherent in Eqs. (\[eqs:3.2\]).
Equation (\[eq:3.3a\]) contains number density fluctuations at all wavelengths. If one restricts the theory to interactions between long-wavelength fluctuations, then this interaction can be rewritten as a sum of parts that includes an interaction between spin-density fluctuations, see Ref. and Appendix \[app:B\]. The basic point is that an interaction between number density fluctuations at large wave numbers can be written as one between spin density fluctuations at small wave numbers. In an effective low-energy theory that contains only fluctuations at wave numbers smaller than some cutoff $\lambda$, Eq. (\[eq:3.3a\]) therefore contains a contribution S\_ = ’ dx [n]{}\_(x) \_(x), \[eq:3.4\] where the prime on the integral indicates that only the small-wave-number contributions (smaller than $\lambda$) to the spin density ${\bm n}_{\text{s}}$ are to be considered in order to avoid overcounting. As has been explained in Ref. , it is convenient to choose the cutoff $\lambda$ as a fixed fraction of the Thomas-Fermi screening wave number, and the spin-triplet interaction amplitude $\Gamma_{\text{t}}$ is a Fermi-surface average over $v_{\,\text{C}}({\bm k}-{\bm p})\,\Theta(\vert {\bm
k}-{\bm p}\vert - \lambda)$, with ${\bm k}$ and ${\bm p}$ pinned to the Fermi surface. The restriction to small wave numbers will be understood from now on, and we will drop the prime on integrals.
$S_{\text{ex}}$ is the exchange interaction between electronic spin-density fluctuations that leads ferromagnetism. For later reference we note that $\Gamma_{\text{t}}$ is dimensionally an energy times a volume which, in this unrenormalized theory, is on the order of a Rydberg times a Fermi volume, or $\Gamma_{\text{t}} \approx e^2/\kF^2$.
### $O(1/c^2)$: Dipole-dipole interaction {#subsubsec:III.B.2}
We now turn to terms of $O(1/c^2)$. We first consider the vector potential ${\bm A}$, which couples to the spin density via the ${\bm B}\cdot{\bm
n}_{\text{s}}$ term in Eq. (\[eq:3.1\]). Since the coupling is directly to the spin channel, no phase space decomposition is necessary and integrating out ${\bm A}$ proceeds as in the classical case, except that the ${\bm
A}$-propagator now is frequency dependent, see Eqs. (\[eqs:3.2\]). We will comment on the consequences of this frequency dependence in Sec. \[subsubsec:III.C.2\] below. Neglecting the dynamical aspects of the dipole-dipole interaction for now, we obtain a contribution to the effective action S\_ &=& 2\^2 dx [n]{}\_(x)\_(x)\
&&0 pt + dxdy (\_x - \_y) n\_\^i(x)d\_[ij]{}([x]{}-[y]{}) n\_\^j(y),\
\[eq:3.5\] with $d_{ij}$ from Eq. (\[eq:2.7b\]). The first term has the same form as the exchange interaction, Eq. (\[eq:3.4\]), but is much smaller, as $\muB^2
\approx \Gamma_{\text{T}}\,(\vF/c)^2$. The second one is the dipole-dipole interaction between the electron spins; if one replaces the electronic spin density by its quantum mechanical and thermal average one recovers the classical dipole-dipole term in Eq. (\[eq:2.4a\]) or (\[eq:2.7a\]).
### $O(1/c^2)$: Dzyaloshinsky-Moriya and related interactions {#subsubsec:III.B.3}
We now return to the effects of integrating out the scalar potential $\varphi$. To $O(1/c^2)$ the relevant contribution comes from the cross-term that multiplies the Coulomb ($\varphi\,n$) term and the spin-orbit (last) term in Eq. (\[eq:3.1\]). Contracting the two scalar potentials, integrating by parts, using Eq. (\[eq:C.2b\]), and keeping only terms that are bilinear in phase-space spin-density fluctuations, we obtain a contribution to the effective action S\_ &=& \_[ilm]{}\_[ijk]{} dxdy [D]{}(x-y) [|]{}\_[\_1]{}(x)\^l\_[\_1\_4]{}\
&&0 pt (\_[\_4]{}(y)) ([|]{}\_[\_3]{}(y)) \^m\_[\_3 \_2]{} \_[\_2]{}(x).\
\[eq:3.6\] After a Fourier transform, Eq. (\[eq:3.6\]) can be written
S\_ &=& ()\^2 \_[q,k,p]{} [D]{}([q]{}) (\_[\_1\_4]{}\_[\_3\_2]{}) . \[eq:3.7\] Here $k = ({\bm k},i\omega_n)$ comprises a wave vector ${\bm k}$ and a fermionic Matsubara frequency $\omega_n = 2\pi T(n+1/2)$, and $p$ and $q$ are used analogously. At this point we generalize to an effective interaction amplitude ${\cal D}_{{\bm k},{\bm p}}({\bm q})$ that depends on ${\bm k}$ and ${\bm p}$ in addition to ${\bm q}$. Such a structure is generated in perturbation theory from the bare theory, where the interaction amplitude is simply given by the gauge field propagator, as we demonstrate in Appendix \[app:D\]. We then have S\_ &=& ()\^2 \_[q,k,p]{} (\_[\_1 \_4]{}\_[\_3 \_2]{})\
&& \_[\_1]{}(k-q/2) [|]{}\_[\_3]{}(p+q/2) \_[\_4]{}(p-q/2) \_[\_2]{}(k+q/2). \[eq:3.8\] Hermiticity requires ${\cal D}^*_{{\bm k},{\bm p}}({\bm q}) = {\cal D}_{{\bm
k},{\bm p}}(-{\bm q})$ (see also Eqs. (\[eqs:B.3\])). We now employ the phase space decomposition explained in Appendix \[app:B\] and focus on the large-angle scattering term, Eq. (\[eq:B.4b\]). Projecting again on the spin density we obtain \[eqs:3.9\] S\_ = \_q [D]{}([q]{}) ([n]{}\_(q)\_(-q)), \[eq:3.9a\] where ([q]{}) &=& \_[[k]{},[p]{}]{} g\_[k]{}g\_[p]{} ([k]{})\
&&+ \_[[k]{},[p]{}]{} g\_[k]{}g\_[p]{} ([k]{}-[p]{})\
&& [D]{}\^[(1)]{}([q]{}) + [D]{}\^[(2)]{}([q]{}). \[eq:3.9b\]
Here g\_[k]{} = (\_[k]{} - ) \[eq:3.10\] is a function that results from the projection onto the spin density and pins ${\bm k}$ and ${\bm p}$ to the Fermi surface.[@projection_footnote] Equation (\[eq:3.9a\]) has the form of the result obtained by Moriya.[@Moriya_1960] In what follows, we discuss the nature of the vector ${\bm D}({\bm q})$ in some more detail.
${\bm D}^{(2)}$ has the form \[eqs:3.11\] \^[(2)]{}([q]{}) = i [q]{}\^[(2)]{} + O([q]{}\^3), \[eq:3.11a\] with ${\bm d}^{\,(2)}$ a real vector given by \^[(2)]{} &=& \_[[k]{},[p]{}]{} g\_[k]{}g\_[p]{} ([p]{}-[k]{}) . \[eq:3.11b\]
${\bm D}^{(1)}$ can be written \[eqs:3.12\] D\^[(1)]{}\_i([q]{}) = i D\_[ij]{}q\_j + O(q\^3), \[eq:3.12a\] with $D_{ij}$ a real rank-2 tensor given by D\_[ij]{} = \_[[k]{},[p]{}]{} g\_[k]{}g\_[p]{}\_[ilm]{}k\_l p\_m[E]{}\_j([k]{},[p]{}), \[eq:3.12b\] where \_i([k]{},[p]{}) &=& d[x]{}d[y]{}(x\_i-y\_i) e\^[-i([k]{}+[p]{})([x]{}+[y]{})/2]{}\
&&-20pt d[z]{} (([k]{}-[p]{}))[D]{}([x]{},[y]{};[z]{}) \[eq:3.12c\] with ${\cal D}({\bm x},{\bm y};{\bm z})$ the Fourier transform of ${\cal
D}_{{\bm k},{\bm p}}({\bm q})$ in analogy to Eq. (\[eq:B.2b\]). $D_{ij}$ has a symmetric part and an antisymmetric part. The latter can be combined with ${\bm D}^{(2)}$ above to form a contribution to ${\bm D}$ that we denote by \[eqs:3.13\] \^[(-)]{}([q]{}) = i[q]{} + O([q]{}\^3), \[eq:3.13a\] where d\_i = d\_i\^[(2)]{} + \_[[k]{},[p]{}]{} g\_[k]{}g\_[p]{} (p\_i k\_j - k\_i p\_j)[E]{}\_j([k]{},[p]{}). \[eq:3.13b\] The symmetric part can be written as a diagonal tensor plus a traceless rank-2 tensor, and the latter can be diagonalized by means of an orthogonal transformation that amounts to a spatial rotation. The symmetric part of $D_{ij}$ we thus can write \[eqs:3.14\] D\^[(+)]{}\_[ij]{} = D \_[ij]{} + a\_i\_[ij]{} \[eq:3.14a\] ([*no*]{} summation convention) where the $a_i$ obey \_i a\_i = 0. \[eq:3.14b\] If desirable, the $a_i$ can be explicitly constructed from Eq.(\[eq:3.12b\]). We thus have a second contribution to ${\bm D}$ that we denote by \^[(+)]{}([q]{}) = D[q]{} + i(
[c]{} a\_x q\_x\
a\_y q\_y\
a\_z q\_z
) + O([q]{}\^3), \[eq:3.15\] and ([q]{}) = [D]{}\^[(+)]{}([q]{}) + [D]{}\^[(-)]{}([q]{}). \[eq:3.16\]
Combining our results, and transforming back to real space, we now have \[eqs:3.17\] S\_ = S\_ + S\_’ + S\_”, \[eq:3.17a\] where S\_ &=& D dx [n]{}\_(x)(\_(x)),\
\[eq:3.17b\]\
S\_’ &=& dx [n]{}\_(x),\
\[eq:3.17c\]\
S\_” &=& dx (\_(x))([d]{}\_(x)). \[eq:3.17d\] $S_{\text{DM}}$ is the Dzyaloshinsky-Moriya interaction that is believed to be responsible for the helimagnetism observed in MnSi and FeGe.[@Bak_Jensen_1980] $S_{\text{DM}}'$ is a closely related term but absent in systems with cubic lattices (e.g., MnSi or FeGe) due to the constraint, Eq. (\[eq:3.14b\]). Finally, $S_{\text{DM}}''$ is another term that is allowed by symmetry, and is generated by the above derivation. All three of these terms are contained in Moriya’s general result,[@Moriya_1960] which takes the form of our Eq. (\[eq:3.9a\]) above, but the effects of $S_{\text{DM}}'$ and $S_{\text{DM}}''$ have, to our knowledge, not been discussed explicitly.
Note that a necessary condition for any of these interactions to be nonzero is that the system is not invariant under parity: Both ${\bm D}^{(1)}$ and ${\bm
D}^{(2)}$ can be nonzero only if ${\cal D}_{{\bm k},{\bm p}}({\bm q})$ is odd under ${\bm q} \to -{\bm q}$, or, equivalently, if ${\cal D}({\bm x},{\bm
y};{\bm z})$ is odd under ${\bm z} \to -{\bm z}$. This implies that the DM interaction requires a lattice that is not invariant under spatial inversion; in any continuum model, where the electron-electron interaction is necessarily even under parity, it vanishes. See Appendix \[app:D\] and Sec. \[sec:IV\] for further discussions of this point.
Fermionic action, and magnetic order parameter {#subsec:III.C}
----------------------------------------------
We now have the following result. After integrating out the photons, the effective fermionic action reads S\_\[|,\] &=& S\_0 + S\_’ + S\_ + S\_ + S\_ + S\_’ + S\_”\
&& S\_0’ + S\_\^. \[eq:3.18\] Here $S_0$ describes non-interacting electrons (either free electrons or band electrons, depending on the model considered), and $S_{\text{int}}'$ contains all interactions between modes other than the spin density, which we have not explicitly considered with the exception of the Coulomb interaction, Eqs.(\[eqs:3.3\]). Collectively we denote these two terms by $S_0'$. $S_{\text{ex}}$, $S_{\text{d-d}}$, $S_{\text{DM}}$, $S_{\text{DM}}$, and $S_{\text{DM}}''$ are the exchange, dipole-dipole, and Dzyaloshinsky-Moriya interactions given by Eqs. (\[eq:3.4\]), (\[eq:3.5\]), and (\[eqs:3.17\]), respectively. $S_{\text{ex}}$ and $S_{\text{d-d}}$ are always present; which, if any, of the terms in $S_{\text{DM}}$ are nonzero depends on the details of the lattice structure and absence of spatial inversion symmetry is a prerequisite for any of them to be nonzero. Collective, we denote the sum of these interactions in the spin-density or triplet channel by $S_{\text{int}}^{\text{t}}$.
### Structure of a magnetic order parameter description {#subsubsec:III.C.1}
For applications such as fermionic cold gases one will want to work directly with the fermionic action. For applications to magnets it is convenient to introduce a composite field ${\bm M}(x)$ whose expectation value is proportional to the magnetization. To this end we write \[eqs:3.19\] S\_\^ = dxdyn\_\^i(x)\_[ij]{}(x-y) n\_\^j(y), \[eq:3.19a\] with \_[ij]{}(x-y) &=& (x-y)\_ + \^2(\_x - \_y)d\_[ij]{}([x]{}-[y]{})\
&&-60pt + (x-y).\
\[eq:3.19b\] Here $(a\partial)_k = a_k\,\partial_k$ (no summation convention). We now decouple Eq. (\[eq:3.19a\]) by means of a Hubbard-Stratonovich transformation with a bosonic field ${\bm M}(x)$. Neglecting a constant contribution to the action, this allows us to write S\_\[[|]{},,[M]{}\] &=& S\_0’\[[|]{},\]\
&& - dxdy M\_i(x)\_[ij]{}(x-y)M\_j(y)\
&& + dxdy.\
\[eq:3.20\]
If one neglects the interacting part of $S_0'$ this action depends only bilinearly on the fermion fields, and one can formally integrate out the fermions in order to obtain a theory entire in terms of the order-parameter field ${\bm M}$. This is a generalization of, and replaces, the Hertz-Millis theory.[@Hertz_1976; @Millis_1993] However, in general this is not a good strategy since it amounts to integrating out soft excitations, which means that any order parameter theory will in general not be well behaved. Physically, these soft quasi-particle excitations can change the nature of the phase transition,[@Belitz_Kirkpatrick_Vojta_2005] or they themselves can become critical.[@Sachdev_2010] In either case they must be treated on equal footing with the order parameter fluctuations. It therefore is technically advantageous, and physically more transparent, to work with the coupled field theory represented by Eq. (\[eq:3.20\]).
### Comments on the magnetization dynamics {#subsubsec:III.C.2}
We conclude this section with a brief discussion of the dynamics of the dipole-dipole interaction, which we neglected in Sec. \[subsubsec:III.B.2\]. If we restore the frequency dependence of the photon propagator and expand in powers of the frequency, then the leading dynamical contribution of the dipole-dipole interaction to Eq. (\[eq:3.20\]) takes the form S\_\^ = \^2 \_k (\_[ij]{} - [k]{}\_i[k]{}\_j) M\_i(k) M\_j(-k). \[eq:3.21\] As long as $\Omega_n$ scales as $\vert{\bm k}\vert$, this scales the same as the $\vert\Omega_n\vert/\vert{\bm k}\vert$ term in the order-parameter theory that is induced by Fermi-liquid effects,[@Hertz_1976] but has a prefactor that is smaller by a factor of $(\vF/c)^2$. However, in classical dipolar magnets the order parameter is known to no longer be conserved.[@Frey_Schwabl_1994] That is, $\Omega$ scales as a constant, and this should be reflected in the quantum theory as well, although it is currently not known how this is realized. This suggests that the contribution shown in Eq. (\[eq:3.21\]) dominates the Fermi-liquid-induced dynamics in a scaling sense, although it has a small prefactor, and will become important at sufficiently long time scales.
Discussion, and Conclusion {#sec:IV}
==========================
We now discuss the significance of various interactions for a number of problems.
Energy scales, and the significance of the dipole-dipole interaction {#subsec:IV.A}
--------------------------------------------------------------------
As we mentioned in Sec. \[subsec:III.A\], the energy scale for the exchange interaction in the bare theory is the atomic scale, or roughly $100,000\,{\text{K}}$. The corresponding length scale is on the order of $1\,\AA$. This is not consistent with the experimental fact that magnetic ordering is observed only at much lower temperatures; e.g., on the order of $1,000\,{\text{K}}$ in Fe and Ni, on the order of room temperature in FeGe, and on the order of $30\,{\text{K}}$ in MnSi. The reason for this discrepancy lies in the fact that the bare theory is renormalized in quantitatively substantial ways, and the corresponding energy and length scales in the properly renormalized theory are consistent with experimental observations.[@Lonzarich_Taillefer_1985]
Equations (\[eq:3.5\]) and (\[eqs:3.17\]) show that the dipole-dipole and DM interactions, within the framework of the bare theory, are weaker than the exchange interaction by a factor of $(\vF/c)^2$, or about $10^{-4}$.[@dipole_strength_footnote] Relative to the bare exchange interaction, this implies an energy scale on the order of $10\,\text{K}$, which is comparable with the ordering temperature in MnSi. On the other hand, the length scale associated with the DM interaction (i.e., the pitch length of the spin helix[@pitch_footnote]) is only about $200\,\AA$ (in MnSi)[@Ishikawa_et_al_1985] to $700\,\AA$ (in FeGe),[@Lebech_Bernhard_Freltoft_1989] or only a factor of $10^{2}$ to $10^{3}$ larger than the atomic length scale.
These observations indicate that there are strong renormalizations, due to band-structure and many-body effects, of all terms in the bare action, and that different terms are renormalized in different ways in different materials. While this makes it hard to make general statements, the bare theory suggests that the DM interactions and the dipole-dipole interaction are generically comparable in strength, and in MnSi, for instance, both are expected to be a substantial fraction of the (greatly reduced by renormalizations) exchange interaction. We thus conclude that there is no a priori reason to neglect the dipole-dipole interaction in any system where the DM interaction is known to be important. This calls for a re-evaluation of a number of interesting problems, some of which we list in the following subsection.
Significance of the dipole-dipole interaction {#subsec:IV.B}
---------------------------------------------
We conclude by discussing a number of problems where the dipole-dipole interaction is either known to be important, or might be important, with an emphasis on low-temperature magnets and other fermion systems.
\(1) Classical Heisenberg ferromagnets. This problem was worked on in great detail by Aharony and Fisher[@Aharony_Fisher_1973] for the static critical behavior and by Frey and Schwabl[@Frey_Schwabl_1994] for the dynamical critical behavior. The renormalization group done by Aharony and Fisher started from a nonlocal order parameter theory, Eqs. (\[eqs:2.7\]), which leads to a somewhat nonconventional renormalization procedure. The nonlocality in Eqs.(\[eqs:2.7\]) is due to integrating out the gauge field fluctuations or photons. It would be interesting to repeat this calculation starting from the coupled local field theory given by our Eqs. (\[eqs:2.2\]) before these fluctuations are integrated out.
\(2) Classical helimagnets. The standard phase transition treatment for helimagnetism due to the DM interaction is due to Bak and Jensen.[@Bak_Jensen_1980] Neglecting the dipole-dipole interaction term they conclude that there is a fluctuation-induced first order phase transition from paramagnetism to helimagnetism. An interesting question is whether or not the dipole-dipole interaction modifies this conclusion. This seems especially relevant for MnSi where the phase transition is at low temperatures, i.e., it is a weak helimagnet.
\(3) It has been shown that in clean itinerant ferromagnets, the ferromagnetic transition is generically of first order at zero temperature.[@Belitz_Kirkpatrick_Vojta_1999; @Belitz_Kirkpatrick_Vojta_2005] This conclusion ignores the effects of the dipole-dipole interaction terms. It would be very interesting to investigate if dipolar interactions can modify this generic conclusion.
\(4) Phase ordering is an important problem in ferromagnets.[@Bray_1994] The dipole-dipole interaction terms has not been included in either the classical or quantum (zero temperature) ferromagnetic phase ordering problems. Simple arguments indicate it will be important.
\(5) Fermionic cold atom systems. Recently there has been a considerable amount of work on gases of fermions with dipolar interactions, see Ref. and references therein. These systems are important for fermions in optical lattices. The dipolar interactions also serve as a mechanism for liquid crystal like phase formation in fermion systems.
\(6) The dipole-dipole interaction term is important in the dynamics of classical antiferromagnets, both in the ordered phase, and near or at the phase transition if the systems is below its upper critical dimension.[@Frey_Schwabl_1994] Simple considerations suggest that they will also be important in low-dimensional (1+1 or 2+1) itinerant quantum antiferromagnets; see Sec. \[subsubsec:III.C.2\] above for one aspect of this problem.
The complete action to $O(1/c^2)$ {#app:A}
=================================
In this appendix we give the complete action for electrons interacting with electromagnetic fields in the weakly relativistic limit, up to and including terms of $O(1/c^2)$. Let $A_{\mu} = (\varphi,-{\bm A})$ be the 4-vector potential, ${\bm\sigma} = (\sigma_1,\sigma_2,\sigma_3)$ the Pauli matrices, and $\muB = e/2\me c$ the Bohr magneton. We use standard relativistic notation, with covariant and contravariant indices related by a Minkowski metric $g_{\mu\,\nu} = (+,-,-,-)$. Expanding the Dirac equation in powers of $1/c$, one obtains, to order $1/c^2$, the following Hamiltonian in first quantization,[@Bjorken_Drell_1964; @Davydov_1976; @higher_order_footnote] \[eqs:A.1\] = [H]{}\_ + [H]{}\_ + [H]{}\_ + [H]{}\_[m]{}. \[eq:A.1a\] Here ${\hat H}_{\text{P}}$ is the Pauli Hamiltonian, \_ &=& (-i-[A]{}([x]{},t))\^2 + e([x]{},t)\
&& - (([x]{},t)), \[eq:A.1b\] and \_ &=& ( ([x]{},t)), \[eq:A.1c\]\
[H]{}\_ &=& \^2 ([x]{},t), \[eq:A.1d\]\
[H]{}\_[m]{} &=& \^4. \[eq:A.1e\] describe the spin-orbit interaction, the Darwin term, and the relativistic mass correction, respectively. Via standard techniques,[@Negele_Orland_1988; @Kaputsa_1989] this theory can be reformulated in terms of an action that depends on fermionic (i.e., Grassmann-valued) field $\psi$ and its adjoint ${\bar\psi}$ as well as the $4$-vector-potential field $A_{\mu}$. For the partition function one obtains Z = Z\_AZ\_, \[eq:A.2\] where Z\_A = D\[[|]{},\]D\[A\_\] e\^[S\_A\[[|]{},;A\_\]]{}, \[eq:A.3\] with an action \[eqs:A.4\] S\_A\[[|]{},;A\_\] = S\_\[[|]{},\] + S\_A\[A\_\] + S\_\[[|]{},;A\_\]. \[eq:A.4a\] Here S\_\[[|]{},\] &=& dx [|]{}\_(x) \_(x) \[eq:A.4b\] describes the electrons with chemical potential $\mu$, and S\_A\[A\_\] &=& dxA\_(x)A\_(x)\
&&-20pt + dx\^2 \[eq:A.4c\] describes the electromagnetic fields, with $\rho\in {\cal R}$ a gauge fixing parameter. We use a 4-vector notation $x \equiv (\tau,{\bm x})$, $\int dx
\equiv \int d\tau\int d{\bm x}$ for space and imaginary time. $S_{\text{c}}$ describes the coupling between the fermions and the electromagnetic field; it contains four separate contributions: S\_\[[|]{},;A\_\] = S\_ + S\_ + S\_ + S\_. \[eq:A.4d\] Here S\_\[[|]{},;A\_\] &=& -ie dx (x)n(x)\
&& + dx [B]{}(x)\_(x) \[eq:A.4e\] is the Coulomb and Zeeman paramagnetic coupling that is included in the Pauli equation, with $n(x) = {\bar\psi}_{\sigma}(x)\,\psi_{\sigma}(x)$ and ${\bm
n}_{\text{s}}(x) = {\bar\psi}_{\sigma_1}(x)\,{\bm\sigma}_{\sigma_1
\sigma_2}\,\psi_{\sigma_2}(x)$. S\_\[[|]{},;A\_\] &=& -i2dx\_(x)[A]{}(x)\_(x)\
&& + idx ((x))n(x)\
&& - dx [A]{}\^2(x)n(x) \[eq:A.4f\] is the Landau diamagnetic coupling; S\_\[[|]{},;A\_\] = dx(\^2(x))n(x) \[eq:A.4g\] is the Darwin term that, in the relativistic hydrogen atom, leads to the so-called [*zitterbewegung*]{}, and S\_\[[|]{},;A\_\] &=& dx\_[\_1]{}(x)\_[\_1 \_2]{}\
&& ((x)) \_[\_2]{}(x) \[eq:A.4h\] is the spin-orbit coupling.
The second factor in Eq. (\[eq:A.2\]) is \[eqs:A.5\] Z\_ = D\[[|]{},\] e\^[-S\_\[[|]{},\]]{} \[eq:A.5a\] Here $\eta$ is a one-component Grassmannian field known as a Fadeev-Popov ghost field, with ${\bar\eta}$ its adjoint, that is governed by an action S\_\[[|]{},\] = dx(x)\_\^(x). \[eq:A.5b\]
Phase-space decomposition of interaction terms {#app:B}
==============================================
For completeness, in this appendix we briefly recapitulate the arguments that lead to the generation of a spin-spin interaction, Eq. (\[eq:3.4\]), from a density-density interaction, Eqs. (\[eqs:3.3\]). For further discussion, see Refs. , and , ; the latter also explain the relation to the work by Shankar.[@Shankar_1994]
Consider an electron-electron interaction with an interaction amplitude $W$. For simplicity we assume that the interaction is purely static, and translationally invariant, but otherwise general. The action has the form
S &=& d[x]{}\_1… d[x]{}\_4 d W([x]{}\_1 -[x]{}\_2, [x]{}\_3 - [x]{}\_4; ([x]{}\_3 + [x]{}\_4 - [x]{}\_1 - [x]{}\_2)/2) \_[\_1\_2\_3\_4]{} [|]{}\_[\_1]{}([x]{}\_1,)[|]{}\_[\_3]{}([x]{}\_3,)\
&& \_[\_4]{}([x]{}\_4,)\_[\_2]{}([x]{}\_2,), \[eq:B.1\] with $\tau$ a general rank-4 tensor. We define Fourier transforms \[eqs:B.2\] \_(k) = \_k e\^[-ikx]{} [|]{}\_(k)&,&\_(k) = \_k e\^[ikx]{}\_(k), \[eq:B.2a\]\
W\_[[k]{},[p]{}]{}([q]{}) &=& d[x]{}d[y]{}d[z]{} W([x]{},[y]{};[z]{}), \[eq:B.2b\] with $kx = {\bm k}\cdot{\bm x} - \omega_n \tau$ where ${\bm k}$ is a wave vector and $\omega_n = 2\pi T (n+1/2)$ a fermionic Matsubara frequency. Hermiticity requires \[eqs:B.3\] W\^\*\_[[k]{},[p]{}]{}([q]{}) &=& W\_[[k]{},[p]{}]{}(-[q]{}), \[eq:B.3a\]\
\^\*\_[\_1 \_2 \_3 \_4]{} &=& \_[\_2 \_1 \_4 \_3]{}. We then have \[eqs:B.4\] S &=& ()\^2 \_[[k]{},[p]{},[q]{}]{} W\_[[k]{},[p]{}]{}([q]{}) \_[\_1\_2\_3\_4]{}\_[\_1]{}(k-q/2)[|]{}\_[\_3]{}(p+q/2)\_[\_4]{}(p-q/2) \_[\_2]{}(k+q/2) \[eq:B.4a\]\
&=& ()\^2 \_[[k]{},[p]{},[q]{}]{} W\_[([k]{}+[p]{}-[q]{})/2,([k]{}+[p]{}+[q]{})/2]{}([p]{}-[k]{})\_[\_1\_2\_3\_4]{}\_[\_1]{}(k-q/2)[|]{}\_[\_3]{}(p+q/2) \_[\_4]{}(k+q/2) \_[\_2]{}(p-q/2)\
\[eq:B.4b\]\
&=& ()\^2 \_[[k]{},[p]{},[q]{}]{} W\_[([p]{}-[k]{}+[q]{})/2,([k]{}-[p]{}+[q]{})/2]{}([p]{}+[k]{})\_[\_1\_2\_3\_4]{} [|]{}\_[\_1]{}(-k+q/2) [|]{}\_[\_3]{}(k+q/2)\
&& \_[\_4]{}(-p+q/2)\_[\_2]{}(p+q/2). \[eq:B.4c\]
As long as all wave vectors are summed over, all three of these expressions are identical. If one restricts the summation in such a way that both $\vert{\bm
q}\vert$ and the modulus of the third argument of $W$ are smaller than a cutoff wave number $\lambda$, then we can represent the action as a sum of all three terms. They represent small-angle scattering, large-angle scattering, and $2\kF$-scattering, respectively. Alternatively, if one is interested in only one of these channels, one can pick the appropriate formulation of $S$, restrict oneself to small wave numbers, and neglect the other channels. For our purposes, we are interested in the large-angle scattering channel, Eq.(\[eq:B.4b\]). By choosing $\tau_{\sigma_1\,\sigma_2\,\sigma_3\,\sigma_4} =
\sigma^0_{\sigma_1\,\sigma_2}\, \sigma^0_{\sigma_3\,\sigma_4}$ and making use of Eq. (\[eq:C.2a\]) we obtain a term that has the structure of the exchange interaction, Eq. (\[eq:3.4\]) (in addition to a contribution to the number-density interaction). The resulting spin-triplet mode is more complicated than a pure spin density, but it has an overlap with the spin density and can be restricted to the latter by the projection technique explained in Ref. . Note that a repulsive Coulomb interaction results in an attractive exchange interaction due to a commutation of fermion fields that is necessary to write the result in the form of Eq. (\[eq:3.4\]).
Similarly, by choosing $\tau_{\sigma_1\,\sigma_2\,\sigma_3\,\sigma_4} =
\sigma^0_{\sigma_1\,\sigma_2}\, {\bm\sigma}_{\sigma_3\,\sigma_4}$ and making use of Eq. (\[eq:C.2b\]) we obtain the structure found in Sec.\[subsubsec:III.B.3\] (in addition to terms that couple the number density and the spin density).
Properties of Pauli matrices {#app:C}
============================
Here we give some properties of the Pauli matrices that were used in Sec.\[sec:III\]. Let ${\bm\sigma} = (\sigma^x,\sigma^y,\sigma^z) \equiv
(\sigma^1,\sigma^2,\sigma^3)$ be the Pauli matrices with the commutator property
\_[ijk]{}\^i\^j = i\^k (i,j,k = 1,2,3), \[eq:C.1\] and $\sigma^0$ the $2\times 2$ unit matrix. Then the following identities hold: \[eqs:C.2\] \^0\_[\_1 \_2]{} \^0\_[\_3 \_4]{} &=& \^0\_[\_1 \_4]{} \^0\_[\_3 \_2]{} + \_[\_1 \_4]{}\_[\_3 \_2]{},\[eq:C.2a\]\
\^0\_[\_1 \_2]{} \_[\_3 \_4]{} &=& \_[\_1 \_4]{} \^0\_[\_3 \_2]{} + \^0\_[\_1 \_4]{} \_[\_3 \_2]{}\
&& + \_[\_1 \_4]{}\_[\_3 \_2]{}. \[eq:C.2b\] Equation (\[eq:C.2a\]) is easily checked by a direct calculation, and Eq.(\[eq:C.2b\]) follows by multiplying Eq. (\[eq:C.2a\]) by ${\bm\sigma}$ from the right and using Eq. (\[eq:C.1\]).
Nonlocal electron-electron interaction {#app:D}
======================================
In this appendix we demonstrate that a general interaction amplitude of the type used in Eq. (\[eq:B.1\]) is generated in perturbation theory from an ordinary two-body interaction. For simplicity, we consider spinless fermions interacting via an interaction potential $V$; spin dependence or gradients, as in the last term of Eq. (\[eq:3.1\]), are easily added. We also consider a translationally invariant model; we will comment on this feature below.
The bare electron-electron interaction is described by a term in the action S = dxdy V(x-y)[|]{}(x)[|]{}(y)(y)(x). \[eq:D.1\] Our goal is to construct an effective interaction of the form S\_ &=& dx\_1 …dx\_4\
&& W(x\_1-x\_2,x\_3-x\_4;(x\_3+x\_4-x\_1-x\_2)/2)\
&& (x\_1) [|]{}(x\_3)(x\_4) (x\_2) \[eq:D.2\] that has a structure necessary for contributing to the tensor $D_{ij}$, Eqs.(\[eq:3.12b\], \[eq:3.12c\]), see Fig. \[fig:D.1\].
-0mm ![Bare interaction $V$ and effective interaction $W$.[]{data-label="fig:D.1"}](fig_D.1.eps "fig:"){width="6.0cm"}
The bare interaction corresponds to W\^[(1)]{}(x,y,;z) = (x)(y)V(z). \[eq:D.3\] This does not contribute to $D_{ij}$ since it enforces $x=y=0$. More generally, contributions to $W$ with the property $W(x,y,;z) = W(y,x;z)$ do not contribute to $D_{ij}$. At second order in $V$, there are several diagrams, both tree-level and one-loop diagrams, that have this property. Consider, however, the diagrams shown in Fig. \[fig:D.2\].
-0mm ![A second-order contribution to $W$.[]{data-label="fig:D.2"}](fig_D.2.eps "fig:"){width="6.0"}
They correspond to
W\^[(2)]{}(x,y;z) &=& (x)V(-y) dx’ G(-x’+y/2) G(x’+y/2) V(-x’+z/2)\
&& + (y)V(-x) dx’ G(-x’+x/2) G(x’+x/2) V(-x’+z/2), \[eq:D.4\]
where $G(x-y) = \langle\psi(x){\bar\psi}(y)\rangle$ is the electron propagator.
In order for $W$ to contribute to $D_{ij}$ we must have $W(x,y;-z) =
-W(x,y;z)$, see Eq. (\[eq:3.12c\]). From Eq. (\[eq:D.4\]) we see that this is the case if and only if $V(x)$ has an odd component $V^-(-x) = -V(x)$. As far as the contribution to $D_{ij}$ is concerned we can thus replace $V$ in Eq. (\[eq:D.4\]) by $V^-$, and this automatically ensures $W(y,x;z) =
-W(x,y;z)$. If $V$ is the propagator of a scalar field, such as the quantity ${\cal D}$ in Sec. \[sec:III\], then these symmetry properties can obviously not be realized. However, in a realistic solid-state model $V$ represents the screened Coulomb interaction, $V(x,y) = v_{\text{C}}({\bm x}-{\bm
y})/\epsilon({\bm x},{\bm y})$, and the dielectric function $\epsilon$ has a contribution from the lattice in addition to an electronic contribution. The latter will only have the symmetry of the space group of the lattice, and on a lattice without inversion symmetry one will have $\epsilon({\bm y},{\bm x})
\neq \epsilon({\bm x},{\bm y})$. Translational symmetry will also be broken, of course, but this is not necessary for making the DM interaction nonzero, as the above example shows. Note that one can consider a coarse-grained continuum theory for the DM interaction, see Eq. (\[eq:2.8\]), but the relevant Landau coefficient [*must*]{} depend on the underlying lattice and vanish if the true continuum limit is taken.
We would like to thank Achim Rosch and Thomas Vojta for discussions. This research was initiated at the Aspen Center for Physics, and supported by the National Science Foundation under Grant Nos. DMR-09-29966 and DMR-09-01907.
[36]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ** (, ), .
, ****, ().
, ** (, ), .
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, .
, ** (, ).
, ** (, ).
, ** (, ).
, , , ** (, ), .
, , , ****, ().
, .
, , , , , ****, ().
, , , ****, ().
, , , ****, ().
, ****, ().
, ** (, ).
, ** (, ).
, ****, ().
, ****, ().
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Measuring metallicity in the nuclear regions of AGNs is difficult because only a few lines are observed and ionization correction becomes a major problem. Nitrogen to carbon ratio has been widely used as an indicator for metallicity, but precise measurements have been lacking. We made such measurements for the first time using a wide baseline of ionization states with observations from FUSE, HST and . observations with FUSE were crucial in this effort. We measured super-solar metallicities in two AGNs and found that N/C does not scale with metallicity. This suggests that chemical enrichment scenario in nuclear regions of galaxies may be different from traditional models of metal enrichment.'
author:
- Smita Mathur
- Dale Fields
title: Metallicity measurements in AGNs
---
[ address=[The Ohio State University]{} ]{}
[ address=[Pierce College]{} ]{}
Metallicity in AGNs
===================
Understanding chemical evolution of galaxies is one of the fundamental problems in astronomy and active galaxies are no exception. We now understand that most, if not all galaxies exhibit an active phase some times in their life. Thus knowing metallicity in AGNs allows us understand the chemical enrichment in galactic nuclei which may differ substantially from the enrichment on galactic scales. AGNs also show signs of outflow. The role of such outflows, especially those from powerful quasars, in enriching the intergalactic medium with metals is unclear. How does the kinematics, dynamics and metal content of AGN outflows compare with those from super-winds on galactic scales? While these are all important questions to answer, we have no knowledge of metallicity in the circumnuclear regions of AGNs.
This is because measuring AGN metallicity is difficult. In the optical, where most of the measurements are made, only a handful of broad emission lines are observed, mostly of hydrogen. At high redshift, the rest-frame UV spectrum is observed with several metal lines such as , , . However, converting the emission line strengths to metallicity is not straight forward as it depends upon ionization balance, temperature, density and geometry. Nonetheless, attempts to estimate metallicities have been made, notably by Hamann & Ferland (1999 and references therein). They suggested the use of /as a metallicity indicator. Because of the secondary C-N-O nucleosynthesis, N/C scales as metallicity Z, so N scales like Z$^2$. Using these method, Hamann & Ferland showed that high redshift quasars have super-solar metallicity.
The use of emission lines as metallicity indicators was questioned by Shemmer & Netzer (2002) who showed that /ratio and /ratio do not give consistent answers for N/C! Given the strong model dependence of emission line strengths, this was not a surprise. Absorption line strengths, on the other hand, are geometry and density independent and so are potentially better tracers of metallicity. Converting the observed column densities of absorption lines to column densities of metals still involves ionization correction; once again this is a difficult quantity to measure because only a handful of lines of different metals are observed in a given band. A long base-line of ionization states is required to make the ionization correction. As discussed below, we performed multi-wavelength observations toward this goal and made the first precise measurement of metallicity in an AGN.
Metallicity measurement in Mrk 1044
===================================
Figure 1 shows the HST STIS spectrum of Mrk 1044. Multiple velocity components of , and are clearly seen. A FUSE spectrum of Mrk 1044 is shown in figure 2. Here we see absorption lines of at the same velocities seen in the HST spectrum. Additionally we detect Ly$\beta$ absorption. We fit a “pseudo continuum” through the emission lines and the AGN continuum and measure the strengths of the absorption lines and determine their column densities (see Fields et al. 2005a,b for details).
The next step is to determine the ionization parameter of the absorption system. To this end we generated a grid of models of ionization parameter $\rm U$ and the total column density $N_H$ for solar metallicity and mixture using CLOUDY and looked for models consistent with observations. This is shown in figure 3; the hatched curves correspond to the locus on the $\log N_H--\log \rm U$ plane corresponding to the observed column density of each ion. The intersection of all the curves, if present, defines the $N_H$ and U of the system. The curves of , and actually do intersect at $\log \rm U =-1.29$. What is noteworthy, however, is that the curves for do not. For the inferred value of $\rm U$, the column density is significantly lower. This implies that the metals are more abundant, or that the metallicity of the system is super-solar. We measure the metallicity to be about 5 times solar.
What is also interesting is that the metal mixture is consistent with solar. In particular, we do not find the locus of displaced from the intersection of and . Thus, we find no evidence for N scaling as $Z^2$.
Metallicity in Mrk 279
======================
Encouraged by the success in measuring metallicity in Mrk 1044, we attempted to apply our technique to another AGN for which multi-wavelength observations existed. Mrk 279 was observed with , HST and FUSE. In contrast to UV/FUV bands, the X-ray region contains hundreds of lines spanning a range of ionization states and so potentially provides powerful diagnostics. We found that the spectrum of Mrk 279 did not have sufficiently high quality to make robust metallicity measurements. However, the data were better fit with a model containing super-solar metallicity.
A consistent picture is now emerging with high resolution grating spectroscopy of X-ray absorbing outflows. The absorber seems to be made up of two or more components in pressure balance with each other. Costantini et al. (2007) had reported that the two components in Mrk 279, the one with low ionization parameter (LIP) and the one with high ionization parameter (HIP) are not in pressure equilibrium. In fact the pressure–temperature curve in Costantini et al. did not have an equilibrium zone. We had shown previously that super-solar metallicity can restore an equilibrium zone in the pressure–temperature curve (Komossa & Mathur 2000). So we re-made a pressure–temperature plot for the absorbers in Mrk 279 with super-solar metallicity as suggested by the fit to the data. As shown in figure 5, this not only restored a equilibrium zone, but now the LIP and the HIP components were found to be in pressure balance. While not conclusive, this make the case for super-solar metallicity in Mrk 279 stronger.
We then compared our X-ray results with the UV data from Gabel et al. (2005). Only when we invoke the model with super-solar metallicity, could we match the X-ray and UV data. Given that the LIP X-ray component is generally found to be consistent with the UV absorber, this gives an additional supporting evidence for super-solar metallicity in Mrk 297 (Fields et al. 2007). Arav et al (2007) independently arrived at the same conclusion.
Conclusions
===========
Studies of Mrk 1044 and Mrk 279 have given strong evidence for super-solar metallicity in these AGNs. The Mrk 1044 result with five times solar metallicity is robust; first such precise measurement for an AGN. While we find metallicity to be overall super-solar, the abundance mixture is consistent with solar. Notably, we do not find any evidence for N/C scaling with metallicity. These studies imply that the chemical enrichment process in galactic nuclei is likely to be very different from traditional models. The Cosmic Origins Spectrograph (COS) to be installed on HST will prove to be invaluable in generalizing these results for a large number of AGNs.
N. Arav, et al. *ApJ*, **658**, 829 (2007).
E. Costantini, et al. *A&A* **461**, 121 (2007).
D. Fields, S. Mathur, R. Pogge, F. Nicastro, and St. Komossa, *ApJ*, **620**, 183 (2005a)
D. Fields, S. Mathur, R. Pogge, F. Nicastro, St. Komossa, and Y. Krongold *ApJ*, **634**, 928 (2005b)
D. Fields, S. Mathur, Y. Krongold, Rik Williams, and F. Nicastro, *ApJ*, **666**, 828 (2007)
F. Hamann & G. Ferland, *ARA&A*, **37**, 487 (1999)
St. Komossa and S. Mathur, *A&A*, **374**, 914 (2001)
O. Shemmer, and H. Netzer, *ApJ*, **567**, L19 (2002)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Does there exist a Lipschitz injection of ${{\mathbb Z}}^d$ into the open set of a site percolation process on ${{\mathbb Z}}^D$, if the percolation parameter $p$ is sufficiently close to $1$? We prove a negative answer when $d=D$, and also when $d\geq 2$ if the Lipschitz constant $M$ is required to be $1$. Earlier work of Dirr, Dondl, Grimmett, Holroyd, and Scheutzow yields a positive answer for $d<D$ and $M=2$. As a result, the above question is answered for all $d$, $D$ and $M$. Our proof in the case $d=D$ uses Tucker’s Lemma from topological combinatorics, together with the aforementioned result for $d<D$. One application is an affirmative answer to a question of Peled concerning embeddings of random words in two and more dimensions.'
address:
- 'Statistical Laboratory, Centre for Mathematical Sciences, Cambridge University, Wilberforce Road, Cambridge CB3 0WB, UK'
- 'Microsoft Research, 1 Microsoft Way, Redmond, WA 98052, USA; and Department of Mathematics, University of British Columbia, 121–1984 Mathematics Road, Vancouver, BC V6T 1Z2, Canada'
author:
- 'Geoffrey R. Grimmett'
- 'Alexander E. Holroyd'
bibliography:
- 'emb.bib'
date: 20 March 2010
title: Lattice embeddings in percolation
---
Introduction and results
========================
Preliminaries {#ss1}
-------------
Let ${{\mathbb Z}}^d$ denote the $d$-dimensional integer lattice. Elements of ${{\mathbb Z}}^d$ are called [[****]{}sites]{}. Let $\|\cdot\|_r$ denote the $\ell^r$-norm on ${{\mathbb Z}}^d$, and abbreviate $\|\cdot\|_1$ to $\|\cdot\|$. We say that a map $f:{{\mathbb Z}}^d\to{{\mathbb Z}}^D$ is $M$-[[****]{}Lipschitz]{}, or $M$-[[****]{}Lip]{}, if $\|f(x)-f(y)\|\leq M$ for all $x,y\in{{\mathbb Z}}^d$ with $\|x-y\|=1$.
For $p \in [0,1]$, consider the site percolation model on ${{\mathbb Z}}^D$. That is, declare each site to be [[****]{}open]{} (or $p$-[[****]{}open]{}) with probability $p$, and otherwise [[****]{}closed]{}, with different sites receiving independent designations. Let $W_p({{\mathbb Z}}^D)$ denote the random set of open sites, and write ${{\mathbb P}}_p$ and ${{\mathbb E}}_p$ for the associated probability measure and expectation operator.
We are interested primarily in the probability $$\label{h1}
L(d,D,M,p):= {{\mathbb P}}_p\Big(
\text{$\exists$ an $M$-Lip injection from ${{\mathbb Z}}^d$ to $W_p({{\mathbb Z}}^D)$}
\Big).$$ Clearly $L$ is increasing in $D$, $M$, and $p$, and decreasing in $d$. Furthermore, $L$ is $\{0,1\}$-valued, since ${{\mathbb P}}_p$ is a product measure and the event in is invariant under translations of ${{\mathbb Z}}^D$. We define the critical probability $${p_{\text{\rm c}}}(d,D,M):=\inf\{p:L(d,D,M,p)=1\},$$ and furthermore $${M_{\text{\rm c}}}(d,D):=\min\{M\geq 1: {p_{\text{\rm c}}}(d,D,M)<1\},$$ (where $\min{\varnothing}:=\infty$). That is, ${M_{\text{\rm c}}}(d,D)$ is the smallest $M$ such that, for some $p<1$, there exists, ${{\mathbb P}}_p$-a.s., an injective $M$-Lip map from ${{\mathbb Z}}^d$ to the open sites of ${{\mathbb Z}}^D$.
Note that $L(1,D,M,p)$ is simply the probability that there exists an open bi-infinite self-avoiding path in the graph with vertex-set ${{\mathbb Z}}^D$ and an edge connecting every pair of sites at $\ell^1$-distance at most $M$. It follows from standard percolation results that ${p_{\text{\rm c}}}(1,D,M)$ is the critical probability for site percolation on this graph (see, for example, [@g2], or [@lyons Proof of Theorem 3.9] for a proof for arbitrary graphs). Therefore, for $M\geq 1$, $${p_{\text{\rm c}}}(1,D,M) \begin{cases} =1 &\text{ if } D=1,\\
\in (0,1) &\text{ if } D\geq 2.
\end{cases}$$ We deduce in particular that ${p_{\text{\rm c}}}(d,D,M)>0$ for all $d,D,M\geq 1$. The problem of interest is to determine for which $d$, $D$, $M$ it is the case that ${p_{\text{\rm c}}}(d,D,M) = 1$.
Main result {#ss2}
-----------
\[main\]Let $d,D,M$ be positive integers.
For all $d$, we have ${p_{\text{\rm c}}}(d,d+1,2)<1$, and hence ${M_{\text{\rm c}}}(d,d+1)\le 2$.
For all $D\geq 2$, we have ${p_{\text{\rm c}}}(2,D,1)=1$, and hence ${M_{\text{\rm c}}}(2,D)>1$.
For all $d\geq 2$ and all $M$, we have ${p_{\text{\rm c}}}(d,d,M)=1$, and hence ${M_{\text{\rm c}}}(d,d)={\infty}$.
It is an elementary observation that if $d>D$ then $L(d,D,M,1)=0$ for all $M$, and hence ${M_{\text{\rm c}}}(d,D)={\infty}$. (To check this, suppose that $f:{{\mathbb Z}}^d \to {{\mathbb Z}}^D$ is an $M$-Lip injection, and let $S_n:= \{x\in {{\mathbb Z}}^d: \|x\|\le n\}$. Then $|S_n|$ has order $n^d$, but $|f(S_n)|$ has order at most $n^D$ ($<n^d$), in contradiction of the injectivity of $f$.) Therefore, the above results suffice to determine the values of ${M_{\text{\rm c}}}$ for all $d$, $D$, as summarized in Table \[table1\]. We note in particular that $$\label{mcoo}
{M_{\text{\rm c}}}(d,D) < {\infty}\quad \text{if and only if} \quad d<D.$$
$d \diagdown D$ 1 2 3 4 5 $\ldots$
----------------- ---------- ---------- ---------- ---------- ---------- ----------------
1 $\infty$ 1 1 1 1 …
2 $\infty$ $\infty$ 2 2 2 …
3 $\infty$ $\infty$ $\infty$ 2 2 …
4 $\infty$ $\infty$ $\infty$ $\infty$ 2 …
5 $\infty$ $\infty$ $\infty$ $\infty$ $\infty$ $\smash\ldots$
$\vdots$ $\ddots$
: The values of ${M_{\text{\rm c}}}(d,D)$ for $d,D \ge 1$.[]{data-label="table1"}
Theorem \[main\](a) is an immediate consequence of a substantially stronger statement proved in [@ddghs], which we state next. For $x=(x_1,\ldots,x_{d-1})\in{{\mathbb Z}}^{d-1}$ and $z\in{{\mathbb Z}}$ we write $(x,z):=(x_1,\ldots,x_{d-1},z)\in{{\mathbb Z}}^d$. Write ${{\mathbb Z}}_+:={{\mathbb Z}}\cap(0,\infty)$.
\[don\] Let $d\geq 2$ and suppose $p>1-(2d)^{-2}$. There exists ${{\mathbb P}}_p$-a.s. a (random) $1$-Lip function $F:{{\mathbb Z}}^{d-1}\to{{\mathbb Z}}_+$ such that for every $x\in{{\mathbb Z}}^{d-1}$, the site $(x,F(x))\in{{\mathbb Z}}^d$ is open.
With $F$ given as in Theorem \[don\], the map $x\mapsto (x,F(x))$ is evidently a $2$-Lip injection, thus establishing Theorem \[main\](a). Other applications of Theorem \[don\] appear in [@DDS; @g-h-sphere]. Further properties of $F$ are explored in [@GH4], where an improved bound on the value of $p$ in Theorem \[don\] is given.
The proof of Theorem \[main\](b) is relatively straightforward and may be found in Section \[nn\] (the proof involves showing that any $1$-Lip injection from ${{\mathbb Z}}^2$ to the full lattice ${{\mathbb Z}}^D$ must satisfy rather rigid conditions). The principal contribution of the current paper is Theorem \[main\](c). Interestingly, our proof of this non-existence result makes use of the above existence result, Theorem \[don\]. Another essential ingredient of this proof is Tucker’s Lemma from topological combinatorics (see [@matousek; @tucker]).
It is immediate from the definition of ${M_{\text{\rm c}}}(d,D)$ that, if ${M_{\text{\rm c}}}(d,D)={\infty}$, then ${p_{\text{\rm c}}}(d,D,M)=1$ for all $M\ge 1$. On the other hand, we have the following result when ${M_{\text{\rm c}}}(d,D)<{\infty}$ (which occurs if and only if $d<D$, as noted in above).
\[other0\] Let $d$, $D$ be positive integers such that ${M_{\text{\rm c}}}(d,D)<{\infty}$. Then ${p_{\text{\rm c}}}(d,D,M)\to 0$ as $M\to\infty$.
Embeddings of words {#ss3}
-------------------
The above results concerning maps from ${{\mathbb Z}}^d$ to the open sites of ${{\mathbb Z}}^D$ have implications in the more general setting of maps that preserve values indexed by ${{\mathbb Z}}^d$, as follows. Let ${\Omega}_d
:=\{0,1\}^{{{\mathbb Z}}^d}$ be the space of percolation configurations, in which the value $1$ ([respectively]{}, $0$) is identified with the state ‘open’ ([respectively]{}, ‘closed’). An [[****]{}embedding]{} of a configuration $\eta\in{\Omega}_d$ into a configuration ${\omega}\in{\Omega}_D$ is an injection $f:{{\mathbb Z}}^d\to{{\mathbb Z}}^D$ such that $\eta(x)={\omega}(f(x))$ for all $x\in{{\mathbb Z}}^d$. We call a configuration $\eta\in{\Omega}_d$ [[****]{}partially periodic]{} if there exist $x\in{{\mathbb Z}}^d$ and $r \in
{{\mathbb Z}}_+$ such that $\eta(x)= \eta(x+ry)$ for all $y \in {{\mathbb Z}}^d$.
\[other\] Let $d$, $D$ be positive integers.
Let $d \ge 2$, $p\in(0,1)$ and $\eta\in{\Omega}_d$. For ${{\mathbb P}}_p$-a.e. ${\omega}\in{\Omega}_D$, there exists no $1$-Lip embedding of $\eta$ into ${\omega}$.
Let $d<D$. For every $p\in(0,1)$, there exists $M\geq 1$ such that: for ${{\mathbb P}}_p$-a.e. ${\omega}\in{\Omega}_D$, it is the case that for every $\eta\in{\Omega}_d$, there exists an $M$-Lip embedding of $\eta$ into $\omega$.
Let $d=D$ and let $\eta\in{\Omega}_d$ be a partially periodic configuration. For every $p\in(0,1)$, $M \ge 1$, and for ${{\mathbb P}}_p$-a.e. ${\omega}\in {\Omega}_D$, there exists no $M$-Lip embedding of $\eta$ into ${\omega}$.
The current work was motivated in part by the problem of Lipschitz embeddings of random one-dimensional configurations (see [@G-demon; @glr]). Proposition \[other\](a) extends Theorem \[main\](b) to more general configurations than the all-$1$ configuration. Part (b) answers affirmatively a question posed by Ron Peled concerning the existence of $M$-Lip embeddings of $d$-dimensional random configurations into spaces of higher dimension; see [@G-demon Sect. 5]. Part (c) leaves unanswered the question of whether or not there exist $d
\ge 1$, $p\in(0,1)$, $\eta\in{\Omega}_d$, and $M <{\infty}$ such that: with strictly positive probability (and therefore probability $1$), there exists an $M$-Lip embedding from $\eta$ into a random configuration ${\omega}\in{\Omega}_d$ having law ${{\mathbb P}}_p$.
Quasi-isometry {#ss4}
--------------
There is a close connection between the existence of embeddings and of quasi-isometries. A [[****]{}quasi-isometry]{} between two metric spaces $(X,\delta)$ and $(Y,\rho)$ is a map $f:X\to Y$ such that: there exist constants $c_i\in(0,{\infty})$ with
$\forall\; x,x'\in X$, $c_1\delta(x,x')-c_2\leq\rho(f(x),f(x'))\leq
c_3\delta(x,x')+c_4,$
$\forall \;y\in Y$, $\exists \; x\in X$ such that $\rho(f(x),y)\leq c_5.$
We call such $f$ a ${\mathbf{c}}$-[[****]{}quasi-isometry]{} when we wish to emphasize the role of the vector ${\mathbf{c}}=(c_1,\ldots,c_5)$. It is not difficult to see that the existence of a quasi-isometry is a symmetric relation on metric spaces. Quasi-isometries of [*random*]{} metric spaces are discussed in [@peled]. A [[****]{}subspace]{} of a metric space $(X,\delta)$ is a metric space $(U,\delta)$ with $U\subseteq X$.
\[quasi\] Let $d$, $D$ be positive integers, and let $E$ be the event that there exists a quasi-isometry between $({{\mathbb Z}}^d,\ell^1)$ and some subspace of $(W_p({{\mathbb Z}}^D),\ell^1)$.
If $d<D$ then for all $p\in(0,1)$ we have ${{\mathbb P}}_p(E)=1$.
If $d\geq D$ then for all $p\in(0,1)$ we have ${{\mathbb P}}_p(E)=0$.
The proofs of Theorem \[main\](b,c) appear respectively in Sections \[nn\] and \[nse\]. The remaining propositions are proved in Section \[ebqi\]. Section \[openp\] contains four open problems.
Nearest-neighbour maps {#nn}
======================
In this section we prove Theorem \[main\](b). A (self-avoiding) [[****]{}path]{} in ${{\mathbb Z}}^d$ is a finite or infinite sequence of distinct sites, each consecutive pair of which is at $\ell^1$-distance $1$. Let $e_1,\ldots,e_d\in{{\mathbb Z}}^d$ be the standard basis vectors of ${{\mathbb Z}}^d$, and let $0$ denote the origin.
\[congruent\] Let $x_1,\ldots,x_k\in{{\mathbb Z}}^D$ be distinct, and let $A=A(x_1,\ldots,x_k)$ be the event that there exists a singly-infinite path $0=y_0,y_1,\ldots$ in ${{\mathbb Z}}^D$ such that the sites $(x_i+y_j: i=1,\ldots,k,\ j=0,1,\ldots)$, are distinct and open. If $p<(2D)^{-1/k}$ then ${{\mathbb P}}_p(A)=0$.
Let $A_n$ be the event that there exists a path $0=y_0,y_1,\ldots,y_n$ of length $n$ in ${{\mathbb Z}}^D$ such that the sites $(x_i+y_j: i=1,\ldots,k,\ j=0,\ldots,n)$ are distinct and open. Note that $A$ is the decreasing limit of $A_n$ as $n\to\infty$. Let $N_n$ be the number of paths $0=y_0,\ldots, y_n$ with the properties required for $A_n$. Then $${{\mathbb P}}_p(A_n)\leq {{\mathbb E}}_p N_n\leq (2D)^n p^{nk}\xrightarrow{n\to\infty}0,
\qquad\text{if }2Dp^k<1.$$ Here, $(2D)^n$ is an upper bound for the number of $n$-step self-avoiding paths $(y_j)$ starting from $0$, while for those paths for which the sites $x_i+y_j$ are distinct, $p^{nk}$ is the probability they are all open.
We must prove that, for any fixed $p<1$ and $D\geq 2$, a.s. there exists no $1$-Lip injection from ${{\mathbb Z}}^2$ to $W_p({{\mathbb Z}}^D)$.
First, suppose $f$ is a $1$-Lip injection from ${{\mathbb Z}}^2$ to the full lattice ${{\mathbb Z}}^D$, and consider the image of a unit square. Specifically, take $(i,j)\in{{\mathbb Z}}^2$ and let $$\begin{aligned}
r_1&=f(i+1,j)-f(i,j), & r_1'&=f(i+1,j+1)-f(i,j+1),\\
r_2&=f(i,j+1)-f(i,j), & r_2'&=f(i+1,j+1)-f(i+1,j).\end{aligned}$$ Note that: the four vectors $r_1$, $r_2$, $r_1'$, $r_2'$ are elements of $\{\pm e_j:j=1,\ldots,D\}$ (by the $1$-Lip property); they satisfy $r_1+r_2'=r_1'+r_2$ (by definition); the pair $r_1$, $r_2$ are neither equal to nor negatives of each other; and similarly for $r_1$, $r_2'$ (a consequence of injectivity). It follows that $r_1=r_1'$ and $r_2=r_2'$. Since this holds for every unit square, for any distinct $i,i'\in{{\mathbb Z}}$, the images under $f$ of the two paths $\{(i,j): j\in{{\mathbb Z}}\}$ and $\{(i',j): j\in{{\mathbb Z}}\}$ are two disjoint self-avoiding paths that are translates of each other. (Another consequence, which we shall not need, is that there exists $\Delta\subset \{1,\ldots,D\}$ such that all horizontal edges have images in $\{\pm e_j:j\in \Delta\}$, and all vertical edges have images in the complement $\{\pm e_j: j \notin \Delta\}$).
Let $B$ be the event that there exist $x_1,x_2,\ldots\in{{\mathbb Z}}^D$ and a self-avoiding path $0=y_0,y_1,\ldots$ in ${{\mathbb Z}}^D$ such that the sites $(x_i+y_j: i\geq 1,j\geq 0)$ are distinct and open. The above argument implies that, if there exists a $1$-Lip injection $f:{{\mathbb Z}}^2\to W_p({{\mathbb Z}}^D)$, then $B$ occurs. We shall now show that ${{\mathbb P}}_p(B)=0$ for all $p<1$ and $d\geq 1$. Let $k$ be large enough that $p<(2D)^{-1/k}$. We define $B_k$ analogously to $B$, except in that we now require the existence of only $k$ sites $x_1,\ldots,x_k$. Lemma \[congruent\] implies that ${{\mathbb P}}_p(B_k)=0$, because $B_k$ is the countable union over all possible $x_1,\ldots,x_k$ of the events $A(x_1,\ldots,x_k)$. Finally, we have $B\subseteq B_k$.
The case of equal dimensions {#nse}
============================
In this section we prove Theorem \[main\](c). We denote integer intervals by ${\llparenthesis}a,b {\rrbracket}:=(a,b]\cap{{\mathbb Z}}$, etc. Fix any $d\geq 2$, $M\geq 1$ and $p\in(0,1)$. We will prove that a.s.there exists no $M$-Lip injection from ${{\mathbb Z}}^d$ to $W_p({{\mathbb Z}}^d)$.
The idea behind the proof is as follows. Suppose that $f$ is such an injection. By a [*hole*]{} we mean a cube of side length $M$ in ${{\mathbb Z}}^d$ all of whose sites are closed (actually, a slightly different definition will be convenient in the formal proof, but this suffices for the current informal sketch). Holes are rare (if $p$ is close to $1$), but the typical spacing between them is a fixed function of $d$, $M$, and $p$. We will consider the image under $f$ of a cuboid ${\llbracket}1,n{\rrbracket}^{d-1}\times{\llbracket}1,m{\rrbracket}\subset{{\mathbb Z}}^d$, where $m\gg n\gg 1$. We will arrange that the images of the two opposite faces ${\llbracket}1,n{\rrbracket}^{d-1}\times\{1\}$ and ${\llbracket}1,n{\rrbracket}^{d-1}\times\{m\}$ are far apart, and separated by a $(d-1)$-dimensional ‘surface of holes’ (at the typical spacing). This implies that image of the interior of the cuboid must pass through this surface, avoiding all the holes. To do so, the image must be in some sense be folded up so as to be locally $(d-1)$-dimensional, and this will give a contradiction to the injectivity of $f$ if $n$ is chosen large enough compared with the spacing of the holes.
In the case $d=2$ (and perhaps for other small values of $d$), the above ideas can be formalized using [*ad hoc*]{} geometric methods, but for general $d$ we need a more systematic approach. The surface of holes will be constructed using Theorem \[don\], and we will augment it with a colouring of the nearby open sites using $d-1$ colours, in such a way that the coloured sites separate the two sides of the surface from each other, but the sites of any given colour fall into well-separated regions of bounded size. Via the map $f$, this colouring will induce a colouring of the cuboid that contradicts a certain topological fact.
The following notation will be used extensively. A [[****]{}colouring]{} of a set of sites $U\subseteq{{\mathbb Z}}^d$ is a map $\chi$ from $U$ to a finite set $Q$. A site $u\in U$ is said to have [[****]{}colour]{} $\chi(u)\in Q$. We introduce the graph $G(U,\ell^r,k)$ having vertex set $U$ and an edge between $u,v\in U$ if and only if $0<\|u-v\|_r\leq k$. An important special case is the [[****]{}star-lattice]{} $G^*=G^*_d:=G({{\mathbb Z}}^d,\ell^\infty,1)$. Given a graph $G$ and a colouring $\chi$ of its vertex set, a $q$-[[****]{}cluster]{} (of $\chi$ with respect to $G$) is a connected component in the subgraph of $G$ induced by the set of vertices of colour $q$. The [[****]{}volume]{} of a cluster is defined to be the number of its sites.
We next state the two main ingredients of the proof: a topological result on colouring a cuboid, and a result on existence of random coloured surfaces in the percolation model.
\[tucker\] Let $d,n,m$ be positive integers, and consider a colouring $$\chi:{\llbracket}1,n{\rrbracket}^{d-1}\times {\llbracket}1,m{\rrbracket}\to\{-\infty,+\infty,1,2,\ldots,d-1\}.$$ If $\chi$ satisfies:
all sites in ${\llbracket}1,n{\rrbracket}^{d-1}\times\{1\}$ have colour $-\infty$;
all sites in ${\llbracket}1,n{\rrbracket}^{d-1}\times\{m\}$ have colour $+\infty$; and
no site of colour $+\infty$ is adjacent in $G^*$ to a site of colour $-\infty$,
then, for some $j\in\{1,2,\ldots,d-1\}$, $\chi$ has a $j$-cluster with respect to $G^*$ of volume at least $n$.
\[surface\] Fix $d \geq 2$, $J\geq 1$, and $p\in(0,1)$. There exist constants $K,C<\infty$ (depending on $d$, $J$, and $p$) such that ${{\mathbb P}}_p$-a.s. there is a (random) colouring $$\lambda:W_p({{\mathbb Z}}^d)\to\{-\infty,+\infty,1,2,\ldots,d-1\}$$ of the open sites of ${{\mathbb Z}}^d$ with the following properties.
No site of colour $+\infty$ is adjacent to a site of colour $-\infty$ in $G(W_p({{\mathbb Z}}^d),\ell^\infty,J)$.
For each $j\in\{1,2,\ldots,d-1\}$, every $j$-cluster with respect to $G(W_p({{\mathbb Z}}^d),\ell^\infty,J)$ has volume at most $K$.
There exists a (random) non-negative real-valued function $g:{{\mathbb Z}}^{d-1}\to[0,\infty)$, with the Lipschitz property that $|g(u)-g(v)|\leq d^{-1} \|u-v\|_1$ for all $u,v\in{{\mathbb Z}}^{d-1}$, such that all open sites in $$S^-:=\big\{ (u,z): u\in{{\mathbb Z}}^{d-1},\; z < g(u)\}$$ are coloured $-\infty$, while all open sites in $$S^+:=\big\{ (u,z): u\in{{\mathbb Z}}^{d-1},\; z>g(u)+C\}$$ are coloured $+\infty$.
In (c) above, note in particular that all open sites in the half-space ${{\mathbb Z}}^{d-1}\times{\llparenthesis}-\infty,0{\rrparenthesis}$ are coloured $-\infty$.
The proof of Theorem \[main\](c) will proceed by playing Propositions \[tucker\] and \[surface\] against one another to obtain a contradiction. The number of permitted colours is crucial — if one colour more were added to $1,\ldots,d-1$ then the conclusion of Proposition \[tucker\] would no longer hold, while with one colour fewer, the conclusion of Proposition \[surface\] would not hold. It should also be noted that the use of the star-lattice $G^*$ is essential in Proposition \[tucker\] — the statement does not hold for the nearest-neighbour lattice $G({{\mathbb Z}}^d,\ell^1,1)$.
The choice of the Lipschitz constant $d^{-1}$ in Proposition \[surface\](c) above is relatively unimportant — the result would hold for any positive constant, while any constant less than $(d-1)^{-1}$ would suffice for our application (see Lemma \[intersection\] below).
Our proof of Proposition \[tucker\] will use Tucker’s Lemma, a beautiful result of topological combinatorics. The general version of [@lefschetz; @tucker] applies to triangulations of a ball, and is a close relative of the Borsuk–Ulam Theorem; see [@matousek] for background. We need only a special case, for the cuboid, which is also proved in [@baker].
For $t\in {\llbracket}1,\infty{\rrparenthesis}^d$, consider the integer cuboid $T=T(t):={\llbracket}0,t_1 {\rrbracket}\times\cdots\times {\llbracket}0,t_d{\rrbracket}\subset{{\mathbb Z}}^d$ with opposite corners $0$ and $t$, and define the boundary $\partial T:=T\setminus [{\llparenthesis}0,t_1 {\rrparenthesis}\times\cdots\times {\llparenthesis}0,t_d{\rrparenthesis}]$. We say that boundary sites $x,y\in\partial T$ are [[****]{}antipodal]{} if $x+y=t$.
\[tucker-cube\] Let $T\subset {{\mathbb Z}}^d$ be a cuboid as above, and suppose $\beta:T\to\{\pm 1,\ldots,\pm d\}$ is a colouring such that for each antipodal pair $x,y\in\partial T$ we have $\beta(x)=-\beta(y)$. Then there exist $u,v\in T$ that are adjacent in $G^*$ (and, in fact, that satisfy $u_i\leq v_i\leq u_i+1$ for all $i$) such that $\beta(u)=-\beta(v)$.
Throughout the proof, adjacency and clusters refer to $G^*$. The ($\ell^\infty$-)[[****]{}diameter]{} of a cluster is the maximum $\ell^\infty$-distance between two of its sites. It suffices to show that for a colouring $\chi$ satisfying the given conditions, there is a $j$-cluster of diameter at least $n$ for some $j\neq\pm\infty$. Suppose that this is false. We will construct a modified colouring that leads to a contradiction.
First define a colouring $\chi'$ of the larger cuboid $T:={\llbracket}0,n+1{\rrbracket}^{d-1}\times {\llbracket}0,m+1{\rrbracket}$ as follows. Let $\chi'$ agree with $\chi$ on $T\setminus\partial T$, except with colour $\infty$ everywhere changed to $d$, and $-\infty$ changed to $-d$. Colour $\partial T$ as follows. For each $i=1,\ldots,d$, let $\chi'$ assign colour $-i$ to the face $\{x\in T:x_i=0\}$, and colour $+i$ to the antipodal face (this rule creates conflicts at the intersections of faces; for definiteness assign such sites the candidate colour of smallest absolute value). Thus $\chi'$ satisfies the condition of Lemma \[tucker-cube\] on the boundary.
Now let $\beta$ be the colouring of $T$ obtained by modifying $\chi'$ as follows. For each $i=1,\ldots,d-1$, recolour with colour $-i$ all $i$-clusters that are adjacent to the face coloured $-i$. Since there were no $i$-clusters of diameter as large as $n$ in $\chi$, this does not affect the colours on $\partial T$. Hence Lemma \[tucker-cube\] applies, so there are adjacent sites $u,v\in T$ with $\beta(u)=-\beta(v)$, which contradicts the manner of construction of $\beta$.
The proof of Proposition \[surface\] relies on Theorem \[don\] concerning Lipschitz surfaces in percolation, together with the following deterministic fact.
\[tile\] For any integers $d\geq 1$ and $R\geq 2d$, there exists a colouring $\alpha:{{\mathbb Z}}^d\to\{0,1,\ldots,d\}$ with the following properties.
The colouring is periodic with period $R$ in each dimension; that is, $\alpha(x+Ry)=\alpha(x)$ for all $x,y\in{{\mathbb Z}}^d$.
For each $j\in\{0,1,\ldots,d\}$, every $j$-cluster with respect to $G^*$ has volume at most $R^d$.
The $0$-clusters with respect to $G^*$ are precisely the cubes\
$Rx+{\llbracket}-(d-1),(d-1){\rrbracket}^d$, for $x\in{{\mathbb Z}}^d$.
(0,0.45)(0,0) (-0.625,0.35)[$0$]{}(-0.585,0.3)[(0,-1)[0.15]{}]{} (-0.3,0.35)[$1$]{}(-0.26,0.3)[(0,-1)[0.15]{}]{}
\
(0,0)(0,0) (0.5,0.1)[$2$]{}(0.45,0.15)[(-1,0)[0.65]{}]{} (0.5,0.32)[$1$]{}(0.45,0.37)[(-1,0)[0.65]{}]{} (0.5,0.54)[$0$]{}(0.45,0.65)[(-2,1)[0.85]{}]{}
\
(0,0)(0,0) (0.04,1.1)[$(3)$]{} (0.1,1.4)[$2$]{}(0.05,1.45)[(-1,0)[0.8]{}]{} (0.1,1.7)[$1$]{}(0.05,1.75)[(-1,0)[0.45]{}]{} (0.1,2.0)[$0$]{}(0.05,2.05)[(-1,0)[0.3]{}]{}
The construction is illustrated in Figure \[slabs\]. Define a [[****]{}slice]{} to be any set of sites of the form $Y=Rx+(I_1\times\cdots\times I_d)$, where $x\in{{\mathbb Z}}^d$ and each $I_i$ is either $\{0\}$ or ${\llbracket}1,R-1{\rrbracket}$. If ${\llbracket}1,R-1{\rrbracket}$ appears $k$ times in this product then we call $Y$ a $k$-[[****]{}slice]{}. The set of all slices forms a partition of ${{\mathbb Z}}^d$. Let $a_k:=d-1-k$. For a $k$-slice $Y$, define the associated $k$-[[****]{}slab]{} to be the set obtained from $Y$ by replacing each occurrence of $\{0\}$ in the product $I_1\times\cdots\times I_d$ with ${\llbracket}-a_k,a_k{\rrbracket}$ (thus ‘thickening’ the slice by distance $a_k$). We now define the colouring: for each site $x$, let $\alpha(x)$ be the smallest $k$ for which $x$ lies in some $k$-slab.
The required properties (a) and (c) are immediate (the cubes in (c) are precisely the $0$-slabs). For (b), note that any $k$-cluster is contained within a single $k$-slab; it is straightforward to check that, for any given $k$, any connection in $G^*$ between two different $k$-slabs is prevented by sites of smaller colours (here it is important that $a_k$ is strictly decreasing in $k$). The volume of a $k$-slab is $(R-1)^k(2a_k+1)^{d-k}< R^d$.
In the following, we sometimes refer to the $d$ coordinate as vertical, with positive and negative senses being up and down respectively, and the other coordinate directions as horizontal.
(0.5,0)(0,0) (0.1,0.35)[$-\infty$]{}(0.05,0.4)[(-1,0)[0.3]{}]{} (0.1,0.81)[$1$]{}(0.05,0.86)[(-1,0)[0.3]{}]{} (0.1,1.27)[$+\infty$]{}(0.05,1.32)[(-1,0)[0.3]{}]{}
\
(0.5,0)(0,0) (0.1,1.9)[$1$]{}(0.05,1.95)[(-1,0)[0.45]{}]{} (0.1,2.3)[$2$]{}(0.05,2.35)[(-1,0)[0.3]{}]{} (0,1)[$(-\infty)$]{} (0,2.9)[$(+\infty)$]{}
See Figure \[fig-surf\] for an illustration of the construction. Let $L$ be a large constant, a multiple of $J$, to be determined later, and let $\alpha$ be the colouring from Lemma \[tile\] with parameters $d-1$ and $R:=L/J$. Let $\alpha'$ be the colouring obtaining by dilating $\alpha$ by a factor $J$, that is, for $u\in{{\mathbb Z}}^{d-1}$, let $\alpha'(u)=\alpha([ u/J])$ (where $[v]$ denotes $v$ with each co-ordinate rounded to the nearest integer, rounding up in the case of ties). Note from property (a) in Lemma \[tile\] that $\alpha'$ has period $L$ in each dimension, while by (b), for $j\in\{0,1,\ldots,d-1\}$, each $j$-cluster of $\alpha'$ with respect to $G({{\mathbb Z}}^{d-1},\ell^\infty,J)$ has volume at most $L^{d-1}$. Write $r:=J(2d-3)$. From Lemma \[tile\](c), the $0$-clusters of $\alpha'$ are $(d-1)$-dimensional cubes of side length $r$ centred (approximately) at the elements of the lattice $L{{\mathbb Z}}^{d-1}$.
For $u\in{{\mathbb Z}}^{d-1}$ we will use $\alpha'(u)$ to determine the colours (other than $\pm\infty$) assigned by $\lambda$ to sites in the vertical column $\{u\}\times{{\mathbb Z}}$. Colours $1,\ldots,d-1$ will be used unchanged, while colour $0$ will be treated in a different way.
We now introduce a renormalized percolation process, starting with certain sets to be used in its definition. For a site $x=(x_1,\ldots, x_d)\in {{\mathbb Z}}^d$, write $\underline{x}:=(x_1,\ldots,x_{d-1})\in{{\mathbb Z}}^{d-1}$ and $\overline{x}:=x_d$, so that $x=(\underline{x},\overline{x})$. Let $\underline{C}_{\underline{x}}\subset{{\mathbb Z}}^{d-1}$ be the $0$-cluster of $\alpha'$ centred at $L\underline{x}$. Let $s:=\lfloor L/d\rfloor$ (where $\lfloor\cdot\rfloor$ denotes the integer part). Let $\overline{C}_{\overline{x}}$ be the interval ${\llbracket}s\overline{x},s(\overline{x}+1){\rrparenthesis}\subset{{\mathbb Z}}$. Define the [[****]{}cell]{} corresponding to $x\in{{\mathbb Z}}^d$ to be the set of sites $C_x:=\underline{C}_{\underline{x}}\times
\overline{C}_{\overline{x}}$. Thus, each cell is a cuboid of height $s$, and side length $r$ in each horizontal dimension. The centres of the cells are spaced at distance $s$ vertically (so that they abut each other), and at distance $L$ horizontally.
Define a [[****]{}hole]{} to be any cube of the form $z+{\llbracket}1,r{\rrbracket}^d$, where $z\in{{\mathbb Z}}^d$, all of whose sites are closed in the percolation configuration. We say that the cell $C_x$ is [[****]{}holey]{} if it contains some hole as a subset. Now we return to the issue of choosing $L$. Since a hole has volume $r^d$ (a function of $J$ and $d$), and a cell has height $s=\lfloor L/d\rfloor$, we may choose $L$ a sufficiently large multiple of $J$ (depending on $J$, $d$, and $p$) so that the probability that a cell is holey exceeds $1-(2d)^{-2}$. For later purposes, ensure also that $L$ is large enough that $s>J$ and $\lfloor(L-r)/2\rfloor>J$. By Theorem \[don\], there exists a.s. a $1$-Lip function $F:{{\mathbb Z}}^{d-1}\to{{\mathbb Z}}_+$, such that all the cells $C_{(u,F(u))}$ for $u\in {{\mathbb Z}}^{d-1}$ are holey.
We specify next a set of sites surrounding each of the holey cells considered above, to be coloured according to $\alpha'$. For any $\underline{x}\in{{\mathbb Z}}^{d-1}$, let $\underline{B}_{\underline{x}}$ be the cube $\{v\in{{\mathbb Z}}^{d-1}:
[v/L]=\underline{x}\}$ (so that these cubes partition ${{\mathbb Z}}^{d-1}$). Let $\overline{B}_{\overline{x}}$ be the interval ${\llbracket}s\overline{x},s\overline{x}+L{\rrparenthesis}$. Define the [[****]{}block]{} corresponding to $x\in{{\mathbb Z}}^d$ to be the set of sites $B_x:=\underline{B}_{\underline{x}}\times
\overline{B}_{\overline{x}}$. Thus $B_x$ is a cube of side $L$ which contains the cell $C_x$ (at its bottom-centre).
Now we define the colouring $\lambda$. For each $u\in
{{\mathbb Z}}^{d-1}$, call the block $B_{(u,F(u))}$ [[****]{}active]{}. To each open site $y\in B_{(u,F(u))}$, assign the colour $\alpha'(\underline{y})$, provided this is one of the colours $1,2,\ldots ,d-1$. For the remaining sites $y$ in the active block (those satisfying $\alpha'(\underline{y})=0$), we proceed as follows. Since the cell is holey, choose one hole $H_u\subset C_{(u,F(u))}$. Since the sites in $H_u$ are closed, they receive no colours. Assign colour $\infty$ to all open sites in the block that lie above the hole $H_u$, and assign colour $-\infty$ to those that lie below $H_u$. (We say that a site $x$ lies [[****]{}above]{} a set $S$ if for all $y\in S$ with $\underline{x}=\underline{y}$ we have $\overline{x}>\overline{y}$; [[****]{}below]{} is defined analogously). We have assigned colours to all open sites lying in active blocks. Finally, assign colour $\infty$ to all open sites that lie above some active block, and colour $-\infty$ to all those that lie below some active block.
Now we must check that the colouring $\lambda$ has all the claimed properties. For property (b), note first that if the function $F$ were constant, then each $j$-cluster for $j=1,\ldots,d-1$ would have volume at most $L^{d-1}\times
L=L^d$, since the colouring $\alpha'$ has merely been ‘thickened’ vertically to thickness $L$. The effect of taking a non-constant $F$ is to displace the active blocks in the vertical direction, and this clearly cannot make these clusters any larger, so we can take $K=L^d$.
Property (c) follows easily from the Lipschitz property of $F$. The constant $d^{-1}$ arises because for $u,v\in{{\mathbb Z}}^{d-1}$ with $\|u-v\|_1=1$, the centres of the corresponding blocks are at horizontal displacement $L$ from each other, and vertical displacement at most $s\leq L/d$. Once the function $g$ is determined for the centres of the blocks, it can be defined elsewhere by linear interpolation.
To check property (a), suppose on the contrary that there exist two sites $x,y$ with respective colours $+\infty,-\infty$ within $\ell^\infty$-distance $J$ of each other. If there is a single active block such that both $x$ and $y$ lie above, below or within it, this contradicts the presence of a hole (which has side length $r>J$) in the corresponding cell. Also, if one of $x,y$ lies within an active block then the other cannot lie above, below or within a different active block, since $\lfloor(L-r)/2\rfloor>J$. Therefore the only other case to consider is that $x$ and $y$ lie respectively above and below two different active blocks, say $B_{(u,F(u))}$ and $B_{(v,F(v))}$, for some $u,v\in{{\mathbb Z}}^{d-1}$. In this case we must have $\|u-v\|_\infty=1$ and therefore $|F(u)-F(v)|\leq
\|u-v\|_1\leq d-1$, so the height intervals $\overline{B}_{F(u)}$ and $\overline{B}_{F(v)}$ overlap by at least $L-(d-1)s\geq s>J$, giving again a contradiction.
To complete the proof of Theorem \[main\](c) we will need the following simple geometric fact in order to find an appropriate separating surface. For a vector $x=(x_1,\ldots,x_d)\in{{\mathbb R}}^d$, write $\widehat{x}_r$ for the $(d-1)$-vector obtained by dropping the $r$-coordinate.
\[intersection\] Let $a_{\pm 1},\ldots ,a_{\pm d}$ be positive constants and define for $i=1,\ldots,d$ the sets $$\begin{aligned}
A_i&:=\big\{x\in{{\mathbb R}}^d: x_i \leq d^{-1} \|\widehat{x}_i\|_1+a_i\big\}; \\
A_{-i}&:=\big\{x\in{{\mathbb R}}^d: x_i \geq -d^{-1} \|\widehat{x}_i\|_1-a_{-i}\big\}.\end{aligned}$$ Then $\bigcap_{i=\pm 1,\ldots,\pm d} A_i$ is bounded.
We may assume without loss of generality that the $a_i$ are all equal, to $a$ say. For $x\in A_i\cap A_{-i}$ we have $|x_i|\leq d^{-1} \|\widehat{x}_i\|_1+a$, hence for $x$ in the given intersection, summing the last inequality over $i$ gives $$\|x\|_1\leq \frac{d-1}{d}\|x\|_1+da,$$ hence $\|x\|_1\leq d^2 a$.
Fix $d\geq 2$, $M\geq 1$ and $p\in(0,1)$, and suppose that $f$ is an $M$-Lip injection from ${{\mathbb Z}}^d$ to $W_p({{\mathbb Z}}^d)$. Let $K$ be the constant from Proposition \[surface\] for the given values of $p$, $d$, and with $J:=dM$. Let $n:=K+1$. Let $N$ be large enough so that the image $f({\llbracket}1,n{\rrbracket}^{d-1}\times\{1\})$ is a subset of ${\llbracket}-N,N{\rrbracket}^d$.
Now apply Proposition \[surface\] (again with parameters $p$, $d$ and $J=dM$), but to the translated lattice having its origin at $(N+1)e_d$, to obtain (a.s.) a colouring of $W_p({{\mathbb Z}}^d)$ in which all open sites in ${{\mathbb Z}}^{d-1}\times
{\llparenthesis}-\infty,N{\rrbracket}$ have colour $-\infty$. Call this colouring $\lambda_d$, and let $S^+_d$ be the set corresponding to $S^+$ in Proposition \[surface\](c) (all of whose open sites are coloured $\infty$). Similarly, for each of the two senses of the $d$ coordinate directions, apply Proposition \[surface\] to the lattice rotated and translated so that the part of the half-axis at distance greater than $N$ from the origin is mapped to the positive $d$-axis. Thus we obtain $2d$ colourings $\lambda_{i}$ of $W_p({{\mathbb Z}}^d)$, with associated sets $S^+_i$, for $i=\pm 1,\ldots,\pm d$, such that all the colourings assign colour $-\infty$ to ${\llbracket}-N,N{\rrbracket}^d$, and $\lambda_i$ assigns colour $\infty$ to sites sufficiently far along the $i$ coordinate half-axis.
For each $i$ as above, let $S_i^{++}$ be the set of sites $y$ such that every site within $\ell^1$-distance $dMn$ of $y$ lies in $S_i^+$. We claim that $$Z:={{\mathbb Z}}^d\setminus \bigcup_{i=\pm 1,\ldots,\pm d} S_i^{++}$$ is a finite set. This follows from Lemma \[intersection\], because ${{\mathbb Z}}^d\setminus S_i^{++}$ lies in a set of the form $A_i$ in the lemma (here it is important the Lipschitz constant in Proposition \[surface\](c) is $d^{-1}$). Since $f$ is injective, it follows that, for some $m>1$, the site $f((1,\ldots,1,m))$ lies outside $Z$, and hence lies in $S_I^{++}$ for some $I$. Since $f({\llbracket}1,n{\rrbracket}^{d-1}\times\{m\})$ has $\ell^1$-diameter at most $dMn$, this implies that $f({\llbracket}1,n{\rrbracket}^{d-1}\times\{m\})$ is a subset of $S_I^+$, and is therefore coloured $\infty$ in $\lambda_I$.
Now define a colouring $$\chi:{\llbracket}1,n{\rrbracket}^{d-1}\times{\llbracket}1,m{\rrbracket}\to\{\infty,-\infty,1,2,\ldots,d-1\}$$ via $\chi:=\lambda_I\circ f$. By the construction, $\chi$ satisfies properties (a) and (b) of Proposition \[tucker\]. Now, if $x,y$ are adjacent sites in $G^*$ then $\|x-y\|_1\leq d$, and therefore the $M$-Lip property gives $$\|f(x)-f(y)\|_\infty \leq
\|f(x)-f(y)\|_1\leq dM=J,$$ so $f(x),f(y)$ are adjacent in $G(W_p({{\mathbb Z}}^d),\ell^\infty,J)$. Hence, property (i) in Proposition \[surface\] implies that $\chi$ has no two adjacent sites in $G^*$ with colours $+\infty$ and $-\infty$, which is property (c) of Proposition \[tucker\]. Therefore by Proposition \[tucker\], for some $j\neq\pm\infty$, $\chi$ has a $j$-cluster of volume at least $n$ with respect to $G^*$. Let $A$ be such a cluster. Since $f$ is injective, $f(A)$ also has volume at least $n$. But by the above observation on adjacency, $f(A)$ is a subset of some $j$-cluster of $\lambda_I$ with respect to $G(W_p({{\mathbb Z}}^d),\ell^\infty,J)$. This contradicts property (b) in Proposition \[surface\] because $n>K$.
Embedding and Quasi-isometry {#ebqi}
============================
We will use the following simple renormalization construction. Fix an integer $r\geq 1$. For a site $x=(x_1,\dots ,x_D)\in
{{\mathbb Z}}^D$ define the corresponding [[****]{}clump]{} (or $r$-[[****]{}clump]{}) to be the set of $r$ sites given by: $$K_x:=\big\{(x_1,\ldots x_{D-1},rx_D+i):i\in{\llbracket}0,r-1{\rrbracket}\big\}.$$ The clumps $(K_x:x\in{{\mathbb Z}}^D)$ form a partition of ${{\mathbb Z}}^D$, with the geometry of ${{\mathbb Z}}^D$ stretched by a factor $r$ in the $D$th coordinate. If $\|x-y\|=k$ then, for all $u\in
K_x$ and $v\in K_y$, we have $\|u-v\|\leq (2r-1)k$.
Let $d<D$, which is to say that ${M_{\text{\rm c}}}(d,D)<{\infty}$. Any given $r$-clump contains one or more open sites with probability $1-(1-p)^r$. If this probability exceeds ${p_{\text{\rm c}}}(d,D,M)$, there exists a.s. an $M$-Lip injection $f:{{\mathbb Z}}^d \to{{\mathbb Z}}^D$ such that, for each $y \in f({{\mathbb Z}}^d)$, the clump $K_y$ contains some open site. By choosing one representative open site in each such clump, we obtain a $(2r-1)M$-Lip injection from ${{\mathbb Z}}^d$ to $W_p({{\mathbb Z}}^D)$. Hence, $${p_{\text{\rm c}}}\big(d,D,(2r-1)M\big)\leq 1-\bigl(1-{p_{\text{\rm c}}}(d,D,M)\bigr)^{1/r}.$$ The claim follows by the monotonicity of ${p_{\text{\rm c}}}$ in $M$.
\(a) We may assume with loss of generality that $d=2\le D$. The proof follows that of Theorem \[main\](b) as presented in Section \[nn\], with one difference. Let $\eta\in{\Omega}_2$. The event $A=A(x_1,\dots,x_k)$ of Lemma \[congruent\] is redefined as the event that there exists a singly-infinite path $0=y_0,y_1,\ldots$ in ${{\mathbb Z}}^D$ such that: the sites $(x_i+y_j: i=1,\ldots,k,\ j=0,1,\ldots)$ are distinct, and ${\omega}(x_i+y_j) = \eta(i,j)$ for all such $i$, $j$. As in the proof of Lemma \[congruent\], ${{\mathbb P}}_p(A)=0$ whenever $\max\{p,1-p\}<(2D)^{-1/k}$. The proof is now completed as for the earlier theorem.
\(b) Let $d<D$ and write $m={M_{\text{\rm c}}}(d,D)<{\infty}$. Given $p\in(0,1)$, choose $r$ sufficiently large that any given $r$-clump contains both an open and a closed site with probability exceeding ${p_{\text{\rm c}}}(d,D,m)$. There exists a.s. an $m$-Lip injection $f : {{\mathbb Z}}^d \to {{\mathbb Z}}^D$ such that, for each $y \in
f({{\mathbb Z}}^d)$, the $r$-clump $K_y$ contains both an open and a closed site. Hence, for any configuration $\eta$, by choosing the open or the closed site as appropriate in each $r$-clump, we obtain a $(2r-1)m$-Lip embedding of $\eta$ into $\omega$.
\(c) Let $d \ge 1$, $M\ge 1$, and let $\eta\in{\Omega}_d$ be partially periodic. Without loss of generality, we may assume, for some $r \in {{\mathbb Z}}_+$ and all $y\in{{\mathbb Z}}^d$, that $\eta(ry) =
\eta(0)=1$. Let ${\omega}\in {\Omega}_d$ and assume there exists an $M$-Lip embedding $f$ from $\eta$ into ${\omega}$. Let $g:{{\mathbb Z}}^d\to{{\mathbb Z}}^d$ be given by $g(x) = f(rx)$. Then $g$ is an $rM$-Lip injection from ${{\mathbb Z}}^d$ into $W_p({{\mathbb Z}}^d)$. By Theorem \[main\](c), such an injection exists only for ${\omega}$ lying in some ${{\mathbb P}}_p$-null set.
\(a) We assume without loss of generality that $D=d+1$. Given $p<1$, take $r$ sufficiently large that a given $r$-clump in ${{\mathbb Z}}^D$ contains some open site with probability exceeding $1-(2D)^{-2}$. By Theorem \[don\], there exists a $1$-Lip map $F:{{\mathbb Z}}^{D-1}\to{{\mathbb Z}}_+$ such that for all $u\in{{\mathbb Z}}^{D-1}$, the $r$-clump $K_{(u,F(u))}$ contains some open site. By choosing an arbitrary open site to represent each such clump, we obtain a quasi-isometry of the required form.
\(b) Let $d\geq D$, and suppose that with positive probability there exists a quasi-isometry from $({{\mathbb Z}}^d,\ell^1)$ to some subspace of $(W_p({{\mathbb Z}}^D),\ell^1)$. We will prove that, for some $p'\in(0,1)$ and $M\ge 1$, there exists an $M$-Lip injection $g: {{\mathbb Z}}^d \to W_{p'}({{\mathbb Z}}^D)$, which will contradict .
Recall the parameters ${\mathbf{c}}=(c_1,c_2,\ldots ,c_5)$ in the definition of a ${\mathbf{c}}$-quasi-isometry, and let $Q_{\mathbf{c}}$ be the event that there exists a ${\mathbf{c}}$-quasi-isometry from $({{\mathbb Z}}^d,\ell^1)$ to some subset of $(W_p({{\mathbb Z}}^D),\ell^1)$. Since each $Q_{\mathbf{c}}$ is invariant under the action of translations of ${{\mathbb Z}}^D$, it has probability $0$ or $1$. Under the above assumption, the event $\bigcup_{\mathbf{c}}Q_{\mathbf{c}}$ has positive probability. By the obvious monotonicities in the parameters $c_i$, this union is equal to the union $\bigcup_{{\mathbf{c}}\in({{\mathbb Q}}\cap(0,\infty))^5} Q_{\mathbf{c}}$ over rational parameters, and hence there exists a [*deterministic*]{} ${\mathbf{c}}$ such that $Q_{\mathbf{c}}$ has probability $1$. We choose ${\mathbf{c}}$ accordingly, and let ${{\mathcal F}}_{\mathbf{c}}$ be the (random) set of quasi-isometries of the required type.
A quasi-isometry $f \in {{\mathcal F}}_{\mathbf{c}}$ is not necessarily an injection, but, by the properties of a ${\mathbf{c}}$-quasi-isometry, there exists $C=C(d,D,{\mathbf{c}})$ such that, for all $y\in W_p({{\mathbb Z}}^D)$ we have $|f^{-1}(y)|\leq C$. Let $r=C$, and take $p'\in(0,1)$ sufficiently large that, with probability at least $p$, every site in any given $r$-clump is $p'$-open. Let $f\in {{\mathcal F}}_{\mathbf{c}}$ be such that: for $y\in f({{\mathbb Z}}^d)$, every site in $K_y$ is $p'$-open. Since the pre-image under $f$ of any $y \in {{\mathbb Z}}^D$ has cardinality $C$ or less, we may construct an injection $g:{{\mathbb Z}}^d \to W_{p'}({{\mathbb Z}}^D)$ such that, for $y \in {{\mathbb Z}}^D$, every $x\in f^{-1}(y)$ has $g(x)\in K_y$, and furthermore distinct elements $x\in f^{-1}(y)$ have distinct images $g(x)$. It is easily seen that $g$ is an $M$-Lip injection for some $M=M(d,D,{\mathbf{c}})$.
Open Questions {#openp}
==============
Derive quantitative versions of Theorem \[main\]. For example, fix $d,M,p$, and let $N=N(n)$ be the smallest integer such that there exists an $M$-Lipschitz injection from the cube $[ 1,n]^d\cap{{\mathbb Z}}^d$ to the open sites of $[1,N]^d\cap{{\mathbb Z}}^d$ with probability at least $\tfrac12$. How does $N$ behave as $n\to\infty$?
For which graphs $G$ and which $M$ is it the case that for $p$ sufficiently close to $1$ there exists an $M$-Lipschitz injection from $V(G)$ to the open sites of $V(G)$ (where $M$-Lipschitz refers to the graph metric of $G$)? Theorem \[main\](c) shows that for ${{\mathbb Z}}^d$, no $M<\infty$ suffices. On the other hand, for the $3$-regular tree, $M=2$ suffices, by the well-known fact that percolation on a 4-ary tree contains a binary tree for $p$ sufficiently close to $1$.
We may interpolate between $1$-Lipschitz and $2$-Lipschitz maps as follows. Let $S$ be a subset of $\{-1,0,1\}^D$, and let $G$ be the graph with vertex set ${{\mathbb Z}}^D$ and an edge between $u,v$ whenever $u-v$ or $v-u$ belongs to $S$. For which $d$, $D$, and $S$ does there exist an injection from ${{\mathbb Z}}^{d}$ to the open sites of ${{\mathbb Z}}^{D}$ that maps neighbours in ${{\mathbb Z}}^d$ to neighbours in $G$?
Does there exist a configuration $\eta\in\{0,1\}^d$ such that, with positive probability there exists a Lipschitz embedding of $\eta$ into the percolation configuration $\omega$ on ${{\mathbb Z}}^d$? When $d=1$, this is related to the main problem of [@glr].
Acknowledgements {#acknowledgements .unnumbered}
================
We thank Olle Häggström and Peter Winkler for valuable discussions. GRG acknowledges support from Microsoft Research during his stay as a Visiting Researcher in the Theory Group in Redmond. This work was completed during his attendance at a programme at the Isaac Newton Institute, Cambridge.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'I show that the predicted densities of the inner dark matter halos in models of structure formation appear to be higher than estimates from real galaxies and constraints from dynamical friction on bars. This inconsistency would not be a problem for the model if physical processes that are omitted in the collisionless collapse simulations were able to reduce the dark matter density in the inner halos. I review the mechanisms proposed to achieve the needed density reduction.'
author:
- 'J. A. Sellwood'
---
Motivation
==========
I was invited to review secular evolution in disk galaxies. Rather than attempt a very superficial review of this vast topic, I here focus on dynamical friction. Several other possible topics could be included in a review of secular evolution, such as: scattering of disk stars, which I reviewed only recently (Sellwood 2008a); mixing and spreading of disks ( Sellwood & Binney 2002; Roskar 2008; Freeman, these proceedings); and the formation of pseudo-bulges (Kormendy & Kennicutt 2004; Binney, these proceedings).
The current paradigm for galaxy formation ( White, these proceedings) makes specific predictions for the dark matter (DM) densities in halos of galaxies. I first argue that halos of some barred galaxies are inconsistent with this prediction, and then consider whether DM halo densities could be lowered by internal galaxy evolution.
Inner Halo Density
==================
Attempts to measure the halo density and its slope in the innermost parts of galaxies are beset by many observational and modeling issues ( Rhee 2004; Valenzuela 2007), while the predictions from simulations in the same innermost region are still being revised, as shown earlier by White. It therefore makes sense to adopt a more robust measure of central density, such as that proposed by Alam, Bullock & Weinberg (2002). Their parameter, $\Delta_{v/2}$, is a measure of the mean DM density, normalized by the cosmic closure density, interior to the radius at which the circular rotational speed due the DM alone rises to half its maximum value. For those more familiar with halo concentrations, it is useful to note that for the precise NFW (Navarro, Frenk & White 1997) halo form, $\Delta_{v/2} =
672 c^3 /[\ln(1+c) - c/(1+c)]$, if $c$ is defined where the mean halo density is 200 times the cosmic closure value; thus $\Delta_{v/2}
\simeq 10^{5.5}$ for a $c=9$ halo. However, a further advantage of $\Delta_{v/2}$ is that it is not tied to a specific density profile.
![Figure reproduced from Macció (2008, with permission), to which I have added the large labeled points that are described in the text. The shaded regions show the $1-$ and $2-\sigma$ ranges of the predicted values of $\Delta_{v/2}$, while the lines show the means, as functions of the maximum circular speed from the DM halos. The small colored symbols show various estimates of these parameters for dwarf and LSB galaxies estimated by Macció from data in de Blok, McGaugh & Rubin (2001), de Blok & Bosma (2002), and Swaters (2003).[]{data-label="Delfig"}](DeltaV2.ps){width="10cm"}
Prediction
----------
Figure \[Delfig\], reproduced from Macció, Dutton & van den Bosch (2008), shows the prediction (shaded) that results when the initial amplitude and spectrum of density fluctuations match the latest cosmic parameters, as determined by the WMAP team (Komatsu 2008). I have added one further predicted point from the [*Via Lactea*]{} model (Diemand, Kuhlen & Madau 2007), which is argued to resemble a typical halo that would host a galaxy such as the Milky Way.
Data from Galaxies
------------------
Macció plot the small colored triangles, squares and pentagons, which show estimates of $\Delta_{v/2}$ culled from the literature. The data are from dwarf and LSB galaxies, that are believed to be DM dominated and the large black circle indicates their mean in both coordinates, with the error bars indicating the ranges. The open circle with error bars shows a revised mean after subtracting a contribution to the central attraction by the estimated baryonic mass in these galaxies. These authors conclude that these data are consistent with the model predictions.
The large cyan circles are estimates for large galaxies: both NGC 4123 (Weiner, Sellwood & Williams 2001) and NGC 3095 (Weiner 2004) have well-estimated halos that are significantly below the predicted range. The inner density estimated for NGC 1365 (Zanmár Sanchéz 2008), on the other hand, is $\Delta_{v/2} \sim 5 \times 10^7$, which is off the top of this plot, though the total halo mass is quite modest; a large uncertainty in the inclination, a possible warp that is very hard to model, together with evidence of a inner disturbance that required us to fit to data only one side of the bar, all conspire to render the inner halo density of this Fornax cluster galaxy quite uncertain.
I also plot two magenta points from different mass models for the Milky Way. The upper point is from Klypin, Zhao & Somerville (2002) while the lower shows the upper bound on the inner halo density estimated by Binney & Evans (2001). Their bound comes from trying to include enough foreground disk stars to match an old estimate (Popowski 2001) of the micro-lensing optical depth to the red-clump stars of the Milky Way bulge; current estimates of this optical depth are somewhat lower (Popowski 2005), suggesting a reanalysis will allow a higher halo density.
It is important to realize that the creation of a disk through condensation, or inflow, of gas into the centers dark halos must deepen the gravitational potential and cause the halo to contract ( Blumenthal 1986; Sellwood & McGaugh 2005). Thus, estimates of the current halo density should be reduced to take account of halo compression. Allowance for compression brings the point for NGC 1365 down by more than one order of magnitude. But making this correction for all the galaxies (which Macció did not do for their open circle point) will move all the data points down, including the well-estimated points for NGC 4123 & NGC 3095 that are already uncomfortably low.
The only real difficulties presented by the comparison with the predictions in Fig. 1 arise from two well-determined low points, which could simply turn out to be anomalous. Additional evidence suggesting uncomfortably low DM densities in real galaxies comes from other rotation curve data ( Kassin, de Jong & Weiner 2006) and the difficulty of matching the observed zero point of the Tully-Fisher relation ( Dutton, van den Bosch & Courteau 2008). However, an independent argument, based on the constraints from dynamical friction on bars, also suggests that the DM density in barred galaxies is generally lower than predicted.
Bar Slow Down {#slow}
-------------
Bars in real galaxies are generally believed to be “fast”, in that the radius of corotation is generally larger than the semi-major axis of the bar by only a small factor, $\cal R$. Indications that $1 \la
{\cal R} \la 1.3$ come from (a) direct measurements in largely gas-free galaxies, summarized by Corsini (2008), (b) models of the gas flow ( Weiner 2001; Bissantz, Englmaier & Gerhard 2003), and (c) indirect arguments about the location of dust lanes ( Athanassoula 1992). Rautiainen, Salo & Laurikainen (2008), and others, claim a few counter-examples from indirect evidence, although they concede that they try to match the morphology of the spiral patterns, which may rotate more slowly than the bar.
After some considerable debate, a consensus seems to be emerging that strong bars in galaxies should experience fierce braking unless the halo density is low (Debattista & Sellwood 1998, 2000; O’Neill & Dubinski 2003; Holley-Bockelmann, Weinberg & Katz 2005; Colín, Valenzeula & Klypin 2006). The counter-example claimed by Valenzuela & Klypin (2003) was shown by Sellwood & Debattista (2006) to have resulted from a numerical artifact in their code. The claims of discrepancies by Athanassoula (2003) are merely that weak, or initially slow bars, are less strongly braked, while she also finds that strong, fast bars slow unacceptably in dense halos.
![Bars in NFW halos. Above: The value of $\cal R$ at the time each simulation was stopped. Below: The number of disk rotations before the bar slowed to the point where ${\cal R}=1.4$.[]{data-label="rplot"}](rplot.ps){width="10cm"}
Thus there is little escape from the conclusion by Debattista & Sellwood (2000), that the existence of fast bars in strongly barred galaxies requires a low density of DM in the inner halo. Our original constraint required near maximal disks, although the halo models in that paper were not at all realistic. Figure \[rplot\] summarizes the results from a new study of exponential disks embedded in NFW halos, computed using the code described in Sellwood (2003) that has greatly superior dynamic range. When scaled to the Milky Way, a rotation period at 3 disk scale lengths in these models is 270 Myr. In all cases, the bar becomes slow by the end of the experiment, although the number of disk rotations needed until ${\cal R} > 1.4$ increases as the concentration index, $c$, is reduced. Thus, friction in NFW halos causes bars to become unacceptably slow in a few disk rotations when $c \ga 10$, but on a time scale $\ga 7.5\;$Gyr when $c
\la 6$ or when the uncompressed $\Delta_{v/2} \la 10^{5.1}$. This conservative bound would exclude well over half the predicted range of halo densities in Fig. \[Delfig\] for an uncompressed $V_{\rm max} =
10^{2.25} \simeq 180\;$km/s.
In the context of this symposium, it would be nice to test Milky Way models for bar slow-down. The halo and disk in the model tested by Valenzuela & Klypin (2003) were selected from the Klypin (2002) models for the Milky Way. The simulations reveal that a very large bar with semi-major axis $\ga 5\;$kpc forms quickly, which slows unacceptably within $\sim 5\;$Gyr. Unfortunately, the absence of a realistic bulge in these experiments crucially prevents these results from being regarded as a test of the Klypin (2002) MW models, since a bulge should cause a much shorter and faster bar to form.
Can the DM Density Be Reduced?
==============================
The model would not be challenged if the present-day DM density in galaxies can be reduced by processes that are neglected in cosmic structure formation simulations, which generally follow the dynamics of collisionless collapse only. Four main ideas have been advanced that might achieve the desired density decrease.
Feedback
--------
This first idea is not a secular effect, and therefore strictly falls outside my assigned topic. However, I discuss it briefly because it should not be omitted from any list of processes that might effect a density reduction.
The basic idea, proposed by Navarro, Eke & Frenk (1997), Binney, Gerhard & Silk (2001), and others is that gas should first collect slowly in a disk at the center of the halo, thereby deepening the gravitational potential well and compressing the halo adiabatically. A burst of star formation would then release so much energy that most of the gas would be blasted back out of the galaxy at very high speed, resulting in a non-adiabatic decompression of the halo, which may possibly result in a net reduction in DM density.
Gnedin & Zhao (2002) present the definitive test of the idea. In their simulations, they slowly grew a disk inside the halo, causing it to compress adiabatically, and then they instantaneously removed the disk. With this artifice, they deliberately set aside all questions of precisely how the star burst could achieve the required outflow, in order simply to test the extreme maximum that any conceivable feedback process could achieve.
They found that the final density of the halo was lower than the initial, confirming that the effect can work, and that mass blasted out from the very center of the potential has greatest effect, presumably because it produces the largest instantaneous change in the gravitational potential. However, density reductions by more than a factor of two required that the disk be unreasonably concentrated, and consequently the baryonic mass has to be blasted out from deep in the potential well.
Bar-Halo Friction
-----------------
The same physical process that slows bars, discussed in §\[slow\], can also reduce the density of the material that takes up the angular momentum, as first reported by Hernquist & Weinberg (1992). This mechanism prompted Weinberg & Katz (2002) to propose the following sequence of events as a means to reconcile halo predictions with bar pattern speed constraints and other data. They argued that a large bar in the gas at an early stage of galaxy formation could reduce the DM density through dynamical friction. The gas bar would then disperse as star formation proceeded, so that were a smaller stellar bar to form later it would not experience much friction. Their idea has been subjected to intensive scrutiny.
Normal Chandrasekhar friction ( Binney & Tremaine 2008, §8.1) is formally invalid in more realistic dynamical systems, such as quasi-spherical halos, because the background particles are bound to the system and will return to interact with the perturber repeatedly. Tremaine & Weinberg (1984) showed that under these circumstances angular momentum exchange occurs at resonances between the motion of the perturber and that of the background particles.
The $N$-body simulations mentioned in §\[slow\] generally did not produce a substantial reduction in halo density, despite the presence of strong friction. This could be because the bars were not strong enough, but Weinberg & Katz (2007a,b) argue that the simulations were too crude and that delicate resonances would not be properly mimicked in simulations unless the number of particles exceeds between $10^7$ & $10^9$, depending on the bar size and strength and the halo mass profile.
Thus two major questions arise: (1) are results from simulations believable? and (2) can realistic bars cause a large density decrease? I addressed both these issues in a recent paper (Sellwood 2008b). While rigid, ellipsoidal bars are not terribly realistic, I used them deliberately in order to test the analysis and to compare with the simulations presented by Weinberg & Katz. Dubinski (these proceedings) presents results of similar tests using fully self-consistent disks that form bars.
### What $N$ Is Enough?
Simple convergence tests reveal that experiments with different numbers of particles converge to an invariant time evolution of both the pattern speed and halo density changes at quite modest numbers of halo particles. I report that $N=10^5$ seemed to be sufficient for a very large bar, while $N\sim 10^6$ was needed for a more realistic bar. I observed no change the results in either case as I increased to $N=10^8$, or when I employed a spectrum of particle masses in order to concentrate more into the crucial inner halo. I found results for different numbers of particles overlaid each other perfectly, with no evidence for the stochasticity that Weinberg & Katz predicted should result if few particles were in resonance.
I also demonstrated that my simulations did indeed capture resonant responses that converged for the same modest particle numbers. I measured the change in the density of particles $F(L_{\rm res})$, where $L_{\rm res}$ is an angular momentum-like variable that depends on orbit precession frequency. Using this variable, I was able to estimate that some 7% – 20% of halo particles participated in resonant angular momentum exchanges with the bar during a short time interval. This fraction is vastly greater than Weinberg & Katz predicted, because they neglected to take into account the broadening of resonances caused by the evolving bar perturbation, that both grows and slows on an orbital timescale. Athanassoula (2002) and Ceverino & Klypin (2007) also demonstrated the existence of resonances in their simulations.
![Results from five different experiments with different bar lengths. The dashed line shows the initial profile, while the solid lines show estimates from the particles of the initial (cusped) and final (cored) density profiles from a series of runs with different bar semi-major axes. The final density lines from the lowest to the highest are for bar lengths, $a/r_s = 1$, 0.8, 0.6, 0.4, & 0.2.[]{data-label="length"}](lengthbw.ps){width="12cm"}
### Density Reduction by Very Strong Bars
Figure \[length\] shows that a strong bar rotating in a halo within a density cusp ($\rho \propto 1/r$) can flatten the cusp to $\sim 1/3$ bar length. The rigid bar needed to accomplish this must have an axis ratio $a/b \ga 3$, a mass $M_b \ga 30\%$ of the halo mass inside $r = a$.
While cusp flattening is a driven response caused by the slowing bar, it is also a collective effect. I find a much smaller change when I hold fixed the monopole terms of the halo self-gravity. Thus it is dangerous, when studying halo density changes, to include any rigid mass component.
The only simulation I am aware of in which a self-consistent bar flattened the inner density cusp is that reported by Holley-Bockelmann (2005). They report a significant density reduction that flattened the cusp to a radius $\sim a/5$. Note that in their model, the initial halo density was not compressed by the inclusion of the disk, since they rederived the halo distribution function that would be in equilibrium in the potential of an uncompressed NFW halo plus the disk.
![Fractional changes, $\mu$, to $\Delta_{v/2}$ in many experiments. The abscissae show the angular momentum given to the halo, expressed as the usual dimensionless spin parameter. Open circles mark results from experiments in which the density profile of the inner cusp was flattened, while squares indicate experiments where cusp flattening did not occur. Filled symbols show results from experiments in which the moment of inertia of the bar was increased by a factor 5 in all cases except the point at the upper right, where the MoI was increased 10-fold (for more details see Sellwood 2008). The changes to $\Delta_{v/2}$ make no allowance for halo compression.[]{data-label="ABW"}](ABW.ps){width="10cm"}
### More Gradual Changes
A number of other simulations in the literature have revealed a modest density reduction caused by angular momentum exchange with a bar in the disk. Debattista & Sellwood (2000) show a reduction in the halo contribution to the central attraction, and something similar can also be seen in Athanassoula’s (2003) simulations. None of these models included very extensive halos.
The inner halo density in fully self-consistent simulations with more extensive and cusped halos can actually rise as the model evolves (Sellwood 2003; Colín 2006). This happens because angular momentum lost by the bar in the disk causes it to contract; the deepening potential of the disk causes further halo compression that overwhelms any density reduction resulting from the angular momentum transferred to the halo.
### Angular Momentum Reservoir
A crucial consideration that limits the magnitude of halo density reduction by bar friction is the total angular momentum available in the baryonic disk. Tidal torques in the early universe lead to halos with a log-normal distribution of spin parameters with a mean $\lambda
\sim 0.05$, where the dimensionless spin parameter is $\lambda =
LE^{1/2}/GM^{5/2}$ as usual. Assuming that the baryons and dark matter are well mixed initially, the fraction of angular momentum in the baryons is equal to the baryonic mass fraction in the galaxy: some 5% – 15%. Thus total angular momentum loss from the disk could increase the halo spin parameter by typically $\delta\lambda \sim
0.005$.
Figure \[ABW\], taken from Sellwood (2008b), shows the factor by which the halo density is reduced, $\mu = \Delta_{v/2,\rm fin} /
\Delta_{v/2,\rm init}$ as the ordinate against the angular momentum gain of halo. A density reduction by a factor of 10 is possible, but the bar must be extreme, having a semi-major axis, $a$, approaching that of the break radius, $r_s$, of the NFW profile and a mass $>30\%$ halo interior to $r = a$. Furthermore, such a density reduction is achieved at the expense of removing a large fraction of the angular momentum of the baryons. It should also be noted that the density changes shown in Fig. \[ABW\] also do not take account of any halo compression that might have occurred as the bar and disk formed.
Baryonic Clumps
---------------
El-Zant, Shlosman & Hoffman (2001) proposed that dynamical friction from the halo on moving clumps of dense gas will also transfer energy to the DM and lower its density. They envisaged that baryons would collect into clumps through the Jeans instability as galaxies are assembled and present somewhat simplified calculations of the consequences of energy loss to the halo. The idea was taken up by Mo & Mao (2004), who saw this as a means to erase the cusps in small halos before they merge to make a main galaxy halo, and by Tonini, Lapi & Salucci (2006).
The proposed mechanism has a number of conceptual problems, however. The model assumes that the settling gas clumps maintain their coherence for many dynamical crossing times without colliding with other clumps or being disrupted by star formation, for example. In addition, calculations ( Kaufmann 2006) of the masses of condensing gas clumps suggest they range up to only $\sim
10^6\;M_\odot$, which is too small to experience strong friction. Larger clumps will probably gather in subhalos, which may get dragged in, but simulations with sub-clumps composed of particles ( Ma & Boylan-Kolchin 2004) indicate that the DM halos of the clumps will be stripped, which simply replaces any DM moved outwards in the halo. Debattista (2008) suggest halo compression is an issue here also, but the essential idea suggested by El-Zant is to displace the DM as the gas settles, which avoids halo compression.
![The changes in density caused by the settling of $N_h$ heavy particles with total mass of $0.1M_{200}$, initially distributed at uniform density within a sphere of radius $4rs$ in an NFW halo. The solid line shows the density measured from the particles at the start while the broken lines show the density after 9 Gyr. Another dashed line shows the corresponding theoretical NFW curve.[]{data-label="collect"}](collectbw.ps){width="12cm"}
### A Direct Test
Setting all these difficulties to one side, Jardel & Sellwood (2008) set out to test the mechanism with $N$-body simulations. As proposed by El-Zant , we divided the entire mass of baryons into $N_h$ equal mass clumps, treated as softened point masses, to which we added isotropic random motion to make their distribution in rough dynamical equilibrium inside an NFW halo composed of $\sim 1\,$M self-gravitating particles. All particles, both light and heavy, experienced the attraction of all others.
Figure \[collect\] shows results after $\sim 9\;$Gyr, when scaled to a $c=15$ halo – the timescale would be even longer for less concentrated halos. We find that some density reduction does occur, but the rate at which the density is decreased is considerably slower than El-Zant predict. We traced this discrepancy to their use of three times too large a Coulomb logarithm in their calculations.
Fig. \[collect\] also shows that the rate of density reduction rises as the baryon mass is concentrated into fewer, more massive particles. Again our result is consistent with that of Ma & Boylan-Kolchin (2004), who employed a mass spectrum of clumps, and who showed that a much smaller density reduction occurred in a separate simulation that omitted the three heaviest clumps. Thus, if this process is to work on an interesting time-scale, it requires a few gas clumps whose masses exceed $1\%$ of the entire halo.
Mashchenko (2006, 2007) argue that the energy input to the halo, mediated by the motion of the mass clumps, can be boosted if the gas is stirred by stellar winds and supernovae – a less extreme form of feedback ([*cf.*]{} §3.1). Their simulations of the effect reveal a density reduction in dwarf galaxies. Peirani (2008), on the other hand, propose AGN activity to accelerate gas clumps. They present simulations that show the cusp can be flattened to $\sim
0.1r_s$ with a clump having mass of $\sim1$% of the galaxy mass, being driven outwards from the center to a distance of half the NFW break radius at a speed of 260 km/s. In both these models, it is unclear how the dense material can be accelerated to the required speed ( MacLow & Ferrara 1999).
Recoiling/Binary BHs
--------------------
In any hierarchical structure-formation model, halos grow through a succession of mergers ( Wechsler 2002). If massive black holes (BHs) have formed in the centers of two galaxies that merge, then one expects both BHs to settle to the center of the merged halo and to form a binary pair of BHs in orbit about each other. The physics of the decay of the orbit is interestingly complicated ( Merritt & Milosavljević 2005).
The star density in the centers of elliptical galaxies can be reduced by star scattering as the BH binary hardens, and also by the separate process of BH recoil if the binary encounters another massive object. Merritt & Milosavljević (2005) point out that the star density can be significantly reduced only within the sphere of gravitational influence of the BHs, which extends to $r \sim r_h$, where $r_h =
GM_\bullet/\sigma^2$, where $M_\bullet$ is the mass of the BH and $\sigma$ is the velocity dispersion of the stars. Generous values for disk galaxies might be $M_\bullet = 10^7 M_\odot$ and $\sigma =
100\;$km/s, yielding $r_h \simeq 4.4\;$pc.
Since none of the processes that affect the star density in the center depend upon the masses of the individual stars, DM particles will be affected in a similar manner, and we must expect the DM density to be depleted also over the same volume. However, because black hole dynamics affects only the inner few parsecs, it can have essentially no effect on bar pattern speed constraints or values of parameters such as $\Delta_{v/2}$.
Conclusions
===========
The inner densities of DM halos of galaxies today continue to challenge the model for galaxy formation. Both direct estimates in a few galaxies, and dynamical friction constraints from bar pattern speeds require inner densities lower than predicted in DM-only simulations by at least a factor of a few.
Several processes that are omitted from DM only simulations can reduce the inner halo density. Feedback has a very slight effect unless a large mass of gas is blasted out from the deepest point in the peotential well. Bar friction requires an extreme bar and removes a large fraction of the angular momentum in the baryons. Dynamical friction on massive gas clumps is too slow, unless moving gas clumps exceed $\sim 1\%$ of total baryonic mass. Scattering by merging or recoiling BHs affects only very center. Thus any of the four suggested mechanism needs to be stretched if it is to cause a significant reduction.
I thank Victor Debattista for comments on the manuscript and the organizers of the conference for travel support. This work was supported by grants AST-0507323 from the NSF and NNG05GC29G from NASA.
References {#references .unnumbered}
==========
Alam, S. M. K., Bullock, J. S. & Weinberg, D. H. 2002, , [**572**]{}, 34
Athanassoula, E. 1992, , [**259**]{}, 345
Athanassoula, E. 2002, , [**569**]{}, L83
Athanassoula, E. 2003, , [**341**]{}, 1179
Binney, J. J. & Evans, N. W. 2001, , [**327**]{}, L27
Binney, J., Gerhard, O. & Silk, J. 2001, , [**321**]{}, 471
Binney, J. & Tremaine, S. 2008, [*Galactic Dynamics*]{} 2nd Ed. (Princeton: Princeton University Press)
Bissantz, N., Englmaier, P. & Gerhard, O. 2003, , [**340**]{}, 949
Blumenthal, G. R., Faber, S. M., Flores, R. & Primack, J. R. 1986, , [**301**]{}, 27
Ceverino, D. & Klypin, A. 2007, , [**379**]{}, 1155
Colín, P., Valenzuela, O. & Klypin, A. 2006, , [**644**]{}, 687
Corsini, E. M. 2008, in [*Formation and Evolution of Galaxy Disks*]{}, eds. J. G. Funes SJ & E. M. Corsini (ASP, to appear)
Debattista, V. P. & Sellwood, J. A. 1998, , [**493**]{}, L5
Debattista, V. P. & Sellwood, J. A. 2000, , [**543**]{}, 704
Debattista, V. P., 2008, , [**681**]{}, 1076
de Blok, W. J. G., McGaugh, S. S. & Rubin, V. C. 2001, , [**122**]{}, 2396
de Blok, W. J. G. & Bosma, A. 2002, , [**385**]{}, 816
Diemand, J., Kuhlen, K. & Madau, P. 2007, , [**667**]{}, 859
Dutton, A. A., van den Bosch, F. C. & Courteau, S. 2008, in [*Formation and Evolution of Galaxy Disks*]{}, eds. J. G. Funes SJ & E. M. Corsini (ASP, to appear) arXiv:0801.1505
El-Zant, A., Shlosman, I. & Hoffman, Y. 2001, , [**560**]{}, 636
Gnedin, O. Y. & Zhao, H.S., 2002, , [**333**]{}, 299
Hernquist, L. & Weinberg, M. D. 1992, , [**400**]{}, 80
Holley-Bockelmann, K., Weinberg, M. & Katz, N. 2005, , [**363**]{}, 991
Jardel, J. & Sellwood, J. A. 2008, , (submitted)
Kassin, S. A., de Jong, R. S. & Weiner, B. J. 2006, , [**643**]{}, 804
Kaufmann, T., Mayer, L., Wadsley, J., Stadel, J. & Moore, B. 2006, , [**370**]{}, 1612
Klypin, A., Zhao, HS. & Somerville, R. S. 2002, , [**573**]{}, 597
Komatsu, E. 2008, arXiv:0803.0547
Kormendy, J. & Kennicutt, R. C. 2004, , [**42**]{}, 603
Ma, C-P. & Boylan-Kolchin, M. 2004, , [**93**]{}, 21301
Macciò, A. V., Dutton, A. A. & van den Bocsh, F. C. 2008, arXiv:0805.1926
MacLow, M-M. & Ferrara, A. 1999, , [**513**]{}, 142
Mashchenko, S., Couchman, H. M. P. & Wadsley, J. 2006, , [**442**]{}, 539
Mashchenko, S., Couchman, H. M. P. & Wadsley, J. 2007, , [**319**]{}, 174
Merritt, D. & Milosavljević, M. 2005, , [**8**]{}, 8 (astro-ph/0410364)
Mo, J. J. & Mao, S. 2004, , [**353**]{}, 829
Navarro, J. F., Eke, V. R. & Frenk, C. S. 1997, , [**283**]{}, L72
Navarro, J. F., Frenk, C. S. & White, S. D. M. 1997, , [**490**]{}, 493
O’Neill, J. K. & Dubinski, J. 2003, , [**346**]{}, 251
Peirani, S, Kay, S. & Silk, J. 2008, , [**479**]{}, 123
Popowski, P. 2001, in [*Astrophysical Ages and Times Scales*]{}, eds. T. von Hippel, C. Simpson, & N. Manset, ASP Conference Series [**245**]{}, p. 358
Popowski, P. 2005, , [**631**]{}, 879
Rautiainen, P., Salo, H. & Laurikainen, E. 2008, arXix:0806.0471
Rhee, G., Valenzuela, O., Klypin, A., Holtzman, J. & Moorthy, B. 2004, , [**617**]{}, 1059
Roskar, R., Debattista, V. P., Stinson, G. S., Quinn, T. R., Kaufmann, T. & Wadsley, J. 2008, , [**675**]{}, L65
Sellwood, J. A. 2003, , [**587**]{}, 638
Sellwood, J. A. 2008a, in [*Formation and Evolution of Galaxy Disks*]{}, eds. J. G. Funes SJ & E. M. Corsini (ASP, to appear) arXiv:0803.1574
Sellwood, J. A. 2008b, , [**679**]{}, 379
Sellwood, J. A. & Binney, J. J. 2002, , [**336**]{}, 785
Sellwood, J. A. & Debattista, V. P. 2006, , [**639**]{}, 868
Sellwood, J. A. & McGaugh, S. S. 2005, , [**634**]{}, 70
Swaters, R. A., Madore, B. F., van den Bosch, F. C. & Balcells, M. 2003, , [**583**]{}, 732
Tonini, C., Lapi, A. & Salucci, P. 2006, , [**649**]{}, 591
Tremaine, S. & Weinberg, M. D. 1984, , [**209**]{}, 729
Valenzuela, O. & Klypin, A. 2003, , [**345**]{}, 406
Valenzuela, O., Rhee, G., Klypin, A., Governato, F., Stinson, G., Quinn, T. & Wadsley, J. 2007, , [**657**]{}, 773
Wechsler, R. H., Bullock, J. S., Primack, J. R., Kravtsov, A. V. & Dekel, A. 2002, , [**568**]{}, 52
Weinberg, M. D. & Katz, N. 2002, , [**580**]{}, 627
Weinberg, M. D. & Katz, N. 2007a, , [**375**]{}, 425
Weinberg, M. D. & Katz, N. 2007b, , [**375**]{}, 460
Weiner, B. J., Sellwood, J. A. & Williams, T. B. 2001, , [**546**]{}, 931
Weiner, B. J. 2004, in IAU Symp. 220, Dark Matter in Galaxies, ed. S. Ryder, D. J. Pisano, M. Walker & K. C. Freeman (Dordrecht: Reidel), p. 35
Zánmar Sánchez, R., Sellwood, J. A., Weiner B. J. & Williams, T. B. 2008, , [**674**]{}, 797
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The relativistic correction of the AdS/CFT implied heavy quark potential is examined within the framework of the potential model. For the typical range of the coupling strength appropriate to heavy-ion collisions, we find the correction is significant in size and lowers the dissociation temperature of quarkonia.'
author:
- Yan Wu
- 'De-fu Hou'
- 'Hai-cang Ren'
title: The relativistic correction of the quarkonium melting temperature with a holographic potential
---
Introduction {#sec:level1}
============
The phase structure of quantum chromodynamics (QCD) remains an active field of researches. At sufficiently high temperature, hadronic matter will evolve into a quark-gluon plasma which has been explored experimentally by relativistic heavy ion collisions (RHIC). The quarkonium dissociation is an important signal of this transition [@Satz].
The quarkonium are bound states composed of a heavy quark $q$ and its antiparticle $\bar{q}$. It is found that, the ground states and the excitation levels of quarkonium are very much smaller than the normal hadrons, and that they are very tightly bounded[@0512217v2]. Theoretically, there are two approaches to study the quarkonium: Lattice QCD and potential models[@method]. From Lattice QCD, we can calculate the spectral function numerically via the quarkonium correlators and identify the quarkonium states with the resonance peaks[@4; @5; @6; @7; @8]. The potential model relies on the small velocity($v<<1$) of the constituent quarks. By solving a non-relativistic Schroedinger equation with a temperature dependent effective potential, we can determine the energy levels and thereby the threshold temperature when the bound state dissolves [@9; @10; @11; @12; @133; @14]. The potential model will be applied in the present work.
AdS/CFT correspondence is a powerful tool to explore the strongly coupled ${\cal N}=4$ super Yang-Mills plasma. The equation of state and viscosity ratio, etc. extracted from the AdS/CFT show remarkable agreement with the Lattice QCD or experimental data from the quark-gluon plasma (QGP) created via Relativistic Heavy Ion Collisions (RHIC). It would be interesting to extend the comparison to a wide range of other quatities, for instance heavy quark dissociation, which are calculable in both ways to assess whether the super Yang-Mills serves an important reference model of the QGP phase of QCD. This is the primary motivation of this paper.
In a previous work[@Hou], we examined the heavy quarkonium dissociation within the potential model with the AdS/CFT implied potential function (holographic potential). We found that the holographic potential can be approximated by a truncated Coulomb potential to a great accuracy. With the typical values of the ’t Hooft coupling constant, $\lambda\equiv\sqrt{N_cg_{\rm
YM}^2}$ considered in the literature [@Gubser], $$5.5 < \lambda < 6\pi,
\label{coupling}$$ our dissociation temperatures are systematically lower, though not far from the lattice prediction. [^1] On the other hand, an estimate of the velocity of the constituent quarks inside the bound state indicates that the non-relativistic approximation may be marginal, especially for $J/\Psi$. This motivates us to examine the relativistic corrections of the holographic potential with the aid of a two-body Dirac equation(TBDE).
While the holographic potential alone is sufficient in the nonrelativistic limit, it does not provide all information necessary for the relativistic corrections even to the order $v^4$ term. Except for the correction brought about by the relativistic kinetic energy, the spin-orbital coupling and the Darwin term depend on how the holographic potential is introduced in the two-body Dirac equation. In addition, gravity dual of spin dependent forces is not available in the literature. Therefore our result remains incomplete at this stage. We would like to comment that the same issues exists for the relativistic corrections of the heavy quark potential extracted from the lattice QCD simulations.
In the next section, the work reported in [@Hou] will be reviewed and the dissociation temperature beyond the truncated Coulomb approximation is presented. The corrections to the dissocation temperature are computed in the section 3 through a Foldy-Wouthuysen (F.W.) transformation of a two-body Dirac Hamiltonian. The kinetic energy contribution and the contribution from the Darwin and the spin-orbit coupling are calculated seperately with the latter obtained simply by replacing the perturbative Coulomb potential in the two-body Dirac Hamiltonian with the holographic potential. The section 4 conclude the paper.
The holographic potential model
===============================
In the conventional potential model of QCD, the non-relativistic wave function of a heavy quarkonium satisfies the Schroedinger equation $$\begin{aligned}
[-\frac{1}{2\mu}\nabla^{2}+U(r,T)]\psi=-E(T)\psi,
\label{schroedinger}\end{aligned}$$ where $E(T)$ is the binding energy and $U(r,T)$ is identified with the internal energy of a pair a static $q$ and $\bar q$ in QGP and is related to the free energy $F(r,T)$ via $$\begin{aligned}
U(r,T)&&=-T^{2}\Big[\frac{\partial}{\partial T}(\frac{F(r,T)}{T})
\Big]_r\end{aligned}$$ The free energy $F(r,T)$ can be extracted from the expectation of a pair of Wilson loops operator according to: $$\begin{aligned}
e^{-\frac{1}{T}F(r,T)}=\frac{tr<W^{\dag}(L_{+})W(L_{-})>}{tr<W^{\dag}(L_{+})><W(L_{-})>}
\label{wilson}\end{aligned}$$ where $L_{\pm}$ stands for the Wilson loop running in Euclidean time direction at spatial coordinates $(0,0,\pm\frac{1}{2}r)$ and is closed with the periodicity $\beta=\frac{1}{T}$ and $$\begin{aligned}
W(L_\pm)=Pe^{-i\oint_{L_\pm}dx^{\mu}A_{\mu}(x)}.
\label{loop}\end{aligned}$$ The spatial coordinates of $L_\pm$ are $(0,0,\pm\frac{1}{2}r)$. The lattice QCD simulation of the expectation value (\[wilson\]) can be found if Ref.[@10; @0746v2].
In case of super Yang-Mills, the holographic principle places the Wilson lines $L_\pm$ on the boundary ($y\to\infty$) of the 5D AdS-Schwarzschild metric [@Maldacena]: $$\begin{aligned}
ds^{2}=\pi^{2}T^{2}y^{2}(fdt^{2}+d\vec{x}^{2})+\frac{1}{\pi^{2}T^{2}y^{2}f}dy^{2}\end{aligned}$$ where $f=1-\frac{1}{y^4}$, $d\vec{x}^{2}=dx_1^{2}+dx_2^{2}+dx_3^{2}$ with the ansatz $x_1=x_2=0$ and $x_3$ a function of $y$.
The free energy $F(r,T)$ of the corresponding super Yang-Mills at large $N_c$ and large ’t Hooft coupling is proportional to the minimum area of the worldsheet in the AdS bulk bounded by $L_+$ and $L_-$, and is given parametrically by [@Maldacena; @Rey]: $$\begin{aligned}
\left\{
\begin{aligned}
F(r,T)&=T \min(I,0)\\
r\;\;\;\;&=\dfrac{2q}{\pi T}\int_{y_{c}}^{\infty}\dfrac{dy}{\sqrt{(y^{4}-1)(y^{4}-y_{c}^{4})}}
\label{raduis}
\end{aligned}
\right.\end{aligned}$$ where $$\begin{aligned}
I=\sqrt{\lambda}[\int_{y_{c}}^{\infty} dy(
\sqrt{\frac{y^{4}-1}{y^{4}-y_{c}^{4}}}-1)+1-y_{c}]
\label{fenergy}\end{aligned}$$ and the parameter $y_c\in (1,\infty)$. Eliminating $y_c$ between (\[raduis\]) and (\[fenergy\]), we find that $$\begin{aligned}
F(r,T)=-\frac{\alpha}{r} \Phi(\rho) \theta(\rho_{0}-\rho)\end{aligned}$$ where $\alpha\doteq0.2285\sqrt{\lambda}$, $\rho=\pi T r$, $\rho_{0}=0.7541$ and $\Phi(\rho)$ is the screening factor. The corresponding internal energy is $$\begin{aligned}
U(r,T)=-\frac{\alpha}{r}[\Phi(\rho)-\rho(\frac{d\Phi}{d\rho})_{y_{c}}-\rho(\frac{d\Phi}{dy_{c}})_{\rho}(\frac{dy_{c}}{d\rho})]\theta(\rho_{0}-\rho)
\label{exact}\end{aligned}$$ and will be substitute into the Schroedinger equation (\[schroedinger\]).
The small $\rho$ expansion of $\Phi(\rho)$ is: $$\Phi(\rho)=1-\frac{\Gamma^{4}(\frac{1}{4})}{4\pi^{3}}\rho+\frac{3\Gamma^{8}(\frac{1}{4})}{640\pi^{6}}\rho^{4}+O(\rho^{8}).
\label{taylor}$$ Within the screening radius $\rho_0$, the first two terms of the series (\[taylor\]) approximate the exact $\Phi$ well as is shown in Fig.1. If we keep only the first two terms, the screening radius $\rho_{0}\simeq0.7359$ and $U(r,T)$ becomes a truncated Coulomb potential $$\begin{aligned}
U=-\frac{\alpha}{r}\theta(\rho_{0}-\rho)\end{aligned}$$ under the approximation.
We define the dissociation temperature $T_d$ as the temperature when the binding energy falls to zero, i. e. $E(T_d)=0$, and the corresponding radial Schrodinger equation reads [@Hou]: $$\begin{aligned}
\frac{d^{2}R}{d\rho^{2}}+\frac{2}{\rho}\frac{d
R}{d\rho}-[\frac{l(l+1)}{\rho^{2}}+V]R=0\end{aligned}$$ where the reduced potential, $V=\frac{mU}{\pi^{2}T^{2}}$ is dimensionless.
The truncated Coulomb potential approximation was employed in [@Hou] and the dissociation temperature of the bound state of $l$-th partial wave and $n$-th radial quantum number is given by $$T_d=\frac{4\alpha\rho_0m}{\pi x_{nl}^2}
\label{dissociation}$$ with $x_{nl}$ the $n$-th nonzero root (ascending order) of the Bessel function $J_{2l}(x)$. The corresponding radial wave function reads $$R(r)=\frac{1}{\sqrt{\rho}}J_{2l+1}\left(x_{nl}\sqrt{\frac{\rho}{\rho_0}}\right).$$ for $\rho\le\rho_0$ and $R(r)={\rm const.}/r^{l+1}$ for $\rho>\rho_0$.
In this work, we have calculated the dissociation temperature with the exact holographic potential (\[exact\]). The comparison with that obtained from the truncated Coulomb potential in [@Hou] for $J/\Psi$ is shown in Table , where we choose $m=1.65$ GeV for the mass of $c$ quarks. From the comparison of these two results we confirmed that the truncated Coulomb approximation is a good approximation and we shall stay with the truncated Coulomb approximation for the rest of this paper.
[ ]{}
-------- ----------- ----------------------- ----------- -----------------------
[Exact]{} [truncated Coulomb]{} [Exact]{} [truncated Coulomb]{}
[1s]{} [142]{} [143]{} [262]{} [265]{}
[2s]{} [27]{} [27]{} [50]{} [50]{}
[1p]{} [31]{} [31]{} [57]{} [58]{}
-------- ----------- ----------------------- ----------- -----------------------
: It lists our melting temperature in MeV’s for $1s$, $2s$ and $1p$ state using the exact potential and compared with the Coulomb potential case. We can find that this two results are very close to each other which confirmed that the Coulomb approximation is excellent.
The relativistic correction of the holographic potential
========================================================
As it is mentioned in the introduction, the velocity of the heavy quarks is not low enough so the relativistic correction may be significant, especially for$J/\Psi$. To explore this correction, one has to go beyond the Schroedinger equation (\[schroedinger\]) and switch to the two body Dirac equation[@Crater; @9612445v1; @0602066v2; @Semay]: $$\begin{aligned}
i\frac{\partial\Psi}{\partial t}=H\Psi\end{aligned}$$ In the center-of-mass frame, the Hamiltonian of the two body Dirac equation is: $$\begin{aligned}
H=\vec{\alpha_{1}}\cdot\vec{p}+\beta_{1}\cdot
m-\vec{\alpha_{2}}\cdot\vec{p}+\beta_{2}\cdot m+U
\label{TBDE}\end{aligned}$$ where, $\vec{\alpha}_{1}=\vec{\alpha}\otimes I $, $
\vec{\alpha}_{2}=I\otimes\vec{\alpha} $, $\beta_{1}=\beta\otimes
I$, $\beta_{2}=I\otimes\beta$, $\vec\alpha$, $\beta$ are usual $4\times4$ Dirac matrix, $\vec p=-i\vec\nabla$ and $U$ is the interaction potential between the two particles. The Hamiltonian $H$ is a $16\times16$ matrix. A quarkonium state corresponds to a bound state of $H$ with the eigenvalue $2m-E(T)$, which goes to $2m$ at the dissociation temperature, i. e. $E(T)=0$. Since we are interested in the leading order relativistic correction of the dissociation temperature for the quarkonium, we have to expand the Hamiltonian to the order $v^{4}$. The sorting of the order in $v$ follows from the rules that $\frac{\vec{p}^2}{m}\sim U\sim v^2$ and $\vec\nabla\sim\frac{1}{r}\sim mv$. Also the expectation values of $\vec\alpha_1$ and $\vec\alpha_2$ are of the order $v$.
In analogous to the one body Foldy-Wouthuysen transformation[@Greiner], we introduce the unitary operator $${\cal U}=e^{iS_{2}'} e^{iS_{1}'} e^{iS_{2}} e^{iS_{1}}$$ where $$\begin{aligned}
&&S_{1}=-\frac{i}{2m}\beta_{1}\cdot O_{1}\\
&&S_{2}=-\frac{i}{2m}\beta_{2}\cdot (-O_{2})\\
&&S_{1}'=-\frac{i}{2m}\beta_{1}\cdot O_{1}'\ \\
&&S_{2}'=-\frac{i}{2m}\beta_{2}\cdot (-O_{2}')\end{aligned}$$ with $$\begin{aligned}
O_{1}=\vec{\alpha}_{1}\cdot\vec{p},\ \ \ \
O_{2}=\vec{\alpha}_{2}\cdot\vec{p}.\nonumber\end{aligned}$$ The transformed Hamiltonian reads $$\begin{aligned}
H_{\rm FW}&=&{\cal U}H{\cal U}^\dagger\nonumber \\
&=&(\beta_{1}+\beta_{2})(m+\frac{\vec{p}^{2}}{2m}-\frac{\vec{p}^{4}}{8m^{3}})+U
+\frac{1}{4m^{2}}\nabla^2{U}+\frac{1}{4m^{2}r}\frac{dU}{dr}(\vec{\sigma}_{1}+\vec{\sigma}_{2})\cdot\vec{L}
\label{Hfw}\end{aligned}$$ where, higher order terms in $v$ have been dropped. The details of the transformation are deferred to the appendix. The non-relativistic wave function, $\Psi=\Psi_{s_1s_2}(\vec{r}_1-\vec{r}_2)$ with subscript $s_1, s_2$ labelling the spin components of the two quarks, corresponding to the sector with $\beta_1=\beta_2=1$ (This wave function can be expanded in the series of products of the orbital wave functions of the preceding section and the spin wave functions). We may stay within this sector for the 1st order perturbation of the $v^4$ term of (\[Hfw\]) with the effective Hamiltonian $H_{\rm eff.}=H_0+H_1$, where $$\begin{aligned}
H_{0}=2m+\frac{\vec{p}^{2}}{m}+U \nonumber\end{aligned}$$ corresponds to the non-relativistic part,\
and $$\begin{aligned}
H_1 &=& -\frac{\vec{p}^{4}}{4m^{3}}+\frac{1}{4m^{2}}\nabla^2{U}+\frac{1}{4m^{2}r}\frac{dU}{dr}(\vec{\sigma}_{1}+\vec{\sigma}_{2})\cdot \vec{L}\nonumber\\
&=& -\frac{\vec{p}^{4}}{4m^{3}}+\frac{1}{4m^{2}}\nabla^2{U}
+\frac{1}{4m^{2}r}\frac{dU}{dr}(J^2-L^2-S^2), \label{h1}\end{aligned}$$ is a relativistic correction. We have introduced the total spin $\vec
S=\frac{1}{2}(\vec\sigma_1+\vec\sigma_2)$ and the total angular momentum $\vec J=\vec L+\vec S$ in the last step. The contribution of $H_1$ is somewhat like the that to the fine structure of the hydrogen atom, and we can see that the first term is the first order of the kinetic energy correction; the second term is Drawin term; and the third term is the spin orbit coupling which can be decomposed into the spin singlet and the spin triplet channels.
Perturbatively, we may write $E(T)=E_0(T)+\delta E(T)$ and $T=T_0+\delta T$, where $E_0(T)$ is the non-relativistic binding energy in (\[schroedinger\]), and $T_0$ are the dissociation temperature given by (\[dissociation\]), and $\delta E(T)$ and $\delta T$ are the $v^4$ corrections. We have $E_0(T_0)=0$. Expanding the dissociation condition $E(T)=0$ to the order $v^4$, we obtain the formula for $\delta T$, i.e. $$\delta T=\delta_1 T+\delta_2 T=-\frac{\delta_1 E(T_0)+\delta_2 E(T_0)}{(\frac{\partial E_0}{\partial T})_{T_0}}.$$ where, $$\begin{aligned}
\delta_1 E(T_0)&=&-<\frac{\vec{p}^{4}}{4m^{3}}>\nonumber\\
\delta_2 E(T_0)&=&\Big[<\frac{1}{4m^{2}}\nabla^2{U}>+<\frac{1}{4m^{2}r}\frac{dU}{dr}(J^2-L^2-S^2)>]\end{aligned}$$ and $$\left(\frac{\partial E_0}{\partial
T}\right)_{T=T_0}=<\frac{\partial H_0}{\partial
T}>=<\frac{\partial U}{\partial T}>,$$ The average $$<O>\equiv\frac{\int d^3\vec r\psi^*(\vec r)O(\vec r)\psi(\vec r)}{\int d^3\vec r\psi^*(\vec r)\psi(\vec r)}.$$ with $\psi(\vec r)$ the non-relativistic wave function. The reason for our separating the contribution from $p^4$, $\delta_1 T$ and that from the Darwin and spin-orbital terms, $\delta_2 T$ is the uncertainty in the representation of the holographic potential in (\[TBDE\]), which does not impact on the $p^4$ correction. We will come to this point in the next section. In the limit of zero binding energy, we find $<\frac{\vec{p}^{4}}{4m^{3}}>=<\frac{1}{4m}U^2>$. For the truncated Coulomb potential, $$\nabla^{2}U=4\pi\alpha\pi^{3}T^{3}\delta^{3}(\vec{\rho})\theta(\rho_{0}-\rho)+\frac{\alpha \pi^{3}T^{3}}{\rho}\delta'(\rho-\rho_{0}).$$
In terms of the radial wave function $R_l(r)$ of $\psi(\vec r)$, $$\begin{aligned}
<-\frac{\vec{p}^{4}}{4m^{3}}>&&=-\frac{\alpha^{2}}{4m\pi T}\int_{0}^{\rho_{0}}R_l(\rho)^{2}d\rho\nonumber \\
<\frac{1}{4m^{2}}\nabla^2{U}>&&=\frac{\alpha}{4m^{2}}\{|R_l(0)|^{2}-2R_l(\rho_{0})R_l'(\rho_{0})\rho_{0}-R_l^{2}(\rho_{0})\}\nonumber \\
<\frac{1}{4m^2r}\frac{d
U}{dr}>&&=\frac{\alpha}{4m^2}\int_0^{\rho_0}\frac{d\rho}{\rho}R^2(\rho)
+\frac{\alpha}{4m^2}R^2(\rho_0)\nonumber \\
<\frac{\partial U}{\partial T}>&&=\frac{\alpha\pi}{\pi^{3}T^{3}}R_l^{2}(\rho_{0})\rho_{0}^{2}\end{aligned}$$ For the $ns$ state, we find that $$\begin{aligned}
&&\delta_1 T=\frac{\pi\alpha T_0^2}{4m\rho_0}\Big[\frac{1}{J_1^2(x_{n0})}-\frac{J_0^2(x_{n0})}{J_1^2(x_{n0})}-1\Big]\nonumber\\
&&\delta_1 T+\delta_2 T=-\frac{\pi\alpha T_0^2}{4m\rho_0}\Big[\frac{J_0^2(x_{n0})}{J_1^2(x_{n0})}+1+\frac{2}{x_{n0}^2}\Big].\end{aligned}$$ For the $np$ state, we can also get analytical expressions, which are more lengthy.\
The numerical values of the corrected temperature $T_0+\delta_1 T$ and $T_0+\delta_1 T+\delta_2 T$ in MeV’s for 1$s$, 2$s$ and 1$p$ states are listed in the Table below.
[ ]{}
[| c | c | c | c | c |]{} &&\
& [$\lambda=5.5$]{} & [$\lambda=6\pi$]{}& [$\lambda=5.5$]{} & [$\lambda=6\pi$]{}\
& [162.54]{} & [387.54]{}&[478.76]{}&[1139.11]{}\
& [29.15]{} & [62.75]{}&[85.67]{}&[184.44]{}\
& [32.04]{} & [62.14]{}&[94.18]{}&[182.66]{}\
[ ]{}
[| c | c | c | c | c |]{} &&\
& [$\lambda=5.5$]{} & [$\lambda=6\pi$]{}& [$\lambda=5.5$]{} & [$\lambda=6\pi$]{}\
& [130.79]{} & [188.65]{}&[385.63]{}&[555.58]{}\
& [130.79]{} & [188.65]{}&[385.63]{}&[555.58]{}\
& [26.71]{} & [48.16]{}&[79.15]{}&[142.59]{}\
& [26.71]{} & [48.16]{}&[79.15]{}&[142.59]{}\
& [31.53]{} & [61.33]{}&[93.54]{}&[180.79]{}\
& [32.65]{} & [68.48]{}&[96.85]{}&[201.80]{}\
& [32.09]{} & [64.90]{}&[95.20]{}&[191.30]{}\
& [30.96]{} & [57.76]{}&[91.89]{}&[170.29]{}\
Discussions
===========
In summary, we have explored the leading relativistic correction to the dissociation temperature of heavy quarkonium state through a F.W.-like transformation of the two body Dirac Hamiltonian with the AdS/CFT implied potential. Among the contributions we considered, the $p^4$ correction of the kinetic energy, being negative, enhances the binding but the Darwin term does the opposite and dominates. Consequently, the dissociation temperature of $s$-state is lowered, leaving the corrected values further below the lattice result. This disagreement can be attributed to the short screening length $r_0=\frac{\rho_0}{\pi T}$, about 0.25fm at $T=200$ MeV, of the AdS/CFT potential and the sharp cutoff nature of the screening. In case of $J/\psi$, the magnitude of the correction ranges from $8\%$ for $\lambda=5.5$ to $30\%$ for $\lambda=6\pi$, indicating significant relativistic towards the high end of the domain (\[dissociation\]).
The potential model, though physically more transparent than spectral function approach, does not provide complete $v^4$ corrections with the holographic potential extracted from the Wilson loop alone. The same deficiency applies the relativistic correction based on the lattice heavy quark potential alone. As the Wilson loop for a nonAbelian theory involves multi-gluon excahnges, its form in the two-body Dirac Hamiltonian (\[TBDE\]) may not be adequate unless the single gluon exchange serves a reasonable approximation. Among the four sectors $\beta_1=\pm 1$ and $\beta_2=\pm 1$, only two of them $\beta_1=\beta_2=\pm 1$ correspond to $q\bar q$ interacting via the holographic potential. The other two sectors with $\beta_1$ and $\beta_2$ correspond to $qq$ or $\bar{q}\bar{q}$ and the interacting potential is unknown. A more general form of the interaction in (\[TBDE\]) without violating the charge conjugation symmetry is to replace $U$ by $$(\Lambda_{++}+\Lambda_{--})U+(\Lambda_{+-}+\Lambda_{-+})U'=U_++\beta_1\beta_2U_-$$ where the projection operator $\Lambda_{\pm 1,\pm 1}\equiv\frac{1\pm\beta_1}{2}
\frac{1\pm\beta_2}{2}$, $U'$ the potential between $qq$ or $\bar q\bar q$ and $U_\pm=\frac{U\pm U'}{2}$. $U'=U$ in the previous section. Repeating the steps of FW transformation in the appendix, we find that the perturbing Hamiltonian (\[h1\]) is replaced by $$H_1=-\frac{\vec{p}^{4}}{4m^{3}}+\frac{1}{4m^{2}}\nabla^2{U_+}
+\frac{1}{4m^{2}r}\frac{dU_+}{dr}(\vec{\sigma}_{1}+\vec{\sigma}_{2})\cdot
\vec{L}
+\frac{1}{8m^2}(\lbrace\vec\sigma_1\cdot\vec\nabla,\lbrace\vec\sigma_1\cdot\vec\nabla,U_-\rbrace\rbrace
+\lbrace\vec\sigma_2\cdot\vec\nabla,\lbrace\vec\sigma_2\cdot\vec\nabla,U_-\rbrace\rbrace$$ with $\lbrace...\rbrace$ an anticommutator. This will modify the potential part of (\[h1\]). Only the correction from the kinetic energy, $p^4$ term of (\[h1\]) is robust, which raise the dissociation temperature.
In addition to the holographic potential considered in this work, there should be spin dependent ones that splits degeracies between spin singlets and spin triplets (e.g. between $\eta_c$ and $J/\psi$). The gravity dual of the latter are unknown in the literature. A first principle derivation of the spin-dependent forces associate them to the expectation value of Wilson loops with operator insertions, $<{\rm tr}{\cal
W}_{\mu\nu}(L_+)^\dagger {\cal W}_{\rho\lambda}(L_-)>$, where $${\cal W}_{\mu\nu}(L)=PF_{\mu\nu}(x)e^{-i\int_Ldx^\mu A_\mu}$$ with $F_{\mu\nu}$ the Yang-Mills field strength and $x$ a point along $L$, ${\cal W}_{\mu\nu}(L)$ is obtained from the Eqs.(\[loop\]) by small distortion of $L$ at $x$. Within the AdS/CFT framework, it corresponds to the perturbation of the Nambu-Goto action of the world shee underlying the holographic potential under a small distortion of its boundary. It is a challenging boundary value problem and we hope to report our progress along this line in future.
Finally, we would like to comment on a phenomenological formulation of the two-body Dirac equation [@wong], which has been applied recently to the same problem addressed in this work [@zhuang]. It amounts to divide the heavy quark potential of Cornell type into a linearly confining term and a single gluon Coulomb term and to generate all spin-dependent forces by the latter. While it is legitimate in vacuum in the weak coupling because of the Lorentz invariance, a direct application to a medium beyond weak coupling remains to be justified, given different screening properties of the electric and magnetic gluons.
Acknowlegdment {#sec:level1}
==============
We thank Prof. Pengfei Zhuang for bringing our attention to their recent paper [@zhuang] prior to uploading to the web. This work is supported in part by NSFC under grant Nos. 10975060, 11135011, 11221504
In this appendix, we shall fill in the details of the F.W. transformation for the two body Dirac equation which we have done in section 4.
Let us recall the Foldy-Wouthuysen transformation of the one body Dirac Hamiltonian, $H=\vec\alpha\cdot\vec p+\beta m+V$. For a 4-component spinor with velocity $v<<1$ and positive energy, one can work in a representation where the upper two component corresponds to the non-relativistic limit, referred to as the large components, while the lower two components are suppressed by a power of $v$, referred to as the small components. An operator is even(odd) if it is diagonal(off-diagonal) with respect to large and small components. For example, $\beta$ is even and $\vec\alpha$ is odd. The Foldy-Wouthuysen transformation amounts to successive unitary transformations that push the odd operators to higher orders in $v$.
The Foldy-Wouthuysen transformation can be easily generalized to the two body case, the Hilbert space of which is spanned by the direct products of two one body spinors. To the leading order relativistic correction, we stop at the $v^4$ terms ignoring all the higher order term. The Hamiltonian of the two body Dirac equation is: $$\begin{aligned}
H=\vec{\alpha_{1}}\vec{p}+\beta_{1}m-\vec{\alpha_{2}}\vec{p}+\beta_{2}m+V\end{aligned}$$ let: $$\begin{aligned}
\vec{\alpha_{1}}\vec{p}=O_{1}\nonumber\\
\vec{\alpha_{2}}\vec{p}=O_{2}\nonumber\end{aligned}$$ so, $$\begin{aligned}
H&&=\underbrace{\beta_{1}m+O_{1}}+\underbrace{\beta_{2}m-O_{2}}+V\nonumber\\
&&=H_{1}+H_{2}+V\nonumber\\\end{aligned}$$ where $H_{1}$ corresponds to the first under-brace, and $H_{2}$ corresponds to the second under-brace. And it is easy to proof the commutation relation that:$[O_{1}, O_{2}]=[\beta_{1}, \beta_{2}]=[O_{1}, \beta_{2}]=[O_{2}, \beta_{1}]=0$, and $O_{1}\sim O_{2}\sim v$.\
We do the first transformation:\
we select: $$\begin{aligned}
S_{1}&&=-\frac{i}{2m}\beta_{1}O_{1}\nonumber\\
S_{2}&&=-\frac{i}{2m}\beta_{2}(-O_{2})\nonumber\end{aligned}$$ so, the transformed Hamiltonian turns into: $H'=\underbrace{e^{iS_{2}}\overbrace{e^{iS_{1}}He^{-iS_{1}}}e^{-iS_{2}}}$. We calculate the mid over-brace first. $$\begin{aligned}
e^{iS_{1}}He^{-iS_{1}}&&=\underbrace{e^{iS_{1}}H_{1}e^{-iS_{1}}}+\underbrace{e^{iS_{1}}H_{2}e^{-iS_{1}}}+\underbrace{e^{iS_{1}}Ve^{-iS_{1}}}\nonumber\\
&&=\overline{H_{1}}+\overline{H_{2}}+V'\nonumber\\\end{aligned}$$ also, $\overline{H_{1}}$ corresponds to the first under-brace, $\overline{H_{2}}$ corresponds to the second, and $V'$ corresponds to the third. $$\begin{aligned}
\overline{H_{1}}&&=H_{1}+[iS_{1}, H_{1}]+\frac{1}{2!}[iS_{1}, [iS_{1}, H_{1}]]+\frac{1}{3!}[iS_{1}, [iS_{1}, [iS_{1}, H_{1}]]]+\frac{1}{4!}[iS_{1}, [iS_{1}, [iS_{1}, [iS_{1}, H_{1}]]]]+\cdots\\
&&[iS_{1}, H_{1}]=-O_{1}+\frac{1}{m}\beta_{1}O_{1}^{2}\nonumber\\
&&\frac{1}{2!}[iS_{1}, [iS_{1}, H_{1}]]=-\frac{1}{2m}\beta_{1}O_{1}^{2}-\frac{1}{2m^{2}}O_{1}^{3}\nonumber\\
&&\frac{1}{3!}[iS_{1}, [iS_{1}, [iS_{1}, H_{1}]]]=\frac{1}{6m^{2}}O_{1}^{3}-\frac{1}{6m^{3}}\beta_{1}O_{1}^{4}\nonumber\\
&&\frac{1}{4!}[iS_{1}, [iS_{1}, [iS_{1}, [iS_{1}, H_{1}]]]]\sim\frac{1}{24m^{3}}\beta_{1}O_{1}^{4}\nonumber\\
\overline{H_{1}}&&=\beta_{1}m+\frac{1}{2m}\beta_{1}O_{1}^{2}-\frac{1}{8m^{3}}\beta_{1}O_{1}^{4}-\frac{1}{3m^{2}}O_{1}^{3}+O(v^{5})\end{aligned}$$ where $v$ is the velocity of the particles. $$\begin{aligned}
\overline{H_{2}}&&=H_{2}+[iS_{1}, H_{2}]+\frac{1}{2!}[iS_{1}, [iS_{1}, H_{2}]]+\frac{1}{3!}[iS_{1}, [iS_{1}, [iS_{1}, H_{2}]]]+\frac{1}{4!}[iS_{1}, [iS_{1}, [iS_{1}, [iS_{1}, H_{2}]]]]+\cdots\\
&&[iS_{1}, H_{2}]=\frac{1}{2!}[iS_{1}, [iS_{1}, H_{2}]]=\frac{1}{3!}[iS_{1}, [iS_{1}, [iS_{1}, H_{1}]]]=\cdots=0\nonumber\\
\overline{H_{2}}&&=H_{2}=\beta_{2}m-O_{2}\nonumber\\
V'&&=V+\frac{1}{2m}\beta_1[O_1,V]-\frac{1}{8m^2}[O_1,[O_1,V]]\end{aligned}$$ so, $$\begin{aligned}
e^{iS_{1}}He^{-iS_{1}}&&=\overline{H_{1}}+\overline{H_{2}}+V'\nonumber\\
&&=\beta_{1}m+\beta_{2}m-O_{2}+V+\frac{1}{2m}\beta_{1}O_{1}^{2}-\frac{1}{8m^{2}}[O_{1}, [O_{1}, V]]-\frac{1}{8m^{3}}\beta_{1}O_{1}^{4}+\frac{1}{2m}\beta_{1}[O_{1}, V]-\frac{1}{3m^{2}}O_{1}^{3}+O(v^{5})\nonumber\\
&&=\overline{H}\end{aligned}$$ that we mark $e^{iS_{1}}He^{-iS_{1}}$ as $\overline{H}$ for convenience. So, $$\begin{aligned}
H'&&=e^{iS_{2}}\overline{H}e^{-iS_{2}}\nonumber\\
&&=\overline{H}+[iS_{2}, \overline{H}]+\frac{1}{2!}[iS_{2}, [iS_{2}, \overline{H}]]+\frac{1}{3!}[iS_{2}, [iS_{2}, [iS_{2}, \overline{H}]]]+\frac{1}{4!}[iS_{2}, [iS_{2}, [iS_{2}, [iS_{2}, \overline{H}]]]]+\cdots\\
&&[iS_{2}, \overline{H}]=O_{2}+\frac{1}{m}\beta_{2}O_{2}^{2}-\frac{1}{2m}\beta_{2}[O_{2}, V]-\frac{1}{4m}\beta_{1}\beta_{2}[O_{2}, [O_{1}, V]]\nonumber\\
&&\frac{1}{2!}[iS_{2}, [iS_{2},\overline{H}]]=-\frac{1}{2m}\beta_{2}O_{2}^{2}+\frac{1}{2m^{2}}O_{2}^{3}-\frac{1}{8m^{2}}[O_{2}, [O_{2}, V]]\nonumber\\
&&\frac{1}{3!}[iS_{2}, [iS_{2}, [iS_{2}, \overline{H}]]]=-\frac{1}{6m^{2}}O_{2}^{3}-\frac{1}{6m^{3}}\beta_{2}O_{2}^{4}\nonumber\\
&&\frac{1}{4!}[iS_{2}, [iS_{2}, [iS_{2}, [iS_{2}, \overline{H}]]]]\sim\frac{1}{24m^{3}}\beta_{2}O_{2}^{4}\nonumber\\
H'=&&\beta_{1}m+\beta_{2}m+V\nonumber\\
&&+\frac{1}{2m}\beta_{1}O_{1}^{2}-\frac{1}{8m^{2}}[O_{1}, [O_{1}, V]]-\frac{1}{8m^{3}}\beta_{1}O_{1}^{4}+\frac{1}{2m}\beta_{1}[O_{1}, V]-\frac{1}{3m^{2}}O_{1}^{3}\nonumber\\
&&+\frac{1}{2m}\beta_{2}O_{2}^{2}-\frac{1}{8m^{2}}[O_{2}, [O_{2}, V]]-\frac{1}{8m^{3}}\beta_{2}O_{2}^{4}-\frac{1}{2m}\beta_{2}[O_{2}, V]+\frac{1}{3m^{2}}O_{2}^{3}\nonumber\\
&&-\frac{1}{4m^{2}}\beta_{1}\beta_{2}[O_{2}, [O_{1}, V]]+O(v^{5})\end{aligned}$$ since, $$\begin{aligned}
H'&&=(\beta_{1}m+V+\frac{1}{2m}\beta_{1}O_{1}^{2}-\frac{1}{8m^{2}}[O_{1}, [O_{1}, V]]-\frac{1}{8m^{3}}\beta_{1}O_{1}^{4}+\frac{1}{2m}\beta_{1}[O_{1}, V]-\frac{1}{3m^{2}}O_{1}^{3})\Longrightarrow mark:H_{1}'\nonumber\\
&&+(\beta_{2}m+\frac{1}{2m}\beta_{2}O_{2}^{2}-\frac{1}{8m^{2}}[O_{2}, [O_{2}, V]]-\frac{1}{8m^{3}}\beta_{2}O_{2}^{4}-\frac{1}{2m}\beta_{2}[O_{2}, V]+\frac{1}{3m^{2}}O_{2}^{3})\Longrightarrow mark:H_{2}'\nonumber\\
&&-\frac{1}{4m^{2}}\beta_{1}\beta_{2}[O_{2}, [O_{1}, V]]\\
&&H_{1}'=\beta_{1}m+\underbrace{V+\frac{1}{2m}\beta_{1}O_{1}^{2}-\frac{1}{8m^{2}}[O_{1}, [O_{1}, V]]-\frac{1}{8m^{3}}\beta_{1}O_{1}^{4}}+\underbrace{\frac{1}{2m}\beta_{1}[O_{1}, V]-\frac{1}{3m^{2}}O_{1}^{3}}\nonumber\\
&&=\beta_{1}m+V_{1}'+O_{1}'\\
&&H_{2}'=\beta_{2}m+\underbrace{\frac{1}{2m}\beta_{2}O_{2}^{2}-\frac{1}{8m^{2}}[O_{2}, [O_{2}, V]]-\frac{1}{8m^{3}}\beta_{2}O_{2}^{4}}-\underbrace{(\frac{1}{2m}\beta_{2}[O_{2}, V]-\frac{1}{3m^{2}}O_{2}^{3})}\nonumber\\
&&=\beta_{2}m+V_{2}'+O_{2}'\\
H'&&=H_{1}'+H_{2}'-\frac{1}{4m^{2}}\beta_{1}\beta_{2}[O_{2},
[O_{1}, V]]\end{aligned}$$ the first under-braces upside in the expression of $H_{1}'(H_{2}')$ correspond to $V_{1}'(V_{2}')$, and the second correspond to $O_{1}'(O_{2}')$ for convenience.\
Then, we do the second transformation:\
We select: $$\begin{aligned}
S_{1}'&&=-\frac{i}{2m}\beta_{1}O_{1}'\nonumber\\
S_{2}'&&=-\frac{i}{2m}\beta_{2}(-O_{2}')\nonumber\end{aligned}$$ and $O'_{1}\sim O'_{2}\sim v^3$\
so, $$\begin{aligned}
H''&&=e^{iS_{2}'}e^{iS_{1}'}H'e^{-iS_{1}'}e^{-iS_{2}'}\nonumber\\
&&=e^{iS_{2}'}\underbrace{e^{iS_{1}'}H_{1}'e^{-iS_{1}'}}e^{-iS_{2}'}+e^{iS_{2}'}\underbrace{e^{iS_{1}'}H_{2}'e^{-iS_{1}'}}e^{-iS_{2}'}-\frac{1}{4m^{2}}e^{iS_{2}'}e^{iS_{1}'}\beta_{1}\beta_{2}[O_{2}, [O_{1}, V]]e^{-iS_{1}'}e^{-iS_{2}'}\nonumber\\
&&=e^{iS_{2}'}\overline{H_{1}'}e^{-iS_{2}'}+e^{iS_{2}'}\overline{H_{2}'}e^{-iS_{2}'}-\frac{1}{4m^{2}}e^{iS_{2}'}e^{iS_{1}'}\beta_{1}\beta_{2}[O_{2}, [O_{1}, V]]e^{-iS_{1}'}e^{-iS_{2}'}\end{aligned}$$ where, the first under-brace upside correspond to $\overline{H_{1}'}$, and the second correspond to $\overline{H_{2}'}$, and we can directly apply the results of the first transformation\
$$\begin{aligned}
\overline{H_{1}'}&&=H_{1}'+[iS_{1}', H_{1}']+O(v^5)\nonumber\\
&&=\beta_{1}m+V_{1}'+\frac{1}{2m}\beta_{1}[O_{1}', V_{1}']\end{aligned}$$ $$\begin{aligned}
e^{iS_{2}'}\overline{H_{1}'}e^{-iS_{2}'}&&=\overline{H_{1}'}+[iS_{2}', \overline{H_{1}'}]+O(v^5)\nonumber\\
&&=\beta_{1}m+V_{1}'+\frac{1}{2m}\beta_{1}[O_{1}',
V_{1}']-\frac{1}{2m}\beta_{2}[O_{2}', V_{1}']\end{aligned}$$ $$\begin{aligned}
\overline{H_{2}'}&&=H_{2}'+[iS_{1}', H_{2}']+O(v^5)\nonumber\\
&&=\beta_{2}m+V_{2}'-O_{2}'+\frac{\beta_{1}}{2m}[O_{1}',V_{2}']-\frac{\beta_{1}}{2m}[O_{1}',O_{2}']\end{aligned}$$ $$\begin{aligned}
e^{iS_{2}'}\overline{H_{2}'}e^{-iS_{2}'}&&=\overline{H_{2}'}+[iS_{2}', \overline{H_{2}'}]+O(v^5)\nonumber\\
&&=\beta_{2}m+V_{2}'-\frac{1}{2m}\beta_{2}[O_{2}',
V_{2}']+\frac{1}{2m}\beta_{1}[O_{1}',
V_{2}']-\frac{1}{2m}\beta_{1}[O_{1}', O_{2}']\end{aligned}$$ Then, we see the last term in $H''$: $-\frac{1}{4m^{2}}\underbrace{e^{iS_{2}'}\overbrace{e^{iS_{1}'}\beta_{1}\beta_{2}[O_{2}, [O_{1}, V]]e^{-iS_{1}'}}e^{-iS_{2}'}}$\
we do the calculation of the over-brace first. $$\begin{aligned}
e^{iS_{1}'}\beta_{1}\beta_{2}[O_{2}, [O_{1}, V]]e^{-iS_{1}'}&&=\beta_{1}\beta_{2}[O_{2}, [O_{1}, V]]+[iS_{1}',\beta_{1}\beta_{2}[O_{2}, [O_{1}, V]]]+O(v^5)\nonumber\\
&&=(\beta_{1}\beta_{2}[O_{2}, [O_{1},
V]]+\frac{1}{4m^{2}}[\beta_{1}[O_{1},V],[O_{2},[O_{1},V]]])\Longrightarrow
mark:A\end{aligned}$$ $$\begin{aligned}
e^{iS_{2}'}Ae^{-iS_{2}'}&&=A+O(v^5)\\
&&=-\frac{1}{4m^{2}}\beta_{1}\beta_{2}[O_{2}, [O_{1}, V]]\end{aligned}$$ considering $O_{1},O_{2}\sim v, O_{1}',O_{2}'\sim v^{3}$ $$\begin{aligned}
H''=&&\beta_{1}m+V_{1}'+\beta_{2}m+V_{2}'-\frac{1}{4m^{2}}\beta_{1}\beta_{2}[O_{2}, [O_{1}, V]]\nonumber\\
=&&\beta_{1}m+\beta_{2}m+V+\frac{1}{2m}\beta_{1}O_{1}^{2}+\frac{1}{2m}\beta_{2}O_{2}^{2}-\frac{1}{8m^{2}}[O_{1},[O_{1},V]]-\frac{1}{8m^{2}}[O_{2},[O_{2},V]]\nonumber\\
&&-\frac{1}{8m^{3}}\beta_{1}O_{1}^{4}-\frac{1}{8m^{3}}\beta_{2}O_{2}^{4}-\frac{1}{4m^{2}}\beta_{1}\beta_{2}[O_{2}, [O_{1}, V]]\\
since,&&O_{1}^{2}=O_{2}^{2}=\vec{p}^{2},O_{1}^{4}=O_{2}^{4}=\vec{p}^{4}\nonumber\\
&&[O_{1},[O_{1},V]]=-\nabla^{2}V-\frac{2}{r}\frac{\partial V}{\partial r}\vec{\Sigma_{1}}\vec{L}\nonumber\\
&&[O_{2},[O_{2},V]]=-\nabla^{2}V-\frac{2}{r}\frac{\partial V}{\partial r}\vec{\Sigma_{2}}\vec{L}\nonumber\\
&&[O_{2},[O_{1},V]]=-\alpha_{1i}\alpha_{2j}\nabla_{i}\nabla_{j}V\nonumber\\
H''=&&\beta_{1}(m+\frac{\vec{p}^{2}}{2m}-\frac{\vec{p}^{4}}{8m^{3}})+\beta_{2}(m+\frac{\vec{p}^{2}}{2m}-\frac{\vec{p}^{4}}{8m^{3}})+V\nonumber\\
&&+\frac{1}{4m^2}\nabla^2V
+\frac{1}{4m^2}\frac{1}{r}\frac{\partial V}{\partial
r}(\vec{\Sigma_{1}}+\vec{\Sigma_{2}})\vec{L}+\frac{1}{4m^{2}}\beta_{1}\beta_{2}\alpha_{1i}\alpha_{2j}\nabla_{i}\nabla_{j}V
\label{last}\end{aligned}$$ The last term in (\[last\]), though of the order of $v^4$, is a direct product of two odd operators and therefore does not contribute to the first order perturbation considered in this paper.
[90]{}
T. Matsui, H. Satz, Phys. Lett. [**B178**]{}, 416 (1986).
Helmut. Satz, arXiv:0512217v2\[hep-ph\].
Agnes Mocsy, arXiv:0811.0337v1\[hep-ph\].
M. Asakawa, T. Hatsuda, and Y. Nakahara, Nucl. Phys. [**A715**]{} (2003) 863; M. Asakawa and T. Hatsuda, Phys. Rev. Lett. [**92**]{} (2004) 012001.
S. Datta, F. Karsch, P. Petreczky, and I.Wetzorke, Phys. Rev. [**D69**]{} (2004) 094507.
T. Umeda, K. Nomura, and H. Matsufuru, Eur. Phys. J. [**C39**]{} (2005) 9; H. Iida, T. Doi, N. Ishii, H. Suganuma.
A. Jakovac, P. Petreczky, K. Petrov, and A. Velytsky, Phys. Rev. [**D75**]{} (2007) 014506.
G. Aarts, C. Allton, M. B. Oktay, M. Peardon, J.-ISkullerud, Phys. Rev. [**D76**]{} 094513 (2007) 094513.
E. V. Shuryak and I. Zahed, Phys. Rev. [**D70**]{} (2004)054507.
W. M. Alberico, A. Beraudo, A. De Pace, and A. Moli-nari, Phys. Rev. [**D72**]{} (2005) 114011.
C. Y. Wong and H. W. Crater, Phys. Rev. [**D75**]{} (2007)034505.
D. Cabrera and R. Rapp, Phys. Rev. [**D76**]{} (2007) 114506.
W. M. Alberico, A. Beraudo, A. De Pace, and A. Moli-nari, Phys. Rev. [**D75**]{} (2007) 074009.
A. Mocsy and P. Petreczky, Phys. Rev. [**D77**]{} (2008)014501.
Defu Hou, Hai-cang Ren, JHEP [**01**]{} (2008) 029.
C. Hoyos, K. Landsteiner and S. Montero, JHEP [**04**]{} (2007) 031.
S. Gubser, Phys. Rev. [**D76**]{}, (2007) 126003.
Agnes Mocsy, arXiv:0908.0746v2\[hep-ph\].
Juan Maldacena, Phys. Rev. Lett. [**80**]{}, (1998) 4859.
S. J. Rey, S. Theisen and J. T. Yee, Nucl. Phys. [**B527**]{}, (1998) 171.
Horace W. Crater, Peter Van Alstine, Phys. Rev. [**D37**]{} (1988) 1982.
Hitoshi Ito, arXiv:9612445v1\[hep-ph\].
Askold DUVIRYAK, arXiv:0602066v2\[math-ph\].
C. Semay, R. Ceuleneer, B. Silvestre-Brac, J. Math. Phys. [**34**]{}, 2215 (1993).
W. Greiner, [*Relativistic Quantum Mechanics: wave equations*]{}, Third Editon, pg 285.
E. Eitchten and F. Feinberg, Phys. Rev. [**D23**]{}, 2724 (1981).
H. W. Crater, J. H. Yoon and C. Y. Wong, Phys. Rev. [**D79**]{}, 034011 (2009).
X. Guo, S. Shi and P. Zhuang, arXiv:1209.5873\[hep-ph\].
[^1]: The holographic spectral function analysis has been done in [@Hoyos]. The dissociation temperature they obtained is inversely proportional to $\lambda$ and higher than the Lattice results for $\lambda$ within the domain (\[coupling\]) .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper presents an occupancy-predicting control algorithm for heating, ventilation, and air conditioning (HVAC) systems in buildings. It incorporates the building’s thermal properties, local weather predictions, and a self-tuning stochastic occupancy model to reduce energy consumption while maintaining occupant comfort. Contrasting with existing approaches, the occupancy model requires no manual training and adapts to changes in occupancy patterns during operation. A prediction-weighted cost function provides conditioning of thermal zones before occupancy begins and reduces system output before occupancy ends. Simulation results with real-world occupancy data demonstrate the algorithm’s effectiveness.'
address: 'Department of Mechanical and Aerospace Engineering, Upson Hall, Cornell University, Ithaca, NY 14853, USA'
author:
- 'Justin R. Dobbs'
- 'Brandon M. Hencey'
bibliography:
- 'library.bib'
- 'mypubs.bib'
title: Model Predictive HVAC Control with Online Occupancy Model
---
model predictive control, MPC, occupancy prediction, on-line training, Markov chains, HVAC
Introduction
============
The long-term increase in energy prices has driven greater interest in demand-based HVAC control. Fixed temperature setpoint schedules and occupancy-triggered operation are commonly used to trim energy consumption, but these approaches have significant drawbacks. First, fixed schedules become outdated; when occupancy patterns change, early or late occupants are left uncomfortable, or the space is conditioned prematurely or for too long. Second, thermal lag limits response speed and thus precludes aggressive temperature set-back. Addressing both schedule inaccuracy and thermal lag requires a stochastic occupancy model and a control scheme that can use it effectively.
Considerable research effort has been directed toward occupancy detection and modeling. Work on detection has focused on boosting accuracy through sensor fusion using probabilistic, neural, or utility networks [@lam2009occupancy; @dodier2006building; @meyn2009sensor; @modelingCountData]. Agent-based models have been used to predict movement within buildings [@liao2012agent; @erickson2009energy], as have Markov chains [@erickson2010occupancy; @page2008generalised; @dong2011integrated]. Erickson and Dong, for example, considered rooms to be Markov states and movements among them to be transitions in order to predict persons’ behavior, while Dong and Lam [@dong2014real] used a semi-Markov model to merge multiple sensor streams into an occupant count estimate. The simpler Page model considered boolean occupancy (occupied or vacant) under a time-heterogeneous Markov chain to generate realistic simulation input data, rather than for on-line forecasting [@page2008generalised].
With the exception of the Page model, the above efforts have found use in heuristic [@erickson2011observe; @selfprogthermostat; @goyal2013occupancy] or model predictive control (MPC) schemes [@dong2014real; @Oldewurtel201215; @goyal2013occupancy], but they face barriers to widespread usage. Most notably, where authors have used MPC, they have also used manually-generated thermal models [@dong2014real; @Oldewurtel201215; @goyal2013occupancy] even though model creation is tedious and time-consuming and therefore expensive. Eager to demonstrate excellent performance, researchers have favored systems with complex topologies and numerous adjustments that yield “one-off” engineering efforts without a clear path to large-scale adoption. The system outlined in [@dong2014real], for instance, uses $\textrm{CO}_{2}$, sound, and light sensors that require carefully set detection thresholds for each room, plus an on-board weather forecasting algorithm in lieu of forecasts already available. We aim, instead, to make occupancy-predicting control accessible to a broader audience by presenting a simple but effective algorithm with a straightforward implementation. For example, we use an automated BIM translation facility outlined in a previous paper [@Greenberg201344], and the core algorithm is industry-standard [MPC]{} with occupancy weighting in the cost function. Each of the very few adjustments serves a clearly-defined purpose, and we have outlined each component’s operation with the practitioner in mind.
Second, recent research has paid little attention to the commissioning and maintenance of occupancy prediction algorithms; model training, if mentioned at all, has been a secondary consideration assumed to by done one time by someone skilled in the art [@page2008generalised; @dong2014real; @erickson2011observe]. Although most training algorithms could be extended to work on-line, ongoing maintenance remains a source of long-term cost neglected by the literature. An occupancy model invariably becomes out-of-date unless it is periodically retrained or can incrementally refine itself with new observations. Our work uses on-line Bayesian inference for stable performance without ongoing manual effort.
The paper progresses as follows. First, we outline the problem formulation. Second, we describe the stochastic occupancy model and its on-line training algorithm. Third, we discuss its integration with model predictive control. Finally, we present simulation results using real-world occupancy data and compare our method’s performance to a correctly set scheduled controller and to an occupancy-triggered controller. Throughout the discussion, the control scenario is kept deliberately simple to emphasize the contribution of occupancy learning and its use with MPC.[^1]
\[sec:Problem-Statement\]Problem Statement
==========================================
We wish to minimize the total energy usage of a building heating (or cooling) system while maintaining occupant comfort. Versus conventional occupancy-triggered or scheduled control, we aim to
- boost comfort by conditioning the space before occupants arrive,
- limit energy consumption by not running the system too early, and
- exploit stored thermal energy by reducing power before occupants leave.
Our approach is based on MPC but uses a cost function weighted by occupancy predictions from a self-training stochastic model (Figure \[fig:ProposedArch\]). At each step, the system measures how much of the previous hour the space was occupied, and the expected occupancy is used to find the best sequence of $N$ future heat inputs to the thermal zone that minimizes the expected cost. The optimization is $$\begin{aligned}\min_{u_{k}\cdots u_{k+N-1}}\quad & \sum_{j=0}^{N-1}{\mathbb{E}\left[g(x_{k+j},u_{k+j},\tau,\Gamma_{k+j})\right]}\\
\textrm{subject to\quad} & \begin{aligned}x_{i+1} & =Ax_{i}+B_{u}u_{i}+B_{w}{\mathbb{E}\left[w_{i}\right]}\quad\forall i\in\mathbb{Z}^{+}\\
0 & \leq u\leq u_{\max}
\end{aligned}
\end{aligned}
\label{eq:firstMpcDef}$$ where
- $k\in\mathbb{Z}^{+}$ is the current time step, and $j\in[0,N-1]$ is the optimization index over the horizon;
- $A\in\mathbb{R}^{n\times n}$ describes the building’s thermal dynamics;
- $x\in\mathbb{R}^{n\times1}$ contains the building’s thermal state;
- $u_{k\ldots k+N-1}$ contains the controller output, constrained within the system’s capacity $u_{\max}$;
- $B_{u}$ is a vector that connects the heat input $u$ to the zone air volume;
- $w_{k}$ is the current weather observation, and $w_{k+1\ldots k+N-1}$ contains an up-to-date weather prediction;
- $B_{w}$ is a vector that connects the weather conditions to the building envelope;
- $\tau$ is the temperature setpoint, which is constant for this study (but can be varied in practice);
- $\Gamma_{k}$ is the latest occupancy measurement, and $\Gamma_{k+1\ldots k+N-1}$ are the predicted occupancies; and
- $g(x,u,\tau,\Gamma)$ is a cost function that penalizes total energy consumption and penalizes discomfort based on the occupancy $\Gamma$.
The expectation operator ${\mathbb{E}\left[g\right]}$ in Equation \[eq:firstMpcDef\] reflects that future values of $g$ require predictions of occupancy and of the weather. The optimization yields an optimal sequence of $N$ power commands to the HVAC system, where positive values are heat and negative are cooling; the first command $u_{k}$ is applied, and the rest are discarded. The previous and current occupancy observations are then used to train the occupancy model, and the entire process repeats the next time step (Figure \[fig:Process-flow-during\]).
![\[fig:Process-flow-during\]Process flow during operation.](process)
Two assumptions are made in this presentation. First, we treat the weather forecast as accurate so that we can later omit the expectation operator from $w$. Second, we use a very simple cost function with constant efficiency and a single linear actuator. These assumptions improve clarity but are not required in practice. Where available, weather uncertainty data can be rolled into the cost function in order to improve robustness [@Oldewurtel201215]. Multiple actuators (e.g. radiant and forced air with vastly different response times) or nonlinear actuation (e.g. variable air volume damper position) can be pulled into the dynamical model and the cost function without undermining the basic approach [@haves2010model; @Oldewurtel201215; @goyal2013occupancy]. Finally, the energy penalty gain can be varied over time to reflect, for example, changing system efficiency or electricity cost.
Building Thermal Model
======================
Thermal model accuracy influences controller performance, so we need a thermal model that closely approximates the dominant dynamics. Here we outline how the state-space building model is generated, and we validate it against EnergyPlus simulation results.
Thermal model creation has historically been a manual process contributing substantially to MPC implementation cost. Research efforts such as the [*Sustain*]{} platform (Figure \[fig:Sustain\]) [@Greenberg201344; @dobbsautomatic] and the Building Resistance-Capacitance Modeling Toolbox [@sturzenegger2014brcm] have arisen to streamline the creation of dynamical equations suitable for MPC. Here we have used a module in [*Sustain*]{} to generate a resistor-capacitor network directly from a CAD model. The thermal model states are the building’s internal temperatures, including zone air plus wall layers and roofing materials that are not normally measured; a state observer can easily estimate these values during operation.[^2] Although not used here, ways to automatically tune the RC network parameters on-line and even estimate disturbances such as solar load have recently been introduced [@radecki_online_2013].
The model used for this study has 41 states: one for zone air and the rest for building structure[^3]. It assumes well-mixed air and uses time-invariant convection coefficients. Fixed coefficients imply that the thermal gradients are always in the same direction, whereas EnergyPlus switches coefficients depending on whether the gradient enhances convection [@EnergyPlusReference2011]. In practice, for improved accuracy, the RC network can be adjusted at each step, or a nonlinear model may be used. We have included limited support for radiant transfer using coefficients from EnergyPlus’ `Simple` and `SimpleCombined` convection algorithms [@EnergyPlusReference2011].
![\[fig:Sustain\]The [*Sustain*]{} modeling and simulation environment. (New York Times building shown.)](sustainShotNyTimes){width="90mm"}
The model accepts the following inputs:
- the outside dry-bulb temperature,
- the ground temperature, and[^4]
- heat injected by the control system to the space (positive or negative).
The state equation of the building is $$x_{k+1}=A_{b}x_{k}+B_{w}w_{k}+B_{u}u_{k},\label{eq:buildingStateEquation}$$ where $k$ is the time step (in hours) and $x_{k}$ is the complete temperature state vector containing the zone temperature $x_{k}^{\textrm{zone}}$. The vector $w_{k}$ contains the weather forecast, and $u_{k}$ is the heat injected into the room by the HVAC system. The sign of $u_{k}$ and its constraints can be made negative, or the sign of the vector $B_{u}$ can be reversed, for cooling.
Let us now validate the model by comparing the zone temperature time response of the RC network to EnergyPlus results under simplified conditions. The goal is not to exactly match EnergyPlus, but rather to show that the dominant response is plausibly close. To do this, we have simulated the building using first the RC network and then EnergyPlus under the following set of conditions:
- a step change in air, ground, and sky infrared temperatures from 10$^{\circ}\textrm{C}$ to 20$^{\circ}\textrm{C}$,
- no wind or humidity, and
- EnergyPlus heat transfer algorithms: `Simple` convection for interior, `SimpleCombined` for exterior, and `CTF` (conduction transfer function) for walls.
The RC network implementation lacks support for sky infrared transfer through windows; by matching the sky radiant temperature to the outside air temperature, we have removed this source of discrepancy from the simulation. Under the simplified conditions, very similar response times (Figure \[fig:rcVersusEp\]) suggest that the RC model is adequate for demonstration.
![\[fig:rcVersusEp\]RC network (solid) and EnergyPlus (dashed) simulation results for a step change in ambient temperature.](comparisonEplusToRC)
{#section .unnumbered}
\[sec:Stochastic-occupancy-model\]Stochastic Occupancy Model
============================================================
The heart of our method is its on-line trained Markov occupancy model that quickly adapts and enables the MPC to predict occupancy. The input is a stream of asynchronous pulses from pyroelectric infrared (PIR) or similar sensors that indicate whether at least one person is in the space. We have chosen the Mitsubishi Electric Research Lab (MERL) motion detector data set [@merldata], which consists of a series of one-second pulses from various motion sensors located throughout hallways and conference rooms in MERL.[^5] The meetings in the Belady conference room show a good balance between repetition and variety to showcase the benefits of on-line learning.
Markov Chain Formulation\[sub:Markov-chain-formulation\]
--------------------------------------------------------
The occupancy model is as a periodic Markov chain updated at every observation. The occupancy at time $k$ is either $\gamma_{k}=$1 (occupied) or $\gamma_{k}=0$ (vacant). The current occupancy state and the time of day determine the probability of future occupancy. We wish to estimate the probabilities $$\begin{aligned}p_{k} & ={\mathbb{P}\left(\gamma_{k+1}=1\mid\gamma_{k}=1\right)},\\
q_{k} & ={\mathbb{P}\left(\gamma_{k+1}=1\mid\gamma_{k}=0\right)}.
\end{aligned}$$ The transition probabilities of this two-state time-varying Markov chain (Figure \[fig:markovChain\]a) are periodic; we have chosen a period $M=24$ hours, so $p_{24}\equiv p_{0}$ and $q_{24}\equiv q_{0}$. To better visualize the periodicity, we unroll the Markov chain into $2M$ states (Figure \[fig:markovChain\]b), where each hour has a $1_{k}$ and $0_{k}$ state. Although $k$ in general grows without bound, its range is limited to $0\leq k\leq M-1$ when dealing with the Markov chain. The choice of $M$ affects how learned patterns relate to subsequent predictions; if space usage patterns vary significantly across the weekdays, one might want to prevent occupancy observations on Monday from influencing control actions on Tuesday, in which case a one-week chain would be more appropriate. For this study, the one-day Markov chain is trained using Monday through Friday occupancy data from the MERL data set, ignoring weekends. In practice, one could switch to a different Markov chain or use occupancy-triggered control over weekends.
Training
--------
In contrast to batch training, which uses a fixed-size history (Figure \[fig:trainingSchematic\]a) , on-line incremental training proceeds without user intervention. It uses observations to update density functions for each of the transition probabilities; the expected values of these density functions in turn populate the Markov chain’s transition matrix (Figure \[fig:trainingSchematic\]b).
![\[fig:trainingSchematic\]Conventional batch training (a) versus the proposed on-line incremental Bayesian training algorithm with forgetting (b).](training)
Boolean occupancy lacks granularity that could otherwise make predictions more accurate. For example, occupancy for the entirety of the previous hour implies different future occupancy compared to just a few minutes. The question we wish to answer is: *Given the space was occupied for a certain fraction of the previous hour, for what portion of subsequent hours do we expect occupancy?* We approach the problem in three steps. First, we explain the simplest case where boolean occupancy is directly observed. Second, we augment the boolean training with forgetting capability. Finally, we refine the approach to use fractional occupancy in order to make predictions more precise.
### Boolean observed occupancy
Each state of the unrolled Markov chain (Figure \[fig:markovChain\]b) has two outgoing transition paths, analogous to a coin toss where the coin’s bias is unknown. The well-known probability function of a biased coin is $$\psi(\theta,N,N_{H})=\binom{N}{N_{H}}\theta^{N_{H}}(1-\theta)^{N-N_{H}},$$ where $\binom{N}{N_{H}}$ is the number of ways to permute $N_{H}$ heads in a sequence of $N$ tosses, and $\theta$ is the heads bias (with 0.5 being a fair coin). This function can be parameterized on $\theta$, $N$, or $N_{H}$ depending on the purpose. With the bias $\theta=\theta_{0}$ known and the number of tosses $N=N_{0}$ fixed, the probability of obtaining $N_{H}$ heads, $\psi(\theta=\theta_{0},N=N_{0},N_{H})$, is a discrete binomial distribution over $N_{H}$. When $N$ and $N_{H}$ are fixed, $\psi(\theta,N=N_{0},N_{H}=N_{H_{0}})$ is the probability density over the bias $\theta$, with $\int_{0}^{1}\psi(\theta,N=N_{0},N_{H}=N_{H_{0}})\,\D\theta=1$.[^6]
Instead of computing $\psi$ using $N$ and $N_{H}$ all at once, we can obtain it iteratively using Bayes’ rule. Suppose we have a sequence of outcomes $x_{j}\in\{1,0\}$ where $1$ means heads. The distribution, now parameterized only on $\theta$, is defined recursively as $$\begin{aligned}\psi_{j}(\theta\mid x_{1\ldots j}) & \sim\psi_{j-1}(\theta\mid x_{1\ldots j-1})\Phi(\theta,x_{j})\\
& =\frac{\psi_{j-1}(\theta\mid x_{1\ldots j-1})\Phi(\theta,x_{j})}{\int_{0}^{1}\psi_{j-1}(\theta\mid x_{1\ldots j-1})\Phi(\theta,x_{j})\D\theta},
\end{aligned}$$ where $$\Phi(\theta,x)=\begin{cases}
\theta & x=1\\
1-\theta & x=0,
\end{cases}$$ and $\psi_{0}(\theta)=1$ is a uniform distribution reflecting no prior knowledge of the bias. The $\sim$ notation means dividing by a constant so that $\int_{0}^{1}\psi_{j}(\theta)\,\D\theta=1$ holds. Our best guess of the bias is ${\mathbb{E}\left[\psi_{j}(\theta)\right]}=\int_{0}^{1}\theta\psi_{j}(\theta)\D\theta$.
Now let us apply this analogy to occupancy prediction. Coin toss outcomes are independent, but occupancy transition probabilities depend on the current state. An any given time there are two possible states, so we need to maintain two distributions per time step. Let $\gamma_{k}\in\{0,1\}$ be the occupancy. The transition probabilities of interest are $$\begin{aligned}p_{k} & ={\mathbb{P}\left(\gamma_{k+1}=1\mid\gamma_{k}=1\right)}={\mathbb{E}\left[f_{k}(p_{k})\right]}\\
q_{k} & ={\mathbb{P}\left(\gamma_{k+1}=1\mid\gamma_{k}=0\right)}={\mathbb{E}\left[g_{k}(q_{k})\right]},
\end{aligned}$$ where the density functions $f_{k}(p_{k})$ and $g_{k}(q_{k})$ are the latest iterations of $f_{k,j}(p_{k})$ and $g_{k,j}(q_{k})$, updated each training instance $j$ using $$\begin{aligned}f_{k,j}(p_{k}\mid\gamma_{1\ldots k+1},\gamma_{k}=1) & \sim f_{k,j-1}(p_{k}\mid\gamma_{1\ldots k})\Phi(p_{k},\gamma_{k+1})\\
f_{k,j}(p_{k}\mid\gamma_{1\ldots k+1},\gamma_{k}=0) & =f_{k,j-1}(p_{k}\mid\gamma_{1\ldots k})\\
g_{k,j}(q_{k}\mid\gamma_{1\ldots k+1},\gamma_{k}=0) & \sim g_{k,j-1}(q_{k}\mid\gamma_{1\ldots k})\Phi(q_{k},\gamma_{k+1})\\
g_{k,j}(q_{k}\mid\gamma_{1\ldots k+1},\gamma_{k}=1) & =g_{k,j-1}(q_{k}\mid\gamma_{1\ldots k}).
\end{aligned}$$ The $\sim$ indicates normalization, and $f_{0}(p_{k})=1$ and $g_{0}(q_{k})=1$ as before. The distribution $f_{k,j}(p_{k})$ does not change from $f_{k,j-1}(p_{k})$ unless the space was occupied, and $g_{k,j-1}(q_{k})$ is also left alone unless the space was vacant. In other words, to update the distributions for a state, a transition out of that state must have been observed.
### Forgetting Factor
As training proceeds, the distributions $f_{k}(p_{k})$ and $g_{k}(q_{k})$ become increasingly narrow and converge toward delta functions, the oldest and newest training data exerting equal but ever-decreasing influence on the model; even the newest training data becomes diluted. This is acceptable for batch training, where the history length is chosen explicitly, but not for incremental training, where eventually the distributions cannot change at all. We introduce a forgetting factor $\lambda$ to gradually discount older training data and allow the Markov chain to retain its flexibility. Linear forgetting is implemented using $$\begin{aligned}f_{k,j}^{\prime}(p_{k}) & =\lambda f_{k,j}(p_{k})+(1-\lambda)f_{0}(p_{k})\\
g_{k,j}^{\prime}(q_{k}) & =\lambda g_{k,j}(q_{k})+(1-\lambda)g_{0}(q_{k}),
\end{aligned}
\label{eq:forgetting}$$ where $f_{0}(p_{k})=1$ and $g_{0}(q_{k})=1$, and $f_{k,j}(p_{k})$ and $g_{k,j}(q_{k})$ are the posterior distributions that have just been trained before application of forgetting.[^7]
There is no direct equivalence between forgetting factors and batch training history length; batch training (Figure \[fig:trainingSchematic\]a) is analogous to a finite impulse response (FIR) filter with a defined memory length, while incremental training (Figure \[fig:trainingSchematic\]b) is structurally reminiscent of an infinite impulse response (IIR) filter where the previous output is fed back into the filter. With batch training, the hand-picked data set may not contain all the transitions of interest, so some transitions may not be trained at all. The incremental approach applies training and forgetting simultaneously, retaining infrequently observed transitions longer.
To illustrate the effect of forgetting on the distributions, we have trained a single state of the Markov chain repeatedly using alternating transitions $\gamma_{0}=0\to\gamma_{1}=1$ and $\gamma_{0}=0\to\gamma_{1}=0$. This is analogous to flipping an unbiased coin numerous times and observing heads every other flip, from which we expect an increasingly narrow distribution for $p_{0}$ peaking near 0.5 (Figure \[fig:forgettingFactor\]a). This result would be preferred if the pattern were never expected to change, but such concentration in the distribution hinders its ability to change and is therefore undesirable. Using 15% forgetting ($\lambda=0.85$) gives a distinctly broader distribution lifted off the horizontal axis (Figure \[fig:forgettingFactor\]b). The distribution—and therefore its expected value—shifts laterally with each alternate observation, even after many iterations; this extra mobility reflects greater adaptability. The value $\lambda=0.85$ is much more forgetful than would be used in practice; Section \[sub:Choosing\] will explore the relationship between $\lambda$ and prediction accuracy.
![\[fig:forgettingFactor\]A particular occupancy transition probability distribution when trained with alternating data using no forgetting ($\lambda=1$) (a) and with considerable forgetting ($\lambda=0.85$) (b). The forgetting factor broadens the distribution and allows it to shift laterally even when extensively trained, reflecting greater ability to adjust to changes in space usage. For each case, the initial distribution is uniform (black horizontal trace). Higher levels of training are shown as brighter color.](trainingDist)
### Using Fractional Occupancy
Measuring the percentage of occupancy over each time makes occupancy predictions more precise. To convert the asynchronous pulses from PIR sensors (Figure \[fig:discretization\]a) into a discrete-time sequence of fractional values, we apply a simple two-step heuristic. First, we merge closely-spaced pulses using a minimum dwell time to get a square wave signal $\gamma(t)$ (Figure \[fig:discretization\]b). Then we superimpose a fixed time grid over the signal and average it over each step to obtain the discrete sequence $$\Gamma_{k}=\frac{1}{t_{k}-t_{k-1}}\int_{t_{k-1}}^{t_{k}}\gamma(t)\,\textrm{d}t\equiv{\mathbb{P}\left(\gamma_{k}=1\right)}\in[0,1],\label{eq:bigGamma}$$ essentially treating $\Gamma_{k}$ (Figure \[fig:discretization\]c) as the duty cycle sequence of the pulse width modulated signal $\gamma(t)$. We subsequently pretend that $\gamma(t)$ is sampled probabilistically through $\Gamma_{k}$ with a distribution over the Markov state space $\pi_{k}\in\mathbb{R}^{1\times2M}$. The statements $\Gamma_{k}=60\%$ and ${\mathbb{P}\left(\gamma_{k}=1\right)}=0.6$ are considered equivalent.
![\[fig:discretization\]Asynchronous sensor pulses (a); derived continuous signal using dwell time (b); resulting discrete-time occupancy percentage (c).](discretization)
From this we estimate the occupancy at time $k+1$ using $$\begin{alignedat}{2}{\mathbb{P}\left(\Gamma_{k+1}=1\mid\Gamma_{k}\right)} & = & \Gamma_{k} & {\mathbb{P}\left(\gamma_{k+1}=1\mid\gamma_{k}=1\right)}\\
& & +\,(1-\Gamma_{k}) & {\mathbb{P}\left(\gamma_{k+1}=1\mid\gamma_{k}=0\right)}\\
& = & \Gamma_{k} & {\mathbb{E}\left[p_{k}\right]}+(1-\Gamma_{k}){\mathbb{E}\left[q_{k}\right]},
\end{alignedat}$$ where the expectation operator reflects the fact that $p_{k}$ and $q_{k}$ are estimated via $f_{k}(p_{k})$ and $g_{k}(q_{k})$. At each step $k$, there are four possible state transitions with associated posterior distributions $$\begin{alignedat}{2}\gamma_{k}=1 & \to\gamma_{k+1}=1:\quad & f_{k,j}^{(1)}(p_{k}) & \sim\Phi(p_{k},1)f_{k,j-1}(p_{k}),\\
\gamma_{k}=1 & \to\gamma_{k+1}=0: & f_{k,j}^{(0)}(p_{k}) & \sim\Phi(p_{k},0)f_{k,j-1}(p_{k}),\\
\gamma_{k}=0 & \to\gamma_{k+1}=1: & g_{k,j}^{(1)}(q_{k}) & \sim\Phi(q_{k},1)g_{k,j-1}(q_{k}),\\
\gamma_{k}=0 & \to\gamma_{k+1}=0: & g_{k,j}^{(0)}(q_{k}) & \sim\Phi(q_{k},0)g_{k,j-1}(q_{k}),
\end{alignedat}$$ where $f_{k,j}^{(1)}$ is the updated posterior distribution as if $\gamma_{k}=1$ and $\gamma_{k+1}=1$ had been observed, $f_{k,j}^{(0)}$ is similar to $f_{k,j}^{(1)}$ but updated as if $\gamma_{k+1}=0$ had been observed, and likewise for $g_{k,j}^{(1)}$ and $g_{k,j}^{(0)}$. To obtain $f_{k,j}(p_{k})$, we blend $f_{k,j}^{(1)}(p_{k})$ and $f_{k,j}^{(0)}(p_{k})$ according to the later observation $\Gamma_{k+1}$. We then weight the training according to $\Gamma_{k}$, which reflects how likely the space was to have started occupied; values of $\Gamma_{k}$ closer to one apply more training to $f_{k}(p_{k})$, while those closer to zero cause heavier training of $g_{k}(q_{k})$. The training for $g_{k,j}(q_{k})$ follows analogously. $$\begin{alignedat}{2}f_{k,j}(p_{k})\, & =\, & \Gamma_{k}\, & \left(\Gamma_{k+1}f_{k,j}^{(1)}(p_{k})+(1-\Gamma_{k+1})f_{k,j}^{(0)}(p_{k})\right)\\
& \ + & (1-\Gamma_{k})\, & f_{k,j-1}(p_{k})\\
g_{k,j}(q_{k})\, & =\ & (1-\Gamma_{k})\, & \left(\Gamma_{k+1}g_{k,j}^{(1)}(q_{k})+(1-\Gamma_{k+1})g_{k,j}^{(0)}(q_{k})\right)\\
& \ + & \Gamma_{k}\, & g_{k,j-1}(q_{k})
\end{alignedat}
\label{eq:fractionalPosteriors}$$ Once the new distributions $f_{k,j}(p_{k})$ and $g_{k,j}(q_{k})$ have been found, forgetting is applied similarly to Equation \[eq:forgetting\], where Equations \[eq:fractionalPosteriors\] are used instead for the posterior distributions. The post-forgetting distributions are then stored.
### Effect of Training on Distribution Shape
To illustrate the connection between training data patterns and the shapes of $f_{k}(p_{k})$ and $g_{k}(q_{k})$, we have trained two Markov chains with the MERL Belady conference room data from March 22 to June 9 and sampled the distributions afterward. In Figure \[fig:Transition-probabilities\], two sets of distributions—one for $\to$ (a) and the other for $\to$ (b)—are shown for both strong forgetting ($\lambda=0.85$, solid) and no forgetting ($\lambda=1.0$, dashed). In Figure \[fig:Transition-probabilities\]a we see that both occupancy and vacancy at strongly imply vacancy at . In other words, early morning occupancy is very uncommon and usually brief. Because occupancy is rare at , the transition $\gamma_{2}=1_{2}\to\gamma_{3}=1_{3}$ (blue) is very weakly trained and has a very flat distribution. In Figure \[fig:Transition-probabilities\]b, we see that occupancy at is more varied, resulting in more typical bell-shaped distributions. The distributions for suggest that meetings are likely to continue into the next hour but are unlikely to start the following hour. The distributions for $\lambda=0.85$ are shaped similarly to those for $\lambda=1.0$ but are markedly subdued with expected values closer to $0.5$.
![\[fig:Transition-probabilities\]Probability densities for the occupied-to-occupied (blue) and vacant-to-occupied (red) at (a) and (b). Two cases are shown: a heavy forgetting factor ($\lambda=0.85$, solid) and no forgetting ($\lambda=1.0$, dashed). ](nightVsDayDistr2)
Transition Matrix and Occupancy Prediction
------------------------------------------
Recall from Section \[sub:Markov-chain-formulation\] and Figure \[fig:markovChain\]b the Markov chain has states $1_{0}\ldots1_{23}$ and $0_{0}\ldots0_{23}$. The probability distribution of the current occupancy state $\pi_{k}\in\mathbb{R}^{1\times2M}$ evolves according to $\pi_{k+1}=\pi_{k}P$. The matrix $P$ can be constructed from the four blocks: $P^{(\textrm{\Romannum{1})}}$ for $1_{k}\to1_{k+1}$, $P^{(\textrm{\Romannum{2}})}$ for $1_{k}\to0_{k+1}$, $P^{(\textrm{\Romannum{3}})}$ for $0_{k}\to1_{k+1}$, and $P^{(\textrm{\Romannum{4}})}$ for $0_{k}\to0_{k+1}$ transitions. The entries for $P^{(\textrm{\Romannum{1}})}$ and $P^{(\textrm{\Romannum{4}})}$ are the expected values of $p_{0\ldots M-1}$ and $q_{0\ldots M-1}$, and the other two matrices are their complements. $$\begin{aligned}
P_{ik}^{\textrm{(\Romannum{1})}} &=
\begin{cases}
\makebox[\foo][l]{${\mathbb{P}\left(\gamma_k=1\mid\gamma_i=1\right)}={\mathbb{E}\left[p_i\right]}$} & (\dagger) \\
0 & \textrm{otherwise}
\end{cases}\\P_{ik}^{\textrm{(\Romannum{2})}} &=
\begin{cases}
\makebox[\foo][l]{${\mathbb{P}\left(\gamma_k=0\mid\gamma_i=1\right)}=1-{\mathbb{E}\left[p_i\right]}$} & (\dagger) \\
0 & \textrm{otherwise}
\end{cases}\\P_{ik}^{\textrm{(\Romannum{3})}} &=
\begin{cases}
\makebox[\foo][l]{${\mathbb{P}\left(\gamma_k=1\mid\gamma_i=0\right)}={\mathbb{E}\left[q_i\right]}$} & (\dagger) \\
0 & \textrm{otherwise}
\end{cases}\\P_{ik}^{\textrm{(\Romannum{4})}} &=
\begin{cases}
\makebox[\foo][l]{${\mathbb{P}\left(\gamma_k=0\mid\gamma_i=0\right)}=1-{\mathbb{E}\left[q_i\right]}$} & (\dagger) \\
0 & \textrm{otherwise}
\end{cases}\end{aligned}$$where $\dagger$ means $k=i+1\ (\bmod M)$. For example, $P^{\textrm{(\Romannum{1})}}$ takes the form $$P^{(\textrm{\Romannum{1}})}={\mathbb{E}\begin{bmatrix}0 & p_{0}\\
& 0 & p_{1}\\
& & \ddots & \ddots\\
& & & 0 & p_{M-2}\\
p_{M-1} & & & & 0
\end{bmatrix}}.$$ The complete matrix is $$P=\left[\begin{array}{c|c}
P^{\textrm{(\Romannum{1})}} & P^{\textrm{(\Romannum{2})}}\\
\hline {\rule{0pt}{10pt}}P^{\textrm{(\Romannum{3})}} & P^{\textrm{(\Romannum{4})}}
\end{array}\right].$$ The expected occupancy $m$ steps in the future given a current estimate $\Gamma_{k}$ is $${\mathbb{E}\left[\Gamma_{k+j}\mid\Gamma_{k}\right]}=\underbrace{\begin{bmatrix}\Gamma_{k}{\mathbb{1}}_{1\times M}^{k} & (1-\Gamma_{k}){\mathbb{1}}_{1\times M}^{k}\end{bmatrix}}_{\pi_{k}}P^{j}\begin{bmatrix}1_{M\times1}\\
0_{M\times1}
\end{bmatrix}\label{eq:occupancyProjection}$$ where ${\mathbb{1}}_{1\times M}^{k}$ is a vector with the $k$th element set to one and all others left zero.
MPC Formulation
===============
To balance competing demands for occupant comfort and low total energy consumption, we need to avoid conditioning the space when vacancy is expected; the level of comfort should scale with occupancy. To simplify the cost function, we have augmented the building’s state space model with the non-changing temperature setpoint and a weather forecast shift-register system, i.e. $$\underbrace{\begin{bmatrix}x_{k+1}\\
\tau_{k+1}\\
\phi_{k+1}
\end{bmatrix}}_{\tilde{x}_{k+1}}=\tilde{A}\underbrace{\begin{bmatrix}x_{k}\\
\tau_{k}\\
\phi_{k}
\end{bmatrix}}_{\tilde{x}_{k}}+\underbrace{\begin{bmatrix}B_{u}\\
0\\
0
\end{bmatrix}}_{\tilde{B}}u_{k},$$ where $x$ is the building’s thermal state (Equation \[eq:buildingStateEquation\]), $\tau$ is the comfort setpoint, and $\phi$ is a shift-register state that iterates through the weather forecast over the MPC horizon. The augmented matrix $\tilde{A}$ connects the weather forecast to the building thermal model internally.[^8] We seek the optimal control law $$\begin{alignedat}{2}u_{k}^{*}(\tilde{x}_{k},\Gamma_{k}) & = & \underset{u}{\arg\min}\quad & {\mathbb{E}\left[\sum_{j=0}^{N-1}g(\tilde{x}_{k+j},u_{k+j},\Gamma_{k+j})\right]}\\
& & \textrm{subject to}\quad & \begin{aligned}\tilde{x}_{i+1} & =\tilde{A}\tilde{x}_{i}+\tilde{B}u_{i}\quad\forall i\in\mathbb{Z}^{+}\\
0 & \leq u\leq u_{\max}
\end{aligned}
\end{alignedat}$$ where $u_{k+j}$ is an individual control action and $\Gamma_{k+j}\in[0,1]$ is an occupancy measurement or prediction. This is standard except that the stage cost adjusts the discomfort weigh using occupancy, i.e. $$\begin{aligned}g(\tilde{x},u,\Gamma) & =\tilde{x}^{\top}\Gamma Q\tilde{x}+r\left|u\right|\\
& =\Gamma\beta\left(x_{\textrm{zone}}-\tau\right)^{2}+r\left|u\right|,
\end{aligned}$$ where
- $\tilde{x}$ is the augmented system state vector and $x_{\textrm{zone}}$ is the zone air temperature being controlled,
- $u$ is the heat input to the zone,
- $\Gamma$ is the observed or predicted occupancy,
- $\tau$ is the comfortable setpoint temperature (constant),
- $Q$ is a matrix that extracts $\beta(x_{\textrm{zone}}-\tau)^{2}$ from $\mbox{\ensuremath{\tilde{x}}}^{\top}Q\tilde{x}$, and
- $\beta$ and $r$ are the discomfort and energy cost gains.[^9]
The many ($2^{N}$) possible occupancy state trajectories, along with the constraints on $u$, make it difficult to find a closed-form solution using exact dynamic programming. (Recall from Figure \[fig:markovChain\] that each occupancy state has two possible outgoing transitions.) If we instead condition all occupancy predictions solely on the present observation, we obtain the approximation $$\begin{aligned}u_{k}^{*}(\tilde{x}_{k},\Gamma_{k}) & \approx\underset{0\leq u\leq u_{\max}}{\arg\min}\Biggl\{ g(\tilde{x}_{k},u_{k},\Gamma_{k})\\
& +\sum_{j=1}^{N-1}g\left(\tilde{x}_{k+j},u_{k+j},\mathbb{E}\left[\Gamma_{k+j}\mid\Gamma_{k}\right]\right)\Biggr\},
\end{aligned}$$ where ${\mathbb{E}\left[\Gamma_{k+j}\mid\Gamma_{k}\right]}$ comes from Equation \[eq:occupancyProjection\].[^10] The optimization is then $$\begin{aligned}\min_{u_{k}\cdots u_{k+N-1}}\quad & \sum_{j=0}^{N-1}\tilde{x}_{k+j}^{\top}{\mathbb{E}\left[\Gamma_{k+j}\mid\Gamma_{k}\right]}Q\tilde{x}_{k+j}+r\left|u_{k+j}\right|\\
\textrm{subject to\quad} & \begin{aligned}\tilde{x}_{i+1} & =\tilde{A}\tilde{x}_{i}+\tilde{B}u_{i}\quad\forall i\in\mathbb{Z}^{+}\\
0 & \le u\le u_{\max}
\end{aligned}
\end{aligned}$$ As with conventional MPC, the controller applies $u_{k}$ to the system and discards $u_{k+1}\ldots u_{k+N-1}$; the solution is repeated at each subsequent step.
The controller never reaches the setpoint $\tau$ for two reasons. First, including energy in the cost function counteracts temperature regulation, with the trade-off tuned through the ratio $\beta/r$. Second, the discomfort cost is weighted by *expected* occupancy, which never reaches 1.0. We have chosen to penalize $\left|u\right|$, rather than $u^{2}$, because quadratic cost suppresses peaks and spreads control action over time; peak suppression inhibits the full system shutdown necessary to save energy during vacancy. When high occupancy is predicted, the discomfort cost (Figure \[fig:Cost-function\]) becomes steeper and causes the temperature to more closely approach the setpoint.
![\[fig:Cost-function\]Discomfort cost for high expected occupancy (blue) and low expected occupancy (red). When high occupancy is predicted, the curve steepens and less deviation from the setpoint is permitted.](costFunction)
Comparison to Conventional Control
==================================
Experimental Setup
------------------
To demonstrate the algorithm’s advantages over conventional control, we have run a simulation under the following conditions:
- MERL occupancy data for the Belady conference room (sensors 452 and 453) from February 12 to April 10, 2007;
- EnergyPlus weather data for Elmira, NY starting March 1 (typical meteorological year) and a three-week warm-up period;
- no un-modeled disturbances;
- one-hour time step;
- system capacity of 8.0kW.
The thermal model is the single-zone building RC network discussed previously. To emphasize the benefit of prediction, we have chosen the weather period to just saturate the control output in typical winter conditions. (These conditions emphasize the need to predict more than one hour out; we could have chosen January and increased the system capacity slightly for the same result.)
Choosing $\lambda$\[sub:Choosing\]
----------------------------------
Before we run the simulation, we need to choose the forgetting factor. Without forgetting ($\lambda=1.0$), consistent occupancy patterns allow predictions to asymptotically approach $\Gamma=0$ and $\Gamma=1$, but the ever-lengthening effective history length hinders adaptation and leads to very large prediction error. At the other extreme, high forgetting ($\lambda\ll1.0$) gives a model easily distracted by irregularities that consistently predicts occupancy near $\Gamma=0.5$, which again leads to high prediction error. Intuition suggests that a minimum prediction error should exist between these limits, and indeed this is the case. Figure \[fig:Influence-of-forgetting\] shows the relationship between $\lambda$ and one-hour prediction error using the simulation occupancy data, with $\lambda=0.974$ giving the best prediction accuracy. Of course, there is no guarantee that the best past value of $\lambda$ will work well in the future; nonetheless, the convexity suggests that $\lambda$ could be calibrated on-line with an extremum-seeking algorithm [@ariyur2003real].
![\[fig:Influence-of-forgetting\]Influence of forgetting factor $\lambda$ on one-hour root-mean-square prediction error. These results were obtained by training the model incrementally over the MERL Belady conference room occupancy data from February 12 to April 10, 2007 and simultaneously comparing each observation to the prediction made in the previous hour.](predErrorVsFF)
Performance Comparison
----------------------
Figure \[fig:Simulation-results\] shows simulation results for three identically-tuned MPC implementations:
1. a purely occupancy-triggered controller,
2. a scheduled controller supplemented with occupancy triggering, and
3. an on-line trained occupancy-predicting controller with one week of pre-training ($\lambda=97.4\%$).
The occupancy-triggered controller (green) maintains $\tau=23^{\circ}\textrm{C}$ during occupied hours and $10^{\circ}\textrm{C}$ during vacant hours. The scheduled controller uses the same setpoint from to and any time the space is occupied. To simplify the simulation, all three controllers ignore occupancy and control to $10^{\circ}\textrm{\textrm{C}}$ ($50^{\circ}\textrm{F}$) over weekends.
### Energy and Comfort
The occupancy-triggered controller consumes by far the least energy because it does not account for thermal lag or expected occupancy and therefore runs the least. Not surprisingly, its comfort performance upon occupant arrival is very poor, with large leading spikes on the discomfort trace in Figure \[fig:Simulation-results\]c and frequent calls for maximum output power in Figure \[fig:Simulation-results\]d. The scheduled controller leaves plenty of margin around the typical occupancy envelope and consequently yields excellent comfort at the expense of energy efficiency. The comfort performance of occupancy predicting MPC lies between these two methods, with peak discomfort slightly worse than scheduled control but without the severe deviations of triggered control. Table \[fig:resultTable\] shows up to 19% energy savings compared to the scheduled controller and significantly lower peak discomfort than the occupancy-triggered controller.
Perhaps more interesting than the discomfort peak is its distribution. Figure \[fig:histogram\] shows how many times various occupancy-weighted discomfort levels occur under each control method. It comes as little surprise that the scheduled controller maintains discomfort within $2^{\circ}\textrm{C}$ at all times. (Clearly, though, an out-of-date schedule would not perform this well, so this is a rather optimistic profile of scheduled control.) The occupancy-predicting controller maintains discomfort less than $2^{\circ}\textrm{C}$ more than 94% of the time with relatively mild outliers. The occupancy-triggered controller trails with 75% incidence of low discomfort and numerous severe violations. In summary, the occupancy predicting control scheme yields comfort performance that rivals that of a properly tuned schedule.
Energy performance is also as expected. The conservative schedule leaves ample time to pre-condition the space along with some margin in the evening. The cost of this performance is 24% more total energy over the simulation than the occupancy-triggered controller. Consumption by the occupancy-predicting controller is moderate, at 12% more than the triggered and 19% less than the scheduled control, and there are very few instances where the system needs to run at maximum power to catch up.
![\[fig:histogram\]Distribution of occupancy-weighted discomfort over a two-month simulation. The properly-tuned schedule shows very little discomfort over the two-month simulation, while occupancy-triggered control produces many severe instances of discomfort. Occupancy predicting control yields a distribution similar to that of scheduled control but shifted slightly to the right.](histogram2)
------------ ------- ------ ---------- ------------- -------------
Total Peak Variance Total (kWh) Savings (%)
Predictive 270 3.73 0.20 2493 19
Triggered 396 7.67 0.72 2237 27
Scheduled 108 1.69 0.04 3088 0
------------ ------- ------ ---------- ------------- -------------
: \[fig:resultTable\]Predictive, triggered, and scheduled control performance summary for two-month simulation.
Conclusion
==========
We have demonstrated the use of model predictive control with a stochastic occupancy model to reduce HVAC energy consumption. Using occupancy predicted by an automatically-trained Markov chain, the algorithm is simplified by approximate dynamic programming where occupancy is projected multiple steps into the future using a current observation. We remark that although our method relies on weather forecasts and a dynamical model of the building, on-line data sources and emerging software tools have made these easy to acquire.
We have made some simplifications to improve clarity. First, we have chosen a rather coarse one-hour time step, even though practical controllers normally operate on a much finer time scale to provide adequate bandwidth; the Markov model may, however, operate on an entirely different time scale from the MPC with only minor implementation changes. Second, our hypothetical system has constant efficiency and operates only in heat mode to simplify the cost function and maintain focus on the paper’s contribution. As long as energy consumption can be controlled and room temperature can be measured, the stochastic occupancy model may be applied to arbitrarily complex [MPC]{} scenarios. Finally, we have used a certainty-equivalence assumption for weather and occupancy predictions; recent research has addressed ways to incorporate uncertainty into the optimization for added robustness. Demonstrating our algorithm without these simplifications is left to future work.
Acknowledgements
================
The authors thank Peter Radecki for his constructive feedback.
References
==========
[^1]: See [@goyal2013occupancy] for a comparison of MPC and heuristic control for a more complex HVAC system.
[^2]: The observability assumption is valid because of the RC network’s construction; the driving sources (exterior and interior conditions) are themselves measurable and there are no hinges in the network. See [@dobbsautomatic] for details on the RC network construction and [@luluControllabilityAndObservability; @dobbsautomatic] Theorem 1 for a proof of observability.
[^3]: The model can include multiple control zones if needed. In practice, one may reduce the model size using balanced truncation, which reconfigures the state space. We have chosen to retain the full-order model to maximize accuracy and preserve physical intuition. See [@dobbscdc2012] for a survey of methods that reduce state space size while preserving structure.
[^4]: Daily ground temperature is available for free through on-line sources such as the U.S. Surface Climate Observing Reference Network [@noaaClimateData].
[^5]: Although we have used the MERL occupancy data, the thermal model is not one of the MERL building.
[^6]: The function $f(\theta)$ is a continuous beta distribution. Once linear forgetting is added, these distributions become prohibitive to maintain analytically because sums of beta distributions are not themselves beta distributions, but rather are complicated piecewise functions [@gupta2004handbook]. Therefore it is more practical to maintain numerical approximations.
[^7]: There are other ways to implement forgetting; see [@generalForgetting] for a survey that compares linear with multiplicative forgetting.
[^8]: Because the forecast is updated at each time step, our implementation adjusts the augmented transition matrix $\tilde{A}$ before each MPC synthesis to reflect the latest prediction. This simplifies the cost function and allows the MPC to be formulated in a compact vectorized form as detailed in Equation 3.8 of [@rossiter2003model].
[^9]: In this simplified formulation, only the ratio between $r$ and $\beta$ matters; together, they constitute a single tuning adjustment. The energy cost gain $r$ can be time-varying if one wishes, for example, to incorporate time-of-day utility pricing.
[^10]: Multi-parametric methods can be used to partition the state space into regions, each with an exact control law parameterized on the entire state at the expense of a more complex MPC formulation [@multiParametricProgramming].
| {
"pile_set_name": "ArXiv"
} |
---
address: |
Department of Computer Science and Software Engineering,\
Swinburne University of Technology,\
Melbourne, VIC-3122, Australia.
author:
- 'Iqbal H. Sarker'
bibliography:
- 'bibfile/PhDThesisBib.bib'
- 'bibfile/mobile-data-science.bib'
- 'bibfile/MyPapers.bib'
title: 'Mobile Data Science: Towards Understanding Data-Driven Intelligent Mobile Applications'
---
Introduction
============
The advancement of mobile computing and the Internet have played a significant role in the development of the current digital age. The Internet has now formed the backbone of modern communication. Nowadays, the use of the Internet, particularly the World Wide Web (WWW) has moved beyond desktop computers to millions of mobile phones. According to ITU (International Telecommunication Union), cellular network coverage has reached 96.8% of the world population, and this number even reaches 100% of the population in the developed countries [@number-of-mobile-phone-users]. With the development of science and technology, the smart phone industry is growing rapidly and the popularity of smart phones has made exponential growth in mobile phone application market [@peng2018personalized]. According to Google Trends [@GoogleTrends2017], users’ interest on *“Mobile Phones”* is more and more than other platforms like *“Desktop Computer”* or *“Tablet Computer”* over time. For instance, Figure \[fig:interest-trends\] shows users’ interests trends over time on mobile phones over desktop and tablet computer for the last five years (December 2012 to December 2017). The data was collected on 13th December 2017 from Google Trends [@GoogleTrends2017].
{width=".8\linewidth"}
In Figure \[fig:interest-trends\], the numbers from 0 to 100 in Y-axis, represent search interests in terms of popularity relative to the highest point on the chart for a given date (e.g., 16/12/2012). For instance, a value of 100 represents the peak popularity for a particular term. A value of 50 means that the term is half as popular. Likewise, a score of 0 means the term was less than 1% as popular as the peak [@GoogleTrends2017]. Based on the popularity of mobile phones over time shown in Figure \[fig:interest-trends\], we are motivated to think about *mobile data science* and corresponding data-driven intelligent mobile applications to assist the end mobile phone users in their daily activities.
Nowadays, mobile phones have become one of the primary ways, in which people around the globe communicate with each other. These devices have transformed over a period of time from merely communication tools to smart and highly personal devices. Such devices are able to assist us in a variety of day-to-day situations in our daily activities. People in the world use mobile phones for various activities, such as for voice communication, Internet browsing, using mobile applications (apps), e-mailing, using online social network, instant messaging etc [@pejovic2014interruptme]. While mobile phones may come in various forms, in this paper they refer to smart mobile phones or Internet accessible mobile phones which have incorporated many advanced features to facilitate better information access and utilization for the benefit of end mobile phone users in their daily life. These smartphones in addition to being used as a communication device are capable of doing things that a computer does because of their computing capability on the devices. For instance, what actually makes the mobile phone “smart”? The answer is it’s ability to handle data, not only the voice calls. In general, these devices can fulfill nearly all communication requirements of the users in recent times. Smart mobile phones are not only used as communication devices but also used in various sectors [@lane2010survey] like health care services, navigation services, business purposes, safety, environmental monitoring, intelligent transport systems, traffic management, intelligent route routing, smart homes, or smart cities etc.
In recent times, the smartphones are becoming more and more powerful in both computing and data storage aspects. Their more and more sensing capabilities have enabled the collection of rich raw contextual information of the user, surrounding real-life environment, and the mobile phones usage records through the device logs, such as phone call logs [@phithakkitnukoon2011behavior] [@sarker2016phone], SMS Logs [@eagle2006reality], mobile application (Apps) usages logs [@zhu2014mining] [@srinivasan2014mobileminer], mobile phone notification logs [@mehrotra2016prefminer], web logs [@halvey2005time], game Logs [@paireekreng2009time], context logs [@cao2010effective] [@zhu2014mining], and smartphone life logs [@rawassizadeh2013ubiqlog] etc. The logged data generated by the smart phones provides a means to get new knowledge about various aspects of the users, like user social interactions, various user activities with the devices, which offers the potential to understand the insight of such mobile phone data.
Typically, the traditional data science helps to solve problems by analyzing data and turns data into data products [@cao2017data] [@han2011data]. However, the concept of mobile data science is not exactly the same thing. The purpose of mobile data science is slightly different, as the smart mobile phones are aware of their user’s real-life surrounding environment, and users’ various types of activities or social interactions, in various contexts in the real world. Thus, *mobile data science* is a term which is concerned with the challenge of collecting user’s real life data in different contexts, such as temporal, spatial, social or others, in a pervasive computing environment, finding data-driven models in order to make dynamic decisions for individual mobile phone users based on relevant contextual information, and use this discovered knowledge for building smart mobile applications to intelligently assist themselves in their daily life. Some examples of such real-life applications are context-aware smart mobile communication, intelligent mobile notification management systems, context-aware smart searching, context-aware mobile recommender system, context-aware app management system, and various context-aware predictive services etc. Thus, the concept of *“mobile data science”* involving users’ real-life activities with mobile phones, corresponding contextual information of the users, and the surrounding dynamic environment, has become interested.
In particular, the contributions of this paper are:
- We explore the potentiality of mobile data science for building various data-driven intelligent mobile applications.
- We review various kinds of mobile phone data and define mobile data science that includes mobile phone data having contextual information and the processing to discover useful knowledge for data-driven applications.
- We discuss the fundamental concepts of mobile data science and highlight the key differences between the traditional data science and mobile data science in various aspects.
The rest of the paper is organized as follows. We briefly review the context-aware mobile computing, which is related to mobile data science, in Section \[Contexts and Context-Aware Mobile Computing: A Background\]. In Section \[Mobile Data Science\], we briefly discuss about mobile data science including the fundamental concepts, and the key modules of mobile data science. Finally, Section \[Conclusion\] concludes this paper.
Context-Aware Mobile Computing: A Background {#Contexts and Context-Aware Mobile Computing: A Background}
============================================
The term *context* can be used with a variety of different meanings in different purposes. In this section, first we briefly discuss about various definitions of contexts in the area of mobile and pervasive computing, and then we discuss about *context-awareness* in mobile computing.
Definitions of Contexts
-----------------------
The notion of context has been used in numerous areas, including Pervasive and Ubiquitous Computing, Human Computer Interaction, Computer-Supported Collaborative Work, and Ambient Intelligence [@dourish2004we]. In the area of Ubiquitous and Pervasive Computing, a number of early works on context-aware computing or context-awareness referred context as the locational contexts, i.e., the location of people and objects [@schilit1994disseminating]. In recent works, context has been extended to include a broader collection of factors in addition to such locational context, such as physical and social aspects of an entity or object, as well as the activities of users [@dourish2004we]. Having examined the definitions and categories of context given by the pervasive and ubiquitous computing community, in this section, we review the definition of contexts within the scope of this paper.
Several studies have attempted to define and represent context from different perspectives. For instance, the user’s location information, the surrounding people and objects around the user, and the changes to those objects are considered as contexts by Schilit et al. [@schilit1994disseminating]. Brown et al. [@brown1997context] also define contexts as user’s locational information, temporal information, the surrounding people around the user, temperature, etc. Similarly, the user’s locational information, environmental information, temporal information, user’s identity, are also taken into account as contexts By Ryan et al. [@ryan1999enhanced].
Other definitions of context have simply provided synonyms for context such as context as the environment or social situation. A number of researchers are taken into account the context as the environmental information of the user. For instance, in [@brown1995stick], the environmental information that the user’s computer knows about are taken into account as context by Brown et al, whereas the social situation of the user is considered as a context in Franklin et al. [@franklin1998all]. On the other hand, a number of other researchers consider it to be the environment related to the applications. For instance, Ward et al. [@ward1997new] consider the state of the surrounding information of the applications as contexts. Hull et al. [@hull1997towards] define context as the aspects of the current situation of the user and include the entire environment. The settings of applications are also treated as context in Rodden et al. [@rodden1998exploiting]. In [@schilit1994context], Schilit et al. claim that the important aspects of context are: (i) where you are, (ii) whom you are with, and (iii) what resources are nearby. The information of the changing environment is taken into account as context in their definition. In addition to the user environment (e.g., user location, nearby people around the user, and the current social situation of the user), they also include the computing environment and the physical environment. For instance, connectivity, available processors, user input and display, network capacity, and costs of computing can be the examples of the computing environment, while the noise level, temperature, the lighting level, can be the examples of the physical environment.
Dey et al. [@dey2001understanding] presents a survey of alternative view of context, which are largely imprecise and indirect, typically defining context by synonym or example. Finally, Dey et al. [@dey2001understanding] offer a definition of context, which is perhaps now the most widely accepted. According to Dey et al. [@dey2001understanding] *“Context is any information that can be used to characterize the situation of an entity. An entity is person, place or object that is considered relevant to the interaction between a user and an application, including the user and the application themselves”*.
**Context Category** **Context Examples**
----------------------- ----------------------
User Identity
Temporal
Spatial
Environmental
Social
Physiological
Psychology
Mobile phone activity
Physical activity
Device-related
App-related
Context-Awareness
-----------------
Typically, a desktop computer application expects a static execution environment, either in office, home or other locations. However, this precondition is generally not applicable for mobile services or systems, as the world around an application is changing frequently. Thus, the nature of mobile applications must change according to the movement of the users. In particular, mobile applications should become more flexible in order to respond in highly dynamic computing environments, which makes the computing more pervasive. Recent advances in smart mobile phones and the pervasive computing environment make it possible in the real-world, where *context-awareness* can be used as the spirit of pervasive computing [@shi2006context]. The primary objective of pervasive computing environment is the creation of an environment saturated with seamlessly integrated devices with computing and communication capabilities having decision-making ability with as little direct user interaction as possible [@anagnostopoulos2005context]. The use of contextual information in mobile applications is able to reduce the amount of human effort and attention that is needed for an application to provide the services according to user’s needs or preferences, in a pervasive computing environment. Thus, context-awareness simply represents the dynamic nature of the applications.
Nowadays, the popularity of context-awareness in mobile computing is increasing because of their adaptation in dynamic environment. Typically, context-awareness is originated as a term from ubiquitous or pervasive computing, which is able to deal with linking changes in the real-world environment with mobile systems. According to Wikipedia [@wikiContextAwareness], context-awareness enables a new class of applications in pervasive computing, which is a property of mobile devices that is defined complementarily to location awareness, i.e., adapting capability in the applications with the movement of mobile phone users. However, as discussed above, various types of contexts can be applied to make the mobile applications more flexible and useful according to users’ overall situations in terms of temporal, spatial or social, and their preferences, as context-awareness is the main spirit of pervasive computing to build adaptive applications. In general, context-aware computing refers to a general class of mobile systems that can sense their surrounding physical environment, and able to adapt their behavior intelligently according to the sensed information in the corresponding applications [@wikiContextAware-computing].
Context-aware systems are a component of a ubiquitous computing or pervasive computing environment [@wikiContextAware-computing]. According to [@wikiContextAware-computing], there are three important aspects of context; these are: where you are; who you are with; and what resources are nearby. In order to make the mobile applications capable of operating in highly dynamic environments demanding on less user attention, different types of contexts might have different impact, shown in Sarker et al. [@sarker2017anapproach]. Hence, we summarize a number of contexts, related to mobile phone users, their activities, and surrounding environment, shown in Table \[tab:context-examples\].
Understanding Mobile Data Science {#Mobile Data Science}
=================================
We are living in the age of data science [@cao2017data]. On the other hand, computing is increasing being carried out using smart mobile phones, which support not only the telephony but also data-driven mobile applications. In this section, we discuss briefly about the mobile data science, ranging from smartphone raw data having contextual information, building data-driven models using machine learning techniques, and corresponding smart mobile applications for the end users.
Motivational Examples
---------------------
The following real-life examples intuitively illustrate the advantage of *mobile data science* in various mobile applications either on single device or in a distributed environment, to assist the users in their daily life. In general, a single device based application is related to the computation on individuals’ devices utilizing their phone log data. On the other hand, application in the distributed environment is designed to allow users of a computer network to access information, and services, as well as to exchange information with others, like a client-server based model.
EXAMPLE 1 (Application Scenario on Single Device). Say, Alice, a smartphone user, is a PhD student. She has installed a large number of mobile applications (apps) on her smartphone. Homescreens of smartphones provide ready access to commonly used apps, which is particularly useful. However, the homescreen of her smartphone is unaware about Alice’s changing contexts, such as her location, and consequently unable to manage apps intelligently according to her needs. Thus, an intelligent mobile app management system is needed for her, which allows the particular app she currently needs to be easily accessible from the mobile homescreen. To build such intelligent applications *mobile data science* can play an important role by predicting her future usages utilizing her app usages log data, that intelligently assists her to use different types of mobile apps, such as Skype, Whatsapp, Facebook, Gmail, Youtube, Linkedin, Microsoft Outlook, etc. according to her current contextual information.
EXAMPLE 2 (Application Scenario in Distributed Environment). Let’s consider a cloud-based mobile service recommendation system for the above user Alice. When she enters her University campus by driving her car, she will automatically get a recommendation for the ‘best’ car parking space to minimize the searching time and cost. In addition, the user will receive on her mobile device detailed driving instructions to reach that space. A distributed cloud-based system is able to accomplish this job automatically without any additional effort of the user. She can get this service using the corresponding client application installed on her mobile phone based on her contextual information and preferences. The cloud-based system is typically a client-server model where in the mobile device based agent (client) extracts the real-time contextual information of the user and communicates with a cloud-based service that houses the corresponding recommender system (server). Unlike the single device based services, where computation is done within the device, the battery life of the device, cpu computational limitations, and user preference modeling limitations, can be ignored in a cloud based services as the server is responsible for all the computations needed. To build such cloud based services *mobile data science* can play an important role to predict the parking availability according to her current contextual information by analyzing the relevant data in the cloud.
The above scenarios show that *mobile data science* is potential to build various data-driven intelligent mobile applications, to intelligently assist the users in various purposes by analyzing the relevant data, not only on a single device but also in a distributed environment.
Understanding Real-Life Smartphone Data
---------------------------------------
Real-life smartphone data having contextual information relevant to individual mobile phone users’ activity is one of the key elements of mobile data science. The reason is that mobile phone data with contextual information is the basis of mobile data science to build data-driven intelligent applications for the users. According to [@cao2017data], we live in the age of data, where everything that surrounds us is linked to a data source and everything in our lives is captured digitally. On the other hand, mobile or cellular phones have become increasingly ubiquitous and powerful. Recent advances in smart mobile phones and their sensing capabilities have enabled the collection of users various activities and corresponding contextual information. Thus, contextual data is available around us to analyze. In this section, we discuss about a variety of mobile phone data containing the associated contextual information of individual users. We have summarized these data in Table \[tab:mobile-phone-data\].
**Smartphone Data** **** **References**
--------------------- ------ ----------------
Call Log
SMS Log
App Usages Log
Notification Log
Web Log
Game Log
Smartphone Life Log
**Key Terms** **Description**
---------------------- -----------------
Context-awareness
Data science
Mobile user activity
Adaptability
**Aspects** **Data Science** **Mobile Data Science**
------------- ------------------- -------------------------
Rationale
Objective
Data
Mechanism Automated
Evaluation Technical metrics
These context-rich historical mobile phone data shown in Table \[tab:mobile-phone-data\], is simply as the collection of the past contexts and user’s actions for the past contexts [@hong2009context]. The main characteristic of such kind of log data is that they contain the actual behavior of the users in different contexts, as the mobile phones automatically record these data. Thus, it is important to study mobile phone data involving contextual information and users diverse activities in different contexts, for the purpose of building data-driven smart and context-aware intelligent mobile applications.
Fundamental Concepts of Mobile Data Science
-------------------------------------------
The traditional data science [@cao2017data] can be treated as a part of mobile data science. Nowadays, many researchers use of the term “data science” to describe the interdisciplinary field of data collection, preprocessing, inferring or making decisions by analyzing the data. According to Cao et al. [@cao2017data], a high-level statement about data science is: “it is the science of data or the study of data”. The outputs of data science are typically data products depending on what type of data are taken into account, which can be a discovery, prediction, service, recommendation, thinking, model, or system. According to the definition of Cao et al. [@cao2017data], the ultimate values of data products are knowledge, intelligence, wisdom, and decision based on relevant data.
On the other hand, computing is moving toward pervasive, ubiquitous environments [@finin2001information]. Thus, the computing devices, software agents, and services are all expected to seamlessly integrate and cooperate in support of human objectives according to their needs and preferences. Mobile devices are able to provide such service in a pervasive computing environment in an anywhere, any-time fashion. Therefore, the next step for pervasive computing is the integration of smart devices that are able to understand the local context of the users and share this information in support of human needs intelligently.
Based on the potentiality of data science and the pervasiveness in computing using mobile devices discussed above, in this paper, we introduce the concept of *mobile data science*, by taking advantages from both, in order to help the researchers and application developers, for the purpose of building data-driven mobile applications utilizing relevant mobile phone data. Our objective is to highlight the importance of working in this area, in order to identify examples of key research issues for the community, thereby acting as a catalyst for new research and development for the benefit of end mobile phone users. We define the field as follows:
*“Mobile data science is research or working area that exists at the intersection of context-aware mobile computing and data science, characterized by a focus on the real-life contextual data collection, data pre-processing, analyzing data using data mining and machine learning techniques, and use of extracted knowledge from the mobile phone data, for the purpose of building smart data-driven mobile applications, in order to intelligently assist the end mobile phone users by providing dynamic decisions in their daily activities, in a context-aware pervasive computing environment.”*.
In Table \[tab:key-terms\], we define some key terms related to *mobile data science*. In a word, mobile data science is related to users’ real-life mobile phone data, mobile user activities, context-aware mobile computing, and the traditional data science, shown in Figure \[fig:overview\]. We also summarize various aspects of *mobile data science* over the traditional data science, highlighted in Table \[tab:mobile-data-science\].
Key Modules of Mobile Data Science
----------------------------------
In this section, we briefly discuss about the key *modules* that are involved in the process of mobile data science, for the purpose of building data-driven intelligent mobile applications.
In the process of mobile data science, first the contextual data is collected from the real world using the devices. After that, a data-driven model based on machine learning techniques according to the target problem, can be built by analyzing and processing the contextual data. Such data-driven models can be used to build various intelligent mobile applications to intelligently assist the end mobile phone users in their daily activities. Based on this process, Figure \[fig:overview-process\] shows a high level overview of mobile data science involving a number of modules, from real-world raw data to applications. In the following, we discuss these modules one by one with examples.
- *Mobile Data Collection*: This is the first module of mobile data science, as mobile phone data is the basis of mobile data science. Thus, this module mainly focuses on collecting user’s real life smartphone data relevant to a particular problem for analysis and making corresponding decisions. Real-world mobile phone data usually comprise features whose interpretation depends on some *contextual information* such as temporal (e.g., in the morning), spatial (e.g., at office), social contexts (e.g., in a meeting) or others. These contextual-sensitive features and their available patterns in the dataset are of high interest to be discovered and analyzed in order to make dynamic decisions intelligently in a pervasive computing environment. The contextual data can be collected from various sources like smartphone sensors, e.g., GPS, Wi-Fi, Blue-tooth, accelerometer, proximity, camera, microphone, light sensors, etc. or using corresponding phone log data, electronic calendars, according to the needs for a particular application. For instance, Sarker et al. [@sarker2016behavior] collect mobile phone call log data including time-series contextual information for their temporal analysis in order to build time-dependent call interruption management system based on users’ unavailability [@sarker2018Unavailability].
- *Data Pre-processing*: This module mainly focuses on preparing the data for modeling. For instance, cleaning the data to address the data quality issues can be a pre-processing task. Real-world mobile phone data may contain noisy or inconsistence instances, missing values, duplicate records, invalid data or outliers. Such noisy instances may have impact on machine learning techniques to build an effective model for a particular problem. Therefore, handling these inconsistency to ensure the quality of data is an important issue for data analysis. Various machine learning techniques can be used to detect such outliers or noisy instances. For instance, in [@sarker2017improved], the authors propose a naive Bayes classifier based noise detection technique to ensure the quality of smart phone data. In addition to the quality of data, to transform the raw data to the formatted or desired data to make it suitable for further analysis is another task in data pre-processing. For instance, the continuous time-series mobile phone data is converted into behavior-oriented segments to capture the behavioral patterns of individual mobile phone users in [@sarker2017individualized]. Like the temporal context, the data-centric social context [@sarker2018DataCentricSocialContext] or others relevant raw data may need to transform into the desired data that can be used for further analysis according to the needs.
- *Contexts and Usages Analysis*: This module mainly focuses on selecting the interesting features or contexts that have an influence on users’ diverse activities to build an effective model. For instance, social contexts, e.g., in a meeting, or social relationship between individuals, e.g., boss, might have an influence to make a call handling decision of an individual mobile phone users [@sarker2018BehavMiner]. Similarly, other contexts such as temporal, or spatial contexts, can play a role in different applications depending on users’ need and preferences [@sarker2017designing]. For instance, building a context-aware model for predicting parking availability in smart cities, temporal and spatial contexts can be used [@zheng2015parking]. Such relevancy of contexts may differ from application-to-application in the real world.
- *Building Data-driven Models*: Once the required contextual information and mobile phone usages are prepared according to a particular problem, this module is responsible to build an intelligent data-driven model to solve the target problem. Thus, this module can be treated as the core module in mobile data science, as intelligent decision making is depended on the corresponding data-driven model. For the purpose of building such models, various popular data mining and machine learning techniques [@han2011data], such as classification analysis (e.g., Decision tree, Naive Bayes, Support vector machine, K-nearest neighbors, Artificial neural network etc.), ensemble learning (e.g., Random forest, AdaBoost, Bootstrap aggregating etc.), clustering (e.g., K-means, Agglomerative, Divisive etc.), regression analysis (e.g., Linear regression, Logistic regression etc.), and association analysis (e.g., AIS, Apriori, FP-growth etc.) or others relevant machine learning techniques [@han2011data] [@witten2005data], can be used for an effective modeling according to the specific needs for a particular problem. The extracted knowledge using such machine leaning techniques then can be used to build various real-life smart mobile applications to intelligently assist the end users. For instance, mining behavioral association rules of individual mobile phone users based on machine learning technique, can be used to effectively predict their future behavior according to their current contextual information [@sarker2018mining].
The modules discussed above are the key in mobile data science for the purpose of building data-driven intelligent mobile applications, in order to assist the end mobile phone users in their daily activities in a context-aware pervasive computing environment.
Conclusion {#Conclusion}
==========
In this paper, we have briefly discussed about the concept of mobile data science by highlighting the characteristics of mobile phone data including users’ various activities in different contexts. In particular, we focused our investigation on context-awareness and real-life mobile phone data, and their proper utilization for building intelligent data-driven models not only on the single devices but also in a distributed environment, to assist them in their daily activities in a pervasive computing environment. Additionally, we have summarized the key differences between the traditional data science and mobile data science in various aspects. We believe this article will help both the researchers and application developers for building various data-driven intelligent mobile applications for the benefit of end mobile phone users utilizing their mobile phone data.
Acknowledgment {#Acknowledgment .unnumbered}
==============
I would like to thank Prof. Jun Han, Swinburne University of Technology, Australia, Dr. Alan Colman, Swinburne University of Technology, Australia, Dr. Ashad Kabir, Charles Sturt University, Australia, for their relevant discussions.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Near-infrared photometry (JHKL$^\prime$M) was obtained for 78 semiregular variables (SRVs) in field \#3 of the Palomar-Groningen survey (PG3, $l\!=\!0\degr$, $b\!=\!-10\degr$). Together with a sample of Miras in this field a comparison is made with a sample of field SRVs and Miras. The PG3 SRVs form a sequence (period & period) with the PG3 Miras, in which the SRVs are the short period extension to the Miras. The field and PG3 Miras follow the same P/(J–K)$_0$ relation, while this is not the case for the field and PG3 SRVs. Both the PG3 SRVs and Miras follow the Sgr I period-luminosity relation adopted from Glass et al. ([@Glass95ea]). They are likely pulsating in the fundamental mode and have metallicities spanning the range from intermediate to approximately solar.'
author:
- 'M. Schultheis'
- 'Y.K. Ng[^1]'
- 'J. Hron'
- 'F. Kerschbaum'
date: 'Received 10 May 1996 / Accepted 22 July 1998'
subtitle: 'II. Near-infrared photometry of semiregular variables[^2]'
title: 'Field \#3 of the Palomar-Groningen survey[^3]'
---
psfig
Introduction {#Introduction}
============
In the last decade much effort has been spent on the studies of Miras (e.g. van der Veen [@vdVeen88], Feast et al. [@Feast89ea], Whitelock et al. [@Whitelock91ea], Blommaert [@Blommaert92], and Glass et al. [@Glass95ea]), OH/IR stars (e.g. Herman [@Herman88], te Lintel-Hekkert [@Lintel90], Lindquist et al. [@Lindquist92ea], van Langevelde [@vLangevelde92], Blommaert et al. [@Blommaert94ea], and Habing [@Habing96]), and carbon stars (e.g. Willems [@Willems87], ChanKwok [@ChanKwok88], Willems [@Willems88a]b, Willemsde Jong [@WillemsdeJong88], de Jong [@deJong89], Stephenson [@Stephenson89], ZuckermanMaddalena [@ZuckMad89], Azzopardi et al. [@Azzopardi91ea], TysonRich [@TR91], Westerlund et al. [@Westerlund91ea], Groenewegen et al. [@Groenewegen92ea], Groenewegen [@Groenewegen93], Marigo et al. [@Marigo96bea], Ng [@Ng97b] & [@Ng98], and Marigo [@Marigo98]).
Systematic investigations of semiregular variables (SRVs) were carried out by KerschbaumHron ([@KH92], [@KH94]; hereafter respectively referred to as KH92KH94) and JuraKleinmann ([@JuraKlein92]) in recent years. From the temperatures, luminosities, mass loss rates and number densities KH92KH94 distinguished a ‘blue’, ‘red’, and a ‘Mira-like’ group among their SRVs. Their result suggests an evolutionary sequence, where the ‘blue’ SRVs evolve towards the thermally pulsing AGB, change into ‘red’ SRVs, and then continue to evolve to the Mira phase.
AGB stars are ideal probes in studies of the galactic bulge, because of their high luminosities. Various photometric (FrogelWhitford [@FW87], Terndrup [@Terndrup88], GeislerFriel [@GeislerFriel92]), spectroscopic (Rich [@Rich90], McWilliamRich [@WR94], Sadler et al. [@Sadler96ea]) and star counts (Ng et al. [@Ng96aea]) studies from Baade’s Window (hereafter referred to as BW; $l\!=\!1\fdg0$, $b\!=\!-3\fdg9$), show a large spread in metallicity ranging from –1.5$<$$<$1.0. The bulge offers a good opportunity to study the evolutionary aspects of AGB-stars in relation with their metallicity, which is one of the major evolutionary parameters.
PG3 (field \#3 of the Palomar-Groningen Variable Star Survey; $l\!=\!0\degr$, $b\!=\!-10\degr$; see Sect. \[PG3\] for details) is well searched for variable stars (Plaut [@Plaut71], Wesselink [@Wesselink87]; the latter will hereafter be referred to as Wess87). Blommaert ([@Blommaert92]; hereafter referred to as Bl92) studied the properties of the PG3 Miras with nIR (near-infrared) and IRAS photometry. The period-colour relation for the PG3 Miras was found to be shifted significantly from the Large Magellanic Cloud (LMC) relation (Feast et al. [@Feast89ea]). With the theoretical relation from Wood et al. ([@Wood91ea]) it was concluded that this must be due to metallicity, which could be about 1.6 times larger for the PG3 Miras (Bl92). Feast ([@Feast96]) obtained similar results for the Miras in the galactic bulge versus those in the LMC. Comparable indications were found by Whitelock ([@Whitelock96]) from the local Miras.
In this study a comparison is made between the pulsation and near-infrared properties of the PG3 SRVs and Miras and the field SRVs and Miras. In Sect. \[Data\] a discussion is given about the SRV data: the observations and the comparison samples. We proceed in Sect. \[Results\] with a description of the results of this analysis, continue with the discussion in Sect. \[Discussion\] and summarize it in Sect. \[Summary\].
Data {#Data}
====
PG3 {#PG3}
---
Four fields were selected in the mid-fifties by Baade and Plaut to search for variable stars (Blaauw [@Blaauw55]; Larsson-Leander [@LarssonLeander59]). The results from the photographic survey, known as the Palomar-Groningen Variable Star survey, were published in a series of six papers (Plaut [@Plaut66], [@Plaut68a]b, [@Plaut70], [@Plaut71], [@Plaut73]). The centre of PG3 is located 10$^\circ$ south of the galactic centre and skims over the edge of the galactic bulge. Important aspects of this field are the large area covered ($6\fdg5\times6\fdg5$) and the low interstellar extinction, which is gradually increasing in the direction towards the galactic centre (Wess87).
With emphasis on the RR Lyrae stars, the variable stars in PG3 were re-examined by Wess87, using UKST B$_{\rm J}$ and R$_{\rm F}$ Schmidt plates. Figure \[fig1\] shows the Fourier spectral window for the epochs of the plates. The figure shows that a good resolution, as intended, is obtained towards the short period variables, but for the long period variables a deficiency of stars with periods between 320 to 500 days could be possible. The three highest alias peaks correspond in decreasing order with one year, one week, and one synodical month.
Run dates year observer(s)
----- -- ---------------- ------ ----------------- --
1 28 Jul – 2 Aug 1990 Ng
2 21 May – 3 Jun 1991 Brown
3 20 – 29 Aug 1991 Ng
4 16 – 21 Jun 1992 Ng & Schultheis
5 29 Jun – 3 Jul 1993 Ng & Schultheis
: Log of the observing runs
The large number of Miras and SRVs discovered in this field makes it very attractive, to subject these stars to a more detailed study. Bl92 studied a sample of Miras and compared them with the IRAS sources in PG3, while we focus on the SRVs.
The SRV sample
--------------
In the GCVS4 catalog (Kholopov et al. [@GCVS4]) the classification of SRVs is based on the shape and the amplitude of the light curve. Generally, the period of the variations ranges from 20 to 2000 days with an amplitude less than . Plaut ([@Plaut71]) distinguished in his classification SRa and SRb type variables, while Wess87 made no distinction. Wess87 based his criteria on the B$_{\rm J}$ and R$_{\rm F}$ amplitudes, smaller than 20, and periods ranging from days. The SRV classification of Wess87 is in most of the cases compatible with a SRV of type ‘a’ (hereafter referred to as SRaV) from Plaut ([@Plaut71]). In general, the lightcurves of the SRaVs resemble those from the Miras. The difference in the classification is merely a consequence of the imposed amplitude limits in the variation.
In this study the SRV stars are selected with the Wess87 classification. Higher priority is given to the observations of stars with a Q=0 quality flag, indicating that there is no doubt about the classification and the period. The SRVs are selected such, that there is no bias to the brightest SRVs and that the sample covers in 25 days intervals the full period range.
Near-infrared Photometry {#NIRphot}
------------------------
Near-infrared photometry (JHKL$^\prime$M) of 78 PG3 SRVs was obtained with the ESO 1-m telescope, La Silla (Chile), equipped with an InSb detector. The observations were carried out under photometric circumstances in the observing seasons 19901993 (ESO N$^o$ 49.5011 & 51.7056). Table 1 shows the log of the observing runs. The 19901991 observations (see Bl92) of the stars S40, S147, S728, S969, S1008, S1016, S1128 and S1204 were carried out as part of the ESO key programme ‘Stellar evolution in the galactic bulge’ (Blommaert et al. [@Blommaert90ea]; ESO N$^o$ 45K.5007). The observations were made in a standard way, through a diaphragm with a 15$\arcsec$ aperture with the chopping and beam-switching technique. An 8Hz sky chopping and a beam-switch throw of approximately 20$\arcsec$ in R.A. was applied.
The JHKL$^\prime$M fluxes are calibrated to the ESO photometric system (Bouchet et al. [@Bouchet91ea]). The typical errors are $\sim$002 in JHK, $\sim$01 in$\rm ^{\null}$ L$^\prime$, and $>$02 in M. For some stars no L$^\prime$ and M photometry was obtained, because they were fainter than the limiting magnitude attainable with the telescope. Table 2 gives the JHKL$^\prime$M magnitudes of the PG3 SRVs together with the period (Wess87).
In general the limiting magnitude was K$_{lim}$$\simeq$10$^m$ for most of our observing nights, which is due to our relatively short integration time and integration sequence. Photometry for some of the fainter stars were obtained in nights when the photometric conditions allowed it. In this respect the DENIS survey (Epchtein et al. [@Epchtein94ea], [@Epchtein97ea]), with , should be able to improve on the photometry and reach significantly deeper limits.
Crowding {#Crowding}
--------
The presence of additional stars in the beam cannot be avoided, because we are dealing with crowded field observations. Two cases should be considered: additional stars in the primary or in the background beam. An additional flux contribution in the primary beam would lead to an increase of the flux from the star, while stars in the background beams would give a background subtraction which is too high. As a consequence the flux of the target star will be underestimated. In first approximation, the number of faint stars in both the primary beam and the background beams are similar.
The stars in the background are in general much fainter and not as red as the stars observed. The induced errors from stars present in the background beams are expected to be less than the errors quoted above. A new position for the background subtraction would have been selected, if a bright star was noticed in one of the background beams, but this was not necessary. The flux of the target star will be overestimated with additional stars in the primary beam. Due to stars surrounding S283 (see finding chart in NgSchultheis [@NS97]), the magnitudes are slightly too bright and the colours too blue for this star.
Interstellar extinction {#extinction}
-----------------------
The procedure described by Bl92 is used, to correct for the interstellar extinction. It is based on the PG3 extinction map, constructed by Wess87 from the colour excess of the RR Lyrae stars at minimum light. In this map the extinction is highest in the plate corner at lowest galactic latitude (A()=114; $b\!=\!-6\degr$) and lowest at the opposite corner (A()=014; $b\!=\!-14\degr$). The extinction is described with a linear relation, because of the smoothness of the gradient in the extinction map. In first approximation we have A()=$0.133\,b + 2.02$, where $b$ is the galactic latitude. For a normal extinction law the corrections for the different infrared passbands can be derived with the standard curve no. 15 of van de Hulst ([@vdHulst49]). All the JHK photometry for the PG3 stars discussed in this study are de-reddened with the procedure outlined above. The photometric data in Table 2 are not de-reddened.
The comparison sample {#comparison}
---------------------
A comparison is made with other near-infrared photometric studies of SRVs and Miras, in order to obtain a better understanding of the evolutionary status of the PG3 SRVs. We use a sample of PG3 Miras (Bl92), well observed O-rich field Miras (Catchpole et al. [@Catchpole79ea]) and a magnitude limited sample of field O-rich SRVs (KH94, Kerschbaum [@Kerschbaum95]). Note, that the field samples mentioned above might not be truly representative in their relative numbers for the local neighbourhood. The field stars were de-reddened with a procedure similar to Feast et al. ([@Feast82ea]). The reddening corrections are small, because the visual absorptions ranges typically from 005020. We further used the Sgr I Miras from Glass et al. ([@Glass95ea]) and the LMC LPVs (long period variables) from Reid et al. ([@Reid95ea]) as an extra indication. For the latter we defined MiraSRV variable groups, following HughesWood ([@HW90]) on basis of the I-band amplitudes. We use the Reid et al. data set instead of the HughesWood data set, because the former are in the same photometric system as the Sgr I Miras.
Photometric transformations {#transformation}
---------------------------
The photometry of the PG3 SRVsMiras and the field SRVs are in the ESO photometric system (Bouchet et al. [@Bouchet91ea], van der Bliek et al. [@vdBliek96ea]). The photometry for the field Miras was obtained in the SAAO photometric system as defined by Glass ([@Glass74]). This photometric system is not identical to the SAAO system in which the photometry for the Sgr I Miras, the LMC LPVs, and the period-luminosity & period-colour relations were obtained (Glass et al. [@Glass95ea]). All the photometry discussed in this paper are in the ESO photometric system (Bouchet et al. [@Bouchet91ea]) or transformed to it from the various SAAO systems. New transformations from Hron et al. ([@Hron98ea]) are used, because existing transformations either refer to the old ESO system (BessellBrett 1988, Carter [@Carter90], Engels et al. [@Engels81ea], Wamsteker [@Wamsteker81]) or do not cover all SAAO systems (BessellBrett [@Bessell88], Bouchet et al. [@Bouchet91ea], Carter [@Carter90], van der Bliek et al. [@vdBliek96ea]). Furthermore, these transformations do not include stars with . Extrapolation of these transformations to the colours of typical AGB stars leads to errors of the same order as e.g. the differences due to metallicity (Hron et al. [@Hron98ea]). The estimated uncertainties in the new transformations are typically in the colours.
Results {#Results}
=======
The observations are presented in Figs. \[fig2\]\[fig8\]. The mean magnitude is used in the figures, when several observations are available for a star. In Figs. \[fig4\]\[fig8\] only those stars are considered which, according to Wess87, are classified correctly and have a good period determined.
Magnitude Distribution {#magdistribution}
----------------------
At present we are mainly interested in the bulge stars. Therefore, disc stars are referred to as the foreground contamination. If the stars were located in the disc, the magnitude distribution of the foreground contamination in PG3 will show a gradually, more or less linear, increase towards fainter magnitude (see for example Fig. 114, Ng et al. [@Ng95ea]). The roughly linear increase is a consequence of the increasing volume in the cone, when sampled towards larger distances. The distribution of stars in the bulge have in first approximation a peaked shape (see for example Fig. 113, Ng et al. [@Ng95ea]). This is a result of the density profile of the bulge stars, which increases towards the galactic centre, is highest near the galactic centre, and decreases afterwards.
Two interpretations are possible for the K-magnitude distributions displayed in Fig. \[fig2\].
1.There is a difference in the distribution of the foreground contamination (K$<$60), between the SRVs and Miras. The foreground contamination of the Miras appears to be significantly larger than for the SRVs. The large fraction of foreground Miras was already noticed by Bl92, especially those which also have an IRAS 12 and . Moreover, Fig. \[fig2\] shows that the peak of the SRVs (65$<$$<$90) is about 1.5 times broader than the Mira peak between 65$<$$<$80. This could indicate that part of the SRVs with 65$<$$<$80 are in fact Mira type variables. They were classified as SRVs, because of the amplitude of the variations in the lightcurves. In this case some of the SRVs with K$>$80 represent a group of intrinsically fainter stars.
2\. the PG3 SRVs represent a group of stars with a distribution similar, but intrinsically fainter, to the Miras. A 10 shift towards brighter magnitudes is an empirical patch for the luminosity difference between SRVs and Miras. With this shift the foreground contamination for the SRVs is slightly smaller, but probably the difference is not significant anymore. In this case the PG3 SRVs are also contaminated with foreground stars.
In Sect. \[foreground\] we argue that the latter possibility is preferred, but that a fraction of the foreground contamination might be associated with the galactic bar.
Periods and Amplitudes {#PeriodAmplitudes}
----------------------
Figure \[fig3\] gives the period distributions of the various samples, used for comparison with the PG3 SRVs. The PG3 SRVs have periods comparable with the short period tail of the PG3 Miras. Their periods further overlap with the ‘red’ field SRVs distinguished by KH92KH94, but the PG3 SRVs have however a longer mean period. There is no significant difference between the period distribution of the field and PG3 Miras, except for a deficiency of PG3 variables with periods larger than 300 days and is due to the spectral window for which the plates (Wess87) were taken, see Sect. \[PG3\].
The mean photo-visual amplitude, estimated from Plaut ([@Plaut71]), for the PG3 SRVs is about 11. This is comparable to the V-amplitude of the field SRVs in the same period range, but much smaller than the amplitudes for the field Miras (55).
Period – K Relation {#PKrelation}
-------------------
Figure \[fig4\] shows the apparent K magnitude versus logP diagram (hereafter referred to as PK$_0$-relation). The PG3 SRVs obey the same PK$_0$-relation as the PG3 Miras (Schultheis et al. [@Schultheis96ea]). This figure suggests a common origin for the two samples. Note that the straight line in Fig. \[fig4\] is not a fit to the data! It shows the PK$_0$-relation for Sgr I (Eq. 5, Glass et al. [@Glass95ea]), transformed from the SAAO to the ESO photometric system. We further adopted an extinction ( instead of ) for the Sgr I field. The resulting PK$_0$-relation for Sgr I in the ESO photometric system is: ${\rm K}_0\!=\!-3.47\log P+15.54$. The dashed lines above and below the Sgr I relation ($\pm\,1^m$) are a combination of the expected scatter due to the depth of the bulge and the dispersion in the magnitudes (not averaged, ) of the PG3 variables. The stars above the dotted line are foreground stars, but see Sect. \[foreground\] for additional comments. A few stars lie under the period-luminosity relation. NgSchultheis ([@NS97]) argue that those stars are located at the edge of the Sagittarius dwarf galaxy. In the further analysis only those stars, which are located between the two dashed lines, are considered.
Period - Colour Relation {#PCrelation}
------------------------
In Figs. \[fig5\]a the period-colour (PC) relations for the PG3 SRVs and Miras are shown for (J–K)$_0$, (J–H)$_0$, and , respectively. The thick straight lines indicate the LMC relation due to Feast et al. ([@Feast89ea]) and Glass et al. ([@Glass95ea]). In Fig. \[fig5\]a all the stars are slightly offset above the relation. In Fig. \[fig5\]b the Miras are below the P/(J–H)$_0$ relation, while the SRVs are located slightly above. The (J–H)$_0$ colour for both the SRVs and Miras appears to be independent of the period. In Fig. \[fig5\]c the PG3 SRVs appear to follow the LMC relation, while the PG3 Miras are offset above. It is also possible that a fraction of the SRVs follows the PG3 Mira P/(H–K)$_0$ relation, which is steeper than the LMC relation. An other fraction of SRVs lies clearly above such a relation.
For the PG3 Miras the mean offset from the LMC PC-relation is . Within the transformation uncertainty of the LMC relation this is comparable to the offset obtained by Bl92. Figure \[fig6\] shows the P/(J–K)$_0$ relation for the field SRVs and the field Miras. The field Miras also follow the LMC relation, although there is a slight offset of towards redder (J–K)$_0$. This offset is not conclusive with regard to possible differences to the LMC or PG3 stars, given the transformation uncertainties (field Mira and the LMC relation). The 005 offset of the PG3 Miras translates with the theoretical period-colour relation from Wood et al. ([@Wood91ea]; assuming comparable masses between the PG3 & the field versus the LMC Miras) in a mean metallicity of the PG3 and field Miras $\sim$1.4 times as high as the LMC.
The majority of the field SRVs appear to follow a different PC-relation with a slope flatter than the field Miras. But this might be an artifact, if the field SRVs are a non-homogeneous sample of fundamental mode pulsators with longer periods and overtone pulsators with shorter periods. Since each mode has its own PC-relation, their combined distribution could well result in the flatter slope.
Colour - Magnitude Diagram {#CMD}
--------------------------
Figure \[fig7\] shows the (K,J–K)$_0$ CMD for the PG3 SRVs and Miras. Isochrones placed at 8 kpc distance for 5 and 10 Gyr old stellar populations with and are displayed in this figure. The isochrones from Bertelli et al. ([@Bertelli94ea]) are used. They converted their isochrones from the theoretical to the observational plane by convolving the near-infrared bands, as provided by BessellBrett ([@Bessell88]), with the spectral energy distributions from ([@Kurucz92]) for temperatures higher then 4000 K. At lower temperatures they used observed spectra as described in Sect. 4 of Bertelli et al. ([@Bertelli94ea]) and they combined the effective temperature scale from Ridgway et al. ([@Ridgway80ea]) for the late M giants with the LançonRocca-Volmerange ([@LRV92]) scale for the early M giants. The lack of very red standards limits the near-infrared colour transformations (BressanNasi [@BressanNasi95]) and causes the colours of the isochrones to ‘saturate’ around . We derived a new, empirical $T_{\it eff}$- colour relation by making a conservative fit through the $T_{\it eff}$ and data available for cool giants (see Ng et al. [@Ng98ea] for details). This relation was adopted to compute the near infrared colours of the isochrones shown in Fig. \[fig7\].
The SRVs and Miras follow the trend indicated by the isochrones. SRVs and Miras with similar age and metallicity, distributed around isochrones with comparable age and metallicity, belong to the same population. Note that variability moves the stars in an almost diagonal direction in the CMD. The upper limits for the variation of the J–K colour around the light cycle is about for a Mira and for a SRV (HronKerschbaum [@HK94]). For the SRVs the amplitudes are too small to explain the scatter, while for the Miras the scatter might be for a large fraction due to their variability. The uncertainties in the interstellar reddening is according to Wess87 in the worst case 017 in B$_{\rm J}$, which translates in in K and in . Interstellar reddening therefore cannot explain the observed spread of in . This spread on the other hand might be related to the intrinsic width of the instability strip. The intrinsic spread of the LMC PC relation provides an upper limit. According to Kanbur et al. ([@Kanbur97ea]) this spread amounts for oxygen rich stars to 011 in . This is again smaller than the observed colour spread.
Independently the variability of the SRVs, the uncertainties in the interstellar reddening and the intrinsic width of the instability strip cannot account for the observed colour spread. In combination even an upper limit for the colour spread, which amounts to 015, is not sufficient to explain the observed scatter.
The isochrones show that the effect of a 5 Gyr age difference results merely in a shift of in the (J–K)$_0$ colour. On the other hand, metallicity differences result in larger shifts in the (J–K)$_0$ colour, e.g. in Fig. \[fig7\]. One might argue that the depth of the bulge would invalidate an analysis with isochrones, but one can verify easily that a similar result is obtained by removing the differential distance effects. For example, through an absolute calibration of the magnitude from the periods of the variables with the PK$_0$-relation at the distance of the galactic centre, i.e. 8 kpc (Wesselink 1987, Reid 1993). The reason for the similarity of the result is that it depends strongly on the colour range covered and less on the magnitude (apparent or absolute) of the stars.
The presence of a large spread in the metallicity could explain the distribution of the stars in the CMD. Due to the large spread in metallicity it is not possible to get a reliable age estimate. The red edge is due to stars around solar metallicity, while the stars at the blue edge have .
Colour – Colour Diagram {#CCD}
-----------------------
The colour-colour diagram in Fig. \[fig8\] demonstrates the difference between the Miras and SRVs in PG3. For comparison the different locations of the Sgr I Miras (Glass et al. [@Glass95ea]; note that we adopted an extinction in agreement with , see also Sect. \[PKrelation\]), the LMC LPVs (Reid et al. [@Reid95ea]) and the field Miras and SRVs (Feast et al. [@Feast89ea], KH92&KH94) are indicated. However, the shape of the LMC box is not well defined due to the small number of stars used present in the region . The PG3 SRVs are shifted with respect to the PG3 Miras to bluer and slightly redder (J–H)$_0$.
From the large similarity in period and amplitude between field ‘red’ SRVs and PG3 SRVs one might expect that the PG3 SRVs will be located in the region of the ‘red’ field SRVs. Although there is some overlap, the PG3 SRVs appear to be on average bluer in both colours than the ‘red’ field SRVs. The PG3 SRVs extend to redder colours than the LMC SRVs
The PG3 Miras are more similar to the Sgr I Miras than to the comparison samples of field and LMC Miras. The PG3 stars do not extend to colours as red as the Sgr I Miras. This could be related with deficiency of PG3 Miras with periods longer than 320 days. Within the uncertainties in the adopted colour transformations PG3 and Sgr I are comparable to each other. For and the LMC is compared to the field and PG3 abundant with relatively blue LPVs/Miras. In addition, the LMC and the field have in contrast to PG3 and Sgr I for and in this region a significant number of LPVs/Miras. In the following section we will argue that this is due to a combination of age and metallicity of the stars.
Discussion {#Discussion}
==========
Miras: PG3 versus field, Sgr I and LMC {#Miras}
--------------------------------------
In Fig. \[fig8\] the PG3 Miras resemble more the Sgr I Miras than the field Miras and LMC LPVs. This is apparently in contradiction with Glass et al. ([@Glass95ea], Fig. 4c) who did not find any significant offset of the Sgr I Miras from the LMC P/(J–K)$_0$ relation. Note however, that the extinction correction is actually the origin of this discrepancy. Using $A_V$=171 the Sgr I Miras will be 003 redder in . The estimated offset is now and within the uncertainties comparable to PG3.
From Figs. \[fig5\]a and \[fig6\] it is not clear if there is a significant offset from the field Miras with respect to either the LMC or PG3, while there is an offset between PG3 and the LMC. Figure \[fig8\] however shows that the field and PG3 Miras are not comparable and that there are noticeable differences between all four groups of Miras: (i) in contrast to PG3 and Sgr I, the field and LMC Miras populate the region with , (ii) the PG3 Miras extend to redder colours than all the other groups, and (iii) only the LMC Miras reach .
The blue limit of the LMC Mira box is only defined by a few objects. The last point above is probably not a real difference but induced by statistical fluctuations. Note that unidentified, hot carbon stars in the Reid et al. ([@Reid95ea]) sample would be located in this area of the colour-colour diagram (see for example Fig. 5 from CostaFrogel [@CostaFrogel96]). However, the Reid et al. stars are LPV’s while the blue carbon stars are not known to be large amplitude variables and nothing is known about the variability of the blue carbon stars. This point can be clarified only by more observational data.
Statistical fluctuations can hardly be responsible for the other two differences between the four groups of Miras. The redder colours of the PG3 Miras could be due to metallicities higher than solar but there is no evidence for this from the colour magnitude diagram. The majority of the field stars with solar metallicity have ages between 1 Gyr and 8 Gyr (see Figs. 3&4 from Ng&Bertelli [@NB98]), while for PG3 our age estimates range from 5 Gyr to 10 Gyr (Sect. \[CMD\], Ng et al. [@Ng95ea]). Thus the redder colour could be due to an older age of the PG3 stars with .
The lack of PG3 stars with and cannot be accounted for by metallicity effects alone, because this region is populated in the field as well as in the LMC. Age differences might again be the reason. In the LMC the last major star formation occurred 68 Gyr ago in some regions, while in other it happened only 23 Gyr ago (Vallenari et al. [@Vallenari96aea]b). For field stars with the age ranges from 210 Gyr (see again Figs. 3&4 from Ng&Bertelli [@NB98]). Thus the lack of PG3 stars could be due to a lack of stars with a metallicity in the range and an age between .
Although age is an attractive parameter to explain the differences between the various groups of Miras we want to emphasize that confirmation through a comparison with isochrones in the colour-colour diagram is still needed. This requires a proper calibration of the colours for the isochrones, possibly combined with an improved description of the AGB-phase (Ng et al. [@Ng98ea]).
All together the data are compatible with a metallicity range spanning from a quarter solar to approximately solar for the field and PG3 stars. The majority of the field stars with metallicities around solar may be considerably younger than their counterparts in PG3. This would explain the smaller colour offset of the field Miras from the LMC relation in comparison to the colour offset for the PG3 Miras. In this respect the apparent bluer colour in Fig. \[fig8\] of Sgr I versus PG3 might be an indication for a slightly younger age for the Miras in the Sgr I area.
SRVs: PG3 versus field and LMC {#SRVs}
------------------------------
The PG3 SRVs are similar to ‘red’ field SRVs in their periods and amplitudes. There are however marked differences: the slope of the PK$_0$ and P/ relations for the PG3 SRVs is similar to that of the PG3 and LMC Miras, while this does not appear to be the case for the field SRVs (see KH92 and Fig. \[fig4\]). In addition, the colours of the PG3 SRVs are slightly bluer than those of the ‘red’ field SRVs. The most plausible explanation for the colour differences is a higher temperature of the PG3 SRVs compared to the field SRVs. A temperature difference could explain the different behaviour in the P/ relations, as outlined in Sect. \[SRVvsMiras\].
The longer mean period of the PG3 stars is probably due to the larger homogeneity of this sample relative to the field stars. This homogeneity concerns both the variability classification and the pulsation mode (see below).
The result that the PG3 SRVs extend to redder colours than the LMC SRVs is due to their higher average metallicity.
SRVs versus Miras {#SRVvsMiras}
-----------------
Most field SRVs with pulsation periods below $\sim$200$^{\rm d}$ are cooler and partly brighter than field Miras with the same period (see Fig. \[fig6\] and KH92). This favours fundamental mode pulsation for the Miras and overtone pulsation for the SRVs. Similar evidence for variables in the LMC was presented by WoodSebo ([@WS96], hereafter WS96). Their results indicate that the fundamental mode pulsation is consistent with stellar masses $\la\!2 M_\odot$. This might be in contradiction with the results obtained by van Leeuwen et al. ([@vLeeuwen97ea]). Their analysis indicate that the majority of the Miras are first overtone pulsators. However, their sample did not include SRVs and is furthermore biased to stars with large radii. For stars in fundamental mode with shorter periods (like the PG3 SRVs) the data for the smaller radii are lacking, due to observational limitations. Therefore, the results from van Leeuwen et al. are not necessarily in disagreement with WS96.
An interpretation of the PG3 stars in terms of pulsation modes has to be in agreement with the behaviour of the SRVs and Miras in the PK$_0$ and diagrams. The PG3 SRVs would be brighter and redder than the Miras at a given period, if the PG3 Miras are fundamental mode pulsators and the PG3 SRVs are overtone pulsators. If Miras and SRVs pulsate in the same mode, a systematically lower metallicity of the SRVs would increase their temperature and make them bluer, but at the same time their periods would be smaller at a given luminosity. This would introduce a luminosity difference relative to the Miras at a given period but would only shift the SRVs along their P/ relation. Therefore, our result that the PG3 SRVs are an extension of the PG3 Mira PK$_0$ and P/(J–K)$_0$ relations, can only be explained by adopting the same metallicity range and pulsation mode for the Miras and SRVs. In view of the similarities between field and PG3 Miras, fundamental mode pulsation is more plausible for the PG3 stars. In addition, we conclude that the PG3 SRVs are [*[not]{}*]{} the analogs of the field SRVs (see Figs. \[fig5\]&\[fig6\]).
*[The metallicity of PG3 Miras and SRVs]{}* {#metallicityPG3}
-------------------------------------------
From star counts and metal-rich globular cluster studies (Ng [@Ng94]; Bertelli et al. [@Bertelli95ea], [@Bertelli96ea]; Minniti [@Minniti95]) one expects a gradient of metal-rich stars towards the galactic centre. BW is located closer to the galactic centre and has a larger number of high metallicity stars with respect to PG3 (Ng et al. [@Ng96aea], [@Ng97ea]). The period-colour relation indicates that the mean metallicity of the PG3 variables is about 1.4 times larger than the LMC mean metallicity. A comparison of both the PG3 and Sgr I (Glass et al. [@Glass95ea]) period-(J–K)$_0$ relation relative to the LMC relation shows that a small trend might be present. But the uncertainties in the extinction correction for Sgr I are larger than in PG3. An uncertainty in the extinction of in one filter would give an error in the colour, i.e. (J–K)$_0$, of a few hundredth of a magnitude. Together with the uncertainties in the transformation, the remaining difference between Sgr I and PG3 period-colour relation is not significant to conclude from the present data, that there is a metallicity gradient towards the galactic centre.
The PG3 SRVs and Miras extend to redder (H–K)$_0$ than the LMC LPVs. This indicates that the mean metallicity for the PG3 LPVs is slightly larger than for the LMC which is also consistent with Fig. \[fig7\]. The bulk of PG3 Miras cover the same position as the Sgr I Miras by Glass et al. ([@Glass95ea]). They should have a similar age and metallicity as the Sgr I Miras, which presumably have solar-type metallicity. This implies that the whole metallicity range from intermediate to solar is possible in PG3 and the LMC, but as mentioned in Sect. \[Miras\] there are perhaps differences in age between PG3 and LMC stars of comparable metallicity.
Hints from galactic structure {#galstruct}
-----------------------------
The field SRVs have a scale height of 230 pc (KH92). This is slightly smaller than the scale height of 260$\pm$30 pc for the PG3 Miras with a detection in IRAS (Bl92). These values are comparable with the 250 pc obtained by Habing ([@Habing88]) for AGB stars and they are consistent with the scale height for disc giants. The whole PG3 Mira sample, on the other hand, has a scale height of 300$\pm$50 pc. Although the previous values are within the uncertainties one has to consider, that differentially there is a noticeable difference between the Miras with and without an IRAS detection. With the ages and scale heights, determined for the stellar populations in the disc (Ng [@Ng94]; Ng et al. [@Ng95ea], [@Ng96aea], [@Ng97ea]), the age of the field SRVs and the PG3 Miras with an detection in IRAS is between with a metallicity ranging from . The age range for the whole PG3 Mira sample is estimated . These ages appears to be in agreement with an age considerably less than 10 Gyr, estimated by HarmonGilmore ([@HarmonGilmore88]) for the bulge IRAS sources with initial masses larger than 1.3 $\cal M_\odot$. But Whitelock et al. ([@Whitelock91ea]) showed, that their lower limit estimate for the initial mass is more likely an upper limit for the majority of the stars. This gives a lower age limit (Bertelli et al. [@Bertelli94ea]), which is consistent with the age deduced from the scale heights.
The very large metallicity spread causes that age estimates are very susceptible. Even ages as old as 16 Gyr estimated by Bl92 are possible. With such an old age the bluest SRVs and Miras should be very metal-poor. This would lead to a contradiction that long period variables cannot be present, because they are not found in old, metal-poor globular clusters. Therefore, the PG3 Miras and SRVs cannot be old and very metal-poor, but this does not rule out the intermediate and solar metallicity cases. The period for the PG3 Miras ranges from 180320 days, which is comparable with the periods of the Miras found in metal-rich globular clusters (FeastWhitelock [@FW87], Whitelock et al. [@Whitelock91ea]). The age of these clusters (Ng et al. [@Ng96cea]d) gives an initial mass of about 1 $\cal M_\odot$, which is consistent with the upper limit obtained from the scale height for these stars.
On the other hand, if the disc density towards the galactic centre is lower than expected from a double exponential density profile (Bertelli et al. [@Bertelli95ea]; Kiraga et al. [@Kiraga97ea]; Ng [@Ng94][@Ng97a]; Ng et al. [@Ng95ea], [@Ng96aea]; PaczyńskiUdalski [@PaczynskiUdalski97]; Paczyński et al. [@Paczynski94ea]), this could imply that the PG3 SRVs and Miras are not due to a disc population. As argued above they are also not related to a very old, metal-poor population. With an upper age of about 8 Gyr, the PG3 Miras are in that case likely related with the ‘bar’ population identified by Ng et al. ([@Ng96aea]b). The age and metallicity spread for this population (t=89 Gyr; ) might imply that the variables in PG3 are located in the outer regions of the ‘bar’. A complication is that one should be careful with the semantics related with the ‘bar’. If the stars of this population can be found in PG3, i.e. $\sim$1.5 kpc out of the galactic plane, this population cannot be originating from a bar as found in bar-like galaxies. Strictly spoken, this should be referred to as a triaxial structure.
The bluest SRVs and Miras might have a metallicity of with an age $\sim$9 Gyr. Although the uncertainty in the ages is between 13 Gyr, the large metallicity spread seems to indicate that all ages between are possible in the metallicity range . The great similarity in Fig. \[fig8\]b between the distributions of the variables from PG3 and the LMC suggests a comparable age and metallicity for both samples. Vallenari et al. ([@Vallenari96aea]b) find indications for enhancements of the star formation rate in the LMC at ages as old as , but in other regions the bulk of star formation has occurred only ago.
This raises the question if stars younger than 5 Gyr are present in the galactic bulge/bar. The bluer colour of the Sgr I Miras with respect to those from PG3 might indicate a younger age for the former, but as pointed out in Sect. \[CCD\] the colour difference is probably due to uncertainties in the colour transformations. The presence of carbon stars could be an indication for a young age, because the work from Marigo et al. ([@Marigo96aea]b and references cited therein) indicates that carbon stars cannot be much older than 4 Gyr. The carbon stars (L199,S283) identified in our sample of LPVs and those identified by Azzopardi et al. ([@Azzopardi91ea]) might therefore be an indication for the presence of stars younger than 5 Gyr in the bulge/bar. Ng& ([@NS97]) argue that S283 is actually related to the Sagittarius dwarf galaxy found by Ibata et al. ([@IGI94]). Possibly the same argument holds for PG3 variable L199. Furthermore it was argued that the sample of carbon stars from Azzopardi et al. ([@Azzopardi91ea]) is probably not associated with the bulge. In the bulge they would be bolometrically too faint, but associated with the Sagittarius dwarf galaxy their luminosities are comparable to carbon stars found in other dwarf galaxies.
If all the bulge carbon stars are related to the Sagittarius dwarf galaxy (Ng [@Ng97b], [@Ng98]) this would imply that they are absent in the bulge. In contrast to the field and LMC sample this implies the absence of a major star formation epoch less than 4 Gyr ago.
PG3 and the foreground stars {#foreground}
----------------------------
As already described in Sect. \[magdistribution\] the foreground contamination of PG3 Miras is either significantly larger or comparable with the PG3 SRVs. Figure \[fig7\] hints that both groups belong to the same population and a similar fraction of foreground stars is therefore expected. This implies that the SRVs are one magnitude fainter than the Miras. The K magnitude distribution of the two groups are in this case comparable. The asymmetry suggests that more stars are found nearby. This could imply a high foreground contamination as suggested by BL92, but most likely it is due to stars located in the nearby side of the triaxial structure mentioned above. Some of the nearby stars in the sample belongs therefore to the same population as the bulge SRVs and Miras and should not have been treated as foreground contamination.
Summary {#Summary}
=======
$\circ$ We have shown that PG3 SRVs are not the analogs to the field SRVs. The comparison of the relation of the two SRV groups shows that they do not obey the same relation. In addition their location in a colour-colour diagram differs slightly. All together this indicates a different nature between the two SRV groups. $\circ$ The PG3 SRVs form a short period extension to the Miras PK$_0$ and PC-relations. This indicates that the PG3 Miras and SRVs are both pulsating in the same mode, possibly the fundamental. $\circ$ The field SRVs (the ‘blue’ and the majority of the ‘red’ group) are likely overtone pulsators. $\circ$ The metallicities of the PG3 SRVs and Miras span the range from intermediate to approximately solar. $\circ$ The age possibly covers a range from . From the absence of LPVs in metal-poor globular clusters it is argued that the PG3 SRVs and Miras in the bulge are likely not older than 10 Gyr. From the upper mass limit of the bulge IRAS sources and the possible absence of bulge carbon stars one obtains a lower age limit of 4 Gyr. $\circ$ Field and PG3 Miras follow the same P/(J–K)$_0$ relation and cover the same region in the (J–H)$_0$ vs (H–K)$_0$ diagram. Therefore, the metallicity of the field and PG3 Miras should overlap each other. The Miras and SRVs in PG3 follow the Sgr I PK$_0$-relation. This confirms independently the work of Whitelock et al. ([@Whitelock91ea]) and Glass et al. ([@Glass95ea]): they found no difference in the PL-relation for different galactic environments.
The following question arises: are there SRVs in PG3, similar to those found in the disc? The presence or absence of these stars might provide an indication of the age of the stars in PG3. The missing SRVs might be hidden among the irregular variables. According to Wess87 those are variable stars with little or no trace of periodicity for which the amplitudes do not exceed 15. For verification, a detailed study of these stars is desired.
Another question which arises concerns the nature of the large spread in metallicity. The large range of ages seems to indicate that both young and old stars can be present with intermediate up to solar metallicity. Even the presence of more massive, young stars, which can be metal-poorer than older stars, is possible. In a closed box model one expects an increasing metallicity towards younger ages. Is this an indication that a closed box model is not applicable? What is the origin of this behaviour?
The authors thank H.J. Habing for his encouragements of this collaboration and the referees (Drs. Feast and Whitelock) for constructive suggestions. P.R. Wood is acknowledged for comments on the pulsation modes of AGB variables. The research of J. Hron, F. Kerschbaum and M. Schultheis is supported by the Austrian Science Fund projects P9638–AST and S7308. Y.K. Ng thanks the Institut für der Universität Wien and the department of Astronomy from the University of Padova, where part of this research was carried out, for their hospitality. The university of Wien provided financial support for Ng’s research visit to Vienna and ANTARES, an astrophysics network funded by the HCM programme of the European Community, supported Ng’s research visits to Padova. At the IAP-CNRS and at the Padova Astronomical Observatory the research of Ng was supported by respectively HCM grant CHRX-CT94-0627 and TMR grant ERBFMRX-CT96-0086 from the EC.
Azzopardi M., Lequeux J., Rebeirot E., Westerlund B.E., 1991, A&AS 88, 265 Bersanelli M., Bouchet P., Falomo R., 1991, A&A 252, 854 Bertelli G., Bressan A., Chiosi C., Fagotto F., Nasi E., 1994, A&AS 106, 275 Bertelli G., Bressan A., Chiosi C., Ng Y.K., Ortolani S., 1995, A&A 301, 381 Bertelli G., Bressan A., Chiosi C., Ng Y.K., 1996, A&A 310, 115
Bessell M.S., Brett J.M., 1988, PASP 100, 1134 Blaauw A., in ‘Coordination of Galactic Research’, IAU symposium 1, 4 Blommaert J.A.D.L., 1992, Ph.D. thesis, Leiden University, the Netherlands (Bl92) Blommaert J.A.D.L., Brown A.G., Habing H.J., et al., 1990, ESO Messenger 62, 6 Blommaert J.A.D.L., Langevelde van H.J., Michiels W., 1994, A&A 287, 479 Bouchet P., Manfroid J., Schmider F.X., 1991, A&AS 91, 409
Bressan A., Nasi E., 1995, private communication
Carter B.S., 1990, MNRAS 242, 1 Catchpole R.M., Robertson B.S.C., Lloyd Evans T.H.H., Feast M.W., Glass I.S., Carter B.S., 1979, South African Astronomical Observatory Circulars 1, 61 Chan S.J., Kwok S., 1988, ApJ 334, 362
Costa E., Frogel J.A., 1996, AJ 112, 2607 de Jong T., 1989, A&A 223, L23
Engels D., Sherwood W.A., Wamsteker W., Schultz G.V., 1981, A&AS 45, 5
Epchtein N., et al., 1994, ApSS 217, 3 Epchtein N., de Batz B., Capoani L., et al., 1997, ESO Messenger 87, 27
Feast M.W., 1996, MNRAS 278, 11 Feast M.W., Robertson B.S.C., Catchpole R.M., 1982, 201, 439 Feast M.W., Whitelock P.A., 1987, in [*‘Late stages of stellar evolution’*]{}, eds. S. Kwok and S.R. Pottasch (Reidel, Dordrecht), p33 Feast M.W., Glass I.A., Whitelock P.A., Catchpole R.M, 1989, MNRAS 241, 375
Frogel J.A., Whitford A.E., 1987, ApJ 320, 199 Geisler D., Friel E.D., 1992, AJ 104, 128
Glass I.S., 1974, MNASSA 33, 53, 71 Glass I.S., Whitelock P.A., Catchpole R.M., Feast M.W., 1995, MNRAS 273, 383 Groenewegen M.A.T., 1993, Ph.D. thesis, University of Amsterdam, the Netherlands Groenewegen M.A.T., de Jong T., van der Bliek N.S., Slijkhuis S., Willems F.J., 1992, A&A 253, 150 Habing H.J., 1988, A&A 200, 40 Habing H.J., 1996, A&AR 7, 97
Harmon R., Gilmore G., 1988, MNRAS 235, 1025 Herman J., 1988, A&AS 74, 133
Hron J., Kerschbaum F., 1994, ApSS 217, 137 Hron J., Kerschbaum F., Ng Y.K., Schultheis M., 1998, A&A [*to be submitted*]{} Hughes S.M.G., Wood P.R., 1990, AJ 99, 784
Ibata R., Gilmore G., Irwin M.J., 1994, Nature 370, 194
Jura M., Kleinmann S.G., 1992, ApJS 79, 105 Kanbur S.M., Hendry M.A., Clarke D., 1997, MNRAS 289, 428
Kerschbaum F., 1993, Ph.D. thesis, University of Vienna, Austria Kerschbaum F., 1995, A&AS 113, 441 Kerschbaum F., Hron J., 1992, A&A 263, 97 (KH92) Kerschbaum F., Hron J., 1994, A&AS 106, 397 (KH94)
Kholopov P.N., Samus N.N., Frolov M.S., et al., 1988, ‘[*General Catalog Of Variable Stars*]{}’, 4$^{th}$ edition, Nauka Publishing House, Moscow.
Kiraga M., Paczyński B., Stanek K.Z., 1997, ApJ 485, 611 Kurucz R.L., 1992, in Proceedings IAU symposium 149, ‘[*The stellar populations of galaxies*]{}’, B. Barbuy and A. Renzini (eds.), 225 Lançon A., Rocca-Volmerange B., 1992, A&AS 96, 593
Larsson-Leander G., 1959, in ‘Second Conference on Coordination of Galactic Research’, IAU symposium 7, 22 Lindquist M., Habing H.J., Winnberg A., 1992, A&A 259, 118
Marigo P., 1998, Ph.D. thesis, Padova University, Italy Marigo P., Bressan A., Chiosi C., 1996a, A&A 313, 545 Marigo P., Girardi L., Chiosi C., 1996b, A&A 316, L1
McWilliam A., Rich R.M., 1994, ApJS 91, 749 Minniti D., 1995, AJ 109, 1663
Ng Y.K., 1994, Ph.D. thesis, Leiden University, the Netherlands Ng Y.K., 1997a, in Proceedings 12th IAP Colloquium [*‘Variable stars and the astrophysical returns from microlensing surveys’*]{}, R. Ferlet and J.-P. Maillard (eds.), 113 Ng Y.K., 1997b, A&A 328, 211 Ng Y.K., 1998, A&A 338, [*in press*]{} Ng Y.K., Bertelli G., 1998, A&A 329, 943 Ng Y.K., Schultheis M., 1997, A&AS 123, 115 Ng Y.K., Bertelli G., Bressan A., Chiosi C., Lub J., 1995, A&A 295, 655 (erratum A&A 301,318) Ng Y.K., Bertelli G., Chiosi C., Bressan A., 1996a, A&A 310, 771 Ng Y.K., Bertelli G., Chiosi C., Bressan A., 1996b, [*‘Spiral Galaxies in the Near-IR’*]{}, D. Minniti and H.-W. Rix (eds.), 110 Ng Y.K., Bertelli G., Chiosi C., 1996c, A&A [*submitted*]{} Ng Y.K., Bertelli G., Chiosi C., 1996d, A&A [*submitted*]{} Ng Y.K., Bertelli G., Chiosi C., Bressan A., 1997, A&A 324, 65 Ng Y.K., Schultheis M., Hron J., 1998, A&A [*to be submitted*]{} Paczyński B., Udalski A., 1997, in Proceedings 12th IAP Colloquium [*‘Variable stars and the astrophysical returns from microlensing surveys’*]{}, R. Ferlet and J.-P. Maillard (eds.), 29 Paczyński B., et al., 1994, AJ 107, 2060
Plaut L., 1966, Bull. Astr. Inst. Neth. Suppl. 1, 105 Plaut L., 1968a, Bull. Astr. Inst. Neth. Suppl. 2, 293 Plaut L., 1968b, Bull. Astr. Inst. Neth. Suppl. 3, 1 Plaut L., 1970, A&A 8, 341 Plaut L., 1971, A&AS 4, 75 (PL71) Plaut L., 1973, A&AS 12, 351
Reid M.J., 1993, ARA&A 31, 345 Reid I.N., Hughes S.M.G., Glass I.S., 1995, MNRAS 275, 331 Rich R.M., 1990, ApJ 362, 604 Ridgway S.T., et al., 1980, ApJ 235, 126 Sadler E.M., Rich R.M., Terndrup D.M., 1996, AJ 112, 171 Schultheis M., Ng Y.K., Hron J., Habing H.J., 1996, in Proceedings IAU symposium 169, [*‘Unsolved problems of the Milky Way’*]{}, L. Blitz (ed.), 347
Stephenson C.B., 1989, Publ. Warner and Swasey Obs., 3, 55 te Lintel-Hekkert P., 1990, Ph.D. thesis, Leiden University, the Netherlands
Terndrup D.M., 1988, AJ 96, 884 Tyson N.D., Rich R.M., 1991, ApJ 367, 547
Vallenari A., Chiosi C., Bertelli G., Ortolani S., 1996a, A&A 309, 358 Vallenari A., Chiosi C., Bertelli G., Aparicio A., Ortolani S., 1996b, A&A 309, 367
van der Bliek N.S., Manfroid J., Bouchet P., 1996, A&AS 119, 547 van de Hulst H.C., 1949, Rech.Astr.Obs.Utrecht, Vol.11 part 2 van der Veen W.E.C.J., 1988, Ph.D. thesis, Leiden University, the Netherlands van Langevelde H.J., 1992, Ph.D. thesis, Leiden University, the Netherlands van Leeuwen F., Feast M. W., Whitelock P. A., Yudin B., 1997, MNRAS 287, 955
Wamsteker W., 1981, A&A 97, 329 Wesselink Th.J.H., 1987, Ph.D. thesis, Catholic University of Nijmegen, the Netherlands (Wess87) Westerlund B.E., Lequeux J. Azzopardi M., Rebeirot E., 1991, A&A 244, 367
Whitelock P.A., 1996, in [*‘Spiral Galaxies in the Near-IR’*]{}, D. Minniti and H.-W. Rix (eds.), 75 Whitelock P.A., Feast M.W., Catchpole R.M., 1991, MNRAS 248, 276
Willems F.J., 1987, Ph.D. thesis, University of Amsterdam, the Netherlands Willems F.J., 1988a, A&A 203, 51 Willems F.J., 1988b, A&A 203, 65 Willems F.J., de Jong T., 1988, A&A 196, 173
Wood P.R., 1990, In: [*[From Miras to Planetary Nebulae: Which Path for Stellar Evolution?]{}*]{}, eds. M.O. Mennessier and A. Omont, Editions Frontières, France, p.67 Wood P.R., Moore G.K.C., Hughes S.M.G., 1991, in ‘[*The Magellanic Clouds’*]{} IAU symposium 148, eds. R. Haynes and D. Milne (Kluwer, Dordrecht), p259 Wood P.R., Sebo, P.R., 1996, MNRAS 282, 958 (WS96) Zuckerman B., Maddalena R.J., 1989, A&A 223, L20
[^1]: [*Present address:*]{} [Space Research Organisation Netherlands (SRON), Sorbonnelaan 2, 3584 CA Utrecht, the Netherlands]{}
[^2]: Table 2 is only available in electronic form at the CDS via anonymous [ftp]{} to [cdsarc.u-strasbg.fr (130.79.128.5)]{} or [WWW]{} at [URL]{} http://cdsweb.u-strasbg.fr/Abstract.html
[^3]: Based on observations obtained at the European Southern Observatory, La Silla, Chile
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The problem of characterizing impacts of data quality on real-time locational marginal price (LMP) is considered. Because the real-time LMP is computed from the estimated network topology and system state, bad data that cause errors in topology processing and state estimation affect real-time LMP. It is shown that the power system state space is partitioned into price regions of convex polytopes. Under different bad data models, the worst case impacts of bad data on real-time LMP are analyzed. Numerical simulations are used to illustrate worst case performance for IEEE-14 and IEEE-118 networks.'
author:
- 'Liyan Jia, Jinsub Kim, Robert J. Thomas, and Lang Tong, '
title: 'Impact of Data Quality on Real-Time Locational Marginal Price [^1]'
---
locational marginal price (LMP), real-time market, power system state estimation, bad data detection, cyber security of smart grid.
Introduction {#sec:intro}
============
deregulated electricity market has two interconnected components. The day-ahead market determines the locational marginal price (LMP) based on the dual variables of the optimal power flow (OPF) solution , given generator offers, demand forecast, system topology, and security constraints. The calculation of LMP in the day-ahead market does not depend on the actual system operation. In the real-time market, on the other hand, an ex-post formulation is often used (by PJM and ISO-New England ) to calculate the real-time LMP by solving an incremental OPF problem. The LMPs in the day-ahead and the real-time markets are combined in the final clearing and settlement processes.
The real-time LMP is a function of data collected by the supervisory control and data acquisition (SCADA) system. Therefore, anomalies in data, if undetected, will affect prices in the real-time market. While the control center employs a bad data detector to “clean” the real-time measurements, miss detections and false alarms will occur inevitably. The increasing reliance on the cyber system also comes with the risk that malicious data may be injected by an adversary to affect system and real-time market operations. An intelligent adversary can carefully design a data attack to avoid detection by the bad data detector.
Regardless of the source of data errors, it is of significant value to assess potential impacts of data quality on the real-time market, especially when a smart grid may in the future deploy demand response based on real-time LMP. To this end, we are interested in characterizing the impact of [*worst case data errors*]{} on the real-time LMP. The focus on the worst case also reflects the lack of an accurate model of bad data and our desire to include the possibility of data attacks.
Summary of Results and Organization
-----------------------------------
We aim to characterize the worst effects of data corruption on real-time LMP. By “worst”, we mean the maximum perturbation of real-time LMP caused by bad or malicious data, when a fixed set of data is subject to corruption. [The complete characterization of worst data impact, however, is not computationally tractable. Our goal here is to develop an optimization based approach to search for [*locally worst data*]{} by restricting the network congestion to a set of lines prone to congestion. We then apply computationally tractable (greedy search) algorithms to find the worst data and evaluate the effects of worst data by simulations.]{}
In characterizing the relation between data and real-time LMP, we first present a geometric characterization of the real-time LMP. In particular, we show that the state space of the power system is partitioned into polytope price regions, as illustrated in Fig. \[fig:geometry\_err\](a), where each polytope is associated with a unique real-time LMP vector, and the price region ${\mathscr{X}}_i$ is defined by a particular set of congested lines that determine the boundaries of the price region.
\[c\][${{\mathscr{X}}}_0$]{} \[c\][${{\mathscr{X}}}_1$]{} \[c\][${{\mathscr{X}}}_2$]{} \[c\][${{\mathscr{X}}}_3$]{} \[c\][${{\mathscr{X}}}_4$]{} \[c\][$\hat{x}$]{} \[c\][$\tilde{x}$]{}
Two types of bad data are considered in this paper. One is the bad data associated with meter measurements such as the branch power flows in the network. Such bad data will cause errors in state estimation, possibly perturbing, as an example, the correct state estimate $\hat{x}$ in ${\mathscr{X}}_0$ to $\tilde{x}$ in ${\mathscr{X}}_3$ (as shown in Fig. \[fig:geometry\_err\](a)). The analysis of the worst case data then corresponds to finding the worst measurement error such that it perturbs the correct state estimation to the worst price region.
The second type of bad data, one that has not been carefully studied in the context of LMP in the literature, is error in digital measurements such as switch or breaker states. Such errors lead directly to topology errors therefore causing a change in the polytope structure as illustrated in Fig. \[fig:geometry\_err\](b). In this case, even if the estimated system state changes little, the prices associated with each region change, sometimes quite significantly.
Before characterizing impacts of bad meter data on LMP, we need to construct appropriate models for bad data. To this end, we propose three increasingly more powerful bad data models based on the dependencies on real-time system measurements: state independent bad data, partially adaptive bad data, and fully adaptive bad data.
In studying the worst case performance, we adopt a widely used approach that casts the problem as one involving an adversary whose goal is to make the system performance as poor as possible. The approach of finding the worst data is equivalent to finding the optimal strategy of an attacker who tries to perturb the real-time LMP and avoid being detected at the same time. By giving the adversary more information about the network state and endowing him with the ability to change data, we are able to capture the worst case performance, sometimes exactly and sometimes as bounds on performance. Finally, we perform simulation studies using the IEEE-14 and IEEE-118 networks. We observe that bad data independent of the system state seems to have limited impact on real-time LMPs, and greater price perturbations can be achieved by state dependent bad data. The results also demonstrate that the real-time LMPs are subject to much larger perturbation if bad topology data are present in addition to bad meter data. While substantial price changes can be realized for small networks by the worst meter data, as the size of network grows while the measurement redundancy rate remains the same, the influence of worst meter data on LMP is reduced. However, larger system actually gives more possibilities for the bad topology data to perturb the real-time LMP more significantly.
Our simulation results also show a degree of robustness provided by the [*nonlinear state estimator.*]{} While there have been many studies on data injection attacks based on DC models, very few consider the fact that the control center typically employs the nonlinear WLS state estimator under the AC model. Our simulation shows that effects of bad analog data designed based on DC model may be mitigated by the nonlinear estimator whereas bad topology data coupled with bad analog data can have greater impacts on LMP.
The rest of the paper is organized as follows. Section \[sec:realtimeLMP\] briefly describes a model of real-time LMP and introduces its geometric characterization in the state space of the power system. Section \[sec:data\_quality\] establishes the bad data models and summarizes state estimation and bad data detection procedures at the control center. In Section \[sec:worstcase\], a metric of impact on real-time LMP caused by bad meter data is introduced. We then discuss the algorithms of finding worst case bad meter data vector in terms of real-time price perturbation under the three different bad data models. Section \[sec:bad\_topo\] considers the effect of bad topology data on real-time LMP. Finally, in Section \[sec:simulation\], simulation results are presented based on IEEE-14 and IEEE-118 networks.
Related Work
------------
Effects of bad data on power system have been studied extensively in the past, see . Finding the worst case bad data is naturally connected with the problem of malicious data. In this context, the results presented in this paper can be viewed as one of analyzing the impact of the worst (malicious) data attack.
In a seminal paper by Liu, Ning, and Reiter , the authors first illustrated the possibility that, by compromising enough number of meters, an adversary can perturb the state estimate arbitrarily in some subspace of the state space without being detected by any bad data detector. Such attacks are referred to as strong attacks. It was shown by Kosut that the condition for the existence of such undetectable attacks is equivalent to the classical notion of network observability. When the adversary can only inject malicious data from a small number of meters, strong attacks do not exist, and any injected malicious data can be detected with some probability. Such attacks are referred to as weak attacks . In order to affect the system operation in some meaningful way, the adversary has to risk being detected by the control center. The impacts of weak attack on power system are not well understood because the detection of such bad data is probabilistic. Our results are perhaps the first to quantify such impacts. Most related research works focused on DC model and linear estimator while only few have addressed the nonlinearity effect .
It is well recognized that bad data can also cause topology errors , and techniques have been developed to detect topology errors. For instance, the residue vector from state estimation was analyzed for topology error detection . Monticelli [@Monticelli:93TPS] introduced the idea of generalized state estimation where, roughly speaking, the topology that fits the meter measurements best is chosen as the topology estimate. The impacts of topology errors on electricity market have not been reported in the literature, and this paper aims to bridge this gap.
The effect of data quality on real-time market was first considered in . In , the authors presented the financial risks induced by the data perturbation and proposed a heuristic technique for finding a case where price change happens. While there are similarities between this paper and , several significant differences exist: (i) This paper focuses on finding the worst case, not only a feasible case. (ii) This paper considers a more general class of bad data where bad data may depend dynamically on the actual system measurements rather than static. (iii) This paper considers a broader range of bad data that also include bad topology data, and our evaluations are based on the AC network model and the presence of nonlinear state estimator.
Structures of Real-Time LMP {#sec:realtimeLMP}
===========================
In this section, we present first a model for the computation of real-time locational marginal price (LMP). While ISOs have somewhat different methods of computing real-time LMP, they share the same two-settlement architecture and similar ways of using real-time measurements. In the following, we will use a simplified ex-post real-time market model, adopted by PJM, ISO New England, and other ISOs . We view this model as a convenient mathematical abstraction that captures the essential components of the real-time LMP calculation. For this reason, our results should be interpreted within the specified setup. Our purpose is not to include all details; we aim to capture the essential features.
In real-time, in order to monitor and operate the system, the control center will calculate the estimated system conditions (including bus voltages, branch flows, generation, and demand) based on real-time measurements. We call a branch congested if the estimated flow is larger than or equal to the security limit. The congestion pattern is defined as the set of all congested lines, denoted as $\hat{\mathscr{C}}$. Note that we use hat ($\hat{\mathscr{C}}$) to denote quantities or sets that are estimated based on real-time measurements. Details of state estimation and bad data detection are discussed in Section \[ssec:se\].
One important usage of state estimation is calculating the real-time LMP. Given the estimated congestion pattern $\hat{\mathscr{C}}$, the following linear program is solved to find the incremental OPF dispatch and associated real-time LMP, $\hat{\lambda}=(\hat{\lambda}_i)$ [@Ott:03TPS]: $$\begin{array}{l l}
\mbox{minimize} & \sum c_{i}^\text{\tiny G} \Delta p_i-\sum c_{j}^\text{\tiny L} \Delta d_j\\
\mbox{subjcet to} & \sum \Delta p_{i}=\sum \Delta d_j \\
& \Delta p_i^{\text{min}} \le \Delta p_{i} \le \Delta p_i^{\text{max}}\\
& \Delta d_j^{\text{min}} \le \Delta d_{j} \le \Delta d_j^{\text{max}}\\
& \sum_i A_{ki}\Delta p_i - \sum_j A_{kj} \Delta d_j \le 0,\text{for all }k \in \hat{\mathscr{C}}, \\
\end{array}
\label{eq:real-timeLMP}$$ where $\Delta d = (\Delta d_j)$ is the vector of incremental dispatchable load, $\Delta p=(\Delta p_i)$ the vector of incremental generation dispatch, $c^\text{\tiny G} = (c_{i}^\text{\tiny G})$ and $c^\text{\tiny L} = (c_{j}^\text{\tiny L})$ the corresponding real-time marginal cost of generations and dispatchable loads, $\Delta p_i^{\text{min}}$ and $\Delta p_i^{\text{max}}$ the lower and upper bounds for incremental generation dispatch, $\Delta d_i^{\text{min}}$ and $\Delta d_i^{\text{max}}$ the lower and upper bounds for incremental dispatchable load, and $A_{ki}$ the sensitivity of branch flow on branch $k$ with respect to the power injection at bus $i$.
The real-time LMP at bus $i$ is defined as the overall cost increase when one unit of extra load is added at bus $i$, which is calculated as $$\hat{\lambda}_i = \eta-\sum_{k \in \hat{\mathscr{C}}}A_{ki}\mu_{k}.
\label{eq:LMPcal}$$ where $\eta$ is the dual variable for the load-generation equality constraint, and $\mu_{k}$ is the dual variable corresponding to the line flow constraint in (\[eq:real-timeLMP\]).
Note that in practice, the control center may use the ex-ante congestion pattern, which is obtained by running a 5 minute ahead security-constrained economic dispatch with the state estimation results and the forecasted loads (for the next five-minute interval) and choosing the lines congested at the dispatch solution . However, to avoid the complication due to ex-ante dispatch calculation, we assume that real-time pricing employs the estimated congestion pattern $\hat{{\mathscr{C}}}$ obtained from state estimation results. By doing so, we attempt to find direct relations among bad data, the state estimate, and real-time LMPs. Notice that once the congestion pattern $\hat{\mathscr{C}}$ is determined, the whole incremental OPF problem (\[eq:real-timeLMP\]) no longer depends on the measurement data. Under the DC model, the power system state, $x$, is defined as the vector of voltage phases, except the phase on the reference bus. The power flow vector $f$ is a function of the system state $x$, $$\centering
f=Fx,
\label{eq:flowequation}$$ where $F$ is the sensitivity matrix of branch flows with respect to the system state.
Assume the system has $n+1$ buses. Then, $x \in \mathscr{X}=[-\pi,\pi]^n$, where $\mathscr{X}$ represents the state space. Any system state corresponds to a unique point in $\mathscr{X}$. From (\[eq:flowequation\]), the branch flow $f$ is determined by the system state $x$. Comparing the flows with the flow limits, we obtain the congestion pattern associated with this state. Hence, each point in the state space corresponds to a particular congestion pattern.
[We note that the above expression in (\[eq:LMPcal\]) appears earlier in where the role of congestion state in LMP computation was discussed. In this paper, our objective is to make explicit the connection between data and LMP. We therefore need a linkage between data and congestion. To this end, we note that the power system state, the congestion state, and LMP form a Markov chain, which led to a geometric characterization of LMP on the power system state space, as shown in the following theorem.]{}
[Assume that the LMP exists for every possible congestion pattern[^2]. Then, the state space ${\mathscr{X}}$ is partitioned into a set of polytopes $\{{\mathscr{X}}_i\}$ where the interior of each ${\mathscr{X}}_i$ is associated with a unique congestion pattern ${\mathscr{C}}_i$ and a real-time LMP vector. Each boundary hyperplane of ${\mathscr{X}}_i$ is defined by a single transmission line.]{} \[thm:partition\]
For a particular congestion pattern ${\mathscr{C}}$ defined by a set of congested lines, the set of states that gives ${\mathscr{C}}$ is given by $${\mathscr{X}}_i\defeq \{x: F_{i\cdot}x \ge T_i^{\text{max}} \text{ } \forall i \in {\mathscr{C}},
F_{j\cdot}x < T_j^{\text{max}} \text{ } \forall j \notin {\mathscr{C}}\},$$ where $F_{i\cdot}$ is the $i$th row of $F$ (see (\[eq:flowequation\])), and $T_{j}^{\mbox{max}}$ the flow limit on branch $j$. Since ${\mathscr{X}}_i$ is defined by the intersection of a set of half spaces, it is a polytope.
[Given an estimated congestion pattern $\hat{{\mathscr{C}}}$, the envelop theorem implies that for any optimal primal solution and dual solution of (\[eq:real-timeLMP\]) that satisfy the KKT conditions, (\[eq:LMPcal\]) always gives the derivative of the optimal objective value with respect to the demand at each bus, which we assume exists, $\ie$ each congestion pattern is associated with a unique real-time LMP vector $\lambda$. Hence, all states with the same congestion pattern share the same real-time LMP, which means each polytope ${\mathscr{X}}_i$ in ${\mathscr{X}}$ corresponds to a unique real-time LMP vector.]{}
Theorem \[thm:partition\] characterizes succinctly the relationship between the system state and LMP. As illustrated in Fig. \[fig:geometry\_err\](a), if bad data are to alter the LMP in real-time, the size of the bad data has to be sufficiently large so that the state estimate at the control center is moved to a different price region from the true system state. On the other hand, if some lines are erroneously removed from or added to the correct topology, as illustrated in Fig. \[fig:geometry\_err\](b), it affects the LMP calculation in three ways[^3]. First, the state estimate is perturbed since the control center employs an incorrect topology in state estimation. Secondly, the price partition of the state space changes due to the errors in topology information. Third, the shift matrix $A$ in (\[eq:real-timeLMP\]), which is a function of topology, changes thereby altering prices attached to each price region.
Data Model and State Estimation {#sec:data_quality}
===============================
Bad Data Model {#ssec:data_quality}
--------------
### Meter data
In order to monitor the system, various meter measurements are collected in real time, such as power injections, branch flows, voltage magnitudes, and phasors, denoted by a vector $z\in{\mathbb{R}}^{m}$. [^4] If there exists bad data $a$ among the measurements, the measurement with bad data, denoted by $z_a$, can be expressed as a function of the system states $x$,$$z_a = z + a=h(x)+w+a,~~a \in {\mathscr{A}},
\label{eq:z=hxwa}$$ where $w$ represents the random measurement noise.
We make a distinction here between the measurement noise and bad data; the former accounts for random noise independently distributed across all meters whereas the latter represents the perturbation caused by bad or malicious data. We assume no specific pattern for bad data except that they do not happen everywhere. We assume that bad data can only happen in a subset of the measurements, $\mathscr{S}$. [We call $\mathscr{S}$ as set of suspectable meters, which means the meter readings with in $\mathscr{S}$ may subject to corruption.]{} If the cardinality of $\mathscr{S}$ is $k$, the feasible set of bad data $a$ is a $k$-dimensional subspace, denoted as ${\mathscr{A}}= \{a: a_i = 0\text{ for all } i \notin \mathscr{S}\}$.
We will consider three bad data models with increasing power of affecting state estimates.
M1. [*State independent bad data*]{}: This type of bad data is independent of real-time measurements. Such bad data may be the replacement of missing measurements.
M2. [*Partially adaptive bad data*]{}: This type of bad data may arise from the so-called man in the middle (MiM) attack where an adversary intercepts the meter data and alter the data based on what he has observed. Such bad data can adapt to the system operating state.
M3. [*Fully adaptive bad data*]{}: This is the most powerful type of bad data, constructed based on the actual measurement $z=h(x)+w$. Note that M3 is in general not realistic. Our purpose of considering this model is to use it as a conservative proxy to obtain performance bounds for the impact of worst case data.
We assume herein a DC model in which the measurement function $h(\cdot)$ in (\[eq:z=hxwa\]) is linear. Specifically, $$\label{eq:DCeq}
z_a=Hx+w+a,~~ a \in {\mathscr{A}},$$ where $H$ is the measurement matrix. Such a DC model, while widely used in the literature, may only be a crude approximation of the real power system. By making such a simplifying assumption and acknowledging its weaknesses, we hope to obtain tractable solutions in searching for worst case scenarios. It is important to note that, although the worst case scenarios are derived from the DC model, we carry out simulations using the actual nonlinear system model.
### Topology data
Topology data are represented by a binary vector $s\in\{0,1\}^{l}$, where each entry of $s$ represents the state of a line breaker ($0$ for open and $1$ for closed). The bad topology data is modeled as $$\label{eq:bad_data_topo}
s_b = s + b~\text{ (mod 2)},~~ b\in{\mathscr{B}},$$ where ${\mathscr{B}}\subset \{0,1\}^{l}$ is the set of possible bad data. When bad data are present, the topology processor will generate the topology estimate corresponding to $s_b$, and this incorrect topology estimate will be passed to the following operations unless detected by the bad data detector.
State Estimation {#ssec:se}
----------------
We assume that the control center employs the standard weighted least squares (WLS) state estimator. Under DC model, $$\label{eq:WLS}
\hat{x}=\arg\min_x (z-Hx)^{\tiny{\text{T}}} R^{-1} (z-Hx) = Kz,$$ where $R$ is the covariance matrix of measurement noise $w$, and $K \triangleq (H^{\tiny{\text{T}}}R^{-1}H)^{-1}H^{\tiny{\text{T}}}R^{-1}$.
If the noise $w$ is Gaussian, the WLS estimator is also the maximum likelihood estimate (MLE) of state $x$. By the invariant property of MLE, from (\[eq:flowequation\]), the maximum likelihood estimate of the branch flows is calculated as $$\label{eq:WLSflow}
\centering
\hat{f} = F\hat{x}=FKz.$$
The congestion pattern used in real-time LMP calculation (\[eq:real-timeLMP\]) is directly from state estimation and consists of all the estimated branch flows which are larger than or equal to the branch flow limits, $\ie$ $$\label{eq:CongestionPattern}
\centering
\hat{{\mathscr{C}}} = \{j: \hat{f}_j \ge T_{j}^{\max} \},$$ where $T_{j}^{\max}$ is the flow limit on branch $j$.
In the presence of bad meter data $a$, the meter measurements collected by control center is actually $z_a=Hx+w+a$. By using $z_a$, the WLS state estimate is $$\hat{x}_a= Kz_a = \hat{x}^* + Ka,
\label{eq:state_move}$$ where $\hat{x}^* = Kz$ is the “correct" state estimate without the presence of the bad data ($a = 0$).
Eq. (\[eq:state\_move\]) shows that the effect of bad data on state estimation is linear. However, because $a$ is confined in a $k$-dimensional subspace ${\mathscr{A}}$, the perturbation on the actual system state is limited to a certain direction.
When bad data exist both in meter and topology data, the control center uses a wrong measurement matrix $\bar{H}$, corresponding to the altered topology data, and the altered meter data $z_{a}$. Then, the WLS state estimate becomes $$\hat{x}_a= \bar{K}z_a = \bar{K}z + \bar{K}a,
\label{eq:state_topo}$$ where $\bar{K} \triangleq (\bar{H}^{\tiny{\text{T}}}R^{-1}\bar{H})^{-1}\bar{H}^{\tiny{\text{T}}}R^{-1}$. Note that unlike the linear effect of bad meter data, bad topology data affects the state estimate by altering the measurement matrix $H$ to $\bar{H}$.
Bad Data Detection
------------------
The control center uses bad data detection to minimize the impact of bad data. Here, we assume a standard bad data detection used in practice, the $J(\hat{x})$-detector in . In particular, the $J(\hat{x})$-detector performs the test on the residue error, $r\triangleq z -H\hat{x}$, based on the state estimate $\hat{x}$. From the WLS state estimate (\[eq:WLS\]), we have $$\label{eq:residue}
\centering
r = \left(I-H(H^{\tiny{\text{T}}}R^{-1}H)^{-1}H^{\tiny{\text{T}}}R^{-1}\right)z = Uz.$$ where $U \defeq (I-H(H^{\tiny{\text{T}}}R^{-1}H)^{-1}H^{\tiny{\text{T}}}R^{-1})$ The $J(\hat{x})$-detector is a threshold detector defined by $$r^{\tiny{\text{T}}}R^{-1}r = z^{\tiny{\text{T}}}Wz \begin{array}{c}
\mbox{bad data}\\
\gtrless\\
\mbox{good data}\\
\end{array}
\tau,$$ where $\tau$ is the threshold calculated from a prescribed false alarm probability, and $W \defeq U^{\tiny{\text{T}}}R^{-1}U$. When the measurement data fail to pass the bad data test, the control center declares the existence of bad data and takes corresponding actions to identify and remove the bad data.
In this paper, we are interested in those cases when bad data are present while the $J(\hat{x})$-detector fails to detect them.
Impact of Bad Data on LMP {#sec:worstcase}
=========================
In this section, we examine the impact of bad data on LMP, assuming that the topology estimate of the network is correct.
[ One thing to notice is that in searching for the “worst” case, we take the perspective of the control center, not that of the attacker. In particular, we look for the worst congestion pattern for the LMP computation, even if this particular congestion pattern is difficult for the attacker to discover. So the focus here is not how easy it is for an attacker to find a locally worst congestion pattern; it is how much such a congestion pattern affects the LMP.]{}
Average Relative Price Perturbation
-----------------------------------
In order to quantify the effect of bad data on real-time price, we need to first define the metric to measure the effect. We define the [*relative price perturbation*]{} (RPP) as the expected percentage price perturbation caused by bad data. Given that LMP varies at different buses, RPP also varies at different locations.
Let $z_a$ be the data received at the control center and $\lambda_i(z_{a})$ the LMP at bus $i$. The RPP at bus $i$ is a function of bad data $a$, given by $$\label{eq:metric}
\text{RPP}_i(a)=\mathbb{E}\left(\left|\frac{\lambda_i(z_a)-\lambda_i(z)}{\lambda_i(z)}\right|\right),$$ where the expectation is over random state and measurement noise. To measure the system-wide price perturbation, we define the [*average relative price perturbation*]{} (ARPP) by $$\label{eq:overall metric}
\text{ARPP}(a)=\frac{1}{n+1}\sum_i \text{RPP}_i (a),$$ where $n+1$ is the number of buses in the system.
The worst case analysis to be followed can be used for other metrics (e.g., price increase ratios or price decrease ratios, which are closely related to the market participants’ gain or loss). Similar results can be showed following the same strategies. However, the comparison among different metrics is beyond the scope of this paper.
Worst ARPP under State Independent Bad Data Model {#ssec:constant}
-------------------------------------------------
First, we consider the state independent bad data model (M1) given in Section \[ssec:data\_quality\]. In this model, the bad data are independent of real-time measurements.
In constructing the state independent worst data, it is useful to incorporate prior information about the state. To this end, we assume that system state follows a Gaussian distribution with mean $x_0$, covariance matrix $\Sigma_x$. Typically, we choose $x_0$ as the day-ahead dispatch since the nominal system state in real-time varies around its day-ahead projection.
In the presence of bad data $a$, the expected state estimate and branch flow estimate on branch $i$ are given by $$\mathbb{E}[\hat{x}] = x_0 + K a.$$ $$\mathbb{E}[f_{i}] = F_{i\cdot} \mathbb{E}[\hat{x}]= F_{i\cdot}x_0 + F_{i\cdot} Ka,$$ where $F_{i\cdot}$ is the corresponding row of branch $i$ in $F$.
Our strategy is to make this expected state estimate into the region with the largest price perturbation among all the possible regions, $\hat{{\mathscr{C}}}^*$. From (\[eq:CongestionPattern\]), this means making all the expected branch flows satisfy the boundary condition of $\hat{{\mathscr{C}}}^*$, $$\begin{tabular}{ll}
$\mathbb{E}[f_{i}] \ge T_i^{\text{max}} $ & $\text{for }i \in \hat{{\mathscr{C}}}^*$ \\
$\mathbb{E}[f_{i}] \le T_j^{\text{max}} $ & $\text{for }j \notin \hat{{\mathscr{C}}}^*$.
\end{tabular}
\label{eq:Ccon}$$
However, due to the uncertainty (from both system state $x$ and measurement noise $w$), the actual estimated state after attack, $\hat{x}$, may be different from $\mathbb{E}[\hat{x}]$. Therefore, we want to make $\mathbb{E}[\hat{x}]$ at the “center” of the desired price region, $\ie$ maximizing the shortest distance from $\mathbb{E}[\hat{x}]$ to the boundaries of the polytope price regions while still holding the boundary constraints. The shortest distance can be calculated as $$\beta = \mbox{min}\{\tilde{\beta}: |\mathbb{E}[f_{i}] - T^{\text{max}}| \ge \tilde{\beta} \text{ for all i}\}.$$
However, the existence of bad data detector prevents the bad data vector $a$ from being arbitrarily large. According to (\[eq:residue\]), the weighted squared residue with $a$ is $$r^{\tiny{\text{T}}}R^{-1}r = z_a^{\tiny{\text{T}}} W z_a = (w+a)^{\tiny{\text{T}}} W (w+a).$$ since $WHx = 0$
Heuristically, since $w$ has zero mean, the term $a^{\tiny{\text{T}}}Wa$ can be used to quantify the effect of data perturbation on estimation residue. Then we use $a^{\tiny{\text{T}}} W a \le \epsilon$ to control the detection probability in the following optimization. Therefore, for a specific congestion pattern $\hat{{\mathscr{C}}}$, the adversary will solve the following optimization problem to move the state estimate to the “center” of the price region $\hat{{\mathscr{C}}}$ and keeping the detection probability low. $$\label{eq:constant}
\begin{tabular}{l l}
$\max_{a\in {\mathscr{A}}, \tilde{\beta} \ge 0}$ & $\tilde{\beta}$ \\
\text{subject to} & $\mathbb{E}[f_{i}] - \tilde{\beta} \ge T_i^{\text{max}},i \in \hat{{\mathscr{C}}}$ \\
& $\mathbb{E}[f_{i}] + \tilde{\beta} < T_j^{\text{max}},j \notin \hat{{\mathscr{C}}}$\\
& $a^{\tiny{\text{T}}} W a \le \epsilon$,\\
\end{tabular}$$ which is a convex program that can be solved easily in practice. We call a region $\hat{{\mathscr{C}}}$ [*feasible*]{} if it makes problem (\[eq:constant\]) feasible.
Among all the feasible congestion patterns, the worst region $\hat{\mathscr{C}}^*$ is chosen as the one giving the largest ARPP. $$\hat{\mathscr{C}}^*=\mbox{arg } \max_{\hat{\mathscr{C}} \in \Gamma} |\tilde{\lambda}_i-\lambda_i(\hat{\mathscr{C}})|,$$ where $\tilde{\lambda}_i$ is the LMP at bus $i$ if the $x_0$ is the system state, and $\Gamma$ the set of all the feasible congestion patterns. Hence, the worst case constant bad data vector is the solution to optimization problem (\[eq:constant\]) by setting the congestion pattern as $\hat{\mathscr{C}}^*$.
Worst ARPP under Partially Adaptive Bad Data
--------------------------------------------
For bad data model M2, only part of the measurement values in real-time are known to the adversary, denoted as $z_{\text{o}}$. The adversary has to first make an estimation of the system state from the observation and prior distribution, then make the attack decision based on the estimation result.
Without the presence of bad data vector, $\ie$ $a = 0$, the system equation (\[eq:DCeq\]) gives $$z_\text{o} = H_{\text{o}} x + w_{\text{o}},$$ where $H_{\text{o}}$ is the rows of $H$ corresponding to the observed measurements and $w_{\text{o}}$ the corresponding part in the measurement noise $w$.
The minimum mean square error (MMSE) estimate of $x$ given $z_{\text{o}}$ is given by the conditional mean $$\mathbb{E}(x|z_{\text{o}})=x_0+\Sigma_x H_{\text{o}}^{\tiny{\text{T}}} (H_{\text{o}} \Sigma_x H_{\text{o}}^{\tiny{\text{T}}})^{-1}(z_{\text{o}}-H_{\text{o}} x_0).$$
Then, the flow estimate on branch $i$ after attack is $$\mathbb{E}[f_{i}|z_{\text{o}}] = F_{i\cdot} \mathbb{E}[\hat{x}|z_{\text{o}}].$$
Still, we want to move the estimation of state to the “center”. On the other hand, the expected measurement value $\mathbb{E}[z_a|z_{\text{o}}] = H\mathbb{E}[\hat{z}|z_{\text{o}}] + a$. Again, we need a pre-designed parameter $\epsilon$ to control the detection probability. Therefore, the solution to the following optimization problem is the best attack given congestion pattern ${\mathscr{A}}$ $$\begin{tabular}{l l}
$\max_{a \in {\mathscr{A}},\tilde{\beta} \ge 0}$ & $\tilde{\beta}$ \\
\text{subject to} & $\mathbb{E}[f_{i}|z_{\text{o}}] - \tilde{\beta} \ge T_i^{\text{max}},i \in \hat{{\mathscr{C}}}$ \\
& $\mathbb{E}[f_{i}|z_{\text{o}}] + \tilde{\beta} < T_j^{\text{max}},j \notin \hat{{\mathscr{C}}}$\\
& $(H\mathbb{E}[z_a|z_{\text{o}}]^{\tiny{\text{T}}}) W (H\mathbb{E}[z_a|z_{\text{o}}]) \le \epsilon$.\\
\end{tabular}
\label{eq:partial}$$
This problem is also a convex optimization problem, which can be easily solved. Among all the $\hat{{\mathscr{C}}}$’s which make the above problem feasible, we choose the one with the largest price perturbation, denoted as $\hat{{\mathscr{C}}}^*$. The solution to problem (\[eq:partial\]) with $\hat{{\mathscr{C}}}^*$ as the congestion pattern is the worst bad data vector.
Worst ARPP under Fully Adaptive Bad Data
----------------------------------------
Finally, we consider the bad data model M3, in which the whole set of measurements $z$ is known to the adversary. The worst bad data vector depends on the value of $z$. Different from the previous two models, with bad data vector $a$, the estimated state is deterministic without uncertainty. In particular $$\hat{x} = K z + K a.$$ And the estimated flow on branch $i$ after attack is also deterministic $$\hat{f}_i = F_{i\cdot} \hat{x} =F_{i\cdot} K z + F_{i\cdot} K a.$$
Similar to the previous two models, congestion pattern is called feasible if there exists some bad data vector $a$ to make the following conditions satisfied: $$\begin{tabular}{l}
$\hat{f}_i \ge T_i^{\text{max}},i \in \hat{{\mathscr{C}}}$ \\
$\hat{f}_i < T_j^{\text{max}},j \notin \hat{{\mathscr{C}}}$\\
$(z+a)^{\tiny{\text{T}}}W(z+a) \le \tau, \quad a \in {\mathscr{A}}$.\\
\end{tabular}
\label{eq:full}$$
Among all the feasible congestion patterns, we choose the one with the largest price perturbation, $\hat{{\mathscr{C}}}^*$. Any bad data vector $a$ satisfying condition (\[eq:full\]) can serve as the worst fully adaptive bad data.
A Greedy Heuristic {#ssec:heuristic}
------------------
[The strategies presented above are based on the exhaustive search over all possible congestion patterns. Such approaches are not scalable for large networks with a large number of possible congestion patterns. We now present a greedy heuristic approach aimed at reducing computation cost. In particular, we develop a gradient like algorithm that searches among a set of likely congestion patterns.]{}
[ First, we restrict ourselves to the set of lines that are close to their respective flow limits and look for bad data that will affect the congestion pattern. The intuition is that it is unlikely that bad data can drive the system state sufficiently far without being detected by the bad data detector. In practice, the cardinality of such a set is usually very small compared with the systems size.]{}
[ Second, we search for the worst data locally by changing one line in the congestion pattern at a time. Specifically, suppose that a congestion pattern is the current candidate for the worst data. Given a set of candidate lines that are prone to congestions, we search locally by flipping one line at a time from the congested state to the un-congested state and vice versa. If no improvement can be made, the algorithm stops. Otherwise, the algorithm updates the current “worst congestion pattern” and continue. The effectiveness of this greedy heuristic is tested in Section \[ssec:sim\_heuristic\]. ]{}
Bad Topology Data on LMP {#sec:bad_topo}
========================
So far, we have considered bad data in the analog measurements. In this section, we include the bad *topology* data, and describe another bad data model.
We represent the network topology by a directed graph ${\mathscr{G}}=({\mathscr{V}},{\mathscr{E}})$ where each $i\in {\mathscr{V}}$ denotes a bus and each $(i,j)\in{\mathscr{E}}$ denotes a *connected* transmission line. For each physical transmission line (a physical line between $i$ and $j$), we assign an arbitrary direction ($(i,j)$) for the line, and $(i,j)$ is in ${\mathscr{E}}$ if and only if bus $i$ and bus $j$ are connected. Bad data may appear in both analog measurements and digital (breaker status) data, as described in Section \[ssec:data\_quality\]: $$\label{eq:both_attack}
\begin{array}{ll}
z_a &= z + a = (Hx + w) + a,~~~~a\in{\mathscr{A}},\\
s_b &= s + b~\text{ (mod 2)},~~~~b\in{\mathscr{B}}.
\end{array}$$
As in Section \[sec:worstcase\], we employ the adversary model to describe the worst case. The adversary alters $s$ to $s_{b}$ by adding $b$ from the set of feasible attack vectors ${\mathscr{B}}\subset \{0,1\}^{l}$ such that the topology processor produces the “target” topology $\bar{{\mathscr{G}}}$ as the topology estimate. In addition, the adversary modifies $z$ by adding $a\in{\mathscr{A}}$ such that $z_a$ looks consistent with $\bar{{\mathscr{G}}}$. In this section, we focus on the worst case when the adversary is able to alter the network topology without changing the state estimate[^5]. We also require that such bad data are generated by an adversary causing undetectable topology change, the bad data escape the system bad data detection. For the worst case analysis, we will maximize the LMP perturbation among the attacks within this specific class. Even though this approach is suboptimal, the simulation results in Section \[sec:simulation\] demonstrate that the resulting LMP perturbation is much greater than the worst case of the bad meter data.
Suppose the adversary wants to mislead the control center with the target topology $\bar{{\mathscr{G}}}=({\mathscr{V}},\bar{{\mathscr{E}}})$, a topology obtained by *removing*[^6] a set of transmission lines ${\mathscr{E}}_{\Delta}$ in ${\mathscr{G}}$ ($\bar{{\mathscr{E}}} = {\mathscr{E}}\setminus{\mathscr{E}}_{\Delta}$). We assume that the system with $\bar{{\mathscr{G}}}$ is observable: the corresponding measurement matrix $\bar{H}$ has full column rank[^7].
Suppose that the adversary changes the breaker status such that the target topology $\bar{{\mathscr{G}}}=({\mathscr{V}},\bar{{\mathscr{E}}})$ is observed at the control center. Simultaneously, if the adversary introduces bad data $a=\bar{H}x-Hx$, then $$z_a =Hx+ a + w = \bar{H}x + w,$$ which means that the meter data received at the control center are completely consistent with the model generated from $\bar{{\mathscr{G}}}$. Thus, any bad data detector will not be effective.
It is of course not obvious how to produce the bad data $a$, especially when the adversary can only modify a limited number of measurements, and it may not have access to the entire state vector $x$. Fortunately, it turns out that $a$ can be generated by observing only a few entries in $z$ without requiring global information (such as the state vector $x$) .
A key observation is that $Hx$ and $\bar{H}x$ differ only in a few entries corresponding to the modified topology (lines in ${\mathscr{E}}_\Delta$) as illustrated in Fig. \[fig:similarity\]. Consider first the noiseless case. Let $z_{ij}$ denote the entry of $z$ corresponding to the flow measurement from $i$ to $j$. As hinted from Fig. \[fig:similarity\], it can be easily seen that $\bar{H}x - Hx$ has the following sparse structure : $$\label{eq:attack_heu}
\bar{H}x - Hx =\displaystyle - \sum_{(i,j)\in {\mathscr{E}}_{\Delta}}\alpha_{ij}m_{(i, j)},$$ where $\alpha_{ij}\in{\mathbb{R}}$ denotes the line flow from $i$ to $j$ when the line is connected and the system state is $x$, and $m_{(i,j)}$ is the column of the measurement-to-branch incidence matrix, that corresponds to $(i,j)$: $m_{(i,j)}$ is an $m$-dimensional vector with $1$ at the entries corresponding to the flow from $i$ to $j$ and the injection at $i$, and $-1$ at the entries for the flow from $j$ to $i$ and the injection at $j$, and $0$ at all other entries. Absence of noise implies that $z_{ij} = \alpha_{ij}$, which leads to $$\label{eq:heuristic_attack}
\bar{H}x - Hx =- \sum_{(i,j)\in {\mathscr{E}}_{\Delta}}z_{ij}m_{(i, j)}.$$
\[c\] \[c\] ![$Hx$ and $\bar{H}x$: Each row is marked by the corresponding meter ($i$ for injection at $i$ and $(i,j)$ for flow from $i$ to $j$).[]{data-label="fig:similarity"}](Similarity "fig:"){width="40.00000%"}
With (\[eq:heuristic\_attack\]) in mind, one can see that setting $a=\bar{H}x - Hx$ and adding $a$ to $z$ is equivalent to the following simple procedure: as described in Fig. \[fig:heuristic\], for each $(i,j)$ in ${\mathscr{E}}_{\Delta}$,
1. Subtract $z_{ij}$ and $z_{ji}$ from $z_{i}$ and $z_{j}$ respectively.
2. Set $z_{ij}$ and $z_{ji}$ to be $0$.
where $z_{i}$ is the entry of $z$ corresponding to the injection measurement at bus $i$.
When measurement noise is present ($z = Hx + w$), the idea of the attack is still the same: to make $a$ approximate $\bar{H}x - Hx$ so that $z_{a}$ is close to $\bar{H}x + w$. Since $z_{ij} = \alpha_{ij} + w_{ij}$, $z_{ij}$ is an unbiased estimate of $\alpha_{ij}$ for each $(i,j)\in{\mathscr{E}}_{\Delta}$, and this implies that $- \sum_{(i,j)\in {\mathscr{E}}_{\Delta}}z_{ij}m_{(i, j)}$ is an unbiased estimate of $- \sum_{(i,j)\in {\mathscr{E}}_{\Delta}}\alpha_{ij}m_{(i, j)} = \bar{H}x - Hx$. Hence, we set $a$ to be $-\sum_{(i,j)\in {\mathscr{E}}_{\Delta}}z_{ij}m_{(i, j)}$, the same as in the noiseless setting, and the attack is executed by the same steps as above.
For launching this attack to modify the topology estimate from ${\mathscr{G}}$ to $\bar{{\mathscr{G}}}$, the adversary should be able to (i) set $b$ such that the topology processor produces $\bar{{\mathscr{G}}}$ instead of ${\mathscr{G}}$ and (ii) observe and modify $z_{ij}$, $z_{ji}$, $z_{i}$, and $z_{j}$ for all $(i,j)\in{\mathscr{E}}_{\Delta}$. The attack is feasible if and only if ${\mathscr{A}}$ and ${\mathscr{B}}$ contain the corresponding attack vectors.
To find the worst case LMP perturbation due to undetectable, state-preserving attacks, let ${\mathscr{F}}$ denote the set of feasible $\bar{{\mathscr{G}}}$s, for which the attack can be launched with ${\mathscr{A}}$ and ${\mathscr{B}}$. Among the feasible targets in ${\mathscr{F}}$, we consider the best target topology that results in the maximum perturbation in real-time LMPs. If ARPP is used as a metric, the best target is chosen as $$\bar{{\mathscr{G}}}^{*}[z]=\mbox{arg } \max_{\bar{{\mathscr{G}}} \in {\mathscr{F}}} \sum_{i}\left|\frac{\lambda_i(z;\bar{{\mathscr{G}}}) - \lambda_i(z;{\mathscr{G}})}{\lambda_i(z;{\mathscr{G}})}\right|. $$ where $\lambda_{i}(z;\bar{{\mathscr{G}}})$ denotes the real-time LMP at bus $i$ when the attack with the target $\bar{{\mathscr{G}}}$ is launched on $z$, and $\lambda_i(z;{\mathscr{G}})$ is the real-time LMP under no attack.
Numerical Results {#sec:simulation}
=================
In this section, we demonstrate the impact of bad data on real-time LMPs with the numerical simulations on IEEE-14 and IEEE-118 systems. We conducted simulations in two different settings: the linear model with the DC state estimator and the nonlinear model with the AC state estimator. The former is usually employed in the literature for the ease of analysis whereas the latter represents the practical state estimator used in the real-world power system. In all simulations, the meter measurements consist of real power injections at all buses and real power flows (both directions) at all branches.
Linear model with DC state estimation {#ssec:sim_linear}
-------------------------------------
We first present the simulation results for the linear model with the DC state estimator. We modeled bus voltage magnitudes and phases as Gaussian random variables with the means equal to the day-ahead dispatched values and small standard deviations. In each Monte Carlo run, we generated a state realization from the statistical model, and the meter measurements were created by the DC model with Gaussian measurement noise. Once the measurements were created, bad data were added in the manners discussed in Section \[sec:worstcase\] and Section \[sec:bad\_topo\]. With the corrupted measurements, the control center executed the DC state estimation and the bad data test with the false alarm probability constraint $0.1$. If the data passed the bad data test, real-time LMPs were evaluated based on the state estimation results. For IEEE-14 and IEEE-118 system, the network parameters[^8] are available in [@IEEEParameter].
We used the number of meter data to be modified by the adversary as the metric for the attack effort. For the 14 bus system, in each Monte Carlo run, we randomly chose two lines, and the adversary was able to modify all the line flow meters on the lines and injection meters located at the ends of the lines. For the 118 bus system, we randomly chose three lines, and the adversary had control over the associated line and injection meters. Both state and topology attacks were set to control the same number of meter data[^9] so that we can fairly compare their impacts on real-time LMPs. [As for the meter data attack, we only considered the lines that are close to their flow limits (estimated flows under M1 and M2, or actual flows under M3) as candidates for congestion pattern search. The threshold is chosen as 10MW in our simulation.]{}
Fig. \[fig:ARPP\_linear\] is the plot of ARPPs[^10] versus detection probabilities of bad data. They show that even when bad data were detected with low probability, ARPPs were large, especially for the fully adaptive bad meter data and the bad topology data.
Comparing ARPPs of the three bad meter data models, we observe that the adversary may significantly improve the perturbation amount by exploiting partial or all real-time meter data (for the partially adaptive case, the adversary observed a half of all meters.) It is worthy to point out that bad topology data result in much greater price perturbation than bad meter data.
Recall the discussion in Section \[sec:realtimeLMP\] and Section \[sec:bad\_topo\] that bad topology data and bad meter data employ different price-perturbing mechanisms: bad topology data perturb real-time LMP by restructuring the price regions without perturbing the state estimate (the line-removal attack introduced in Section \[sec:bad\_topo\] does not perturb state estimate) whereas bad meter data perturb real-time LMP by simply moving the state estimate to a different price region. Therefore, the observation implies that restructuring the price regions has much greater impact on real-time LMP than merely perturbing the state estimate.
Nonlinear model with AC state estimation
----------------------------------------
The simulations with the nonlinear model intend to investigate the vulnerability of the real-world power system to the worst adversarial act, designed based on the linear model. The simulations were conducted on IEEE-14 and IEEE-118 systems in the same manner as the linear case except that we employed the nonlinear model and the AC state estimation.
[Fig. \[fig:ARPP\_nonlinear\] is the plot of ARPPs versus detection probabilities. The result shows that the proposed methodology can affect the system to some extent even when nonlinear estimator is used, especially when the bad data are present in the topology data, although the nonlinear estimator makes this effect relatively less significant compared with the linear case results. ]{}
Performance of the greedy search heuristic {#ssec:sim_heuristic}
------------------------------------------
[We also conducted simulation based on the proposed greedy search technique in Section \[ssec:heuristic\]. The simulation was based on 118 bus system, and all parameters were the same as those presented in Section \[ssec:sim\_linear\]. We compared the performance and computation time of the greedy heuristics with exhaustive search benchmark, as shown in Table \[tab:greedy\]. Notice here the exhaustive search and greedy search are both over the lines that are close to their flow limits (estimated flows under M1 and M2, or actual flows under M3), the same as in Section \[ssec:sim\_linear\]. In Table \[tab:greedy\], the second column (average search time) is the average searching time for worst congestion pattern over 1000 Monte Carlo runs, and the third column (accuracy) is the percentage that the greedy search find the same worst congestion pattern as the exhaustive search. From the result, we can see that using greedy heuristic can give us much faster processing algorithm without losing much of the accuracy.]{}
method average search time accuracy
------------------- --------------------- ----------
exhaustive search 1.23s -
greedy search 0.51s 97.3%
: Performance of greedy search method
\[tab:greedy\]
[ ]{}
Conclusion
==========
We report in this paper a study on impacts of worst data on the real-time market operation. A key result of this paper is the geometric characterization of real-time LMP given in Theorem \[thm:partition\]. This result provides insights into the relation between data and the real-time LMP; it serves as the basis of characterizing impacts of bad data.
Our investigation includes bad data scenarios that arise from both analog meter measurements and digital breaker state data. To this end, we have presented a systematic approach by casting the problem as one involving an adversary injecting malicious data. While such an approach often gives overly conservative analysis, it can be used as a measure of assurance when the impacts based on worst case analysis are deemed acceptable. We note that, because we use adversary attacks as a way to study the worst data, our results have direct implications when cyber-security of smart grid is considered. Given the increasing reliance on information networks, developing effective countermeasures against malicious data attack on the operations of a future smart grid is crucial. See for discussion about countermeasures.
[From a practical viewpoint, our result can serve as the guideline to the real-time operation. Following the methodology in our paper, worst effect of a specific set of meters on real-time LMP can be checked. Once a huge potential perturbation is detected, alarm should be made and the operator needs to check the accuracy of these specific data, add protection devices, or even add more redundant meters.]{}
Although our findings are obtained from academic benchmarks involving relatively small size networks, we believe that the general trend that characterizes the effects of bad data is likely to persist in practical networks of much larger size. In particular, as the network size increases and the number of simultaneous appearance of bad data is limited, the effects of the worst meter data on LMP decrease whereas the effects of the worst topology data stay nonnegligible regardless of the network size. This observation suggests that the bad topology data are potentially more detrimental to the real-time market operation than the bad meter data.
Acknowledgement {#acknowledgement .unnumbered}
===============
The authors wish to acknowledge comments and suggestions from the anonymous reviewers that help to clarify a number of issues and improve the presentation.
[^1]: L. Jia, J. Kim, R. J. Thomas, and L. Tong are with the School of Electrical and Computer Engineering, Cornell University, Ithaca, NY 14853, USA. Email: [(lj92, jk752, rjt1, ltong)@cornell.edu]{}. Part of this work was presented at HICSS 2012 and PES General Meeting 2012.
[^2]: This is equivalent to assuming that the derivative of the optimal value of (\[eq:real-timeLMP\]) with respect to demand at each bus exists
[^3]: In addition to these, the change in topology will affect contingency analysis. Such effect will appear as changes in contingency constraints in real-time LMP calculation (\[eq:real-timeLMP\]) [@Ott:03TPS]. However, dealing with contingency constraints will significantly complicate our analysis and possibly obscure the more direct link between bad data and real-time LMP. Hence, we consider only line congestion constraints in (\[eq:real-timeLMP\]).
[^4]: Notice here both conventional measurements and PMU measurements can be incorporated. Although PMU data seem to have more direct impact on state estimation and real-time LMP calculation, we won’t differentiate the types of measurements in the following discussion.
[^5]: In general, the adversary can design the worst data to affect both the state estimate and network topology. It is, however, much more difficult to make such attack undetectable.
[^6]: Line addition by the adversary is also possible . However, compared to line removal attacks, line addition attacks require the adversary to observe a much larger set of meter measurements to design undetectable attacks. In addition, the number of necessary modifications in breaker data is also much larger: to make a line appear to be connected, the adversary should make all the breakers on the line appear to be closed. Please see [@KimTong:13JSAC] for the detail.
[^7]: Without observability, the system may not proceed to state estimation and real-time pricing. Hence, for the adversary to affect pricing, the system with the target topology has to be observable.
[^8]: In addition to the network parameters given in [@IEEEParameter], we used the following line limit and real-time offer parameters. In the IEEE-14 simulation, the generators at the buses 1, 2, 3, 6, and 8 had capacities 330, 140, 100, 100, and 100 MW and the real-time offers 15, 31, 30, 10, and 20 $\$/\text{MW}$. Lines (2, 3), (4, 5), and (6, 11) had line capacities 50, 50, and 20 MW, and other lines had no line limit. In the IEEE-118 simulation, the generators had generation costs arbitrarily selected from $\{20, 25, 30, 35, 40 ~\$/\text{MW}\}$ and generation capacities arbitrarily selected from $\{200, 250, 300, 350, 400~\text{MW}\}$. Total 16 lines had the line capacities arbitrarily selected from $\{70, 90, 110~\text{MW}\}$, and other lines had no line limit. To handle possible occurrence of price spikes, we set the upper and lower price caps as 500$\$/\text{MW}$ and -100$\$/\text{MW}$ respectively. Total 1000 Monte Carlo runs were executed for each case.
[^9]: Topology attacks need to make few additional modifications on breaker state data such that the target lines appear to be disconnected to the topology processor. However, for simplicity, we do not take into account this additional effort.
[^10]: The detection probabilities for the fully adaptive bad meter data and the bad topology data cases were less than $0.1$ in all the simulations. In the figures, we draw ARPPs of those cases as horizontal lines so that we can compare them with other cases.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study a hybrid system consisting of a narrowband atomic optical resonance and the long-range periodic order of an opaline photonic nanostructure. To this end, we have infiltrated atomic cesium vapor in a thin silica opal photonic crystal. With increasing temperature, the frequencies of the opal’s reflectivity peaks shift down by $> 20\%$ due to chemical reduction of the silica. Simultaneously, the photonic bands and gaps shift relative to the fixed near-infrared cesium D$_1$ transitions. As a result the narrow atomic resonances with high finesse ($\omega/\Delta \omega = 8 \cdot 10^5$) dramatically change shape from a usual dispersive shape at the blue edge of a stop gap, to an inverted dispersion lineshape at the red edge of a stop gap. The lineshape, amplitude, and off-resonance reflectivity are well modeled with a transfer-matrix model that includes the dispersion and absorption of Cs hyperfine transitions and the chemically-reduced opal. An ensemble of atoms in a photonic crystal is an intriguing hybrid system that features narrow defect-like resonances with a strong dispersion, with potential applications in slow light, sensing and optical memory.'
author:
- 'Philip J. Harding'
- 'Pepijn W.H. Pinkse'
- 'Allard P. Mosk'
- 'Willem L. Vos'
date: 'Prepared on Sept 9th, 2014'
title: Nanophotonic hybridization of narrow atomic cesium resonances and photonic stop gaps of opaline nanostructures
---
Introduction
============
There is great and ongoing interest in nanophotonics to create arrays of resonant systems with high quality factors $Q$ in order to control the propagation and emission of light at a deep and fundamental level. New states occur when the state from an individual resonator interacts with the others in the array, such as hybrid collective states or states typical for tight-binding systems [@Ashcroft1976]. It has been predicted that periodic arrays of coupled cavities will form waveguides with unusual and tunable dispersion properties [@Yariv1999]. If a composite quantum resonance is considered that consists of a cavity strongly coupled to a single quantum dot, such a system will exhibit extreme non-linear behavior in response to even a single energy quantum; arrays of such resonating systems have been predicted to exhibit unusual repulsive boson behavior for individual photons [@Hartmann2006; @Greentree2006]. Experimentally, great advances have recently been made in the design and fabrication of high-$Q$ cavities in photonic crystals [@Akahane2003; @Kuramochi2006; @Dharanipathy2013] that may be assembled into arrays of resonators [@Altug2004; @Notomi2008]. It is very difficult, however, to achieve resonant transport in such an array, since all resonators must mutually be tuned to within their linewidth. This entails a tuning to typically one part in a million, which is extremely challenging with the state-of-the-art in nanotechnology. Therefore, we propose a new hybrid system exploiting alkali atoms in the gas phase as high-Q resonators in the photonic crystal environment.
Atomic resonances hold several advantages, firstly they exhibit strong and extremely narrow resonances with relative inverse linewidths - or effective quality factors $Q$ - in the range of $10^7$ and even higher, and secondly their physical properties are extremely well understood. Thirdly and most importantly, all atoms (of the same isotope) are identical, hence they do not need to be individually tuned. This advantage is in contrast to engineered resonant nanostructures - such as cavity arrays - where each individual unit differs from all other ones [@Koenderink2005].
![(color online) Illustration of the concept of our study: We study an atomic vapor with a sharp (dipole-allowed) resonance at frequency $\omega_{\rm atom}$ and concomitant dispersion. The vapor is introduced in a photonic crystal with a photonic gap such as an opal. As a result, we anticipate narrowband resonances inside the photonic gap, shown as the red trough. []{data-label="fig:concept"}](Fig1_Harding.eps "fig:"){width="1.0\columnwidth"}\
To have resonating atoms interact with the optical properties of the host photonic crystal, an ensemble of many atoms must be infiltrated in the crystal, see figure \[fig:concept\]. The relative dielectric function $\epsilon(\omega) = \epsilon'(\omega) + i \epsilon''(\omega)$ of the atomic medium changes drastically around a resonance frequency $\omega_{a}$, $\epsilon'(\omega)$ exceeding unity below resonance, and being below unity above the resonant frequency $\omega_a$. To achieve $\epsilon'(\omega) < 1$, the atomic transitions must be isolated. For this reason, we have chosen to study alkali atoms. It is desirable for the atomic medium to be dense, in order to have $\epsilon'(\omega)$ change considerably near the resonance frequency $\omega_{\rm atom}$. Therefore, we have chosen cesium, as it has the highest vapor density. Since the imaginary part of $\epsilon(\omega)$ - associated with extinction - is related to the dispersion by the Kramers-Kronig relations [@Bohren1983], $\epsilon'(\omega)$ can considerably differ from unity, while $\epsilon''(\omega) << 1$ for a select frequency range outside resonance, see Fig. \[fig:concept\]. Thus, atomic systems have a large modulation of $\epsilon(\omega)$, which warrants a study of their effect on photonic materials.
It has already been realized that hybrid atom-photonic systems offer exciting prospects: The creation of new modes in the photonic band gap has been predicted [@Coevorden1996; @Sivachenko2001], as well as induced transparency [@Ye2008; @Tidstrom2010], nonlinearities [@Soljacic2004], and modified pulse propagation [@Camacho2007]. Previously, only few experiments have been reported of composite systems of photonic crystals with embedded resonant media. It was observed that a strongly absorbing dye in a photonic crystal leads to intriguing braggoriton resonances [@Eradat2002]. Unfortunately, however, the linewidth of dye is limited by quenching interactions with other dye molecules that increase with density [@Imhof1999], thus strongly limiting resonant effects. While semiconductor quantum dots are being pursued as internal probes inside photonic crystals, their broad inhomogeneous width (e.g., $5 \%$ for CdSe [@Norris1996]) precludes large changes in $\epsilon'(\omega)$. To reduce inhomogeneous width effects, van Coevorden [*et al.*]{} theoretically studied a photonic crystal consisting of cold atoms on a lattice [@Coevorden1996]. Such systems have indeed been studied, although the low lattice site occupancy of $<1 \%$ limited the atomic density in the pioneering studies [@Weidemuller1995; @Birkl1995]. A 3D band gap is predicted to open when the resonance wavelength equals the lattice spacing [@Coevorden1996]. The required condition of one atom per lattice site has been realized [@Greiner2002], although photonic crystal properties have not been reported to date. Recently, sub-Doppler features were reported in the spectra of Cs infiltrated in opal nanostructures [@Ballin2013], and a preliminary study of photonic crystals infiltrated with resonant alkali atoms has been reported [@Harding2008]. Here we report on an experiment in which we infiltrate an alkali gas - a saturated cesium vapor - in a thin opal photonic crystal.
We have decided to study a dense and hot vapor of $^{133}$Cs as a strongly polarizable resonant medium, as it has only one electron in the outer shell in the $6^2S_{1/2}$ state. The well-known D-transition from the ground state to the $6P$ manifold is strong, in particular the D$_1$ transitions of $^{133}$Cs with a frequency of 335.116 THz (wavelength $\lambda = 895$ nm) that consist of four isolated hyperfine transitions. The excited state has a narrow intrinsic linewidth with $\Gamma_0 = 2 \pi \times 4.56$ MHz, and a large polarizability at resonance [@Young1994; @Rafac1998; @Rafac1999]. In the ground state, the well-known hyperfine splitting of 9.193 GHz defines the unit of time.
Experimental section
====================
![(color online) The heart of the experimental setup is the cesium cell which consists of a Cs reservoir and the measuring cell which contains an opal grown on a vacuum viewport. These two cells are situated in a large vacuum chamber (not shown) to prevent convection. Valves V$_1$ and V$_2$ connect the reservoir with the cell and the cell with the pump, respectively. []{data-label="fig:setup"}](Fig2_Harding.eps "fig:"){width="1.0\columnwidth"}\
The need for a high vapor density requires heating of liquid Cs. Therefore, a cell was built in which Cs can both be heated as well as spatially confined, see Fig. \[fig:setup\]. Heating is accomplished by two heating elements (thermocoax) brazed at the top and the bottom into the cylindrical reservoir. Upon the opening of valve V$_1$ (Swagelok SS-4BG-V51, heated), hot Cs vapor diffuses into the experimental cell, which is also heated by two heating elements. Because Cs reacts strongly with both oxygen and water, the reservoir and experimental cell are kept in a vacuum of $<10\,\mu$bar, that is detection limited by the low conductance (0.08l/s) of the valve and the thin tube. To remove moisture and oxygen, the assembly is baked and evacuated at $150\,{^{\circ}}$C via valve V$_2$ by a turbo-molecular pump for at least 24 hours before admitting Cs. Both the reservoir and the experimental cell are independently temperature controlled. The temperature is measured by ring J-type thermocouples. Heating tape is wound around the tubes and valves to prevent cold spots that would otherwise determine the vapor pressure of Cs. The experimental cell, the valves, and the tubes are kept at a temperature 30 to 50K above that of the reservoir to prevent condensation. The experimental cell is placed in a vacuum tank kept at a pressure below $10^{-3}\,$mbar to prevent convection and thus stabilize the temperature.
To avoid light absorption from Cs vapor between the photonic crystal and the window, [SiO$_2$]{} opals were grown directly on the viewport windows by vertical controlled drying [@Jiang1999; @Hartsuiker2008]. To this end, the viewports (VacGen, UK) are cleaned overnight in a solution of 30g NaOH, 30ml of ultrapure water, and 200ml of ethanol. All glassware is rinsed in deionized water and in ethanol, and dried in a stream of nitrogen. An ethanol suspension of 0.1 to 0.5 vol% [SiO$_2$]{} colloidal nanospheres with a diameter $D = 460$ nm is placed in a vial. The viewport is inserted at an angle between $40{^{\circ}}$ and $60{^{\circ}}$ into the vial, which is heated in an oven between $30{^{\circ}}$C and $60{^{\circ}}$C. Upon evaporation of the suspension, the colloids move towards the area of fastest evaporation, which is the liquid meniscus, and are deposited on the substrate. The structure of such an array is close packed, typically face-centered cubic (fcc) [@Miguez1997; @Megens2001]. As the evaporation proceeds, ordered arrays of spheres grow until finally there is no suspension left and the growth stops [@Jiang1999; @Hartsuiker2008]. Despite consistently scanning the parameter space in temperature, angle, and colloid density, the opals on the viewports were of moderate quality, with a reflectivity of $2 {R_{\rm glass}}\simeq 10 \%$, where ${R_{\rm glass}}$ is the reflectivity of viewport window. In contrast, similar opals grown on microscope slides showed peaks in excess of $13 {R_{\rm glass}}\simeq 55\%$. Thus the quality of the opals was limited by the surface quality of the viewports. From optical microscope images, the structures were estimated to be between 1 and 3 lattice spacings (i.e., fcc 111 spacings) thick. Since we observed greenish opalescence from the deposited nanostructures, likely related to second-order Bragg diffraction (see below), we refer to such a structure as opal.
The light for the high-resolution measurements is focused at an angle of $\theta = (27\pm2){^{\circ}}$ onto the experimental cell, as shown in Fig. \[fig:setup\]. At this angle the stop bands of the opal are blue shifted to match with the Cs resonance. Moreover, this procedure avoids the 2mm thick window to act as a Fabry-Pérot etalon, with a fringe spacing of $1/(2\times 0.2\,{\rm cm}\ n_{\rm glass}) = 1.7\,$cm$^{-1}$, or 50GHz, which is comparable to the scan range of the laser.
![(color online) Optical setup with the optical paths indicated in red. The narrowband tunable laser (MBR) emits linearly polarized light that is modulated by a chopper. The laser power is set by polarizing beamsplitters (PBS) and a $\lambda/2$ waveplate (WP), and is monitored by diode D$_2$. Light reflected by the sample in the experimental cell is detected by a photomultiplier tube (PMT) and lock-in amplifier. To determine the optical frequency of the laser, a coarse measurement is performed with a wavemeter, and a high-resolution frequency is measured by saturated absorption spectroscopy in a Cs cell with diode D$_1$. Light from a white halogen lamp is used to measure the broadband optical properties of the sample. The sample cell is placed in an evacuated tank (volume $0.037\,$m$^{3}$) to minimize air convection around all heated elements. []{data-label="fig:setup_large"}](Fig3_Harding.eps "fig:"){width="1.0\columnwidth"}\
Light for the high-resolution experiments is generated by a narrowband single-mode cw laser (Coherent MBR-110), pumped by a frequency doubled Nd:YAG laser (Coherent Verdi V-10), see Fig. \[fig:setup\_large\]. The laser can be scanned over 40GHz near $\lambda = 895$ nm, sufficient to excite all four Cs hyperfine transitions in one scan. The laser light is linearly polarized, modulated by a chopper, and the power is set by polarization means. The beam is focused by an achromatic lens ($f=300\,$mm) at an angle of $\theta = (27\pm2){^{\circ}}$ onto the experimental cell containing the cesium vapor and the opal. The elastically reflected light is collected by a photomultiplier tube (Hamamatsu R928). The photomultiplier is essentially insensitive to fluorescence from the sample due to its very small solid angle of $\Omega\approx 10^{-6}\,$sr. The photocurrent of the photomultiplier is measured by phase-sensitive detection. The laser frequency is coarsely monitored by a wavemeter (Burleigh WA-10L) with a resolution of 0.1cm$^{-1}$. The precise frequency is measured by Doppler-free spectroscopy [@Demtroder] with a resolution of 26 MHz using a Cs reference cell [@Harding2008]. Broadband photonic crystal properties are probed with a spatially-filtered halogen white light source and the reflectivity is measured with a 2048-channel spectrometer (Ocean Optics, USB 2000) in the visible and near infrared range.
Results
=======
Opal features
-------------
![(a) Reflectivity spectrum of an SiO$_2$ opal grown on the viewport normalized to the reflectance $R_{\rm glass}$ of the viewport, at oblique incidence ($27 \pm 2 ^{\circ}$, internally $20 \pm 2 ^{\circ}$). The first peak at $10800$ cm$^{-1}$ is identified with the first-order (111) stop gap of a SiO$_2$ opal. The second reflectivity peak is attributed to the range of flat bands and higher-order stop gaps. The horizontal dashed line indicates the resonance frequency of the Cs D$_1$ transition. (b) Bandstructure from U to L in the irreducible part of the Brillouin zone of an fcc opal photonic crystal. The gray horizontal bar indicates the forbidden gap at an angle of $20 \pm 2^\circ$ inside the crystal. []{data-label="fig:Reflectivity_and_bandstructure"}](Fig4_Harding.eps "fig:"){width="\columnwidth"}\
Figure \[fig:Reflectivity\_and\_bandstructure\](a) shows the broadband reflectivity spectrum of a thin opal grown on a viewport. A broad peak is observed at $10800$ cm$^{-1}$ (324THz). The large relative bandwidth of 48 $\%$ is attributed to finite-size effects of the thin opal [@Bertone1999], also in view of the low reflectivity. At $11178$ cm$^{-1}$, the Cs D$_1$ resonance is indicated, which is slightly blue shifted compared to the first stop band, as aimed for. A second high-frequency peak is observed near $18500$ cm$^{-1}$ or 555THz. The high-frequency peaks seems to be split into 3 separate peaks that likely originate from multiple stop-gap interferences and concomitant flat bands, observed earlier in opals and inverse opals [@Vos2000; @Romanov2001; @Galisteo-Lopez2004]. The bandstructure from U to L for a SiO$_2$ opal (n’ = 1.45) is shown in Fig. \[fig:Reflectivity\_and\_bandstructure\](b). The peak at $10800$ cm$^{-1}$ corresponds well to the frequency range of the directional forbidden stop gap at an angle of $20 \pm 2^\circ$ inside the crystal. The broad high-frequency peak at $18500$ cm$^{-1}$ corresponds well with the range of second-order Bragg diffraction that includes complex stop gaps and flat bands to which external light cannot couple.
![The stopbands of the opal shift with the increasing temperature because of the progressing reduction of the [SiO$_2$]{}. On the left the reflection of the opal is shown as a function of probe frequency for four different temperatures, where the first line of spectra is a reference at room temperature without Cs. The dashed arrow indicates the red-shifting second-order stop band. On the right, the corresponding transfer-matrix calculations are shown with the used index of the opal material. The inset is an extreme zoom in into the region where Cs resonances are expected. The temperature was increased by 10 to 20$\,{^{\circ}}$C per hour. []{data-label="StopbandsFig"}](Fig5_Harding.eps "fig:"){width="\columnwidth"}\
Broadband reflectivity spectra are shown in Fig. \[StopbandsFig\] for increasing temperature and thus cesium vapor pressure. At first contact with Cs (not shown), even before spectra are observed, the first-order opal peak has shifted down from 324THz to 299THz (9.5% relative shift), and the second peak has shifted from 555THz to 489THz (12%). At a cell temperature of $T = 370$ K, the first-order peak has already shifted out of the instrumental spectral range, and the second peak has shifted down to 460THz. At a temperature of $T = 400$ K, the second reflectivity peak has further shifted to 447THz, or 20% down compared to the starting state. At higher temperatures the peaks cannot be identified unambiguously. The overall reflectivity decreases with increasing temperature, which is caused by a thin oil film that settled on the optics in the vacuum tank. The shifting of the reflectivity peaks is attributed to the chemical reduction of the [SiO$_2$]{} nanospheres [@Jahier2001]; since Cs is strongly caustic, it reduces [SiO$_2$]{} to [SiO$_x$]{} [@Pascal1958]. If $x$ varies from $2$ to $0$, the refractive index of the nanospheres will increase from $n^\prime = 1.45$ for [SiO$_2$]{} to at most $n^\prime = 3.5$ in case of pure Si [@Andalkar2002]. If a homogeneous reduction to SiO occurs, the refractive index of the nanospheres will be equal to that of SiO ($n_{\rm SiO} = 2.0\pm0.1$). We surmise that the colloidal nanospheres consist of a mixture of [SiO$_x$]{} with different $x$, possibly mixed with Cs oxides. By substituting the increased average refractive index of the opal layer into Bragg’s law [@Vos1996], the peak shifts to a lower frequency.
{width="2.0\columnwidth"}
To get more insight in the shape and the shifting of the observed resonances, a transfer-matrix calculation was performed on a single crystal layer of nanospheres surrounded by an atomic vapor, and probed at an angle of $27{^{\circ}}$. At the right hand side of Fig. \[StopbandsFig\], a corresponding transfer-matrix calculation of a single crystal layer is shown. The nanosphere’s refractive index is chosen to give the best overal match. For the reference spectrum, before contact with Cs, the agreement is reasonable (first row). The D$_1$ resonance is just blue of the first peak. On first contact with Cs, both the first and second peak red shift (not shown). At $T = 370\,$K, the second peak has shifted to 460THz, which is matched by the calculation for a nanosphere refractive index $n = 1.57$. The D$_1$ resonance is at the blue foot of the first-order peak. At $T = 400$K, the D$_1$ resonance is exactly in the trough between two peaks, obtained for $n = 1.65$. As the temperature is increased even further, $n = 2.15$ is fitted, shifting the second peak so that the D$_1$ resonance is now on its red side.
Atomic features
---------------
We see that the increase in cesium density (moderated by the increasing temperature) and progression in time increase the nanosphere refractive index. As a result, the opal reflectivity peaks are tuned with respect to the atomic resonances. Compared to the broadband reflectivity, we increased the spectral resolution by $10^4$ to resolve the Cs hyperfine structure. Therefore, intermixed with the measurements of the broadband spectra we have scanned the narrowband excitation laser and recorded the reflected intensity, as displayed in Fig. \[fig:ExplanationFig\], that zoom in on the Cs D$_1$ transitions. At a low Cs density (at $370$ K), the background reflection of the opal is low and the Cs resonances appear as red-shifted peaks in reflectivity. To determine the linewidth, we modeled the peaks with a Gaussian function and obtain about $\Gamma = 2\pi\times 0.4$ GHz FWHM. From earlier work [@Harding2008], we deduced that the peaks are both collision and Doppler broadened. If we consider the atoms to behave as tiny resonators, the width correspond to an effective quality factor $Q=\omega/\Delta\omega=8\times10^{5}$. While similar - or even higher - quality factors have been reported [@Akahane2003; @Kuramochi2006; @Dharanipathy2013], the advantageous feature of the Cs atoms studied here is that their center frequency varies by less than 1 part in $840.000$ (= 335 THz/0.4 GHz), which is beyond state-of-the-art in nanocavity fabrication. Therefore, atoms warrant attention as resonant systems in nanophotonic arrays.
Figure \[fig:ExplanationFig\](a) shows measured high-resolution reflectivity spectra at four moments in time while the temperature and hence the Cs density was raised. The two hyperfine resonances $F=4\rightarrow F^\prime=3$ and $F=4\rightarrow F^\prime=4$ are clearly observed. At $T = 370$ K, the background reflectivity of the opal is around $0.065 R_{\rm{glass}}$, or $0.26\%$. Close to the two resonances, the reflectivity shows two clear peaks. Below the $4 \rightarrow 3'$ transition frequency, the spectrum shows a short tail with a trough, and beyond the resonances the spectrum shows a broad and asymmetric tail. When the second spectrum is taken, both peaks have become more pronounced. The low-frequency tail has expanded, and the low-frequency trough has disappeared. The background reflectivity of the opal has decreased to 0.013. The third spectrum has dramatically changed: after the first peak a trough has appeared that clearly dives below the opal’s background. Similarly, a second trough with a minimum below the background has appeared at the second resonance. The initial trough below resonance has completely disappeared, and the baseline has recovered to 0.02. The last spectrum, at $T = 420$ K, shows another striking change: the peaks seen at the earlier spectra have completely vanished and given way to two marked troughs. The spectra is completely without peaks, exactly the opposite of the second spectrum. Moreover, the opal background has recovered even further to 0.05.
Our observations can be qualitatively understood by considering a polarizable atomic medium inside a photonic structure, whose reflectivity peaks shift due to a changing opal refractive index. These features have been incorporated in our transfer-matrix model, whose results are shown in Fig. \[fig:ExplanationFig\](b). We consider the real and imaginary parts of the atomic refractive index $n^\prime$ and $n^{\prime\prime}$, respectively, and their frequency relative to that of the opal’s stopband (see Fig. \[fig:ExplanationFig\]). We distinguish 4 characteristic cases as illustrated in Fig. \[fig:ExplanationFig\](c).
1. [If the atomic resonance is at the blue edge of the reflectivity peak, an increase in $n^\prime$ from unity below resonance will effectivly cause the stop gap to red shift at that frequency. A red shift at the blue edge of a peak reduces the reflectivity. Above resonance, the opposite happens since $n^\prime$ is below unity. The behavior of $n^\prime$ versus frequency across the resonance thus results in an decreasing and increased reflectivity, as seen at $T = 370$ K. On resonance, the enhanced $n^{\prime\prime}$ will destroy the destructive interference, thereby increasing the reflectivity. ]{}
2. [For resonances at the trough between two opal reflectivity peaks, the increase in $n^{\prime\prime}$ on resonance will dominate the reflectivity and cause the two peaks, as seen on the second row where we fitted n’=1.58. The gradient of the opal’s reflectivity vanishes, and any change in $n^\prime$ will hardly change $R(\omega)$, since the opal’s reflectivity peaks will be little shifted. ]{}
3. [At the red edge of the second opal peak, the increase in $n^\prime$ below resonance will red shift the opal peak. As a result the reflectivity will increase. Above resonance, the decrease in $n^\prime$ will blue shift the opal peak, resulting in a decrease of the reflectivity. Hence, the measured reflectivity has a positive correlation with $n^\prime$. The result of increase in $n^{\prime\prime}$ depends on whether the resonance is at the foot of the opal reflectivity peak or near the maximum: in the former case, the reflectivity increases, while in the latter case the constructive Bragg interference is suppressed, resulting in a reflectivity trough. ]{}
4. [An atomic resonance near the center of the reflectivity peak will lead to following behavior: Since the gradient of the reflectivity peak is small at its maximum, the induced shift by the change in $n^\prime$ will have little effect on the opal’s reflectivity. In contrast, the absorption expressed by the $n^{\prime\prime}$ will remove the constructive Bragg interference. Hence an increase in $n^{\prime\prime}$ will correlate with a trough at resonance, as seen in the last spectrum, at $T = 420$ K. ]{}
It appears that the resonance frequencies in our experiments are systematically lower from calculated ones. The difference is attributed to a pressure shift of remnant N$_2$ in the reservoir. Using the shift due to a [N$_2$]{} buffer gas of -10.97MHz/mbar from Ref. [@Andalkar2002], a partial pressure of 27mbar explains the difference. This partial pressure agrees excellently with a measurement of $27\pm10\,$mbar at evacuation after the experiment.
From the good match between the measured spectra and the theory curves we conclude that the system is well described by an atomic vapor filling the voids around the opal spheres. Moreover, from the agreement we deduce that the real part of the refractive index must have been less than unity for probe frequencies above resonance. To the best of our knowledge, this is the first observation of sub-unity refractive index in a photonic crystal, inside a stop band.
Discussion
==========
The striking spectral features reported above reveal that a hybridization occurs between the optical properties of the cesium atoms and the opal structure. The result of the hybridization depends strongly on the tuning of the atomic resonances relative to the opal’s gap structure. The simple case anticipated in Fig. \[fig:concept\] occurs when the atomic resonances are centred in the gap. This case corresponds to the observations at $T = 420$ K in Fig. \[fig:ExplanationFig\]. In addition, intriguing hybridizations are observed when the atomic resonances are tuned to either the red or the blue edge of a stop gap. In these cases, the real parts of the atomic dispersion come into play, leading to asymmetric reflectivity features in the observed spectra. This situation differs from arrays of cavities in photonic crystals. In the latter case, the cavities must necessarily be tuned deep within the gap, in order to profit from the confinement offered by the gaps. Conversely, a cavity resonance tuned to the edge of a gap will obviously be lossy, since the cavity resonance is derived from the surrounding photonic crystal, in contrast to an atomic resonance, that obviously also exists outside a photonic crystal.
In contrast to the one-dimensional (1D) cases of long hollow optical fibers [@Muller2000; @Ghosh2006; @Light2007; @Slepkov2008] or waveguides on a chip [@Yang2007], three-dimensional (3D) photonic crystals have the demonstrated potential to significantly alter the local density of states (LDOS) [@Lodahl2004; @Noda2007; @Leistikow2011], allowing to control the excited-state lifetime of emitters more in 3D [@Vats2002] than in 1D [@Hooijer1999]. Our results demonstrate that the required high atomic densities can be achieved. Several improvements merit investigation in future studies. First of all, other combinations of gases and photonic crystals will have to be more stable against chemical interactions. Another option is to coat the opal with a thin protective layer or to use light-induced adsorption to increase the alkali pressure. Once these issues are resolved, we can investigate the influence of the LDOS on the lifetime of excited states. The high-$Q$ resonances in a photonic crystal could even allow to study the thermalization of a photon gas [@Klaers2010]. Finally, our experiment could offer an alternative experimental route to complex experiments that are being done with atomic lattices [@Bloch2008].
The measured peaks are somewhat broader than those calculated with our model. The model does not take into account collisional broadening or transit time broadening. While collisional self-broadening [@Chen1968] does not play a significant role at our experimental conditions, the N$_2$ background broadens a delta-function line to about 400MHz FWHM [@Andalkar2002]. In addition, the atoms between the opal spheres will have only a small interaction time with the light before hitting a material wall or before being outside the optical focus, leading to transit-time broadening [@Demtroder]. Since the magnitude strongly depends on geometry, it is hard to estimate, yet it could easily amount to a few 100MHz. In view of these broadening mechanisms, the observed resonances are surprisingly narrow, suggestive of selective reflection effects [@SelectivereflectionLit; @Wang1997]. In selective reflection, slow atoms with only a small velocity component parallel to the propagation direction of light contribute disproportionately to the signal, leading to sharp spectral features. The prospect of exploiting selective reflection to yield even sharper features in future makes this system even more attractive.
Conclusions
===========
We report on an experiment in which we infiltrate a dense alkali gas - a saturated cesium vapor - in a thin opal photonic crystal. The stop bands of the nanostructure are tuned to overlap with the narrow atomic resonances of the alkali atoms, leading to an intriguing hybridization of the atomic resonances and the crystal spectrum. As the temperature is increased, we observe a shift of the peak opal reflectivity frequency in excess of $20\,\%$, which is attributed to the chemical reduction of the silica by the cesium. As a result, the frequency of the photonic bands are scanned through the near-infrared Cs D$_1$-transition. Simultaneously, the Cs resonances undergo marked changes in strength, off-resonance reflectivity, and lineshape. When the atomic resonance is tuned to the red edge of the stop gap, the reflectivity reveals a striking inversely dispersive shape. We successfully interpret all observed spectral features by using a transfer-matrix model that includes the dispersion and absorption of two of the hyperfine transitions of Cs, as well as the ongoing change of the opal’s refractive index due to the chemical reduction. Our study is a step forward in the understanding of strongly resonant systems inside photonic crystals, and shows that a hybrid combination of atomic and condensed matter holds interesting physics.
Acknowledgments
===============
It is a pleasure to thank Cock Harteveld, and Leon Woldering for expert technical support. We thank Javier Garc[í]{}a de Abajo, Tom Hijmans, Ad Lagendijk, and Gerhard Rempe for encouragements on the project. This work is part of the research program of the ’Stichting voor Fundamenteel Onderzoek der Materie’ (FOM) that is financially supported by the ’Nederlandse Organisatie voor Wetenschappelijk Onderzoek’ (NWO), notably the PPOM program and “Inrichting leerstoelpositie” grant (02ILP012). We also thank ERC (grant no 279248), NWO (Vici), and STW for support.
[17]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
N. W. Ashcroft and N. D. Mermin, *Solid state physics* (Holt Rinehart Winston, New York, 1976).
A. Yariv, Y. Xu, R. K. Lee, and A. Scherer, Opt. Lett. **24**, 711 (1999). M. J. Hartmann, F. G. S. L. Brandao, and M. B. Plenio, Nat. Phys. **2**, 849 (2006). A. D. Greentree, C. Tahan, J. H. Cole, and L. C. L. Hollenberg, Nat. Phys. **2**, 856 (2006) Y. Akahane, T. Asano, B.-S. Song, and S. Noda, Nature (London) **425**, 944 (2003).
E. Kuramochi, M. Notomi, S. Mitsugi, A. Shinya, T. Tanabe, and T. Watanabe, Appl. Phys. Lett. **88**, 041112 (2006). U. P. Dharanipathy, M. Minkov, M. Tonin, V. Savona, and R. Houdr[é]{} http://arxiv.org/abs/1311.0997 H. Altug and J. Vučković, Appl. Phys. Lett. **84**, 161 (2004). M. Notomi, E. Kuramochi, and T. Tanabe, Nature Photon. **2**, 741 (2008). A. F. Koenderink, A. Lagendijk, and W. L. Vos, Phys. Rev. B **72**, 153102 (2005). Y. Takahashi, H. Hagino, Y. Tanaka, B.-S. Song, T. Asano, and S. Noda, Opt. Expr. **15**. 17206 (2007). C. F. Bohren and D. R. Huffman, *Absorption and Scattering of Light by Small Particles* (Wiley, New York, 1983).
D. V. van Coevorden, R. Sprik, A. Tip, and A. Lagendijk, Phys. Rev. Lett. **77**, 2412 (1996). A. Y. Sivachenko, M. E. Raikh, and Z. V. Vardeny, Phys. Rev. A **64**, 013809 (2001). Y. Q. Ye, J. Q. Shen, and Y. Jin, Appl. Phys. A **93**, 505 (2008). J. Tidström, C. W. Neff, and L. M. Andersson, J. Opt. **12** 035105 (2010). M. Soljacic and J. D. Joannopoulos, Nat. Mater. **1**, 3 (2004). R. M. Camacho, M. V. Pack, J. C. Howell, A. Schweinsberg, and R. W. Boyd, Phys. Rev. Lett. **98**, 153601 (2007). N. Eradat, A. Y. Sivachenko, M. E. Raikh, Z. V. Vardeny, A. A. Zakhidov, and R. H. Baughman, Appl. Phys. Lett. **80**, 3491(2002). A. Imhof, M. Megens, J. J. Engelberts, D. T. N. de Lang, R. Sprik, and W. L. Vos, J. Phys. Chem. B **103**, 1408 (1999). D. J. Norris and M. G. Bawendi, Phys. Rev. B **53**, 16338 (1996). M. Weidemüller, A. Hemmerich, A. Görlitz, T. Esslinger, and T. W. Hänsch, Phys. Rev. Lett. **75**, 4583 (1995). G. Birkl, M. Gatzke, I. H. Deutsch, S. L. Rolston, and W. D. Phillips, Phys. Rev. Lett. **75**, 2823 (1995). M. Greiner, O. Mandel, T. Esslinger, T. W. Hänsch, and I. Bloch, Nature (London) **415**, 39 (2002). P. Ballin, E. Moufarej, I. Maurin, A. Laliotis, and D. Bloch, Appl. Phys. Lett. **102**, 231115 (2013). P. J. Harding, [*Photonic crystals modified by optically resonant systems*]{} (Ph.D. thesis, University of Twente, 2008), available from www.photonicbandgaps.com.
L. Young, W. T. Hill III, S. J. Sibener, S. D. Price, C. E. Tanner, C. E. Wieman, and S. R. Leone, Phys. Rev. A **50**, 2174 (1994). R. J. Rafac and C. E. Tanner, Phys. Rev. A **58**, 1087 (1998). R. J. Rafac, C. E. Tanner, A. E. Livingston, and H. G. Berry, Phys. Rev. A **60**, 3648 (1999). H. Miguez, F. Meseguer, C. López, A. Mifsud, J.S. Moya, and L. Vazquez, Langmuir **13**, 6009 (1997). M. Megens and W.L. Vos, Phys. Rev. Lett. **86**, 4855 (2001). P. Jiang, J. Bertone, K. Hwang, and V. Colvin, Chem. Mater. **11**, 2132 (1999). A. Hartsuiker and W. L. Vos, Langmuir **24**, 4670 (2008). J. F. Bertone, P. Jiang, K. S. Hwang, D. M. Mittleman, and V. L. Colvin, Phys. Rev. Lett. **83**, 300 (1999)
W. L. Vos and H. M. van Driel, Phys. Lett. A **272**, 101 (2000). S. G. Romanov, T. Maka, C. M. Sotomayor Torres, M. Müller, R. Zentel, D. Cassagne, J. Manzanares-Martinez, and C. Jouanin, Phys. Rev. E **63**, 056603 (2001). J. F. Galisteo-López and C. López, Phys. Rev. B **70**, 035108 (2004). E. Jahier, J. Guéna, P. Jacquier, M. Lintz, and M.A. Bouchiat, Europhys. J. D **13**, 221 (2001).
P. Pascal [*Nouveau traité de chimie minérale*]{}, Tome III (Masson, Paris, 1958).
A. Andalkar and R. B. Warrington, Phys. Rev. A **65**, 032708 (2002). W. L. Vos, R. Sprik, A. van Blaaderen, A. Imhof, A. Lagendijk, and G. H. Wegdam, Phys. Rev. B. **53**, 16231 (1996). C. L. Chen and A. V. Phelps, Phys. Rev. **173**, 62 (1968). W. Demtröder, [*Laser Spectroscopy*]{}, (Springer, Berlin, 2003), 3rd ed.
J. P. Woerdman and M. F. H. Schuurmans, Opt. Commun. **14**, 248 (1975); M. F. H. Schuurmans, J. Phys. (France) **37**, 469 (1976); A. L. J. Burgmans and J. P. Woerdman, J. Phys. (France) **37**, 677 (1976); G. Nienhuis, F. Schuller, and M. Ducloy, Phys. Rev. A, **38**, 5197 (1988); M. Ducloy and M. Fichet, J. Phys. (France) II **1**, 1429 (1991); M. Chevroller, M. Fichet, M. Oria, G. Rafmat, D. Bloch, and M. Ducloy, J. Phys. (France) II **2**, 631 (1992). P. Wang, A. Gallagher, and J. Cooper, Phys. Rev. A **56** 1598 (1997). D. Müller, E. A. Cornell, D. Z. Anderson, and E. R. I. Abraham, Phys. Rev. A, **61**, 033411 (2000). S. Ghosh, A. R. Bhagwat, C. K. Renshaw, S. Goh, A. L. Gaeta, and B. J. Kirby, Phys. Rev. Lett. **97**, 023603 (2006). P. S. Light, F. Benabid, F. Couny, M. Maric, and A. N. Luiten, Opt. Lett. **32**, 1323 (2007). A. D. Slepkov, A. R. Bhagwat, V. Venkataraman, P. Londero, and A. L. Gaeta, Opt. Expr. **16**, 18976 (2008). W. Yang, D. B. Conkey, B. Wu, D. Yin, A. R. Hawkins, and H. Schmidt, Nat. Photon. **1**, 331 (2007). P. Lodahl, A. F. van Driel, I. S. Nikolaev, A. Irman, K. Overgaag, D. Vanmaekelberg, and W. L. Vos, Nature (London) **430**, 654 (2004). S. Noda, M. Fujita, and T. Asano, Nat. Photon. **1**, 449 (2007).
M. D. Leistikow, A. P. Mosk, E. Yeganegi, S. R. Huisman, A. Lagendijk, and W. L. Vos, Phys. Rev. Lett. **107**, 193903 (2011).
N. Vats, S. John, and K. Busch, Phys. Rev. A **65**, 043808 (2002). C. Hooijer, D. Lenstra, and A. Lagendijk, Opt. Lett. **25**, 22, 1666 (2000) J. Klaers, F. Vewinger, and M. Weitz, Nat. Phys. **6**, 512 (2010). I. Bloch, J. Dalibard, and W. Zwerger, Rev. Mod. Phys. **80**, 885 (2008).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show that the decay mode ${B \to D \tau \nu_{\tau}}$ is competitive with and complementary to ${B \to \tau \nu_\tau}$ in the search for charged–Higgs effects. Updating the relevant form factors, we find that the differential distribution in the decay chain $\bar{B}\to D\bar{\nu}_{\tau}\tau^-[\to\pi^-\nu_{\tau}]$ excellently discriminates between Standard–Model and charged–Higgs contributions. By measuring the $D$ and $\pi^-$ energies and the angle between the $D$ and $\pi^-$ three-momenta one can determine the effective charged–Higgs coupling including a possible CP–violating phase.'
author:
- Ulrich Nierste
- Stéphanie Trine
- Susanne Westhoff
title: |
[TTP08-06SFB/CPP-08-11]{}\
Charged–Higgs effects in a new $B\to D \tau \nu_{\tau}$ differential decay distribution
---
Introduction
============
The $B$ factories BABAR and BELLE have accumulated enough statistics to probe extensions of the Higgs sector of the Standard Model. Notably, the decay ${B^+ \to \tau^+ \nu_\tau}$ allows us to place useful constraints on the parameters $\tan \beta$ and $M_{H^+}$ of the two–Higgs–doublet model (2HDM) of type II [@btau]. Here $\tan\beta$ is the ratio of the two Higgs vacuum expectation values and $M_{H^+}$ is the mass of the physical charged Higgs boson $H^+$ in the model. Since the couplings of $H^+$ to $b$’s and $\tau$’s grow with $\tan\beta$, ${B^+ \to \tau^+ \nu_\tau}$ probes large values of $\tan\beta$. Earlier (but less powerful) constraints on the 2HDM were obtained by the OPAL collaboration, which found $\tan\beta/M_{H^+}< 0.53{\,\mbox{GeV}}{}^{-1}$ from ${\cal B}(\bar{B}\to X\tau\bar{\nu}_{\tau})$ [@opal1] and $\tan\beta/M_{H^+}< 0.78{\,\mbox{GeV}}{}^{-1}$ from ${\cal B}(\tau \to \mu \bar{\nu}_{\mu} \nu_{\tau})$ [@opal2] at the 95% CL. The direct search for a charged Higgs boson through $t \to b H^+$ at the Tevatron has yielded slightly stronger bounds: $M_{H^+}> 125{\,\mbox{GeV}}$ for $\tan \beta=50$ and $M_{H^+}> 150{\,\mbox{GeV}}$ for $\tan \beta=70$ [@tev]. In the low and intermediate $\tan\beta$ regions, the most constraining bound currently comes from the FCNC–induced process $b\to s\gamma$, which yields $M_{H^+}> 295{\,\mbox{GeV}}$ independently of $\tan \beta$ [@bsg]. At tree–level the Higgs sector of the Minimal Supersymmetric Standard Model (MSSM) coincides with the type–II 2HDM. The coupling of $H^+$ to fermions can be modified by a factor of order one due to $\tan\beta$–enhanced radiative corrections [@ar03; @ltb], yet this introduces only a few additional supersymmetric parameters and the access to the Higgs sector in (semi-)leptonic $B$ decays is not obfuscated like in many other modes, such as the loop–induced $b\to s\gamma$ decay. This explains the great theoretical interest in the experimental ranges for ${\cal B} ({B^+ \to \tau^+ \nu_\tau})$ [@bpexp].
The decay ${B \to D \tau \nu_{\tau}}$ provides an alternative route to charged–Higgs effects \[9–15\]. As we will show in the following, this mode is not only competitive with $B^+\to \tau^+\nu_\tau$, but also opens the door to a potential CP–violating phase in the Yukawa couplings of the $H^+$ to $b$ and $\tau$. ${B \to D \tau \nu_{\tau}}$ compares to ${B \to \tau \nu_\tau}$ as follows:
i\) ${\cal B} ({B \to D \tau \nu_{\tau}})$ exceeds ${\cal B} ({B \to \tau \nu_\tau})$ by roughly a factor of 50 in the Standard Model.
ii\) ${B \to D \tau \nu_{\tau}}$ involves the well–known element $V_{cb}$ of the Cabibbo–Kobayashi–Maskawa (CKM) matrix. The uncertainty on $|V_{ub}|$ entering ${B \to \tau \nu_\tau}$ is much larger.
iii\) ${\cal B} ({B \to \tau \nu_\tau})$ is proportional to two powers of the $B$ decay constant $f_B$, which must be obtained with non–perturbative methods. Current lattice gauge theory computations are struggling with chiral logarithms and $f_B^2$ can only be determined with an uncertainty of 30% or more [@latt07]. ${B \to D \tau \nu_{\tau}}$ involves two form factors, one of which can be measured in ${B \to D \ell \nu_\ell}$ ($\ell=e,\mu$) decays [@b01; @HFAG]. The other one is tightly constrained by Heavy Quark Effective Theory (HQET) \[19–22\], so that hadronic uncertainties can be reduced to well below 10% once the measurement of ${B \to D \ell \nu_\ell}$ is improved.
iv\) Unlike ${B \to \tau \nu_\tau}$ the three–body decay ${B \to D \tau \nu_{\tau}}$ permits the study of decay distributions which discriminate between $W^+$ and $H^+$ exchange [@gh; @ks; @miki02].
v\) The Standard Model (SM) contribution to ${B \to \tau \nu_\tau}$ is (mildly) helicity–suppressed, so that the sensitivity of ${\cal B}
({B \to \tau \nu_\tau})$ to $H^+$ is enhanced. For ${B \to D \tau \nu_{\tau}}$ a similar effect only occurs near the kinematic endpoint, where the $D$ moves slowly in the $B$ rest frame [@gh]: While the transversely polarized $W^+$ contribution suffers from a P–wave suppression, the virtual $H^+$ recoils against the $D$ meson in an unsuppressed S–wave.
Items iv) and v) strongly suggest to study differential decay distributions in ${B \to D \tau \nu_{\tau}}$. The $\tau$ in the final state poses an experimental challenge, because it does not travel far enough for a displaced vertex and its decay involves at least one more neutrino. In particular, the $\tau$ polarization, known as a charged-Higgs analyzer [@t], is not directly accessible to experiment. To our knowledge, the only theory papers which address the question of the missing information on the $\tau$ momentum are [@gh; @ks], where a study of the $D$ meson energy spectrum is proposed. Another straightforward way to deal with the missing information on the $\tau$ kinematics, which in addition retains information on the $\tau$ polarization, is to consider the full decay chain down to the detectable particles stemming from the $\tau$. We have studied the decays $\tau^- \to \pi^- \nu_\tau$, $\tau^- \to
\rho^- \nu_\tau$, and $\tau^- \to \ell^-\bar{\nu}_\ell \nu_\tau$ and assessed the sensitivity of the decay distributions to $H^+$ effects. We find that the decay chain $\bar{B}\to D \bar{\nu}_{\tau}\tau^-[\to \pi^-
\nu_{\tau}]$ discriminates between $W^+$ and $H^+$ exchange in an excellent way. In this Letter we present the results for the differential decay rate as a function of the $D$ and $\pi^-$ energies and the angle between the $D$ and $\pi^-$ three–momenta for this decay chain. Our result greatly facilitates the determination of the effective coupling $g_S$ governing $H^+$ exchange, including a potential complex phase, if e.g. a maximum likelihood fit of the data to the theoretical decay distribution given below is employed. A conventional analysis combining Monte Carlo simulations of $\bar{B}\to D
\bar{\nu}_{\tau} \tau^-$ and $\tau^- \to \pi^- \nu_\tau$ decays would be very cumbersome, because the ${B \to D \tau \nu_{\tau}}$ differential distributions strongly depend on the a–priori unknown value of $g_S$.
$B\rightarrow D$ Form Factors
=============================
The effective hamiltonian describing $B\rightarrow(D)\tau\nu_{\tau}$ transitions mediated by $W^+$ or $H^+$ reads (with $q=u\,(c)$) $$\label{eq:1}
\begin{split}
H_{\textrm{eff}} & =\frac{G_{F}}{\sqrt{2}} V_{qb}\text{\thinspace}
\big\{\left[ \overline{q}\gamma^{\mu}(1-\gamma_{5})b\right] \text{\thinspace}
\left[ \overline{\tau}\gamma_{\mu}(1-\gamma_{5})\nu_{\tau}\right] \\
& -\frac{{\overline{m}}_{b}m_{\tau}}{m_{B}^{2}}\text{\thinspace}\text{\thinspace}
\overline{q}\left[ g_{S} + g_P \gamma_5\right] b \text{\thinspace}
\left[\overline{\tau}(1-\gamma_{5})\nu_{\tau}\right]\big\}\,+\,\mbox{h.c.}
\end{split}$$ The effective coupling constant $g_P$ only enters the ${B \to \tau \nu_\tau}$ decay, while ${B \to D \tau \nu_{\tau}}$ is only sensitive to $g_S\,$. The $B^+$ meson mass $m_B$ is introduced in Eq. (\[eq:1\]), so that ${\cal B} ({B \to \tau \nu_\tau})$ vanishes for $g_P=1$. The above operators as well as ${\overline{m}}_b$ are defined in the ${\overline{\rm MS}}$ scheme. In the MSSM, which is our main focus, one has $g_S=g_P$.
The analysis of ${B \to D \tau \nu_{\tau}}$ requires the knowledge of the form factors $F_V$ and $F_S$ which parametrize the vector and scalar current matrix elements: $$\begin{aligned}
\langle D(p_D)|\bar{c}\gamma^{\mu}b|\bar{B}(p_B)\rangle \! =\,&
F_V(q^2)\!
\left[p_B^{\mu}+p_D^{\mu}-m_B^2\frac{1-r^2}{q^2}q^{\mu}\right]{\nonumber \\}+\,& F_S(q^2)\,m_B^2\frac{1-r^2}{q^2}q^{\mu}\,,{\nonumber \\}\langle
D(p_D)|\bar{c}b|\bar{B}(p_B)\rangle=\,& \frac{m_B^2\,(1-r^2)}{{\overline{m}}_b-{\overline{m}}_c}F_S(q^2)\,,\label{eq:2}\end{aligned}$$ where $p_B$ and $p_D$ denote the meson four–momenta, $q=p_B-p_D$, and $r=m_D/m_B\,$. It is convenient to introduce the normalized form factors $V_1\equiv F_V\,2\sqrt{r}/(1+r)$ and $S_1\equiv F_S(1+r)/(2\sqrt{r})$, as well as the kinematic variable $$\label{eq:3}
w\equiv(1+r^2-q^2/m_B^2)/2r.$$ In the limit of infinitely heavy quark masses $m_Q=m_b,\,m_c$ (which are properly infrared–subtracted pole masses), both $V_1(w)$ and $S_1(w)$ reduce to the universal Isgur–Wise function $\xi(w)$, normalized to $\xi(1)=1$. At the kinematic endpoint $w=1$, corrections to this limit read $$\label{eq:4}
V_1(1) =\eta_v-\frac{1-r}{1+r}\big(\delta_{\textrm{rad}}+\delta_{1/m_Q}\big)\,,\quad
S_1(1) =\eta_v\,,$$ up to $\mathcal{O}(\alpha_s^2,1/m_Q^2)$. Here $\eta_v$ denotes radiative corrections in the limit of equal heavy meson masses, and $\delta_{\textrm{rad}}\,(\delta_{1/m_Q})$ are the first order radiative $(1/m_Q)$ corrections to the function $\xi_-$ defined in [@n94]. The $\delta_{1/m_Q}$ term depends on the subleading function $\xi_3(w=1)=\bar{\Lambda}\eta(w=1)$ and on the HQET parameter $\bar{\Lambda}$. We take $\bar{\Lambda}=0.5\pm
0.1\,\textrm{GeV}$, $\eta(1)=0.6\pm 0.2$ [@lnn], $\eta_v$ and $\delta_{\textrm{rad}}$ to $\mathcal{O}(\alpha_s)$ from Ref. [@n92], and add a $5\%$ error to the form factors at $w=1$ to account for higher order corrections. We obtain $V_1(1)=1.05\pm0.08$ and $S_1(1)=1.02\pm0.05$.
The semileptonic decay into light leptons ${B \to D \ell \nu_\ell}$ depends solely on the vector form factor $V_1(w)$. The measured quantity $|V_{cb}|V_1(w)$ was fitted by the BELLE collaboration [@b01] to a two–parameter ansatz $V_1(w,V_1(1),\rho_1^2)$ [@cn] derived from dispersion relations and heavy quark spin symmetry [@bgl]. The fitted curve, however, suffers from large statistical and systematic uncertainties: $|V_{cb}|V_1(1)=(4.11\pm0.44\pm0.52)\%$, $\rho_1^2=1.12\pm0.22\pm0.14$ [@b01]. We thus take $V_1(1)$ from HQET instead, use $|V_{cb}|=(4.17\pm0.07)\%$ from inclusive semileptonic $B$ decays [@pdg], and only fix the form factor at large recoil $w=1.45$ from the data, including the dominant systematic errors in a conservative way: $|V_{cb}|V_1(1.45)=(2.63\pm0.51)\%$. The form factor over the whole kinematic range is then obtained using a two–parameter description $F_{V}(w,a_0^{V},a_1^{V})$, which uses a conformal mapping $w\to z(w)$ resulting in an essentially linear dependence of $F_V$ on $z$ [@rh]. This linearity in $z(w)$ is confirmed by the fact that fitting the ${B \to D \ell \nu_\ell}$ data with both $F_V$ parametrizations without further theoretical constraints essentially gives the same result (see Fig. \[ff\]). The sets of parameters corresponding to the minimal and maximal form factors satisfying the HQET constraint at $w=1$ are displayed in Tab. \[par\] for both parametrizations $V_1(w,V_1(1),\rho_1^2)$ and $F_V(w,a_0^V,a_1^V)$. They delimit the dark gray area in Fig. \[ff\]. We stress that the large error band in Fig. \[ff\] at large $w$ is not due to theory uncertainties but rather to the large systematic error on $|V_{cb}|V_1(1.45)$ from [@b01].
We choose to use only the most recent set of experimental data for our numerical analysis. The HFAG [@HFAG] treats systematic errors in a different way and, including the older CLEO and ALEPH data, finds smaller uncertainties at large recoil (see light gray band in Fig. \[ff\]. The corresponding minimal and maximal curves are given in good approximation by the parameters in the first two lines of Tab. \[tab:HFAG\] for $w$ inside the ${B \to D \tau \nu_{\tau}}$ phase space). The vector form factor has also been studied on the lattice. Computations with quenched Wilson [@FvQuenched] and dynamical staggered [@FvStaggered] fermions, however, both suffer from potentially large systematic errors, which are not fully controlled. In the end, the improvements in the measurements of the ${B \to D \ell \nu_\ell}$ and ${B \to D \tau \nu_{\tau}}$ modes will go together, and $|V_{cb}|F_V$ will most likely be best determined from experimental data alone. For the time being, we will proceed with the conservative estimation of Tab. \[par\].
Parameters min. $|V_{cb}|F$ max. $|V_{cb}|F$ centr. $|V_{cb}|F$
------------------------------------ ------------------ ------------------ --------------------
$\{|V_{cb}|V_1(1),\rho_1^2\}$ $\{0.040,1.47\}$ $\{0.048,1.06\}$ $\{0.044,1.24\}$
$|V_{cb}|\{a_0^V,a_1^V\}[10^{-5}]$ $\{0.94,-5.7\}$ $\{1.28,-2.2\}$ $\{1.11,-3.9\}$
$|V_{cb}|\{a_0^S,a_1^S\}[10^{-4}]$ $\{1.62,-1.1\}$ $\{2.14,-3.2\}$ $\{1.88,-6.8\}$
: Parameters $\{|V_{cb}|V_1(1),\rho_1^2\}$ for $|V_{cb}|F_V$ [@cn] and $\{|V_{cb}|a_0^{V,S},|V_{cb}|a_1^{V,S}\}$ for $|V_{cb}|F_{V,S}$ (see [@rh], $Q^2=0$, $\eta=2$, subthreshold poles: $m(1^-)=6.337,\,6.899,\,7.012\,\textrm{GeV}$ and $m(0^+)=6.700,\,7.108\,\textrm{GeV}$ [@poles]). $F_V$ is displayed in dark gray in Fig. \[ff\].[]{data-label="par"}
Parameters min. $|V_{cb}|F$ max. $|V_{cb}|F$ centr. $|V_{cb}|F$
------------------------------------ ------------------ ------------------ --------------------
$\{|V_{cb}|V_1(1),\rho_1^2\}$ $\{0.038,1.01\}$ $\{0.047,1.30\}$ $\{0.042,1.17\}$
$|V_{cb}|\{a_0^V,a_1^V\}[10^{-5}]$ $\{1.03,-1.3\}$ $\{1.17,-4.8\}$ $\{1.10,-3.0\}$
$|V_{cb}|\{a_0^S,a_1^S\}[10^{-4}]$ $\{1.78,-5.7\}$ $\{2.00,-7.6\}$ $\{1.89,-6.6\}$
: Parameters $\{|V_{cb}|V_1(1),\rho_1^2\}$ for $|V_{cb}|F_V$ from HFAG [@HFAG], and $\{|V_{cb}|a_0^{V,S},|V_{cb}|a_1^{V,S}\}$ for $|V_{cb}|F_{V,S}$. $F_V$ is displayed in light gray in Fig. \[ff\].[]{data-label="tab:HFAG"}
In a similar way, the scalar form factor $F_{S}(w,a_0^{S},a_1^{S})$ is constrained by HQET at $w=1$, while its value at large recoil is fixed from the relation $F_S(q^2=0)=F_V(q^2=0)$. The resulting parameters are displayed in the third line of Tab. \[par\] (or Tab. \[tab:HFAG\] if $F_V$ is taken from [@HFAG]). As expected from the heavy-quark limit, the normalized form factor $S_1$ is quite close to $V_1$ on the whole $w$ range, with slightly smaller errors.
Charged–Higgs Effects
=====================
The MSSM is a well–motivated new–physics scenario in which charged scalar current interactions occur at tree–level. Resumming the dominant $\tan\beta$-enhanced loop corrections to all orders, the couplings $g_{S,P}$ in Eq. (\[eq:1\]) specify to [@yuk; @itoh] $$\label{eq:5}
g_{S}=g_{P}=\frac{m_{B}^{2}}{M_{H^+}^{2}}\,
\frac{\tan^{2}\beta}{ (1+\widetilde \epsilon_0\,\tan\beta)
(1+\epsilon_{\tau}\,\tan\beta)}\,.$$ This particular form holds in MSSM scenarios with Minimal Flavor Violation (MFV). The loop factor $\widetilde \epsilon_0$ arises from the quark Yukawa sector and depends on ratios of superparticle masses, resulting in a sizable non–decoupling effect $\widetilde
\epsilon_0\tan\beta={\cal O}(1)$ for $\tan\beta={\cal O}(50)$. $\epsilon_{\tau}$ comprises the corresponding effect for the $\tau$ lepton. $\widetilde
\epsilon_0$ and $\epsilon_\tau$ can receive sizable complex phases from the Higgsino mass parameter $\mu$, if first–generation sfermions are sufficiently heavy to soften the impact of the bounds on electric dipole moments on $\arg \mu$. Beyond MFV also phases from squark mass matrices will easily render $g_S$ complex. It is therefore mandatory to constrain – and eventually measure – both magnitude and phase of $g_S$. The type–II 2HDM is recovered by setting $\widetilde \epsilon_0=\epsilon_{\tau}=0$.
The ${B \to D \tau \nu_{\tau}}$ branching ratio has recently been measured by the BABAR collaboration [@babarbdt]: $$\label{eq:6}
R^{\textrm{exp}}\equiv\frac{\mathcal{B}({B \to D \tau \nu_{\tau}})}{\mathcal{B}({B \to D \ell \nu_\ell})}=(41.6\pm 11.7\pm 5.2)\%\,.$$ The normalization to $\mathcal{B}({B \to D \ell \nu_\ell})$ reduces the dependence on the vector form factor $F_V$ and thus tames the main theoretical uncertainties. In the presence of charged–Higgs contributions, the theoretical ratio is approximated to 1% by $$R^{\textrm{th}}=\frac{ 1.126 + 0.037\, r_V + r_0^2\, ( 1.544 + 0.082\, r_S + N_{H^+} ) }{ 10 - 0.95\, r_V }\ ,$$ $$\label{eq:7}
\begin{split}
N_{H^+}=&- r_{cb}\,Re[g_S]\,( 1.038 + 0.076\, r_S )\\
&+ r_{cb}^2\,|g_S|^2\,( 0.186 + 0.017\, r_S ),
\end{split}$$ with $r_V=(a_1^V/a_0^V)/(-3.4)$, $r_S=(a_1^S/a_0^S)/(-3.5)$, $r_0=(a_0^S/a_0^V)/17$, and $r_{cb}=0.8/(1-{\overline{m}}_c/{\overline{m}}_b)$. The dependence on the slope parameters $a_1^{V,S}$ appears to be quite mild. In Fig. \[bdt\] we compare $R^{\textrm{th}}$ (right–hand side) as well as $\mathcal{B}(B\to\tau\nu)$ (left–hand side) to their one–sigma measurements for positive $g_S$ and $g_P$. For $R^{\textrm{th}}$, we also display the less conservative theoretical prediction obtained from the HFAG vector form factor in Tab. \[tab:HFAG\] (light gray band). In particular, we obtain the SM estimates $${\cal B} (B^- \to D^0 \tau^- \bar{\nu}_{\tau})^{\textrm{SM}}
= (0.71\pm0.09)\%$$ and $${\cal B} (\bar{B}^0 \to D^+ \tau^- \bar{\nu}_{\tau})^{\textrm{SM}}
= (0.66\pm 0.08)\%.$$ (Error sources: $|V_{cb}|F_V(w),\ S_1(1),\ |V_{cb}|$). We cannot reproduce the small errors of Ref. [@cg06].
The ${B \to D \tau \nu_{\tau}}$ branching fraction is promising to discover – or constrain – charged–Higgs effects, but not to measure $g_S$ with good precision, as the dependence in Fig. \[bdt\] is too flat. The differential distribution in the decay chain $\bar{B}\to D \bar{\nu}_{\tau}\tau^-[\to\pi^-\nu_{\tau}]$ is better suited for that purpose. The experimentally accessible quantities are the energies $E_D$ and $E_\pi$ of the $D$ and $\pi^-$ mesons, respectively, and the angle $\theta$ between the three–momenta $\vec p_D$ and $\vec p_\pi$. We define these quantities in the $B$ rest frame, which can be accessed from the $\Upsilon(4S)$ rest frame thanks to full $B$ reconstruction [@babarbdt]. We integrate over the phase space of the two unobserved neutrinos in the final state. Our formulae contain the full spin correlation between the production and decay of the $\tau$, which is important to discriminate between SM and charged–Higgs contributions. This approach further facilitates the rejection of backgrounds from neutral particles escaping detection, as in $\bar{B}\to D
D^-[\to \pi^- \pi^0]$ with an undetected $\pi^0$: If the mass of the undetected particle is $m$, this background can be suppressed by cuts excluding the region around $$\begin{aligned}
\cos\theta = \frac{(m_B-E_D-E_{\pi})^2-2(E_D^2-m_D^2)-m^2}{2(E_D^2-m_D^2)}\,. \label{eq:8}\end{aligned}$$
We obtain the differential distribution $$\begin{aligned}
& \frac{d\Gamma(\bar{B}\to D\bar{\nu}_{\tau}\tau^-
[\to\pi^-\nu_{\tau}])}{dE_D\,dE_{\pi}\,d\cos\theta}=G_F^4f_{\pi}^2|V_{ud}|^2|V_{cb}|^2\tau_{\tau}\label{eq:9}\\
&
\times\left[C_W(F_V,F_S)-C_{WH}(F_V,F_S)\,Re[g_S]+C_H(F_S)|g_S|^2\right]\nonumber\end{aligned}$$ with form–factor–dependent functions of $E_D$, $E_{\pi}$, and $\cos\theta$ for the SM ($C_W$), interference ($C_{WH}$), and Higgs ($C_H$) contributions, given as follows for vanishing $m_\pi$ (this approximation, which is good to 1%, is not used in our numerical analysis), $$\begin{split}
C_W& =\kappa \frac{m_{\tau}^4}{2}\frac{l^2}{p_{\pi}\cdot
l} \,
\bigg\{ P^2 (b-1) + (P \cdot l)^2 \frac{2b}{l^2} + \\
& \Bigg[
\frac{l^2 (P\cdot p_\pi)^2 }{ (p_{\pi}\cdot l)^2}
- \frac{2 (P\cdot l)(P\cdot p_{\pi})
}{p_{\pi}\cdot l} \Bigg] (3 b - 1)
\bigg\}\, , \\
C_{WH} & = 2\kappa \, m_{\tau}^4 \,
\frac{(1-r^2)F_S}{1-{\overline{m}}_c/{\overline{m}}_b} \, b \,
\bigg[ P \cdot l -\frac{l^2 P\cdot p_\pi }{p_{\pi}\cdot l} \bigg]\, , \\
C_H &=\kappa\,m_{\tau}^6 \frac{(1-r^2)^2
F_S^2}{(1-{\overline{m}}_c/{\overline{m}}_b)^2}
\Big(1-\frac{m_{\tau}^2}{2\,p_{\pi}\cdot l} \Big)\, , \\
\end{split} \label{defc}$$ where ${\overline{m}}_c$ and ${\overline{m}}_b$ must be evaluated at the same scale so that ${\overline{m}}_c/{\overline{m}}_b=0.20\pm 0.02$ [@run], and $$\begin{aligned}
P \!\!& =&\! F_V (p_B+p_D) -
(F_V-F_S)\frac{m_B^2(1-r^2)}{q^2} (p_B-p_D) \,, {\nonumber \\}\kappa \!\!& =&\!
\frac{E_{\pi}\sqrt{E_D^2-m_D^2}}{
128\,\pi^4\,m_B m_{\tau}}\,,
\quad b = \frac{m_\tau^2}{p_\pi \cdot l}
\Big( 1 - \frac{m_{\tau}^2}{2\,p_{\pi}\cdot l} \Big)\,, {\nonumber \\}l \!\!& =&\! p_B-p_D-p_\pi \,,\quad
q^2 = (p_B-p_D)^2.\label{defpkb}\end{aligned}$$ The dot products appearing in [Eqs. (\[defc\]) and (\[defpkb\])]{} are related to the energies, momenta, and the angle $\theta$ measured in the $B$ rest frame as $p_B\!\cdot\! l=m_B (m_B-E_D-E_\pi)$, $p_D\!\cdot\! l = E_D (m_B -E_D-E_\pi) +
|\vec p_D|^2+|\vec p_D| E_\pi \cos\theta$, $p_\pi\!\cdot\! l = E_\pi (m_B-E_D)
+ |\vec p_D| E_\pi \cos\theta$, and $p_B\!\cdot\! p_D = m_B E_D$. Further $\tau_{\tau}=(290.6\pm 1.0)\times 10^{-15}s$ is the $\tau$ lepton lifetime, $f_{\pi}=(130.7\pm 0.1\pm 0.36)\,\textrm{MeV}$ the pion decay constant, and the CKM matrix elements are $|V_{ud}|=0.97377\pm
0.00027$ and $|V_{cb}|=(41.7\pm 0.7)\times 10^{-3}$, the latter being well determined from inclusive semileptonic $B$ decays [@pdg]. Remarkably, one can probe a CP–violating phase of $g_S$ by exploiting the shape of the distribution in Eq. (\[eq:9\]), which is not possible from the branching fraction of either ${B \to D \tau \nu_{\tau}}$ or ${B \to \tau \nu_\tau}$.
For illustration, we show the differential decay distribution including charged–Higgs effects in comparison with the SM for the meson energies $E_D=2\,\textrm{GeV}$ and $E_{\pi}=1\,\textrm{GeV}$, so that the whole range of $\cos\theta$ is kinematically accessible. In this particular region of phase space the SM rate is strongly suppressed for $\cos\theta=-1$. For a large scalar coupling $g_S=2$ (Fig. \[bdp\], left), the Higgs contribution dominates the rate at this point (dark gray band), so that we can clearly distinguish it from the SM (light gray band). The experimental information from $\mathcal{B}({B \to \tau \nu_\tau})$ constrains $|1-g_P|$. For real $g_P$ this permits a range near $g_P=0$ and another range around $g_P=2$. In the MSSM situation with $g_P=g_S$, the case $g_S=2$ therefore is in agreement with ${B \to \tau \nu_\tau}$, but can be confirmed or ruled out by measuring our distribution. The discrimination potential for the phase of $g_S$ shows up in the light gray band: It corresponds to a complex $g_S=1+i$, which yields the same $\mathcal{B}({B \to \tau \nu_\tau})$ as $g_S=0,2$. The ${B \to D \tau \nu_{\tau}}$ branching ratio alone may also help to distinguish between these solutions, depending on the future experimental value of $\mathcal{B}({B \to D \tau \nu_{\tau}})$, see Fig. \[bdt\]. For general $g_S$ values, a fit to the triple differential distribution in Eq. (\[eq:9\]) would excellently quantify charged–Higgs effects, especially once better experimental information on the form factors is available, as we illustrate with fixed $D$ and $\pi^-$ energies in Fig. \[bdp\] (right–hand side) for $g_S=0.5$. Such a fit would combine information from different parts of the phase space, and thus resolve much smaller $g_S$ values. A more precise quantitative analysis would require the fit to actual data, and thus goes beyond the scope of this paper. Still, keep in mind that even with more precise ${B \to \tau \nu_\tau}$ experimental data and improved estimates of $f_B$ and $|V_{ub}|$, a value of $g_P\simeq0.2-0.3$ will be very difficult to exclude with $\mathcal{B}({B \to \tau \nu_\tau})$. ${B \to D \tau \nu_{\tau}}$ is thus definitely competitive.
As mentioned in the Introduction, a similar analysis was performed for the other $\tau$ decay channels $\tau^-\to\rho^-\nu_\tau$ and $\tau^-\to\ell^-\bar{\nu}_\ell \nu_\tau$, which together with $\tau^-\to\pi^-\nu_\tau$ constitute more than 70% of the $\tau$ branching fraction. Ultimately, a combined analysis of all these modes is desirable in order to exploit the available and forthcoming experimental data in an optimal way.
Conclusions
===========
We have studied charged–Higgs effects in a differential distribution of the decay chain $\bar{B}\to D\bar{\nu}_{\tau}\tau^-[\to\pi^-\nu_{\tau}]$, which has the following advantages over the branching fractions $\mathcal{B}({B \to \tau \nu_\tau})$ and $\mathcal{B}({B \to D \tau \nu_{\tau}})$:
i\) The Higgs coupling constant $g_S$ can be determined from the *shape* of the distribution in sensitive phase space regions. This analysis should be possible with current $B$ factory data.
ii\) The dependence on both $|g_S|$ and ${\mathrm{Re}\,}[g_S]$ allows to quantify a possible CP–violating phase. Since our decay distribution is a CP–conserving quantity, the phase of $g_S$ is determined with a two–fold ambiguity. In the MSSM such a phase stems from the $\mu$ parameter or the soft breaking terms and enters through $\tan\beta$–enhanced loop factors. $B\to D\tau \nu$ complements collider studies of these phases [@ccp].
The main uncertainties stem from the form factors. One can gain a much better accuracy with better data on the vector form factor $F_V$. The recent ${B \to D \ell \nu_\ell}$ measurement by BABAR [@babarbdl] furnishes promising data for a new fit.
Within the MSSM, one will be able to place new constraints on the $\tan\beta-M_{H^+}$ plane, once our results are confronted with actual data from the $B$ factories. If $\tan\beta/M_{H^+}$ is indeed large, there is a fair chance to reveal charged–Higgs effects ahead of the LHC.\
*Acknowledgments*: The authors acknowledge helpful discussions with Richard Hill and Michael Feindt. This work is supported in part by the DFG grant No. NI 1105/1–1, by the DFG–SFB/TR9, and by the EU Contract No. MRTN-CT-2006-035482, FLAVIAnet”.
[00]{}
W. Hou, Phys. Rev. D [**48**]{}, 2342 (1993). OPAL Collaboration, Phys. Lett. B [**520**]{}, 1 (2001). OPAL Collaboration, Phys. Lett. B [**551**]{}, 35 (2003). J. Nielsen \[CDF and D0 Collaborations\], Nucl. Phys. Proc. Suppl. [**177**]{}, 224 (2008). M. Misiak *et al.*, Phys. Rev. Lett. [**98**]{}, 022002 (2007).
L. Hall, R. Rattazzi and U. Sarid, Phys. Rev. D [**50**]{}, 7048 (1994); T. Blazek, S. Raby and S. Pokorski, Phys. Rev. D [**52**]{}, 4151 (1995). M. Carena *et al.*, Nucl. Phys. B [**577**]{}, 88 (2000). For earlier work on $t\to b H^+$ without resummation of $\tan\beta$–enhanced radiative corrections, see R.A. Jimenez and J. Sola, Phys. Lett. B [**389**]{}, 53 (1996).
A. Akeroyd and S. Recksiegel, J. Phys. G [**29**]{}, 2311 (2003). BELLE Collaboration, Phys. Rev. Lett. [**97**]{}, 251802 (2006); BABAR Collaboration, Phys. Rev. D [**76**]{}, 052002 (2007); average: D. Monorchio, talk presented at HEP2007, http://www.hep.man.ac.uk/HEP2007/.
B. Grzadkowski and W. Hou, Phys. Lett. B [**283**]{}, 427 (1992). M. Tanaka, Z. Phys. C [**67**]{}, 321 (1995). K. Kiers and A. Soni, Phys. Rev. D [**56**]{}, 5786 (1997). T. Miki, T. Miura and M. Tanaka, \[hep-ph/0210051\].
H. Itoh, S. Komine and Y. Okada, Prog. Theor. Phys. [**114**]{}, 179 (2005). C.-H. Chen and C.-Q. Geng, JHEP [**0610**]{}, 053 (2006).
J.F. Kamenik and F. Mescia, \[0802.3790\]\[hep-ph\].
M. Della Morte, PoS LAT2007, 8. BELLE Collaboration, Phys. Lett. B [**526**]{}, 258 (2002). Heavy Flavor Averaging Group, \[0704.3575\]\[hep-ex\], and online update at http://www.slac.stanford.edu/xorg/hfag/.
N. Isgur and M.B. Wise, Phys. Lett. B [**232**]{}, 113 (1989); [**237**]{}, 527 (1990); B. Grinstein, Nucl. Phys. B [**339**]{}, 253 (1990); E. Eichten and B. Hill, Phys. Lett. B [**234**]{}, 511 (1990); H. Georgi, Phys. Lett. B [**240**]{}, 447 (1990).
M. Neubert, Phys. Rept. [**245**]{}, 259 (1994). C.G. Boyd, B. Grinstein and R.F. Lebed, Phys. Rev. D [**56**]{}, 6895 (1997).
I. Caprini, L. Lellouch and M. Neubert, Nucl. Phys. B [**530**]{}, 153 (1998). Z. Ligeti, Y. Nir and M. Neubert, Phys. Rev. D [**49**]{}, 1302 (1994). M. Neubert, Phys. Rev. D [**46**]{}, 2212 (1992). W.-M. Yao *et al.* \[Particle Data Group\], J. Phys. G [**33**]{}, 1 (2006). R. Hill, \[hep-ph/0606023\]. G. M. de Divitiis, R. Petronzio and N. Tantalo, JHEP [**0710**]{}, 062 (2007). M. Okamoto *et al.*, Nucl. Phys. Proc. Suppl. [**140**]{}, 461 (2005).
E. Eichten and C. Quigg, Phys. Rev. D [**49**]{}, 5845 (1994). A. Buras *et al.*, Nucl. Phys. B [**659**]{}, 3 (2003). BABAR Collaboration, Phys. Rev. Lett. [**100**]{}, 021801 (2008). CKMfitter, fit inputs for summer 2007, http://www.slac.stanford.edu/xorg/ckmfitter/.
K. Chetyrkin, J. Kühn and M. Steinhauser, Comput. Phys. Commun. [**133**]{}, 43 (2000). Input values: J. Kühn, M. Steinhauser and C. Sturm, Nucl. Phys. B [**778**]{}, 192 (2007). A. Bartl, K. Hohenwarter-Sodek, T. Kernreiter, O. Kittel and M. Terwort, \[0802.3592\]\[hep-ph\], and references therein.
BABAR Collaboration, \[0708.1738\]\[hep-ex\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We consider a repeated Matching Pennies game in which players have limited access to randomness. Playing the (unique) Nash equilibrium in this $n$-stage game requires $n$ random bits. Can there be Nash equilibria (or $\varepsilon$-Nash equilibria) that use less than $n$ random coins?
Our main results are as follows
- We give a full characterization of approximate equilibria, showing that, for any $\gamma \in [0,1]$, the game has a $\gamma$-Nash equilibrium if and only if both players have $(1 - \gamma)n$ random coins.
- When players are bound to run in polynomial time with $n^\delta$ bits of randomness, approximate Nash equilibria can exist if and only if one-way functions exist.
- It is possible to trade-off randomness for running time. In particular, under reasonable assumptions, if we give one player only $O(\log n)$ random coins but allow him to run in arbitrary polynomial time with $n^\delta$ bits of randomness and we restrict his opponent to run in time $n^k$, for some fixed $k$, then we can sustain an $\varepsilon$-Nash equilibrium.
- When the game is played for an infinite amount of rounds with time discounted utilities, under reasonable assumptions, we can reduce the amount of randomness required to achieve a $\varepsilon$-Nash equilibrium to $n^\delta$, where $n$ is the number of random coins necessary to achieve an approximate Nash equilibrium in the general case.
author:
- |
Michele Budinich[^1]\
[**[email protected]**]{}
- |
Lance Fortnow\
[**[email protected]**]{}
title: |
Repeated Matching Pennies\
with Limited Randomness[^2]
---
Introduction
============
In the classical setting of Game Theory, one of the core assumptions is that all participating agents are “fully rational”. This amounts not only to the fact that an agent must be able to make optimal decisions, given the other players’ actions, but also to the fact that he must understand how these actions will affect the behavior of all other participants. If this is the case, a Nash equilibrium can be viewed as a set of strategies in which each agent is simply computing his best response given his opponents’ actions. However, in real world strategic interactions, people often behave in manners that are not fully rational. There are many reasons behind non-rational behavior, we focus on two: limitations on computation and limitations on randomness.
Since the work of Herbert Simon [@simon1955behavioral], much research has focused on defining models that take computational issues into account. In recent years, the idea that the full rationality assumption is often unrealistic has been formalized using tools and ideas from computational complexity. It is in fact easy to come up with settings, as in Fortnow and Santhanam [@fortnow2010bounding], in which simply computing a best response strategy involves solving a computationally hard problem. Furthermore there is strong evidence that, in general, the problem of finding a Nash equilibrium is computationally difficult for matrix games (Daskalakis, Goldberg and Papadimitriou [@DBLP:journals/siamcomp/DaskalakisGP09], Chen and Deng [@chen2006settling]).
Traditionally bounded rationality has focused on two computational resources: time and space. In this paper we focus on another fundamental resource: randomness.
It is a basic fact that games in which agents are not allowed to randomize might have no Nash equilibrium. In this sense, randomness is essential in game theory. We focus on a simple two player zero-sum game that captures this: Matching Pennies (Figure \[fig:game\]).
![The payoff bimatrix for the Matching Pennies game.[]{data-label="fig:game"}](matchingpennies.eps){width="1.8in"}
Specifically, we consider the repeated version of Matching Pennies, played for $n$ rounds. In this game, the unique Nash equilibrium is the one in which, at every round, both players choose one of their two strategies uniformly at random. The algorithm that implements this strategy requires $n$ random coins, one for each round. The main question we address in this paper is: can there be Nash equilibria if the amount of randomness available to both players is less than $n$?
First we show that, in general, the game cannot have a Nash equilibrium in which both players have only a fraction of $n$ random coins. In particular, when we give both players $n(1
-\gamma)$ random coins, we can only achieve a $\gamma$-Nash equilibrium. This turns out to be tight, in the sense that we can show that any game with a $\gamma$-Nash equilibrium both players must have at least $n(1 - \gamma)$ coins.
The proof of this fact, however, relies on the players’ ability to implement a strategy that runs in exponential time. We then consider games in which the players’ strategies are polynomially-bounded. Using ideas developed in cryptography and computational complexity we show that, in this setting, $\varepsilon$-Nash equilibria that use only $n^{\delta}$ coins exist if and only if one-way functions exist.
We also show that the amount of randomness can be “traded” for time. If we allow one of the players to run in arbitrary polynomial time, but use only $O(\log n)$ bits, we can still achieve a $\varepsilon$-Nash equilibrium if we restrict his opponent to run in time $n^k$ for some fixed $k > 0$, while giving him $n^{\delta}$ random bits.
Finally we consider an infinitely repeated game with time discounted utilities. In this case, in general, for any discount factor $\delta$ and approximation $\varepsilon$, we can always achieve a $\varepsilon$-Nash equilibrium with only $n$ random coins, if $n$ is large enough. When we limit players’ strategies to polynomial size circuits, we can reduce the amount of randomness to $n^{\delta}$, for any $\delta > 0$.
### Related Work {#related-work .unnumbered}
There are many recent approaches to bounded rationality using a computational complexity perspective. For instance Halpern and Pass [@halpernpass2010] study games in which players’ strategies are Turing machines. The idea of considering randomness as a costly resource in game theory has received only limited attention. Kalyanaraman and Umans [@kalyanaraman2007algorithms] study zero-sum games, and give both an efficient deterministic algorithm for finding $\varepsilon$-Nash equilibria, as well as a weaker, but more general, result in the spirit of our Lemma \[lemma:unbounded1\], giving a randomness-efficient adaptive on-line algorithm for playing repeated zero-sum games. Hu [@hu2009complexity] also considers a similar setting but he is concerned with computability rather than complexity. He considers infinitely repeated plays of 2 player zero-sum games that have no pure strategy Nash equilibrium, and in which players have a set of feasible actions, which represents both the strategies they can play and the strategies they can predict. In this setting Hu gives necessary and sufficient conditions for the existence of Nash equilibria. Finally Gossner and Tomala [@gossner2008entropy], give entropy bounds on Bayesian learning in a game theoretic setting, in a more general framework then this paper. Their results applied to Matching Pennies do not achieve the tight bounds we get in Lemma \[lemma:unbounded1\].\
The rest of the paper is organized as follows. In Section \[sec:background\] we introduce the notation and known results used. Section \[sec:info\] presents an information theoretic impossibility result. Section \[sec:eff\] considers players whose strategies are limited to polynomial sized Boolean circuit families, while Sections \[sec:nw\] and \[sec:infinite\] give extensions of the main results to complexity pseudorandom number generators and infinitely repeated versions of the game.
Background and Definitions {#sec:background}
==========================
Game Theory Notation
--------------------
Throughout the paper we consider a repeated game of Matching Pennies. We focus on this game because it captures one of the fundamental aspects of randomness in game theory. Studying such a simple game also allows us to get tight bounds. However, variations of our results extend to other similar 2 person zero-sum repeated games.
The payoffs at each round are shown in Figure \[fig:game\]. Let $h: \{\text{H,T}\} \times \{\text{H,T}\} \to \{-1, 1\}$, be the payoff to Player 1 (P1), and $-h$ the payoff for P2. When we allow the players to randomize, we denote as $S_i$ a randomized strategy on $\Delta\{H, T\}$ for player $i$. P1’s expected payoff in one round is $${\mathbf{E}\left[ h(S_1, S_2) \right]} = \sum_{s_1,s_2 \in \{\text{H,T}\}} {\text{Pr}\left( S_1 = s_1 \right)} {\text{Pr}\left( S_2 = s_2 \right)} h(s_1, s_2).$$ Let $u:\{\text{H,T\}}^n \times \{\text{H,T\}}^n \to \{-n,
\cdots, n\}$ be P1’s cumulative payoff when the game is played for $n$ rounds. In the repeated game, mixed strategies can be viewed as distribution over sequences of length $n$ that are dependend on the opponent’s strategy. Given the adversary’s strategy, let $R_i =
(r_i^1, \cdots, r_i^n) \in \Delta\{\text{H,T\}}^n$ denote a randomized strategies for player $i$. P1’s expected cumulative payoff is $${\mathbf{E}\left[ u(R_1, R_2) \right]} = \sum_{t = 1}^n {\mathbf{E}\left[ h(r_1^t, r_2^t) \right]}.$$ Finally we define the expected average payoff to P1 for the $n$-round game as $${\mathbf{E}\left[ U(R_1, R_2) \right]} = \frac{{\mathbf{E}\left[ u(R_1, R_2) \right]}}{n},$$ and consequently P2’s expected payoff is $-{\mathbf{E}\left[ U(R_1, R_2) \right]}$. To denote player’s $i$ payoff we will sometimes use the standard notation ${\mathbf{E}\left[ U(R_i, R_{-i}) \right]}$, where $R_i$ is player’s $i$ mixed strategy and $R_{-i}$ is his opponent’s mixed strategy.
A pair of mixed strategies $(R_1, R_2)$ is a Nash equilibrium for the $n$-stage Matching Pennies game if, for $i = 1,2$: $${\mathbf{E}\left[ U(R_i, R_{-i}) \right]} \ge {\mathbf{E}\left[ U(R_i', R_{-i}) \right]} \quad \text{for all } R_i' \in \Delta\{\text{H,T}\}^n.$$
In some cases we will consider a relaxed notion of equilibrium, namely $\varepsilon$-Nash equilibrium.
A pair of mixed strategies $(R_1, R_2)$ is a $\varepsilon$-Nash equilibrium for the $n$-stage Matching Pennies game if, for $i
= 1,2$: $${\mathbf{E}\left[ U(R_i, R_{-i}) \right]} \ge {\mathbf{E}\left[ U(R_i', R_{-i}) \right]} - \varepsilon \quad \text{for all } R_i' \in \Delta\{\text{H,T}\}^n.$$
Complexity and Pseudorandomness
-------------------------------
We give a brief description of the pseudorandomness tools we need for this paper. For more details we recommend the textbooks of Arora and Barak [@Arora:1253363] and Goldreich [@goldreich2004foundations].
The model of computation used throughout most of the paper is based on Boolean Circuits. We consider circuits with and gates, and denote by $C_n$ a circuit with $n$ input nodes. A circuit family $\{C_i\}_{i \in
\mathbb{N}}$ is an infinite collection of circuits, intuitively one for each input length.
The size of a circuit $|C_n|$ is the number of gates. A circuit family is polynomial sized if there is a $k >0$ such that, for all $n$, $|C_n| \le n^k$. The class of languages recognizable by families of polynomial sized circuits is called . Any language that can be decided in polynomial time by a deterministic or randomized Turing machine is also in . Formally $\textsf{P} \subseteq \textsf{BPP}
\subseteq \textsf{P/Poly}$.
A function is one-way if it is easy to compute and hard to invert.
A one-way function is a polynomial-time computable function $f:\{0,1\}^n \to \{0,1\}^n$ such that, for all polynomial size circuits $D$ and $y = f(x)$ (where $x$ is chosen uniformly at random on $\{0,1\}^n$), $\Pr(f(D(y)) = f(x)) < n^{-c}$ for all $c
> 0$ and sufficiently large $n$.
Informally, two objects are indistinguishable if no polynomial sized circuit family can tell them apart with noticeable probability.
Let $X,Y$ be two random variables on $\{0,1\}^n$. We say that $X$ and $Y$ are computationally indistinguishable if for every family of polynomial size circuits $\left\{ C_i \right\}_{i \in
\mathbb{N}}$, every $c > 0$ and for sufficiently large $n$ $$\Big| {\text{Pr}\left( C_n\left( X \right) = 1 \right)} - {\text{Pr}\left( C_n\left( Y \right) = 1 \right)} \Big| < \frac{1}{n^c}.$$
A cryptographic pseudorandom number generator (PRNG) is a deterministic algorithm whose output is computationally indistinguishable from the uniform distribution, provided that it’s input is truly random. We will denote by $U_k$ a random variable uniformly distributed on $\{0,1\}^k$.
A cryptographic pseudorandom number generator is a deterministic polynomial time algorithm $G: \{0,1\}^{l(n)} \to
\{0,1\}^{n}$, where $l(n)<n$ is a polynomial time computable function, such that $G(U_{l(n)})$ and $U_{n}$ are computationally indistinguishable.
There are two basic properties about pseudorandom number generators that we will use. One relates the notion of pseudorandomness to the notion of predictability.
Let $G: \{0,1\}^{l(n)} \to \{0,1\}^{n}$ be a polynomial time algorithm, and $G(x) = (y_1, \cdots, y_n)$. We call $G$ unpredictable if for every family of polynomial size circuits $\left\{ D_i \right\}_{i \in \mathbb{N}}$, all $c>0$ and for sufficiently large $n$ $${\text{Pr}\left( D_i\left( y_1, \dots, y_{i-1} \right) = y_{i} \right)} \le \frac{1}{2} + \frac{1}{n^c}.$$
Intuitively a pseudorandom number generator must be unpredictable, otherwise we could easily build a test for it by using the predictor circuits. In 1982 Yao [@yao1982theory] proved the opposite implication, thus establishing the following theorem.
\[yaothm\] A polynomial time algorithm $G: \{0,1\}^{l(n)} \to \{0,1\}^{n}$ is unpredictable if and only if $G$ is a pseudorandom number generator.
Hstad, Impagliazzo, Luby and Levin in 1999 [@hastad1999pseudorandom] showed how to construct pseudorandom number generators with polynomial expansion based on one-way functions.
\[hillthm\] One way functions exist if and only if for every $\delta >0$ there is a pseudorandom number generator with $l(n) = n^\delta$.
Cryptographic pseudorandom number generators’ main power lies in the ability to fool any polynomial sized adversary, while running in polynomial time. However, in other areas of complexity, such as derandomization, the crucial issue is having a smaller seed.
A complexity pseudorandom number generator is a $2^{l(n)}$ time computable function $G: \{0,1\}^{l(n)} \to \{0,1\}^{n}$, such that for any circuit $C$ of size $n$ $$\Big| {\text{Pr}\left( C\left( U_{l(n)} \right) = 1 \right)} -
{\text{Pr}\left( C\left(G\left( U_n \right)\right) = 1 \right)} \Big| < n^{-1}.$$
The essential difference with cryptographic pseudorandom number generators is the order of quantifiers. A cryptographic pseudorandom number generator fools circuits of an arbitrary polynomial size. The complexity pseudorandom number generator fools circuits only of a fixed polynomial size but under the right assumptions requires far fewer random bits.
Impagliazzo and Wigderson [@DBLP:journals/jcss/ImpagliazzoW01] building on a series of paper starting with Nisan and Wigderson [@nisan1994hardness] characterize when complexity pseudorandom number generators exist.
\[thm:nw\] There exists $L \in \textsf{DTIME}(2^{O(l(n))})$ and $\varepsilon > 0$ such that no circuit of size at most $2^{\varepsilon l(n)}$ can compute $L$ if and only if there exists a complexity pseudorandom number generator with $l(n) = C \log n$ for some $C >
0$.
Information Theoretic Bounds {#sec:info}
============================
In this section we make no computational assumptions on the players, and show that there can be no Nash equilibrium if we limit the amount of randomness available to both players.
\[lemma:unbounded1\] For any $\gamma \in [0,1]$, if P2 has less than $n(1-\gamma)$ random bits, then P1 has a deterministic strategy $A$ that achieves an expected average payoff of at least $\gamma$.
We will give a strategy $A = (a^1, \cdots, a^n)$ for P1 that achieves a high payoff against any strategy $B = (b^1, \cdots,
b^n)$ from P2.
Player 1 will enumerate all of P2’s possible coin flips, and will obtain a set of $2^{n(1-\gamma)}$ possible strategies, one of which is the one being used by P2. After each play by P2, P1 can eliminate all the strategies that do not play that action at that round. Let $S^t$ be the set of strategies that are consistent with P2’s plays up to round $t$. Initially the set $S^1$ contains $2^{n(1 - \gamma)}$ strategies, and, for all $t$, $|S^{t+1}| \le |S^t|$.
The strategy $A$ for P1 is straightforward: at round $t$, P1 will play based on the most likely event: he will consider all strategies in $S^t$ and play H if the majority of strategies in $S^t$ use H at round $t$ and play T otherwise. Let $p^t$ be the exact fraction of strategies that are the majority at round $t$, $$\label{eq:pt}
p^t = \frac{\max\{ |\{b^t ~|~ b^t = H\}|,|\{b^t ~|~ b^t = T\}| \}}{|S^t|},$$ so that $p^t \in [1/2, 1]$. P1’s expected payoff at round $t$ is: $${\mathbf{E}\left[ h(a^t, b^t) \right]} = p^t - (1 - p^t) = 2 p^t - 1 \ge 0.$$ Thus P1’s average expected payoff is at least 0. To show that P1 can actually achieve an average expected payoff of $\gamma$ we need to consider the amount of information P1 gains at each round. We define the following potential function $\phi: \{1,
\dots, n\} \to \mathbb{R}$: $$\phi(t) = \sum_{k = 1}^{t-1} h(a^k, b^k) - \log |S^t|,$$ which considers both the accumulated payoff for P1 and the log-size of the set of consistent strategies. At time $t = 1$ there are $2^{n(1 - \gamma)}$ possible strategies for P2, so $\phi(1) = -n(1 - \gamma)$. We will now lower bound the expected increase in $\phi$ at each round. We can express this as $$\begin{aligned}
{\mathbf{E}\left[ \phi(t+1) - \phi(t) \right]} &= \left( {\mathbf{E}\left[ \sum_{k = 1}^{t} h(a^k, b^k) \right]} - {\mathbf{E}\left[ \sum_{k = 1}^{t-1} h(a^k, b^k) \right]}\right) \\
& \qquad \qquad- \left( {\mathbf{E}\left[ \log |S^{t+1}| \right]} - {\mathbf{E}\left[ \log|S^t| \right]} \right) \\
&= {\mathbf{E}\left[ h(a^t, b^t) \right]} - \left( {\mathbf{E}\left[ \log |S^{t+1}| - \log|S^t| \right]} \right) \\
&= 2p^t - 1 - \left( {\mathbf{E}\left[ \log |S^{t+1}| - \log|S^t| \right]} \right).
\end{aligned}$$ Now consider ${\mathbf{E}\left[ \log|S^{t+1}| \right]}$. When P1 looses he can eliminate a $p^t$ fraction of strategies, thus the new set $S^{t+1}$ will contain a $(1 - p^t)$ fraction of the strategies in $S^t$. On the other hand, when P1 wins, $|S^{t+1}| =
p^t|S^t|$. To complete the analysis we have to consider two cases, since if $p^t = 1$ then ${\mathbf{E}\left[ \log|S^{t+1}| \right]}$ is not well defined.
First assume $p^t = 1$. This happens when all feasible strategies for P2 have the same action at round $t$. In this case P1 will win with probability 1, and the size of the set of feasible strategies will stay the same. So, overall, the increase in $\phi$ will be 1.
Now assume $p^t \in [1/2, 1)$. Then, the expected size of the set $S^{t+1}$ is: $$\label{eq:st+1}
{\mathbf{E}\left[ \log |S^{t+1}| \right]} = \left( (1 - p^t) \log (1 - p^t) |S^t| + p^t \log p^t |S^t| \right).$$ The expected change in $\log |S^t|$ does not depend on $S^t$, but only on $p^t$, since $$\begin{aligned}
{\mathbf{E}\left[ \log |S^{t+1}| - \log |S^{t}| \right]} &= \left[ (1 - p^t) \log (1 - p^t) |S^t| \right. \\
& \left. \quad \qquad + p^t \log p^t|S^t| \right] - \log |S^t| \\
&= (1 - p^t) \log (1-p^t) + p^t \log p^t.
\end{aligned}$$ So that the overall change in potential when $p^t < 1$ is $${\mathbf{E}\left[ \phi(t+1) - \phi(t) \right]} \ge
2p^t - 1 - ( 1 - p^t) \log (1-p^t) - p^t \log p^t,$$ which is always at least $1$ for $p^t \ge 1/2$.
So, for all $p^t \in [1/2, 1]$, at each step the potential function $\varphi$ increases by at least $1$. Thus, after $n$ rounds we have that $$\begin{aligned}
{\mathbf{E}\left[ \phi(n) \right]} &\ge \phi(1) + \min_t \{ n {\mathbf{E}\left[ \phi(t+1) - \phi(t) \right]} \} \\
&\ge -n(1 - \gamma) + n = n\gamma.
\end{aligned}$$ Since P1’s expected payoff is at least ${\mathbf{E}\left[ \phi(n) \right]}$, this completes the proof.
This result immediately implies that, without any computational assumption, there can be no equilibrium with less than $n$ random coins.
For all $\gamma_1, \gamma_2 \in [0,1]$ such that $\gamma_1 +
\gamma_2 > 0$, if P1 and P2 have, respectively, $n(1 -
\gamma_1)$ and $n(1 - \gamma_2)$ random coins, then there can be no Nash equilibrium in the $n$-stage Matching Pennies repeated game.
Assume, by contradiction, that $(S_1, S_2)$ is such a Nash equilibrium. By Lemma \[lemma:unbounded1\], P1’s expected payoff must be ${\mathbf{E}\left[ U(S_1, S_2) \right]} \ge \gamma_2$ and P2’s payoff $-{\mathbf{E}\left[ U(S_1, S_2) \right]} \ge \gamma_1$, otherwise they would be better off by using the majority strategy in the proof of Lemma \[lemma:unbounded1\]. Summing the two inequalities we get $\gamma_1 + \gamma_2 \le 0$, a contradiction, since we assume $\gamma_1 + \gamma_2 > 0$.
However, if we limit the amount of randomness available to both players, we are still able to achieve an $\varepsilon$-Nash equilibrium. Furthermore, if the game has a $\varepsilon$-Nash equilibrium, then both players must have at least $(1 -
\varepsilon)n$ random coins.
\[lemma:delta-eq-tight\] Let $\gamma \in [0,1]$. The game has a $\gamma$-Nash equilibrium if and only if both players have $n(1 - \gamma)$ random coins.
For simplicity we assume $\gamma$ is the same for both players, however a similar result holds even in the case where the two players have a different amount of random coins.
To show the “only-if” implication, consider, by way of contradiction, a game that has a $\gamma$-Nash equilibrium $(S_1, S_2)$ but in which both players have less than $n(1 -
\gamma)$ random bits. Thus there must be a $\gamma' > \gamma$ such that they have exactly $n(1 - \gamma')$ random bits. Since $(S_1, S_2)$ is a $\gamma$-Nash equilibrium, it must be the case that, for any strategy $S_1'$ for P1 $${\mathbf{E}\left[ U(S_1, S_2) \right]} \ge {\mathbf{E}\left[ U(S_1', S_2) \right]} - \gamma.$$ By Lemma \[lemma:unbounded1\] we know that both players have a strategy that achieves a payoff of at least $\gamma' >
\gamma$, so that the above implies ${\mathbf{E}\left[ U(S_1, S_2) \right]} > 0$. Applying the same argument to P2, we get $-{\mathbf{E}\left[ U(S_1, S_2) \right]} >
0$, a contradiction.
The “if” part follows from Lemma \[lemma:delta-eq\] below.
\[lemma:delta-eq\] Let $\gamma \in [0,1]$. If both players have $n(1 - \gamma)$ random coins, then the game has a $\gamma$-Nash equilibrium.
For simplicity we assume $\gamma n$ is even.
Consider the following strategies: both player use their random coins to play uniformly at random for the first $n(1 - \gamma)$ rounds. Thereafter P1 will always play $H$, while P2 will alternate between $H$ and $T$, playing $H, T, \dots$. We claim that this is a $\gamma$-Nash equilibrium.
First notice that no player can improve his payoff in the first $n(1 - \gamma)$ rounds, given his opponent’s strategy. Let’s consider the remaining $n\gamma$ rounds. P1 could improve his payoff by playing $H,T,H,T\dots$, however this only increases his payoff by $\gamma$. This holds also for P2, that could play $T, T, \dots$, however gaining only $\gamma$.
Computationally Efficient Players {#sec:eff}
=================================
The proof of Lemma \[lemma:unbounded1\] in the previous section relies heavily the fact that we make no computational assumptions. In particular, to implement the majority strategy and compute $p^t$ in requires solving $\#$ hard problems, by reduction from $\#$. If we restrict the players to run in time polynomial in $n$ this particular strategy likely becomes unfeasible. In this setting, under reasonable complexity assumptions, it is possible to greatly reduce the amount of randomness and, at the same time, achieve a $\varepsilon$-Nash equilibrium.
We consider players’ whose actions are polynomial size Boolean circuits. A strategy is thus a circuit family $\{C_i\}_{i \in
\mathbb{N}}$, such that circuit $C_{l(n)}$ takes as input $l(n)$ random coins and outputs the $n$ actions to be played. Notice that this definition implies that each agent can simulate any of his opponent’s strategies.
We consider equilibria that use $n^{\delta}$ random coins for any $\delta >0$. Theorem \[thm:nash-iff-one-way\] shows that such $\varepsilon$-Nash equilibria exist if and only if one-way functions exist.
\[thm:nash-iff-one-way\] If players are bound to run in time polynomial in $n$, then, for all $\delta > 0$ and sufficiently large $n$, $\varepsilon$-Nash equilibria that use only $n^\delta$ random coins exist, where $\varepsilon = n^{-k}$ for all $k > 0$ and sufficiently large $n$’s, if and only if one-way functions exist.
The if part is Lemma \[one-way-to-nash\], while the only-if part is Lemma \[nash-to-one-way\].
As a preliminary result we show that, in our setting, the expected utility when at least one player uses a pseudorandom number generator can’t be too far from the expected utility when playing uniformly at random.
\[lemma\] Assume one-way functions exist, and let $G$ be the strategy corresponding to the output of a pseudorandom number generator. For any strategy $S$ that runs in time polynomial in $n$, for all $k > 0$ and sufficiently large $n$, $$\Big|{\mathbf{E}\left[ U(G, S) \right]} \Big| \le n^{-k} \text{ and } \Big|{\mathbf{E}\left[ U(S, G) \right]} \Big| \le n^{-k}.$$
We prove only the first inequality, the proof for the second one being symmetric.
Proof by contradiction. Assuming there is a $k > 0$ such that $\Big|{\mathbf{E}\left[ U(G, S) \right]}\Big|> n^{-k}$ for infinitely many $n$’s, we will construct a test $T$ for $G$, and show that $$\Big|
{\text{Pr}\left( T\left( G\left( U_{l(n)} \right) \right) = 1 \right)} -
{\text{Pr}\left( T\left( U_n \right) = 1 \right)}\Big| > n^{-c}
\label{eq:break}$$ for some $c > 0$ and infinitely many $n$’s, thus contradicting the assumption that $G$ is a pseudorandom number generator.
First consider the $n$ random variables $A_i(G, S)$, for $i = 1, \cdots, n$, where $A_i(G, S)$ is simply P1’s payoff at round $i$. Since $\sum_{i = 1}^n A_i(G, S) = n U(G, S)$, $$\sum_{i = 1}^n {\mathbf{E}\left[ A_i(G, S) \right]} > n^{1 - k}.$$ This implies that there must be an $i$ such that ${\mathbf{E}\left[ A_i(G, S) \right]}
> n^{-k}$. Fix that $i$.
The test $T$ takes as input an $n$-bit sequence $x$ and generates a sequence of plays $s$ according to strategy $S$. Now $T$ simulates an $n$-stage repeated Matching Pennies game with strategies $(x,s)$. If P1 wins the $i$-th round then it will output 1, otherwise the output will be 0. In other words, $T$ outputs 1 if and only if $A_i(x,s) = 1$. Notice that $T$ runs in time polynomial in $n$.
When $x$ is drawn from the uniform distribution, P1 will win with probability $1/2$, or ${\text{Pr}\left( T(U_n) = 1 \right)} = 1/2$.
Now notice, that since $A_i \in \{-1, 1\}$, ${\text{Pr}\left( A_i = 1 \right)} =
\frac{{\mathbf{E}\left[ A_i \right]} + 1}{2}$. This implies that $${\text{Pr}\left( T(G(U_{l(n)})) = 1 \right)} = \frac{{\mathbf{E}\left[ A_i \right]} + 1}{2} > \frac{1}{2n^{k}} + \frac{1}{2}.$$ Thus $$\Big|
{\text{Pr}\left( T\left( G_1\left( U_{l(n)} \right) \right) = 1 \right)} -
{\text{Pr}\left( T\left( U_n \right) = 1 \right)}\Big| > \frac{1}{2n^{k}},$$ which proves the lemma.
\[one-way-to-nash\] If one-way functions exist and players are bound to run in time polynomial in $n$, then for every $\delta, k > 0$ and for sufficiently large $n$, the $n$-stage has an $n^{-k}$-Nash equilibrium in which each player uses at most $n^\delta$ random bits.
Assume, by contradiction, that one-way functions exist but there are values $\delta > 0$ and $k > 0$ such that the game has no $n^{-k}$-Nash equilibrium in which players use at most $n^\delta$ random coins.
Since we assume one-way functions exist, by Theorem \[hillthm\] there exist pseudorandom number generators that use $n^\delta$ coins. Assume both players use the output of such pseudorandom number generators as their strategies (which we call, respectively, $G_1$ and $G_2$). Since we are assuming that this is not a $n^{-k}$-Nash equilibrium, one of the players, say Player 1, must have a strategy $A$ such that $$\label{eq:nonash}
{\mathbf{E}\left[ U\left( A, G_2 \right) \right]} > {\mathbf{E}\left[ U\left( G_1, G_2 \right) \right]} + n^{-k}.$$ By Lemma \[lemma\] we can choose a $k' > 0$ such that $$\label{eq:nonash1}
{\mathbf{E}\left[ U\left( A, G_2 \right) \right]} > -n^{-k'} + n^{-k}.$$ Pick $c = k' = k+1$, so that ${\mathbf{E}\left[ U\left( A, G_2 \right) \right]} >
n^{-c}$ for $n > 2$. This contradicts Lemma \[lemma\], proving the claim.
We now prove the opposite direction, that is that the existence of Nash equilibria that use few random bits implies the existence of one-way functions.
\[nash-to-one-way\] If for every $\delta > 0$ there is a Nash equilibrium in which each player uses $n^\delta$ random bits and runs in time polynomial in $n$, then one-way functions exist.
Let $(A,B)$ be such a Nash equilibrium and assume, by contradiction, that one-way functions don’t exist. This implies that pseudorandom number generators can’t exist (Goldreich [@goldreich2004foundations]), and so, $A$ and $B$ can’t be sequences that are computationally indistinguishable from uniform.
Thus, by Yao’s theorem (Theorem \[yaothm\]), we know that there are polynomial size circuit families $\left\{ C_i
\right\}_{i \in \mathbb{N}}$ and $\left\{ D_i \right\}_{i \in
\mathbb{N}}$ such that $$\begin{aligned}
{\text{Pr}\left( C_i(y_1, \dots, y_{i}) = y_{i+1} \right)} &> 1/2 + \delta_1 \\
{\text{Pr}\left( D_i(z_1, \dots, z_{i}) = z_{i+1} \right)} &> 1/2 + \delta_2,
\end{aligned}$$ for some $\delta_1, \delta_2 > 0$, where $A(x_1, \dots,
x_{l_1(n)}) = (y_1, y_2, \dots, y_{n})$ and $B(x_1, \dots,
x_{l_2(n)}) = (z_1, z_2, \dots, z_{n})$.
To get a contradiction it is sufficient to show that players are better off by using the predictor circuits $C$ and $D$. Consider Player 1: using $D$, at each round he can guess, given the previous history, the opponent’s next move with probability $\frac{1}{2} + \delta_1$. Thus his expected payoff at any round $t$ is $${\mathbf{E}\left[ h(d^t, b^t) \right]} > \frac{1}{2} + \delta_1 - \frac{1}{2} + \delta_1 = 2\delta_1,$$ where the expectation is over the internal coin tosses of the predictor circuit $D$. The overall expected payoff is $${\mathbf{E}\left[ U(D, B) \right]} > \frac{1}{n} \sum_{t = 1}^n 2\delta = 2\delta_1.$$ Now, let $w = {\mathbf{E}\left[ U(A, B) \right]}$ be the value of the expected payoff when players play $(A,B)$. Consider the following cases:
1. $w \le 0$: this implies that Player 1 could gain $2\delta_1$ by using strategy $D$,
2. $w > 0$: by definition Player 2’s expected payoff is $-{\mathbf{E}\left[ U(A, B) \right]} < 0$, so Player 2 can achieve a higher payoff by using his predictor circuit $C$,
In both cases we see that $(A, B)$ can’t be a Nash equilibrium, a contradiction.
Exchanging Time for Randomness {#sec:nw}
==============================
In this section we determine conditions under which a $\varepsilon$-Nash equilibrium can arise, given that one of the players has only a logarithmic amount of randomness and his opponent must run in time $n^k$ for some fixed $k$. This shows how we can trade off randomness for time; the player with $O(\log n)$ random bits runs in time polynomial in $n$, while the player with more random bits runs in fixed polynomial time.
\[thm:logn-vs-poly\] Assume there exists $f \in \textsf{DTIME}(2^{O(l(n))})$ and $\varepsilon > 0$ such that no circuit of size at most $2^{\varepsilon l(n)}$ can compute $f$ and that one-way functions exist. Let Player 1’s strategies be circuits of size at most $n^k$ that use at most $n^{\delta}$ random bits for some $k > 2 + c \delta$, where $c \ge 1$ is a constant related to the implementation of a cryptographic pseudorandom number generator. Assume Player 2 has access to only $M \log n$ random bits. As long as $M > Ck$, where $C$ is the constant in Theorem \[thm:nw\], then for all $\varepsilon > 0$ and sufficiently large $n$ there is a $\varepsilon$-Nash equilibrium.
Let $G_1$ be the cryptographic pseudorandom number generator available to Player 1 and $G_2$ be the complexity pseudorandom number generator used by player 2. Furthermore let $\mathcal{S}_1$ be the set of all possible strategies for P1 (for all $\delta > 0$ circuits of size at most $n^k$ that use $n^{\delta}$ random bits), and $\mathcal{S}_2$ the set of strategies available to P2 (polynomial size circuit families and $M \log n$ random coins). We will show that for all $\varepsilon > 0$ and sufficiently large $n$, $(G_1, G_2)$ is a $\varepsilon$-Nash equilibrium with the required properties.
First we argue that, for all $\gamma > 0$ and sufficiently large $n$, $|{\mathbf{E}\left[ U(G_1, G_2) \right]}| < \gamma$. The proof of this fact is similar to the proof of Lemma \[lemma\], showing by way of contradiction, that if $|{\mathbf{E}\left[ U(G_1, G_2) \right]}| \ge \gamma$ then we can build a test for the cryptographic pseudorandom number generator $G_1$.
Now we show that, for the appropriate setting of parameters, $G_1$ fools $\mathcal{S}_2$ and $G_2$ fools $\mathcal{S}_1$. For any $k$, Player 2 can fool circuits of size $n^k$ by using $C \log n^k = Ck \log n$ random bits. So, for $M > Ck$, $G_2$ fools $\mathcal{S}_1$. Notice also that since Player 2 runs in time $O(n^{Ck})$, the cryptographic pseudorandom number generator $G_1$ fools $\mathcal{S}_2$. Let $h$ be the one-way permutation used by the pseudorandom number generator $G_1$, and assume $h$ us computable in time $n^c$ for some $c > 0$. Given $h$, $G_1$ is defined as follows: let $x,y \in
\{0,1\}^{\frac{n^{\delta}}{2}}$, and let $(x,y)$ be $G_1$’s seed (notice that $|(x,y)| = n^\delta$), then $$G_1(x, y) = \left( f^n(x) \odot y, f^{n-1}(x) \odot y, \dots, f(x) \odot y \right),$$ where $x \odot y = \sum_i x_i y_i \mod 2$. There are $O(n^{2})$ applications of $h$, so $G_1$ runs in time $O(n^{2 +
c\delta})$. So, for $k \ge 2 + c\delta$, $G_1$ fools $\mathcal{S}_2$.
At this point we’re almost done. As in Lemma \[one-way-to-nash\] assume, by contradiction, that the assumptions in the theorem hold but $(G_1, G_2)$ is not a $\varepsilon$-Nash equilibrium for some $\varepsilon > 0$. This implies that at least one of the two players can improve his expected payoff by more than $\varepsilon$ by switching to some other strategy. First consider P2, and assume there is a strategy $S_2 \in \mathcal{S}_2$ such that $${\mathbf{E}\left[ U(G_1, S_2) \right]} > {\mathbf{E}\left[ U(G_1, G_2) \right]} + \varepsilon.$$ As in Lemma \[one-way-to-nash\] this implies that $S_2$ would be a test for the cryptographic pseudorandom number generator $G_1$, contradicting the fact that $G_1$ fools $\mathcal{S}_2$. Similarly, assume P1 has a strategy $S_1 \in \mathcal{S}_1$ such that $${\mathbf{E}\left[ U(S_1, G_2) \right]} > {\mathbf{E}\left[ U(G_1, G_2) \right]} + \varepsilon.$$ Again, $S_1$ can easily be made into a test for $G_2$, contradicting the fact that $G_2$ fools $\mathcal{S}_1$.
Infinite Play {#sec:infinite}
=============
We now consider an infinitely repeated game of Matching Pennies, and show that, if utilities are time discounted, we can always achieve a $\varepsilon$-Nash equilibria using a large enough (but finite) amount of random coins. First we determine the least amount of randomness required to achieve a $\varepsilon$-Nash equilibrium in the general, i.e. computationally unbounded, case.
\[lemma:infinite1\] For all discount factors $\delta \in (0,1)$ and all $\varepsilon > 0$, there is an $\varepsilon$-Nash equilibrium in which the players use only $n$ random bits, for $n >
\frac{\log \varepsilon (1- \delta)}{\log \delta}$.
Given $\delta, \varepsilon > 0$ consider the following strategies: both players play the Nash equilibrium strategy for the first $n$ rounds. After this P1 will always play $H$, while P2 will play $H$ and $T$ alternatively. The overall expected payoff is 0. However, after round $n$, both players could switch to a strategy that always wins, achieving a total expected payoff of $0 + \sum_{t = n}^{\infty} \delta^t =
\frac{\delta^n}{1 - \delta}$. To ensure that our strategies are indeed a $\varepsilon$-Nash equilibrium we just need to make sure that $$0 > \frac{\delta^n}{1 - \delta} -\varepsilon.$$ Rearranging and taking logarithms we get $n > \frac{\log
\varepsilon (1- \delta)}{\log \delta}$.
Now we consider players’ whose strategies are families of polynomial size Boolean circuits (as in Section \[sec:eff\]), and assume one-way functions exist. We first give a version of Lemma \[lemma\] for time discounted utilities on a finite number of rounds.
\[lemma:discounted\] Assume one-way functions exist, and let $G = (g^1, \dots, g^n)$ be the strategy corresponding to the output of a cryptographic pseudorandom number generator. Let $S = (s^1, \dots, s^n)$ be any strategy. For all $\delta \in (0,1)$, $k > 0$ and for sufficiently large $n$ $$\Big| {\mathbf{E}\left[ \sum_{t = 1}^n \delta^t h(g^t, s^t) \right]} \Big| \le n^{-k} \text{ and } \Big| {\mathbf{E}\left[ \sum_{t = 1}^n \delta^t h(s^t, g^t) \right]} \Big| \le n^{-k}.$$
Again we give the proof only for the first inequality. Assume, by contradiction, that $| {\mathbf{E}\left[ \sum_{t = 1}^n \delta^t h(g^t,
s^t) \right]} | > n^{-k}$ for some $\delta$, $k$ and infinitely many $n$’s. Consider the random variables $A_1(c_1^1, c_2^1), \dots,
A_n(c_1^n, c_2^n)$, defined as $A_t(c_1^t, c_2^t) = 1$ if P1 wins round $t$ when playing according to $(C_1, C_2)$ and 0 otherwise. Let $A(C_1, C_2) = \sum_t \delta^t A_t(c_1^t,
c_2^t)$, so that ${\mathbf{E}\left[ A(G, S) \right]} \ge |{\mathbf{E}\left[ \sum_{t = 1}^n \delta^t
h(g^t, s^t) \right]}| > n^{-k}$. This implies that there is a $t$ such that $\delta^t {\mathbf{E}\left[ A_t(g^t, s^t) \right]} > n^{-k-1}$, which implies ${\mathbf{E}\left[ A_t(g^t, s^t) \right]} > n^{-k-1}$. Fix that $t$. As in Lemma \[lemma\], consider the test $T$ that, given a sequence of plays $x$, generates a play $s$ from $S$ and outputs 1 if P1 wins round $t$ and outputs 0 otherwise.
When $x$ is drawn uniformly at random, ${\text{Pr}\left( T(U_n) = 1 \right)} = 1/2$. On the other hand, when $x$ is $G$’s output, $${\text{Pr}\left( T(G(U_{l(n)})) = 1 \right)} = {\mathbf{E}\left[ A_t \right]} > n^{-k-1}.$$ Now $$\Big| {\text{Pr}\left( T(G(U_{l(n)})) = 1 \right)} - {\text{Pr}\left( T(U_n) = 1 \right)} \Big| > \frac{1}{2} - \frac{1}{n^{k+1}} = \frac{1}{n^c},$$ for $c > - \frac{\log(1/2 - n^{-1-k})}{\log n} \ge 0$, contradicting the assumption that $G$ is a pseudorandom number generator.
Using the above Lemma we can show that, for all discount factors, we can greatly reduce the amount of random coins needed to get an $\varepsilon$-Nash equilibrium.
\[lemma:infinite2\] For all discount factors $\delta \in (0,1)$, all $\varepsilon >
0$ and all $\xi > 0$, there is a $n^{-k}$-Nash equilibrium in which players use only $n^\xi$ random coins, for sufficiently large $n$’s.
As in the proof of Lemma \[lemma:infinite1\] we consider the following strategy for both players: for the first $n$ rounds play the output of a cryptographic pseudorandom number generator $G$, with seed length $n^\xi$. Thereafter P1 will always play $H$, while P2 will alternate between $H$ and $T$. Pick any $k > 0$, we now show that this is a $n^{-k}$-Nash equilibrium. By Lemma \[lemma:discounted\] we can pick $k' = k/2$ such that the expected utility in the first $n$ rounds lies in the interval $[-n^{-k'}, n^{-k'}]$. To ensure that this is a $n^{-k}$-Nash equilibrium we just need to show that $$-n^{-k'} > n^{-k'} + \frac{\delta^n}{(1 - \delta)} - n^{-k},$$ or $$-2n^{-k'} + n^{-2k'} > \frac{\delta^n}{(1 - \delta)}.$$ Now, for any $c > 0$, if we set $k' = \frac{\log\left(
(\sqrt{c+1} - 1)/c \right)}{\log n}$, then the left hand side of the above inequality is $c$, so that it always holds for sufficiently large $n$’s.
Thus, given any $n$ that satisfies the conditions in Lemma \[lemma:infinite1\], there can be a $n^{-k}$-Nash equilibrium using $n^{\delta}$ coins, for any $\delta > 0$. To see this, consider an $m$ sufficiently large so that Lemma \[lemma:infinite2\] holds and pick $\xi = \delta \frac{\log
n}{\log m}$.
Conclusions
===========
We have shown how, in a simple setting, reducing the amount of randomness available to players affects Nash equilibria. In particular, if we make no computational assumptions on the players, there is a direct tradeoff between the amount of randomness and the approximation to a Nash equilibrium we can achieve. If, instead, players are bound to run in polynomial time, we can get very close to a Nash equilibrium with only $n^{\delta}$ random coins, for any $\delta > 0$.
Some directions for future research include:
- Is it possible to extend Lemma \[lemma:unbounded1\] to $m$ player games, for $m > 2$? Notice that the strategy used in that proof does not generalize to this setting.
- Under what circumstances is it possible to further reduce the amount of randomness available (say to $O(\log n)$ for both players)?
- Is it possible to extend these results to general zero-sum games or even non zero-sum games?
#### Acknowledgments
We wish to thank Tai-Wei Hu, Peter Bro Miltersen, Rahul Santhanam and Rakesh Vohra for fruitful discussions.
[10]{}
S. Arora and B. Barak. . Cambridge University Press, 2009.
X. Chen and X. Deng. . In [*47th FOCS*]{}, 2006.
C. Daskalakis, P. W. Goldberg, and C. H. Papadimitriou. The complexity of computing a nash equilibrium. , 39(1):195–259, 2009.
L. Fortnow and R. Santhanam. . In [*1st ICS*]{}, 2010.
O. Goldreich. . Cambridge University Press, 2004.
O. Gossner and T. Tomala. . , 44(1):24–32, 2008.
J. Y. Halpern and R. Pass. Game theory with costly computation: Formulation and application to protocol security. In [*1st ICS*]{}, pages 120–142, 2010.
J. Hastad, R. Impagliazzo, L. Levin, and M. Luby. . , 28(4):1364–1396, 1999.
T.-W. Hu. . <http://bit.ly/e4N8cN>, 2010. .
R. Impagliazzo and A. Wigderson. Randomness vs time: Derandomization under a uniform assumption. , 63(4):672–688, 2001.
S. Kalyanaraman and C. Umans. . In [*15th ESA*]{}, 2007.
N. Nisan and A. Wigderson. . , 49(2):149–167, 1994.
H. Simon. . , 69(1):99–118, 1955.
A. Yao. . In [*23rd FOCS*]{}, 1982.
[^1]: This work was done while the author was visiting Northwestern University, Department of Electrical Engineering and Computer Science.
[^2]: This research was partially supported by NSF grants CCF-0829754 and DMS-0652521.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We report theoretical and numerical evaluations of the phase diagram for a model of patchy particles. Specifically we study hard-spheres whose surface is decorated by a small number $f$ of identical sites (“sticky spots”) interacting via a short-range square-well attraction. We theoretically evaluate, solving the Wertheim theory, the location of the critical point and the gas-liquid coexistence line for several values of $f$ and compare them to results of Gibbs and Grand Canonical Monte Carlo simulations. We study both ordered and disordered arrangements of the sites on the hard-sphere surface and confirm that patchiness has a strong effect on the phase diagram: the gas-liquid coexistence region in the temperature-density plane is significantly reduced as $f$ decreases. We also theoretically evaluate the locus of specific heat maxima and the percolation line.'
author:
- Emanuela Bianchi
- Piero Tartaglia
- Emanuela Zaccarelli
- Francesco Sciortino
bibliography:
- './A712051JCP.bib'
title: 'Theoretical and numerical study of the phase diagram of patchy colloids: ordered and disordered patch arrangements'
---
INTRODUCTION
============
Patchy particles are particles interacting via a limited number of directional interactions. The anisotropy of the interaction leads to collective behaviors different from those of simple liquids. Gelation [@Trappe; @Cipelletti; @zaccajpcm; @advances; @genova], gas-liquid phase separation [@Zacca1; @bian], crystallization [@Wilber-2006; @doye; @octahedral] and clustering are strongly affected by patchiness [@Glotz_Solomon_natmat; @zhangglotzer; @Zhang_03; @simone; @bastiaan]. Recently, a new generation of colloidal particles with chemically or physically patterned surfaces has been designed and synthesized in the attempt to provide valence to colloids [@Manoh_03; @mohovald; @Blaad06; @Cho_05; @Glotz_Solomon; @Glotz_Solomon_natmat]. This relevant synthesis effort aims to generate super atoms — atoms at the nano and micro-scopic level — in order to reproduce and extend the atomic and molecular behavior on larger length scale. It also offers the possibility to export the supra-molecular chemistry ideas [@leibler; @lehn1; @lehn2] to new colloidal materials, opening the new field of supra-particle colloidal physics. Thus, a general effort to develop a deeper understanding of self-assembly and to construct a more unified theoretical underpinning for this technologically and scientifically important field is crucial. The outcome of this effort may also have an impact in our understanding of the phase behavior of protein solutions, due to their intrinsic patchy character [@Lomakin; @Sear_99; @KumarJCP07; @McManus].
Our recent work [@bian] has shown that the Wertheim theory [@Werth1; @Werth2] describes rather well the critical properties of particles decorated on their surface by a predefined number of attractive sites. The Wertheim theory is a thermodynamic perturbation theory introduced to describe association under the hypothesis of a single-bond per patch, which means that an attractive site on a particle cannot bind simultaneously to two (or more) sites on another particle. The single-bond per patch condition can be naturally implemented in colloids by choosing an appropriate small ratio between the range of the attractive patches and the particles size. The single-bonding condition results also from DNA complementarity [@dna; @dnastarr] as well as from complementary “lock-and-key” interactions associated to biological specificity [@hiddessen1; @hiddessen2]. These types of interactions provide a versatile way of controlling inter-particle binding. An extension of the Wertheim thermodynamic perturbation theory to interpret and/or predict the behavior of a wide range of substances with potential industrial applications is provided by the statistical associating fluid theory (SAFT) [@Chap1; @Chap2] and by its developments [@Saft-vr; @Saft-vre].
In previous works we have shown [@Zacca1; @bian; @23] that for patchy colloidal particles with a small number of sticky sites the critical point of the gas-liquid phase separation moves towards small packing fraction ($\phi$) and temperature ($T$) with decreasing the number of patches. According to this study, liquid phases of vanishing density can be generated once a small fraction of polyfunctional particles is added to a system of bifunctional ones. Indeed, the study of binary mixtures of patchy particles with different functionalities allows to explore also the range of non-integer valence down to $2$. This means that with the new generation of non-spherical sticky colloids [@Manoh_03; @Cho_05], it should be possible to realize “empty liquids” [@bian], i.e. states with an arbitrarily small occupied packing fraction at temperature lower than the liquid-gas critical temperature. The shift with valence of the critical point, both in density and temperature, leads to substantial changes in phase behavior with branching: the reduction of the number of bonded nearest neighbours is accompanied by an enlargement of the region of stability of the liquid phase in the $(T,\phi)$ plane. This fact could favor the establishment, at low $T$ and at small $\phi$, of homogeneous disordered materials, i.e. equilibrium disordered states in which particles are interconnected in a persistent network. At such low $T$, the bond-lifetime will become comparable to the experimental observation time. Under these conditions, it should be possible to approach dynamical arrested states continuously from equilibrium and to generate a state of matter as close as possible to an ideal gel [@genova; @zaccajpcm].
In this article we extend the preliminary study of Ref. [@bian] reporting a Monte Carlo investigation of the $f$-dependence of the critical point location for a model with a disordered arrangement of patches. The present study confirms the trend discussed in Ref. [@bian] for the corresponding ordered case. To evaluate the role of the valence on the coexistence region, we also numerically investigate the shape of the gas-liquid binodal line for the ordered case and compare it with theoretical predictions based on the Wertheim theory. Finally, we analytically calculate several equilibrium properties, such as the energy per particle, the specific heat, the extent of polymerization and the percolation line, to get insights on their $f$-dependence.
We find that the reduction of valence is accompanied by a significant shift of the coexistence curve towards low temperature and density. The percolation line is always found to lie above the critical point, merging with the gas-liquid spinodal at low density $\rho$. The liquid state is thus always characterized by an infinite spanning network. This confirms the possibility of observing, for large attraction strengths, dynamical arrested states driven by bonding (as opposed to packing) in single phase conditions, i.e. homogenous arrested states at low density.
We also provide in the Appendix A a physical insight of the Wertheim theory by showing that the theoretical expression for the free energy in Ref. [@Werth1; @Werth2] is formally equivalent to the free energy of a system of non-interacting clusters distributed according to the Flory-Stockmayer cluster size distribution [@flory].
The model {#sec:model}
=========
We focus on a system of particles modeled as hard-spheres of diameter $\sigma$, whose surface is decorated by $f$ bonding sites at fixed locations. Sites on different particles interact via a square-well potential. The interaction $V({\bf 1,2})$ between particles [**1**]{} and [**2**]{} is $$V({\bf 1,2})=V_{HS}({\bf r_{12}})+\sum_{i=1}^f\sum_{j=1}^f V_{SW}({\mathbf r}^{_{ij}}_{_{12}})$$ where the individual sites are denoted by $i$ and $j$, $V_{HS}$ is the hard-sphere potential, $V_{SW}(x)$ is a square-well interaction (of depth $-u_0$ for $x \leq \delta$, 0 otherwise) and ${\bf r_{12}}$ and ${\mathbf r}^{_{ij}}_{_{12}}$ are respectively the vectors joining the particle-particle and the site-site (on different particles) centers. Geometric considerations for a three touching spheres configuration show that the choice $\delta=0.5(\sqrt{5-2\sqrt{3}}-1) \sigma \approx 0.119 \sigma$ guarantees that each site is engaged at most in one bond. Hence, with this choice of $\delta$, each particle can form only up to $f$ bonds. We note that in this model bonding is properly defined: two particles are bonded when their pair interaction energy is -$u_0$. Distances are measured in units of $\sigma$. Temperature is measured in units of the potential depth (i.e. Boltzmann constant $k_B=1$).
We study two different arrangements of the $f$ sites on the particles surface. In the first case sites are arranged in a regular structure (see Fig. 1 of Ref. [@bian]). In the second case, the distribution of the sites is random and different for each particle. In this latter case, the only constraint on the site position is formulated on the basis of a minimum distance $d_{min}$ criterion between different sites on the same particle: the choice of $d_{min}$ aims to minimize the possibility of double bonding between the same pair of particles as well as the shading of a bonding site by the presence of a nearby bonded site. We choose $d_{min}=0.4$.
The theory
==========
The first-order thermodynamic perturbation Wertheim theory [@Werth1; @Werth2; @Hansennew] provides an expression for the free energy of particles with a number $f$ of attractive sticky sites on their surface, independently from the specific geometric arrangement of the sites. The theory assumes that all sites have the same probability of forming bonds and that the correlation between adjacent sites is missing.
The Helmholtz free energy of the system is written as a sum of the hard-sphere reference free energy, $F^{HS}$, plus a bond contribution, $F^{bond}$. The Helmholtz free energy due to bonding derives from a summation over certain classes of relevant graphs in the Mayer expansion [@Hansennew]. In the sum, closed loops graphs are neglected. The fundamental assumption of the Wertheim theory is that the conditions of steric incompatibilities are satisfied: (i) no site can be engaged in more than one bond and (ii) no pair of particles can be double bonded. These steric incompatibilities are satisfied in both our models thanks to (i) the small $\delta$ chosen for the short-ranged square-well attraction and to (ii) the location of the sticky sites on the hard-sphere particles surface. In the formulation of Ref. [@Chap1], the bond free energy density of a system of $f$-functional particles is $$\label{eq:f}
\frac{\beta F^{bond}}{V} = \rho \ln (1-p_b)^f +\frac{1}{2}\rho f p_b$$ where $\beta=1/k_BT$, $\rho=N/V$ is the particle number density and $p_b$ is the bond probability. Since we assume equal reactivity for all sites, the bonding process can be seen as a chemical reaction between two unsaturated sites in equilibrium with a pair of bonded sites. In this respect one can write $$\label{eq:pb}
\frac{p_b}{(1-p_b)^2}= \rho e^{-\beta {\cal F}_b}$$ where ${\cal F}_b$ is the site-site bond free-energy, i.e. the free energy difference between the bonded and the unbonded state.
The Wertheim theory predicts an expression for ${\cal F}_b$ in term of liquid state correlation functions and spherically averaged Mayer functions. More precisely $$\label{eq:deltaf}
e^{-\beta {\cal F}_b} = f \Delta$$ where $\Delta$ refers to a single site-site interaction (since all bonding sites are identical) and it is defined as $$\label{eqn:Deltabis}
\Delta= 4 \pi \displaystyle \int_{\sigma}^{\sigma+\delta} {g_{HS}(r_{12})\langle f(12)\rangle _{\omega_{1},\omega_{2}} r_{12}^{2} d r_{12}}.$$ Here $g_{HS}(r_{12})$ is the reference hard-sphere fluid pair correlation function, the site-site Mayer function is $f(12)=\exp[-V_{SW}({\mathbf r}^{_{ij}}_{_{12}})/k_{B}T]-1$, and $\langle f(12)\rangle _{\omega_{1},\omega_{2}}$ represents an angle-average over all orientations of particles 1 and 2 at fixed relative distance $r_{12}$. Since the Wertheim theory is insensitive to the location of the attractive sites, the number of interacting sites on each particle is encoded only in the factor $f$ before $\Delta$ in Eq. \[eq:deltaf\]. For a site-site square-well interaction, the Mayer function can be calculated as [@Werth5] $$\langle f(12)\rangle _{\omega_{1},\omega_{2}} = \left [
\exp(\beta u_0)-1 \right ] S(r)$$ where $S(r)$ is the fraction of solid angle available to bonding when two particles are located at relative center-to-center distance $r$ $(r\equiv r_{12})$, i.e. $$S(r)=\frac{(\delta + \sigma -r)^2 (2 \delta - \sigma +r )}{ 6 \sigma^2 r}.$$ The evaluation of $\Delta$ requires only an expression for $g_{HS}(r)$ in the range where bonding occurs ($\sigma<r<\sigma+\delta$). We use the linear approximation [@Nez_90] $$g_{HS}(r)= \frac{1-0.5 \phi}{(1- \phi)^3}-\frac{9}{2}\frac{\phi (1+\phi)}{(1- \phi)^3} \left[\frac{r-\sigma}{\sigma}\right]
% \hspace{1cm} (\mbox{where} \hspace{0.1cm} \phi=\frac{\pi}{6} \sigma^3 \rho),
\label{eq:ghsr}$$ (where $\phi=\frac{\pi}{6} \sigma^3 \rho$), which provides the correct Carnahan-Starling [@CS_69] value at contact. This gives $$\begin{aligned}
\label{eq:deltamodel}
\Delta=
%\\ \nonumber
\frac{V_b (e^{\beta u_0} - 1)}{ (1-\phi )^3} \times \\ \nonumber
\left[1-\frac{5}{2}\frac{\left(3\sigma^2+8 \delta\sigma +3 \delta ^2\right) }{ \sigma(15\sigma+4 \delta ) }\phi -\frac{3}{2}\frac{\left(12 \delta \sigma +5 \delta
^2\right) }{\sigma (15\sigma +4 \delta ) }\phi ^2\right]
%\\ \frac{[( 0.0928861 \rho+1)( 0.431176 \rho-1)]}{(0.5236\rho - 1)^3}.\end{aligned}$$ where we have defined the spherically averaged bonding volume $V_b \equiv 4 \pi \int_\sigma^{\sigma+\delta} S(r) r^2 dr =\pi \delta^4 (15\sigma +4 \delta)/30\sigma^2$. We note that the above expression of $\Delta$ simplifies in the low density limit. Indeed, when $\rho \rightarrow 0$, the hard-sphere pair correlation function tends to the ideal gas limit value $g_{HS}(r) \approx 1$. In this limit $\Delta$ doesn’t depend on $\rho$, i.e. $\Delta=V_b (e^{\beta u_0} - 1)$. We note that bonding takes approximatively place when $\exp(\beta u_0) \gg 1$. Indeed bond formation arises from a balance between the energetic gain of forming a bond ($\Delta U_b=-u_0$) and an entropy loss ($\Delta S_b$), which is expressed in the theory as logarithm of the ratio between $V_b$ and the volume per bonding site, $V/(f N)$ [@M2]. Since $V_b \ll V/(fN)$, bonding becomes relevant when $\beta u_0 \gg 1$. In the following we will thus approximate ($e^{\beta u_0} - 1$) with $e^{\beta u_0}$ to simplify the theoretical expressions.
Once the free energy is known, it is possible to derive various equilibrium properties of the system through thermodynamic relations. We find expressions for the energy per particle, the specific heat maxima, the extent of polymerization and the pressure of the system in terms of $p_b$, which is a function of $T$ and $\rho$ from Eq. \[eq:pb\]. The potential energy per particle $E/N$ is given by $$\label{eqn:energy}
\frac{E}{N}= \frac{\partial (\frac{\beta F^{bond}}{N})}{\partial \beta}=-\frac{1}{2} f u_0 p_b$$ i.e. it is exactly the fraction of bond times $-u_0 f/2$. The constant volume specific heat $C_V$ can be calculated as $$\label{eq:cv}
C_V= \frac{\partial (\frac {E}{N})}{\partial T}= \frac{1}{2}f \frac{u_0^2}{T^2} \frac{p_b(1-p_b)}{1+p_b}.$$ At each $\rho$, the specific heat has a maximum (whose amplitude increases with $f$) at finite $T$, which defines a line of specific heat extrema in the $(T,\rho)$ plane. The $C_V^{max}$ line can be used as an estimate of the polymerization transition line [@Greer88; @Greer96; @Greer02; @Milchev98; @douglas; @Dudo_03].
In the characterization of the self-assembly of particles, experimentalists often consider a quantity called extent of polymerization $\Phi(t)$, which is normally measured by spectroscopy. $\Phi$ is defined as the fraction of particles bonded in clusters. This quantity plays the role of order parameter in the polymerization transition: it changes continuosly form the value zero at high $T$, when all particles are in the monomeric state, to the value one at low $T$, when particles are bonded in clusters. This crossover becomes sharper and sharper on decreasing $\rho$. Since the monomer density is simply obtained by the observation that all of the sticky spots on each particle must be unbonded, i.e. $\rho_{1} = \rho (1-p_b)^f$, the extent of polymerization is given by $$\Phi= \frac{\rho -\rho_{1}}{\rho} = 1- (1-p_b)^f.$$ As a function of $p_b$, the branched polymerization transition becomes sharper and sharper on increasing the functionality of the system.
The pressure $P$ of the system can be evaluated by deriving, respect to the volume, the Wertheim free energy, i.e. $\beta P=- (\partial \beta F/\partial V)_T$. The bonding contribution to $P$ is thus $$\label{eq:Pbond}
\frac{\beta P^{bond}}{\rho}= \rho f \left[\frac{1}{2}-\frac{1}{1-p_b}\right]\frac{\partial p_b}{\partial \rho}.$$ In the low $\rho$ limit ($g_{HS}(r) \approx 1$), it is possible to neglect the $\rho$ dependence of $\Delta$ and $\beta P^{bond}/ \rho$ becomes equal to $ -\frac{1}{2} f p_b$. Appendix A provides a physical understanding of this expression. The hard-sphere contribution to the pressure can be evaluated via the Carnahan-Starling equation of state [@CS_69] $$\label{eq:betaPhs}
\frac{\beta P^{HS}}{\rho}= \frac{(1 + \phi + \phi^2 - \phi^3)}{(1 - \phi)^3}.$$
From the resulting $V$ and $T$ dependence of $P$, it is possible to evaluate the liquid-gas coexistence region in the phase diagram, by solving the following set of equations $$\begin{aligned}
\label{eq:coex}
\nonumber
T_g = T_{l} \equiv T^* \\
P_g = P_{l} \equiv P^* \\
\nonumber
\int_{V_l}^{V_g}[P(V,T^*)-P^*] dV = 0,\end{aligned}$$ where $T_g, P_g, V_g$ and $ T_l, P_l, V_l$ are respectively the temperature, the pressure and the volume of the two coexisting phases. The third equation corresponds to the Maxwell construction.
The main assumption of the Wertheim theory is that molecules (or particles) cluster in open structures without closed bond loops. The hypothesis of absence of closed bonding loops is also at the heart of the Flory-Stockmayer theory, developed to model aggregation in chemical gelation. The Flory-Stockmayer theory [@flory] provides expressions for the number density of clusters of $n$ particles, $\rho_n$, as a function of the bond probability (the extent of the reaction in the Flory-Stockmayer language). For functionality $f$ $$\begin{aligned}
\label{eq:flory}
\rho_n &=& \rho (1-p_b)^f \left[ p_b (1-p_b)^{f-2}\right]^{n-1} \omega_n\\
\nonumber
\omega_n &=& \frac{f (f n-n)!}{(fn-2n+2)!n!}\end{aligned}$$ where $\rho \equiv \sum_{n} n \rho_n= N/V$ is the total number density. In Appendix A, we show that the Wertheim free-energy of Eq. (\[eq:f\]) is equivalent to the free energy of a system of non-interacting clusters distributed according to Eq. \[eq:flory\]. Here we make use of the Flory-Stockmayer theory for providing an expression, to be used in conjunction with the bond probability derived using the Wertheim theory, to evaluate the location in the $(T,\rho)$ plane of the percolation line. The bond probability at percolation, $p_b^p$, is $$\label{eq:perc}
p_b^p=\frac{1}{f-1}.$$
Fig. \[fig:pdWtot\] shows the resulting phase diagram evaluated according to the Wertheim theory for three different values of $f$. More specifically it shows the relative location of the phase coexistence line, the percolation and the maxima of specific heat line. According to the Wertheim theory, the coexistence region becomes wider on increasing $f$. For the case $f=5$, at low $T$ the gas coexists with a liquid with number density $\rho \approx 0.8$, a value significantly smaller than the one commonly observed for particles interacting via spherical potentials. The percolation line merges into the coexistence curve on the left of the critical point, confirming that a spanning cluster of bond is a pre-requisite for the development of a critical phenomena [@coniglio]. For the shown $f$ values, the locus of $C_V^{max}$ is located below the corresponding percolation line, in agreement also with recent findings for a spherical model with $f=4$ [@prossimo]. However in the limit where $f \rightarrow 2$, realized via a mixture of $f =2$ and $f=3$ particles with average functionality $2.055$ [@23], the percolation line lies below the $C_V^{max}$ line. The intersection of the $C_V^{max}$ line with the coexistence curve progressively moves from the left to the right of the critical point on increasing $f$. Already for $f=5$ the density at which the $C_V^{max}$ line meets the coexistence line is more than twice the critical density. While it is not reasonable to extend the Wertheim theory to large $f$ values, it is tempting to speculate that, on further increasing $f$, the intersection point will keep on moving to larger densities so that in the spherical limit case ( with analogous range of interaction) the entire $C_V^{max}$ line lies in a physically inaccessible region (due to packing-driven kinetic arrest).
Monte Carlo Simulation
======================
We perform simulations of the first model discussed in Sec. \[sec:model\] (in which the sticky spots location is regular) with the aim of evaluating the gas-liquid coexistence lines. We aim to provide a definitive proof that reducing valence generates a region of thermodynamic stability of the liquid phase down to vanishing temperatures in a wide range of densities. Previous studies of the same models were indeed focused only on the location of the critical point [@bian]. We perform Gibbs Ensemble Monte Carlo simulations (GEMC) in order to evaluate the phase coexistence region of one component systems with functionality $f$. The GEMC method [@GEMC] allows us to study coexistence in the region where the gas-liquid free-energy barrier is sufficiently high to avoid crossing between the two phases. We simulate for about 5 million MC steps, where a MC step is defined as $N_{\Delta}=10^5$ attempts to translate and rotate a randomly chosen particle, $N_{N}= 10^3$ attempts to swap a particle between the gas and the liquid boxes and $N_V=100$ attempts to modify the volumes. A translational/rotational move is defined as a displacement in each direction of a random quantity distributed uniformly between $\pm ~0.05~\sigma$ and a rotation around a random axis of random angle distributed uniformly between $\pm 0.1$ radiant. The choice of such a large ratio between translation/rotation and swap attempts, $N_{\Delta}/N_N=100$, is dictated by the necessity of ensuring a proper equilibration. In the case of particles with short-range and highly directional interactions this choice is relevant, since the probability of inserting a particle with the correct orientation and position for bonding is significantly reduced as compared to the case of spherical interactions.
We also study the model in which the sticky spots are non-regularly distributed on the surface (see Sec. \[sec:model\]). In particular we focus on the location of the critical point, since the values of critical temperature and density for the corresponding ordered arrangement have already been studied [@bian]. To assess the effect of the randomness on the location of the critical point we perform standard Grand Canonical (GCMC) simulations. In this ensemble, the chemical potential $\mu$, the temperature $T$ and the volume $V$ are fixed. MC moves include insertion and deletion of particles as well as particle translation and rotations. Translational and rotational moves are identical to the one described above for GEMC. In each particle insertion move, a particle with a different realization of the location of the spots is placed in the box. GCMC simulations are extremely helpful in the study of the behavior of the system close to the critical point, since they allow for a correct exploration of the large range of densities experienced by systems in the vicinity of a critical point. To locate the critical point we perform simulations at fixed $T$, $\mu$ and $V$, and we tune $T$ and $\mu$ until the simulated system shows ample density fluctuations, signaling the proximity to the critical point. Once a reasonable guess of the critical point in the $(T,\mu)$-plane is reached, we start at least 8 independent GCMC simulations to improve the statistics of the fluctuations in the number of particles $N$ in the box and of the potential energy $E$. The location of the critical point is performed through a fitting procedure associated to histogram reweighting [@histrew] and a comparison of the fluctuation distribution of the ordering operator $\mathcal{M}$ at the critical point with the universal distribution characterizing the Ising class [@Wilding_96]. The ordering operator $\mathcal{M}$ of the gas-liquid transition is a linear combination $\mathcal{M}\sim \rho +s u$, where $\rho$ is the number density, $u$ is the energy density of the system, and $s$ is the field mixing parameter. Exactly at the critical point, fluctuations of $\mathcal{M}$ are found to follow the Ising model universal distribution [@Wilding_96].
Numerical Results and Comparison with Wertheim predictions
==========================================================
We first focus on the effect of patchiness on the phase coexistence region when the particles functionality $f$ is small. Fig. \[fig:pdW\] shows the numerical phase coexistence curves for systems with a number $f=3$,$4$,$5$ of attractive sites geometrically distributed on the particles surfaces. The figure clearly shows a strong reduction of the phase separation region, i.e. an extention of the region of stabilty of the liquid phase on decreasing $f$. Similar behavior is shown by the Wertheim predictions, despite the agreement gets worse on increasing $f$. Hence, both theory and simulations confirm [@bian; @Zacca1] that a region of densities which is not affected by phase separation is a characteristic of patchy interacting particles systems. The reduction of the valence is thus crucial for suppressing the low temperature ubiquitous process of separation in a dense and dilute solution of particles always observed with spherical potentials. Fig. \[fig:pdW\] also suggets that the small functionality of the particles makes it possible to observe chains and clusters in long-lived thermodynamic equilibrium. In other words, patchiness offers a way of sampling equilibrium homogeneous states in a large region of intermediate and small densities, where packing is not any longer the leading driving force controlling the structure of the system. The shrinking of the unstable region explains why particles interacting via a limited number of functional groups tend to form, at low temperature, open homogeneous structures, which are stabilized by an extended network of long-lived bonds.
To assess if the shape of the coexistence does depend on $f$ we show in Fig. \[fig:pdscaled\] the same data of Fig. \[fig:pdW\] represented as a function of reduced variables, $T/T_c$ and $\rho/\rho_c$. While the Wertheim theory predicts a scaled width of the gas-liquid coexistence that shrinks with $f$, numerical data show that, far from the critical point, the curve for $f=3$ appears to be significantly wider than the one for $f=4$ and $5$. Instead, close to the critical point the shape in reduced units appears to be rather insensitive on $f$ (in agreement with previous findings [@KumarJCP07]).
Next we focus on the differences between a geometric and a random distribution of the patches and in particular on the $f$ dependence of $\rho_c$ and $T_c$ in the two different cases. In the disordered case we vary $f$ from 4 to 6. The results of the GCMC simulations are reported in Tab. \[table:total\], together with corresponding quantities previously calculated for the geometric case [@bian]. The results for the critical point location in the two models are also graphically illustrated in Fig. \[fig:tcphic\]. The same trend with $f$ is shown by both models.
It is interesting to observe that, keeping $f$ constant, $T_c$ and $\rho_c$ both decrease on moving from the geometric to the random arrangement of the sticky sites. This decrease suggests that the propagation of the connectivity is less efficient in the random patches case, speaking for (i) a waste of bond formation possibilities and/or (ii) a failure in the development of long range paths of bonds. Concerning point (i), we note that a random distribution of patches on the particle surface may introduce correlation in the formation of adjacent bonds. Indeed the presence of a bonded interaction may induce a screening effect (and hence a decrease in the probability of forming bonds) on sites closeby located, due to excluded volume interactions. Concerning (ii), we note that a random distribution of sites may also favor the formation of closed loops of bonds, due to a increase in the number of angular possibilities which satisfy short ring structures, which are known to suppress the critical phenomena [@Kindt]. These observation can also explain why the Wertheim theory predictions (which are based on the assumption of both independent bonds and absence of ring structures) are closer to the geometric case model (see Tab. \[table:teo\]).
We note that Tab. \[table:total\] and \[table:teo\] also report the reduced values of the second virial coefficient at the critical point $B_2^c/B_2^{HS}$. The analytical expression of $B_2/B_2^{HS}$ is the following $$\frac{B_2}{B_2^{HS}}= 1 - f^2 \frac{3}{4\pi} (e^{\beta u_0}-1)\frac{V_b}{\sigma^3}$$ where $B_2^{HS}=2/3 \pi \sigma^3$ is the hard-sphere virial coefficient.
We also evaluate the bond probability at the critical point, $p_b^c$, on varing $f$ in both the geometric and random patches models and we compare (see Tab. \[table:pb\]) the two sets of values with the Wertheim theoretical predictions. As previously observed, the Wertheim predictions are closer to the regular case, even if the theory is insensitive to the patches distribution on the particles surface. We also note that the critical bond probabilities in the geometric model are comparable with the ones recently calculated in Ref. [@FoffiKern] for particles with the same bonding geometry interacting via the Kern-Frenkel potential [@Kern_03]. On the other hand, $p_b^c$ for the random model is significantly larger than for the ordered case, supporting our scenario of a less efficient propagation of connectivity in the random case as compared to the geometric one.
Finally, we report in Fig. \[fig:rw5\] the critical fluctuations distributions $P(\mathcal{M})$ of the order parameter $\mathcal{M}$ in both the geometric and random patches models with $f=5$. The calculated distributions are compared to the expected fluctuations at the critical point for systems in the Ising universality class [@Wilding_96]. The comparison provides evidence that the transition belongs to the Ising universality class in both studied cases. This is true for each studied value of $f$. The inset of Fig. \[fig:rw5\] shows the corresponding density fluctuations distributions $P(\rho)$ at the estimated $T_c$ and critical chemical potential $\mu_c$. The distribution becomes more asymmetric on decreasing $\rho_c$, signaling an increasing role of the mixing field $s$ (see also Tab. \[table:total\]). This means that, at equal $f$, the density fluctuation distributions are more asymmetric in the random case rather than in the geometric one.
Discussion and Conclusions
==========================
We study the $f$-dependence of the critical behavior in two different patchy models of $f$-functional particles. In both models, the patchy particles are hard-spheres decorated on their surface by a small number of identical sticky sites, interacting via a short-range square-well attraction. The difference between the two studied models is the arrangement of the attractive sites on the particles surface. In the first case sites are arranged on a regular structure (see Fig. 1 of Ref. [@bian]) in the same geometry of recently synthesized patchy colloidal particles [@Manoh_03; @mohovald; @Blaad06]. In the second case, the distribution of the sites is random and different for each particle.
We compare numerical results and predictions of the thermodynamic perturbation theory developed by Wertheim [@Werth1; @Werth2]. This theory assumes the condition of single-bond per patch and neglects the possibility of forming loops of bonded particles. As previously suggested in Ref. [@Chap2], the free-energy expression provided by the Wertheim theory can be interpreted as the free energy of non-interacting clusters. We show in the Appendix A that the corresponding cluster size distribution is the one provided by Flory and Stockmayer in their seminal work on chemical gelation [@flory; @colby]. In this respect, our study provides an effective expression for describing the density and temperature dependence of the free energy in self-assembly of branched structures and networks. The theory of equilibrium association for systems that form branched structures is receiving particular attention in the last years [@safran; @karlcn; @douglas; @Workum_06; @23; @Kindt], since these systems are found in many technological and biomedical applications, as well as in many biological processes. It is thus crucial to provide a general approach for describing the thermodynamics of the branched polymer self-assembly over the whole range of temperatures, extending to branched system the work developed in the last decades for the case of self-assembling chains and wormlike micelles [@cates1; @Milchev98; @milchevdyn].
We explicitly solve the Wertheim theory for the chosen site-site interaction and we evaluate lines of specific heat maxima (a signature of the presence of a specific bonding process) and the the gas-liquid coexistence lines for $3 \leq f \leq5$. Thanks to the mapping between the Wertheim theory and the Flory-Stockmayer approach we also provide expressions for the dependence on $f$ of the percolation line. We find that, for all studied $f$, the percolation line merges into the phase separation curve on the left hand side of the critical point, while the intersection between the $C_V^{max}$ line and the coexistence curve moves from the left to the right of the critical point on increasing $f$. Even if the Wertheim basic assumptions can not be extended to high valence cases, we speculate that, on further increasing $f$, the intersection between coexistence and $C_V^{max}$ line will further shift to larger densities. In this respect the absence of a $C_V$ maximum in the spherical case could be due to the fact that the entire $C_V^{max}$ line lies in a region of large densities, made physically inaccessible by the progressive slowing down of the dynamics on approaching the glass transition. Indeed, on increasing $f$, the patchy potential tends to a spherical square-well model with analogous range of interaction. For spherical potentials, it has been shown that the glass line — which provides the large-density limit of stability of the liquid state [@nuovotartaglia; @sastryprl] — intersects the coexistence line at a finite temperature and density.
The Wertheim predictions for the the gas-liquid coexistence curve are compared to results of Gibbs Monte Carlo simulations of the regular patches arrangement model. We find that the reduction of the number of patches is accompanied by an enlargement of the region of stability of the liquid phase in the $(T,\rho)$ plane, confirming the scenario suggested in Ref. [@bian].
Both the Wertheim theory and the simulations show that in models of reduced valence, states with $u_0\gg k_{B}T$ can be approached in equilibrium and reversibly. Thus in the presence of patchy interactions it becomes possible in a wide range of densities to cool down the system progressively via a sequence of equilibrium homogeneous states. This is not possible in spherically interacting particles for which phase separation always destabilizes the formation of a homogeneous arrested system at low $T$. Exploring homogeneous states at low temperature opens the way for sampling thermodynamic states characterized by bonds with very long lifetime. When the bond-lifetime becomes comparable to the experimental observation time, a dynamic arrest phenomenon at small packing fraction takes place. The reduction of the valence thus makes it feasible to approach dynamic arrest continuously from equilibrium and to generate a state of matter as close as possible to an ideal gel [@genova; @zaccajpcm]. The relation between arrest in limited valence patchy colloidal particles and arrest in strong network forming molecular and atomic liquids have been recently discussed in Ref. [@simone; @cristianosilica; @statphys; @Zacca2].
Finally, we also study through Grand Canonical Monte Carlo simulations the location of the critical point for disordered arrangements of sites on the hard-sphere surface. Even in this case, we find that $T_c$ and $\rho_c$ moves towards lower temperatures and densities on decreasing the number of the patches. This confirms that the maximum number of bonds per particle plays an important role in controlling the stability of the liquid phase. The fact that the shift with valence of the critical point towards lower temperature and densities can be accomplished with either a geometric or random arrangement of patches could be particularly significant to experimentalists, since it indicates that ordered arrangements of patches are not absolutely necessary to achieve interesting assembly effects.
We observe that, even if the Wertheim theory is insensitive to the arrangement of the sticky sites, the theoretical predictions for the critical point location in the phase diagram are closer to the geometric case model rather than the random one, suggesting that the propagation of the connectivity is less efficient in the random patches case. We recall that the theory is based on the assumption of (i) independent bonds and (ii) absence of closed loops of particles. Hence, the reduced connectivity of the random model, at equal temperature and density, could be related to (i) a reduction of bond formation possibilities, induced by a correlation between nearby sites, and/or to (ii) an increase in possibilities of ring strucures formation, which disfavor the development of branched bonding patterns.
As a side remark, we add in Appendix B the demonstration that, within the Wertheim theory, when particles interact only via bonds and no hard-core repulsion is present, the thermodynamic stability line (spinodal) coincides with the percolation line.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank Jack F. Douglas and Julio Largo for fruitful and continuous discussions. We acknowledge support from MIUR-Prin and MCRTN-CT-2003-504712.
Appendix A
==========
Here we provide a physical insight of the Wertheim theory, by discussing two equivalent alternative derivations of the Wertheim bond free-energy. Both derivations assume the system of associating particles to be formally equivalent to a system of non-interacting clusters, in thermodynamic equilibrium. For simplicity, we assume a De Broglie length $\Lambda=\sigma=1$ in both derivations.
In the first derivation, we assume that the cluster-size distribution is provided by the Flory-Stockmayer expressions (see Eq. \[eq:flory\]), i.e. that the system of $N$ $f$-functional associating particles aggregates in clusters characterized by the absence of closed bonding loops. Bonds are also assumed to be uncorrelated so that to each bond is associated the same single-bond free energy ${\cal F}_b$. In the absence of loops, the number of bonds in a cluster of size $n$ is exactly $(n-1)$, since each new bond adds one new particle. Hence the bond free energy of the cluster is $(n-1) {\cal F}_b$. If clusters do not interact, the system free energy $F$ can be written as sum of the ideal-gas free energy of each distinct bonding topology cluster-type (accounting for the translational center of mass degrees of freedom) and a sum over the cluster bond free energies. Defining $\rho_n^k$ as the number density of clusters with size $n$ and with bonding topology $k$ one can write $$\label{eqn:Fini}
\frac{\beta F} {V}= \sum_{n}\sum_{k}\rho_n^k \left[\ln \rho_n^k-1\right] + \sum_{n}\sum_{k}\rho_n^k (n-1) \beta {\cal F}_b.$$ Here $V$ is the volume, $\beta \equiv 1/k_B T$ (with $k_B$ the Boltzmann constant) and the sum on $n$ runs over all the possible cluster sizes, from one (monomers) to infinity, while the sum over $k$ includes all $\omega_n$ distinct bonding topology of clusters of size $n$. Since clusters with the same size but different bonding pattern are equiprobable, then $\rho_n^k \equiv \rho_n / \omega_n$ (see Eq. \[eq:flory\]). Thus Eq. \[eqn:Fini\] becomes $$\label{eq:Fmiddle}
\frac{\beta F} {V}= \sum_n \rho_n \left[\ln \frac{\rho_n}{\omega_n}-1\right] + \sum_{n}\rho_n (n-1) \beta {\cal F}_b.$$ Substituting Eq. \[eq:flory\] in Eq. \[eq:Fmiddle\] and summing over $n$ one obtains, for $p_b < (f-1)^{-1}$ (which express the condition that all clusters are finite [@flory]), the following expression for the system free energy in term of $p_b$ and bond free energy $$\begin{aligned}
\frac{\beta F} {V} &=& \rho \ln\rho - \rho + \\
\nonumber
&+& \rho \ln (1-p_b)^f - \frac{\rho f p_b}{2} \left[ \ln \frac{\rho (1-p_b)^2}{p_b} e^{- \beta {\cal F}_b} -1\right].\end{aligned}$$ Such expression can be seen as a high temperature contribution [@Hansennew] ($\rho \ln\rho - \rho$) plus a remaining bonding term. The bonding free energy coincides with the Wertheim expression [@Werth1; @Werth2; @Hansennew] ($\rho \ln (1-p_b)^f + \frac{\rho f p_b}{2} $) when the connection between $p_b$ and $ {\cal F}_b$ is given by Eq. \[eq:pb\]. This simple derivation can be also extended to binary mixtures.
An even simpler derivation has been suggested in Ref. [@Chap2] and it is here reported for completeness. Also this derivation assumes that the system is an ideal gas of clusters and hence that the product $\beta PV$ is identical to the number of clusters $N_c$. The evaluation of $N_c$ is straightforward for $p_b$ values smaller than the percolation threshold $p_b^p$, since, in the absence of closed bond loops, $N_{c}$ is equal to the number of particles minus the number of bonds $N_b$. Calling $N_b^{max}= \frac{Nf}{2}$ the maximum number of possible bonds and noting that $p_b$ is the ratio between $N_b$ and $N_b^{max}$, one finds $$N_c=N-N_b=N \left(1 - \frac{f}{2} p_b\right )
\label{eq:nc}$$ and $$\beta P = \rho \left(1 - \frac{f}{2} p_b\right ).
\label{eq:pvnc}$$ Since the system is in dynamic equilibrium, the particle chemical potential is independent from the cluster to which the particle belongs to. Hence, it is identical to the chemical potential $\mu$ of the monomer. The ideal-gas hypothesis implies that the activity of the monomer $z\equiv \exp(\beta \mu)$ is related to the monomer number density by $z=\rho_1=\rho (1-p_b)^f$. Hence the system free energy density can be immediately written as $$\beta F = \rho \beta \mu -\beta P = \rho \ln \left[\rho (1-p_b)^f\right ] - \rho \left(1 - \frac{f}{2} p_b\right )$$ which coincides with the Wertheim expression when the reference system is the ideal gas.
We note in conclusions that the above relations are valid only before percolation. Indeed, the sums over $n$ in Eq. \[eq:Fmiddle\] as well as Eq. \[eq:nc\] are valid only for $p_b< p_b^p$. Hence, the region of validity of the Wertheim theory should be strictly limited to non-percolating states. Nevertheless we observe that the theory works well even below the percolation threshold [@bian; @23]. It could be in principle possible to extend the formalism to $p_b> p_b^p$ by accounting correctly for the $p_b$ dependence of the number of clusters (which is always possible, analytically for small $f$ and numerically above), but it is not clear how to handle the free energy contribution of the percolating cluster.
Appendix B
==========
In this Appendix we examine the thermodynamic stability of a system composed of non-interacting clusters, described by the free energy of Eq. \[eq:Fmiddle\]. A stable system is characterized by a negative volume derivative of the pressure. The region of stability is delimited by the so-called spinodal line, defined as the locus of points such that $\left(\partial \beta P/\partial V\right)_T =0$. The volume derivative of the pressure, under the ideal gas approximation $g_{HS}(r) = 1$, is controlled only by the volume derivative of $p_b$ (see Eq. \[eq:pvnc\]). Interestingly, it gives for the bond probability at the spinodal line $p_b^s$ $$p_b^s=\frac{1}{f-1}$$ i.e. the same condition that defines percolation. Hence $p_b^s=p_b^p$. The system is thus mechanically stable only in the non-percolating region. In other words in the ideal gas approximation no dense stable states are possible and the system exists only in the gas-phase. In the Wertheim theory the existence of a liquid phase is generated by the significant increase of the pressure at low $V$ introduced by the hard-sphere reference contribution. Fig. \[fig:p\] provides an example of the effect of the hard-sphere contribution on the pressure for the case $f=3$. We conclude noting that the absence of the hard-core repulsion appears to be essential in formally associating the percolation line with the spinodal line, providing an analytic simple example of the possibility of interpreting critical phenomena in term of percolation [@coniglio; @padua].
$f$ $T_c$ $\rho_c$ $\mu_c$ $s$ $L$ $B_2^c/B_2^{HS}$
------------- ------- ---------- --------- ------ ----- ------------------
3 geometric 0.094 0.141 -0.471 0.46 9 -28.772
4 geometric 0.118 0.273 -0.418 0.08 7 -5.080
5 geometric 0.132 0.351 -0.410 0 7 -2.866
4 random 0.102 0.208 -0.531 0.46 8 -21.978
5 random 0.118 0.258 -0.500 0.25 8 -8.500
6 random 0.133 0.310 -0.482 0.22 7 -4.258
: Values of the relevant parameters at the critical point for the geometric [@bian] $(f=3,4,5)$ and random $(f=4,5,6)$ cases: $T_c$ is the critical temperature, $\rho_c$ is the density of the critical point, $\mu_c$ is the critical chemical potential, $s$ is the field mixing parameter and $L$ indicates the largest studied box size. $B_2^c/B_2^{HS}$ is the value of the reduced second virial coefficient at the critical point.[]{data-label="table:total"}
$f$ $T_c$ $\rho_c$ $B_2^c/B_2^{HS}$
---------- -------- ---------- ------------------ -- -- --
3 theory 0.0925 0.086 -34.378
4 theory 0.1121 0.154 -8.498
5 theory 0.1275 0.212 -4.052
6 theory 0.1411 0.261 -2.414
: Critical values of the temperature and density for $3 \leq f \leq 6$ evaluated through the Wertheim theory. We also report the corresponding values of the reduced second virial coefficient.[]{data-label="table:teo"}
$f$ $p_b^c$ $theory$ $p_b^c$ $geometric$ $p_b^c$ $random$
----- ------------------ --------------------- ------------------ -- -- --
3 0.633 0.728
4 0.488 0.640 0.737
5 0.417 0.577 0.615
6 0.360 0.539
: Critical values of the bond probability $p_b$ for $f$ varying from $3$ to $6$. Theoretical values are obtained solving Eq. \[eq:pb\] at the critical point, while numerical ones are obtained as the ratio between the potential energy at the critical point and the energy of the fully bonded system.[]{data-label="table:pb"}
![[]{data-label="fig:pdWtot"}](A712051JCPFig1.eps){width="10cm"}
![[]{data-label="fig:pdW"}](A712051JCPFig2.eps){width="11cm"}
![[]{data-label="fig:pdscaled"}](A712051JCPFig3.eps){width="10cm"}
![[]{data-label="fig:tcphic"}](A712051JCPFig4.eps){width="10cm"}
![[]{data-label="fig:rw5"}](A712051JCPFig5.eps){width="10cm"}
![[]{data-label="fig:p"}](A712051JCPFig6.eps){width="10cm"}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Motivated by applications of distributed storage systems to key-value stores, the multi-version coding problem was formulated to efficiently store frequently updated data in asynchronous decentralized storage systems. Inspired by consistency requirements in distributed systems, the main goal in the multi-version coding problem is to ensure that the latest possible version of the data is decodable, even if the data updates have not reached some servers in the system. In this paper, we study the storage cost of ensuring consistency for the case where the data versions are correlated, in contrast to previous work where data versions were treated as being independent. We provide multi-version code constructions that show that the storage cost can be significantly smaller than the previous constructions depending on the degree of correlation, despite the asynchrony and the decentralized nature. Our achievability results are based on Reed-Solomon codes and random binning. Through an information-theoretic converse, we show that our multi-version codes are nearly-optimal, within a factor of $2$, in certain interesting regimes.'
author:
- 'Ramy E. Ali and Viveck R. Cadambe [^1]'
bibliography:
- 'IEEEabrv.bib'
- 'Nulls.bib'
title: 'Harnessing Correlations in Distributed Erasure-Coded Key-Value Stores'
---
Acknowledgement {#acknowledgement .unnumbered}
===============
The authors would like to thank A. Tulino and J. Llorca for their helpful comments.
Conclusion {#Conclusion}
==========
In this paper, we have proposed multi-version codes to efficiently store correlated updates of data in a decentralized asynchronous storage system. These constructions are based on Reed-Solomon codes and random binning. An outcome of our results is that the correlation between versions can be used to reduce storage costs in asynchronous decentralized systems, even if there is no single server or client node who is aware of all data versions, in applications where consistency is important. In addition, our converse result shows that these constructions are within a factor of $2$ from the information-theoretic optimum in certain interesting regimes. The development of practical coding schemes for the case where $\delta_K$ is known a priori is an open research question, which would require non-trivial generalizations of previous code constructions for the Slepian-Wolf problem [@pradhan2003distributed; @schonberg2004distributed].
\[sec:appendix\]
[^1]: Ramy E. Ali (email: [email protected]) and Viveck R. Cadambe (email: [email protected]) are with the School of Electrical Engineering and Computer Science, The Pennsylvania State University, University Park, PA 16802. This work is supported by NSF grant No. CCF 1553248 and is published in part in the Proceedings of the 2016 IEEE Information Theory Workshop [@ali2016consistent].
| {
"pile_set_name": "ArXiv"
} |
\
\
ABSTRACT. Measuring a strength of dependence of random variables is an important problem in statistical practice. In this paper, we propose a new function valued measure of dependence of two random variables. It allows one to study and visualize explicit dependence structure, both in some theoretical models and empirically, without prior model structure. This provides a comprehensive view of association structure and makes possible much detailed inference than based on standard numeric measures of association. We present theoretical properties of the new measure of dependence and discuss in detail estimation and application of copula-based variant of it. Some artificial and real data examples illustrate the behavior and practical utility of the measure and its estimator.\
[KEY WORDS:]{} Copula; Dependence measures; Graphical method; Local correlation; Nonparametric association; Rank tests.
Teresa Ledwina, Institute of Mathematics, Polish Academy of Sciences, ul. Kopernika 18, 51-617 Wroc[ł]{}aw, Poland ( E-mail: [*[email protected]*]{}). Research was supported by the Grant N N201 608440 from the National Science Center, Poland. The author thanks Dr. Grzegorz Wy[ł]{}upek for his help in preparing figures and tables of this article and for useful remarks.\
[**1. Introduction**]{}\
Measuring a strength of dependence of two random variables has long history and wide applications. For brief overview see Jogdeo (1982) and Lancaster (1982). More detailed information can be found in Drouet Mari and Kotz (2001) as well as Balakrishnan and Lai (2009), for example. Most of measures of dependence, introduced in vast literature on the subject, are scalar ones. Such indices are called global measures of dependence. However, nowadays there is strong evidence that an attempt to represent complex dependence structure via a single number can be misleading. To overcome this drawback, several local indices have been proposed; see Section 6.3 of Drouet Mari and Kotz (2001) for details. Many of these indices were introduced in the context of regression models or survival analysis. Some local dependence functions have been introduced as well. In particular, Kowalczyk and Pleszczyńska (1977) invented function valued measure of monotonic dependence, based on some conditional expectations and adjusted to detect dependence weaker than the quadrant one. Next, Bjerve and Doksum (1993), Bairamov et al. (2003) and Li et al. (2014), among others, introduced local dependence measures based on regression concepts. See the last mentioned paper for more information. Holland and Wang (1987) defined the local dependence function, which mimics cross-product ratios for bivariate densities and treats the two variables in a symmetrical way. This function valued measure has several appealing properties and received considerable attention in the literature; cf. Jones and Koch (2003) for discussion and references. However, on the other hand, this measure has some limitations: it is not normalized, requires existence of densities of the bivariate distribution, and is intimately linked to strong form of dependence, the likelihood ratio dependence. Recently, Tj[ø]{}stheim and Hufthammer (2013) extensively discussed the role and history of local dependence measures in finance and econometrics. They also proposed the new local dependence measure, the local correlation function, based on approximating of bivariate density locally by a family of Gaussian densities. Similarly as the measure of Holland and Wang (1987), this measure treats both variables on the same basis. Though the idea behind the construction of this measure is intuitive one its computation and estimation is a difficult and complex problem. The asymptotic theory developed in Tj[ø]{}stheim and Hufthammer (2013) treats in detail the problems mentioned above, in a scope that covers some time series models. In Berentsen et al. (2013) this theory is applied to describe dependence structure of different copula models.
In this paper, we propose the new function valued measure of dependence of two random variables $X$ and $Y$ and present its properties. The measure has simple form and its definition exploits cumulative distribution functions (cdf’s), only. In particular, we do not assume existence of a density of the observed vector. The measure takes values in \[-1,1\] and treats both variables in a symmetrical way. The measure preserves the correlation order, or equivalently the concordance order, which is the quadrant order restricted to the class of distributions with fixed marginals. In particular, it is non-negative (non-positive) if and only if $X$ and $Y$ are positively (negatively) quadrant dependent. Quadrant dependence is relatively weak, intuitive and useful dependence notion, widely used in insurance and economics; see Dhaene et al. (2009) for an evidence and further references. The new measure obeys several properties formulated in the literature as useful or desirable. We introduce two variants of the measure. In Section 2 we consider general case, assuming that the vector $(X,Y)$ has joint cdf $H$ and marginals $F$ and $G$, respectively. In Section 3 we discuss its copula-based counterpart which corresponds to some cdf $C$ on $[0,1]^2$ with uniform marginals. Both variants allows for readable visualization of departures from independence. We focus our presentation on the copula-based variant. Simple and natural estimator of the copula-based measure in the i.i.d. case is proposed and its appealing properties are discussed. The estimator can be effectively exploited to assess graphically underlying bivariate dependence structure and to build some formal local and global tests. Some illustrative examples are given in Section 4 to support utility of new solution. Section 6 concludes.\
[**2. General case**]{}\
Consider a pair of random variables $X$ and $Y$ with cdf’s $F(x)=P(X\leq x)$ and $G(y)=P(Y\leq y)$, respectively and a joint cdf $H(x,y)=P(X\leq x,Y\leq y)$. Set $\mathbbm{D} =\{(x,y): 0<F(x)<1,0<G(y)<1\}$ and define $$q(x,y)=q_H(x,y)=\frac{H(x,y)-F(x)G(y)}{\sqrt{F(x)G(y)[1-F(x)][1-G(y)]}}\;\;\; \mbox{for}\;\;(x,y)\in \mathbbm{D}.
\eqno(1)$$ From (1) it is seen that $q$ treats both variables $X$ and $Y$ symmetrically and a knowledge of $q$ and the marginal distributions allows one to recover $H$. The measure $q$ fulfills the following properties, motivated by the axioms formulated in Schweitzer and Wolff (1981) and updated in Embrechts et al. (2002).\
[**Proposition 1.**]{}\
1. $q$ is defined for any $X$ and $Y$.\
2. $-1 \leq q \leq 1$ for all $(x,y) \in \mathbbm{D}$.\
3. If the variables $X$ and $Y$ are exchangeable then $q(x,y)=q(y,x)$ for $(x,y) \in \mathbbm{D}$.\
4. $q(x,y)\equiv 0$ if and only if $X$ and $Y$ are independent.\
5. $q$ is non-negative (non-positive) if and only if $(X,Y)$ are positively (negatively) quadrant dependent.\
6. $q$ is maximal (minimal) if and only if $Y=f(X)$ and $f$ is non-decreasing (non-increasing) a.s. on the range of $X$.\
7. $q$ respects concordance ordering, i.e. for cdf’s $H_1$ and $H_2$ with the same marginals, $H_1(x,y) \leq H_2(x,y)$ for all $(x,y) \in \mathbbm{R}^2$ implies $q_{H_1}(x,y) \leq q_{H_2}(x,y)$ for all $(x,y) \in \mathbbm{D}$.\
[**Proof**]{}. Most of the above mentioned properties are obvious. The property 6 is an immediate consequence of Fréchet-Hoeffding bounds and their properties. To justify 2 it is enough to show that $q_H(x,y)$ is the correlation coefficient of some random variables. For this purpose, for $(x,y) \in \mathbbm{D}$ and $(s,t) \in \mathbbm{R}^2$ set $$\phi_x(s)=-\sqrt{\frac{1-F(x)}{F(x)}}\mathbbm{1}_{(-\infty ,x]}(s) + \sqrt{\frac{F(x)}{1-F(x)}}\mathbbm{1}_{(x, +\infty)}(s),$$ $$\psi_y(t)=-\sqrt{\frac{1-G(y)}{G(y)}}\mathbbm{1}_{(-\infty ,y]}(t) + \sqrt{\frac{G(y)}{1-G(y)}}\mathbbm{1}_{(y, +\infty)}(t).$$ Then, by an elementary argument one gets $$q(x,y)=q_H(x,y)=E_H\phi_x(X)\psi_y(Y)=Cov_H\phi_x(X)\psi_y(Y)=Corr_H\phi_x(X)\psi_y(Y).
\eqno(2)$$ $\square$\
[**Remark 1.**]{} The last expression in (2) shows that the function $q$ is based on aggregated local correlations. Moreover, note that $\int_{\mathbbm{R}}\phi_x(s)dF(s)=\int_{\mathbbm{R}}\psi_y(t)dG(t)=0 $ and $\int_{\mathbbm{R}}\phi_x^2(s)dF(s)=\int_{\mathbbm{R}}\psi_y^2(t)dG(t)=1 $, for all $(x,y) \in \mathbbm{D}$. Therefore, the value $q(x,y)$ can be interpreted as the Fourier coefficient of the cdf $H(s,t)$ pertaining to the quasi-monotone function $\phi_x(s) \psi_y(t),\;(s,t) \in \mathbbm{R}^2$.\
Given random sample $(X_1,Y_1),\ldots,(X_n,Y_n)$ from cdf $H$ with marginals $F$ and $G$, set $H_n$, $F_n$, and $G_n$ for respective empirical cdf’s. A natural estimator $\hat {q}_H$ of $q_H(x,y)$ results by plugging these empirical cdf’s into (1). This, given $(x,y)$, yields rank statistics. Note that the values $\chi_{ni}$ given by $\hat {q}_H(X_i,Y_i),\; i=1,\ldots,n$, have been already introduced in Fisher and Switzer (1985), as one of the two components of the so-called chi-plots, designed to investigate possible patterns of association of two random variables.
We shall not study the estimator $\hat {q}_H$ in this paper. In the next section we comment in more detail on special case of (1), and the related problems, in the case when the role of $H$ is played by the pertaining copula.\
[**3. Copula-based measure of dependence**]{}\
In this section, to avoid technicalities and to concentrate on the main idea, we restrict attention to cdf’s $H$ with continuous marginals $F$ and $G$. Under such a restriction there exists a unique copula $C$ such that $H(x,y)=C(F(x),G(y))$. In other words, $C$ is the restriction to the unit square of the joint cdf of $(F(X),G(Y))$. The copula captures the dependence structure among $X$ and $Y$, irrespective of their marginal cdf’s. This is important in many applications. For the related discussion see Póczos et al. (2012).\
[*3.1. The form and further properties of q*]{}\
We have $$q(u,v)=q_{C}(u,v)=\frac{C(u,v)-uv}{\sqrt{uv(1-u)(1-v)}} = w(u,v)[C(u,v)-uv], \;\;\;\;(u,v) \in (0,1)^2,
\eqno(3)$$ where $$w(u,v)=\frac{1}{\sqrt{uv(1-u)(1-v)}}.
\eqno(4)$$ The interpretation of $q$ in terms of correlations, given in (2), is still valid with some obvious adjustment. Namely, now for $u \in (0,1)$ and $s\in (0,1)$ we consider $$\phi_u(s)=\psi_u(s)=-\sqrt{\frac{1-u}{u}}\mathbbm{1}_{[0,u]}(s)+\sqrt{\frac{u}{1-u}}\mathbbm{1}_{(u,1]}(s).$$
[**Proposition 2.**]{} The copula based measure of dependence $q$, given by (3), additionally to 1-7, has the following properties.\
8. $q$ is invariant to strictly increasing a.s. on ranges of $X$ and $Y$, respectively, transformations.\
9. If $X$ and $Y$ are transformed by strictly decreasing a.s. functions then $q(u,v)$ transforms to $q(1-u,1-v)$.\
10. If $f$ and $g$ are strictly decreasing a.s. on ranges of $X$ and $Y$, respectively, then $q$’s for the pairs $(f(x),Y)$ and $(X,g(Y))$ take the forms $-q(1-u,v)$ and $-q(u,1-v)$, accordingly.\
11. The equation $q(u,v)\equiv c$, $c$ a constant, can hold true if and only if $c=0$.\
12. If $(X,Y)$ and $(X_n,Y_n),\;n=1,2,\ldots,$ are pairs of random variables with joint cdf’s $H$ and $H_n$, and the pertaining copulas $C$ and $C_n$, respectively, then weak convergence of $\{H_n\}$ to $H$ implies $q_{C_n}(u,v) \to q_C(u,v)$ for each $(u,v) \in (0,1)^2$.\
[**Proof.**]{} Properties 8-10 follow from Theorem 3 in Schweizer and Wolff (1981). The convergence in 12 is due to continuity of $C$. To justify 11 observe that the equation is equivalent to $C(u,v)=C_c(u,v)=uv+c\sqrt{uv(1-u)(1-v)}$. Since $C$ is quasi-monotone, then $C_c(u,v)$ should also possess such a property. Since $C_c(u,v)$ is absolutely continuous then quasi-monotonicity is equivalent to $\frac{\partial^2}{\partial u \partial v} C_c(u,v) \geq 0$ for almost all $(u,v) \in [0,1]$ (in the Lebesgue measure); cf. Cambanis et al. (1976). However, $\frac{\partial^2}{\partial u \partial v} C_c(u,v) = 1 + c[u-1/2][v-1/2]w(u,v)$ and for $c\not= 0$ this expression can be negative on the set of positive Lebesgue measure.$\square$\
[**Remark 2.**]{} The properties 4 and 8-10 provide some compromise to too demanding postulates P4 and P5 discussed in Embrechts et al. (2002).\
[**Remark 3.**]{} The property 11 is very different from respective property of the local dependence function of Holland and Wang (1987) which is constant for the bivariate normal distribution and some other models; cf. Jones (1998) for details.\
[**Remark 4.**]{} By Fréchet-Hoeffding bounds for copulas, the property 2 saying that $q(u,v) \in [-1,1]$ can be further sharpened to $$B_*(u,v) \leq q_C(u,v) \leq B^*(u,v),\;\;\;(u,v) \in (0,1)^2$$ where $B_*(u,v)=w(u,v)[\max\{u+v-1,0\} -uv]$ and $B^*(u,v)=w(u,v)[\min\{u,v\} -uv]$. Note that $B_*(u,1-u) = -1$ for $u \in (0,1)$ while $-1 < B_*(u,v) \leq 0$ otherwise. Similarly, $B^*(u,u)=+1$ for $u \in (0,1)$ and $0 \leq B^*(u,v)<1$ in the remaining cases. In Figure 1 we show these bounds on the the regular grid $$\mathbbm{G}_{16} = \{(u,v): u=i/16, v=j/16,\; i,j=1,\ldots,15\}.$$
[*Fig. 1.*]{} Left panel: lower bound $B_{*}(u, v)$ of $q_C(u, v)$ on the grid $\mathbbm {G}_{16}$; right panel: upper bound $B^{*}(u, v)$ of $q_C(u, v)$ on the grid $\mathbbm {G}_{16}$.\
[**Remark 5.**]{} Similarly as $B_*$ and $B^*$, the measure $q_C$ can be displayed in a graphical form. This is useful feature helping to visualize a structure of departures from independence. It is also worth emphasizing that though we focused of cdf’s with continuous marginals we still do not assume that $H$ or $C$ posses a density. In Figures 2 and 3 we show plots of $q_C$ pertaining to classical Marshall-Olkin and recently introduced Mai-Scherer (2011) extreme value copulas. The displays are accompanied by scatter plots of simulated data and pertaining heat maps. Therefore, we introduce first a natural estimator of $q_C$ and present all illustrative examples in Section 4.\
[*3.2. Estimates of $q_C$*]{}\
Let $(X_1,Y_1),\ldots,(X_n,Y_n)$ be a random sample from cdf $H$. Furthermore, let $R_i$ be the rank of $X_i,\;i=1,\ldots,n$, in the sample $X_1,\ldots,X_n$ and $S_i$ the rank of $Y_i,\;i=1,\ldots,n,$ within $Y_1,\ldots,Y_n$. Simple estimate of $q_C$ has the form $$\hat q_C (u,v)=w(u,v)[{D_n(u,v)-uv}], \;\;\;\;(u,v) \in (0,1)^2,$$ where $D_n$ is rank-based empirical copula estimator of $C$, i.e. $$D_n(u,v)=\frac{1}{n}\sum_{i=1}^n \mathbbm{1}\Bigl(\frac{R_i}{n}\leq u,\frac{S_i}{n}\leq v\Bigr), \quad
(u,v) \in [0,1]^2.
\eqno(5)$$ The paper of Swanepoel and Allison (2013) provides exact mean and variance of $D_n(u,v)$. Ledwina and Wy[ł]{}upek (2014) have shown that $D_n(u,v)$ preserves the quadrant order. More precisely, if a copula $C_1$ has larger quadrant dependence than a copula $C_{2}$ then, under any fixed $(u,v) \in [0,1]^2$, any $c \in \mathbb{R}$ and any $n$, it holds that $$P_{C_1}\Bigl(D_n(u,v) \geq c\Bigr) \geq P_{C_{2}}\Bigl(D_n(u,v)\geq c\Bigr).
\eqno(6)$$ An obvious consequence of (6) is analogous order preserving property of $\hat q_C (u,v)$.
Asymptotic properties of the empirical copula process $Z_n(u,v)=\{\sqrt n [D_n(u,v)-uv],\;u,v \in [0,1]\}$ have been studied by many authors; cf. Fermanian et al. (2004) for the related results and references. Since $uv(1-u)(1-v)$ is the asymptotic variance of $Z_n(u,v)$ under independence, $\sqrt n \hat q_C (u,v)$ coincides with natural weighted empirical copula process while Theorem 3 of Fermanian et al. (2004) implies that, under independence, $\sqrt n \hat q_C (u,v)$ is asymptotically $N(0,1)$ for each $u,v \in (0,1)$. The estimate $D_n(u,v),$ in a series of papers originated by Deheuvels (1979), has been called empirical dependence function.
In application oriented papers, more popular variant of rank-based estimator of $C$ is $$C_n(u,v)=\frac{1}{n}\sum_{i=1}^n \mathbbm{1}\Bigl(\frac{R_i}{n+1}\leq u,\frac{S_i}{n+1}\leq v\Bigr), \quad
(u,v) \in [0,1]^2.
\eqno(7)$$ The variables $(R_i/(n+1),S_i/(n+1)),\; i=1,\ldots,n$ are called pseudo-observations in the literature. Obviously, the finite sample and basic asymptotic properties of $C_n(u,v)$ are inherited after $D_n(u,v)$. Therefore, we shall consider the following estimator of $q_C$ $$Q_n(u,v)=w(u,v)[{C_n(u,v)-uv}] = \frac{C_n(u,v)-uv}{\sqrt{uv(1-u)(1-v)}}, \;\;\;\;(u,v) \in (0,1)^2,
\eqno(8)$$ which, similarly as $\hat q_C (u,v)$, is the rank statistic. Moreover, we set $$L_n(u,v)=\sqrt n Q_n(u,v)
\eqno(9)$$ for the standardized version of this estimate. Simple algebra yields that for any $(u,v) \in (0,1)^2$ it holds $$L_n(u,v)=\frac{1}{\sqrt n} \sum_{i=1}^n\phi_u(\frac{R_i}{n+1})\phi_v(\frac{S_i}{n+1}) + O(\frac{1}{\sqrt n}).
\eqno(10)$$ So, up to deterministic term of the order $O(1/\sqrt n)$, the standardized estimator $L_n(u,v)$ is linear rank statistic with the quasi-monotone score generating function $\phi_u \times \phi_v$. Moreover, the definition of $L_n$ and (6) yield that $$P_{C_1}\Bigl(L_n(u,v) \geq c\Bigr) \geq P_{C_{2}}\Bigl(L_n(u,v)\geq c\Bigr)
\eqno(11)$$ for any $(u,v) \in (0,1)^2$, any $c$, any $n$, and any two copulas $C_1$ and $C_2$ such that $C_1$ has larger quadrant dependence than $C_2$. Summarizing the above mentioned results, let us note that under independence $L_n(u,v)$ is distribution free. So, given $n$, under independence, the significance of the obtained values of this statistic can be easily assessed on a basis of simple simulation experiment. For large $n$ one can rely on asymptotic normality of $L_n(u,v)$. Due to (11), similar conclusions follow if one likes to verify hypothesis asserting that $q_C(u,v) \geq 0.$ Moreover, (11) implies that different levels of strength of quadrant dependence of the underlying $H$’s shall be adequately quantified by order preserving $L_n(u,v)$’s. These results make the values of $L_n(u,v), \;(u,v) \in (0,1)^2,$ a useful diagnostic tool. For example, since quadrant dependence is relatively weak notion, significantly negative values of $L_n(u,v)$ for some $(u,v)$’s make questionable positive quadrant dependence, and many other forms of positive dependence of the data at hand, as well.
To close, note that, given $u$ and $u$, the score generating function $\phi_u \times \phi_v$, appearing in (10), is not smooth one and takes on at most four possible values, only. This causes that, under independence, the convergence of $L_n(u,v)$ to the limiting $N(0,1)$ law is not very fast. Moreover, the rate of convergence is expected to depend on $u$ and $v$, with the least favorable situation when $(u,v)$ is close to the vertices of the unit square. We illustrate these aspects in Table 1, where simulated critical values of the test rejecting independence for large values of $|L_n(u,v)|$ are given under some $(u,v)$’s, five different sample sizes, and two selected significance levels $\alpha$. In cases when finite sample distribution of $L_n$ is far from continuous one, the problem of uniqueness of sample quantiles arises. Through, to calculate sample quantiles we apply Gumbel’s approach, described in Hyndman and Fan (1996) by Definition 7.\
Table 1. Simulated critical values of the test rejecting independence for large values of $|L_n(u,v)|$ for selected $(u,v)$, versus $n$ and $\alpha$.\
------------------------------- ------- ------- ------- ------- ------- ------- ------- ------- ------- -------
$(u,v)$
200 300 400 500 600 200 300 400 500 600
$(\frac{1}{2},\frac{1}{2})$ 2.546 2.540 2.600 2.504 2.613 1.980 1.848 2.000 1.968 1.960
$(\frac{1}{12},\frac{1}{12})$ 2.520 2.960 2.800 2.716 2.583 1.594 1.575 1.782 1.968 2.049
$(\frac{1}{16},\frac{1}{16})$ 2.753 2.879 2.933 2.349 2.591 1.546 1.894 2.080 1.586 1.894
$(\frac{1}{20},\frac{1}{20})$ 2.233 2.735 3.158 2.589 3.008 2.233 1.519 2.105 1.648 2.149
------------------------------- ------- ------- ------- ------- ------- ------- ------- ------- ------- -------
\
The sample sizes in Table 1 are close to that we shall consider in Section 4. The results exhibit that, for this range of sample sizes, the simulated quantiles of $|L_n(1/2,1/2)|$ are reasonably close to the limiting ones. We studied some selection of regular grids $\{(i/g,j/g),\;i,j=1,\ldots,g-1\}$, for some $g$’s, the related behavior of $L_n$’s and reported in Table 1 the results on $L_n(1/g,1/g)$ for $g=12, 16, 20$. They illustrate the conclusion that too dense grid shall imply too much inaccuracy in the simulated quantiles while too conservative choice not necessarily provides much progress. In view of these observations, we decided to apply the regular grid $\mathbbm{G}_{16}$ through. With this choice, for the sample sizes under consideration, the significance of observed values of $|L_n(i/g,j/g)|$ can be easily approximately evaluated looking at the heat maps, that we provide in each case. For more precise evaluation extra simulations are needed.
Note also that there are available some smooth nonparametric estimators of copulas. See Janssen et al. (2012) and Omelka et al. (2009) for recent contributions and extensive overview. However, we do prefer to insert the estimator $C_n$, since it is naturally linked to local correlations defined in the paper and obeys the property (11), which is crucial in finite sample inference on dependence. Counterparts of (11) for more general rank statistics are not available, according to the best our knowledge. For some related discussion see Ledwina and Wy[ł]{}upek (2014), Section 3.\
[**4. Illustration**]{}\
[*4.1. Example 1: Extreme value copulas*]{}\
We start with two simulated data sets of size $n=500$ from Marshall-Olkin and Mai-Scherer (2001) copulas given by $C(u,v)=C_{a,b}^{MO}(u,v)=\min\{u^{1-a}v,uv^{1-b}\},\;a=0.50, b=0.75$, and $C(u,v)=C_{a,b}^{MS}(u,v)=\min\{u^a,v^b\}\min\{u^{1-a},v^{1-b}\},\;a=0.9, b=0.5$; cf. Nelsen (2006), p. 53 and Mai-Scherer (2001), p. 313, respectively. Both copulas possess a singular part. In Figures 2 and 3 we show dependence functions $q_C(u,v)$ for these models. The functions are accompanied by scatter plots of pseudo-observations $(R_i/(n+1),S_i/(n+1)), i=1,\ldots,500$, from the simulated samples. The scatter plots nicely exhibit the singularities and show some tendencies in the data. Right panels in these figures display respective heat maps of standardized correlations $L_n(u,v)$’s calculated on the grid $\mathbbm {G}_{16}$. Each square of size $0.0625\times0.0625$ represents the respective value of $L_n$ in its upper-right corner. To simplify reading, each heat map is accompanied with two numbers $$L_* = \min_{1\leq i,j \leq 15} L_n(i/16,j/16) \;\;\;\mbox{and}\;\;\; L^*=\max_{1\leq i,j \leq 15} L_n(i/16,j/16).
\eqno(12)$$ Both copulas represent positively quadrant dependent distributions. Under such dependence large values of $U$ tend to associate to large values of $V$ and similar pattern applies to small values. This tendency is nicely seen in the figures. Intuitively, the tendency is stronger for $C_{0.9,0.5}^{MS}$ than for $C_{0.50,0.75}^{MO}$ and this is indeed well reflected by the heat maps. The points of the grid $\mathbbm{G}_{16}$ in which the estimated correlations $Q_n$ are significant on the levels 0.05 and 0.01 can be easily identified; cf. Table 1. Some possibility of testing for positive local and/or global dependence is sketched in Section 4.2.\
Marshall-Olkin copula, $\alpha = 1/2$, $\beta = 3/4$\
[*Fig. 2.*]{} Left panel: dependence function $q_C(u, v)$ for the Marshall-Olkin copula; middle panel: scatter plot of $(R_i/(n+1),S_i/(n+1))$, $i = 1,\ldots,n$, $n = 500$, of simulated observations from the copula; right panel: standardized estimator $L_n(u,v)$ of $q_C(u, v)$ on the grid $\mathbbm {G}_{16}$. $L_{*} = -0.2$, $L^{*} = 12.2$.\
Mai-Scherer copula, $a = 0.9$, $b = 0.5$\
[*Fig. 3.*]{} Left panel: dependence function $q_C(u, v)$ for the Mai-Scherer copula; middle panel: scatter plot of $(R_i/(n+1),S_i/(n+1))$, $i = 1,\ldots,n$, $n = 500$, of simulated observations from the copula; right panel: standardized estimator $L_n(u,v)$ of $q_C(u, v)$ on the grid $\mathbbm {G}_{16}$. $L_{*} = 1.5$, $L^{*} = 16.1$.\
Next examples follow similar pattern. They concern three real data sets considered earlier by Jones and Koch (2003). In each example of our paper we use the same scale of intensity of colors in the heat maps. This allows one to compare how different degrees of association are reflected by our estimators.\
[*4.2. Example 2: Automobile data*]{}\
We shall consider two data sets of size $n=392$ available through [www/http://lib.stat.\
cmu.edu/datasets/cars]{}. This is 1983 [*ASA Data Exposition*]{} data set, collected by Ernesto Ramos and David Donoho.
The first sample collects observations of engine power (variable $X$), measured in horsepower, and fuel consumption (variable $Y$). This example was already investigated by Hawkins (1994), who fitted to the original data points decreasing regression function. Strong negative association is also clearly manifested by the scatter plot, which is based of transformed observations.\
, $n = 392$\
[*Fig. 4.*]{} Left panel: scatter plot of $(R_i/(n+1),S_i/(n+1))$, $i = 1,\ldots,n$, $n = 392$; middle panel: estimator $Q_n(u,v)$ of $q_C(u, v)$ on the grid $\mathbbm {G}_{16}$; right panel: standardized estimator $L_n(u,v) = \sqrt{n}\, Q_n(u,v)$ on the grid $\mathbbm {G}_{16}$. $L_{*} = -16.0$, $L^{*} = -1.2$.\
The heat map indicates visible negative trend and strong negative dependence in most of the points of the grid $\mathbbm {G}_{16}$. Only for some points close to the edges (0,0) and (1,1) the correlations are small in absolute value. To allow for some immediate quantitative analysis we give in Table 2 simulated quantiles of $L_n(u,v)$ for $n=392$ and two choices of $(u,v)$’s.\
Table 2. Simulated $\alpha$-quantiles of $L_n(u,v)$ for two selected $(u,v)$ and $n=392$.\
------------------------------- -------- -------- -------- ------- ------- -------
$(u,v)$ 0.01 0.05 0.10 0.90 0.95 0.99
$(\frac{1}{2},\frac{1}{2})$ -2.424 -1.616 -1.212 1.212 1.616 2.222
$(\frac{1}{16},\frac{1}{16})$ -1.266 -1.266 -1.266 1.320 2.182 3.044
------------------------------- -------- -------- -------- ------- ------- -------
\
The local correlations can be used to test independence and to verify local negative and positive dependence, as well. In particular, observe that except five values close to the vertices (0,0) and (1,1), where the standardized empirical local correlations are in \[-1.212,-2.000), the remaining ones are strictly less than -2.000. This, along with the rough information contained in Table 2, allow one to expect that all local correlations (in the grid points $\mathbbm {G}_{16}$) shall be accepted to be non positive on the standard level $\alpha$=0.05.
Obviously, the local correlations can be also used to form a new test statistic on global negative dependence. A reasonable candidate is the test rejecting such hypothesis for large values of $L^*$, see (12). By Theorem 1 of Ledwina and Wy[ł]{}upek (2014) such statistic preserves the correlation order. For the data under consideration, simulated in 10 000 MC runs, $p$-value of this test is equal to 1.
Our conclusion is that the heat map displayed in Figure 4 of this paper supports more simple picture of the overall dependence structure than this one presented in Figure 4 of Jones and Koch (2003) and obtained via kernel methods applied to the original data.\
The second sample of automobile data consists of observations of acceleration time ($X$) and fuel consumption ($Y$). Both, the original data display in Jones and Koch (2003) and the scatter plot provided in Figure 5 of our paper, suggest some not very strong positive dependence. The plot of $Q_n$ supports this suggestion. The heat map visualizes standardized correlations and gives better insight into the strength of this dependence. The strongest local correlations are observed close to the lower tails of both (transformed) variables and the strength of the dependence is getting weaker towards the upper tails. Again our look at the data reveals a simpler structure of the dependence that this one provided in Figure 5 of Jones and Koch (2003). In particular, we do not notice zero local dependence between moderately large values of both variables. Test rejecting positive local dependence for small values of $L_n(i/16,j/16), i,j=1,\ldots,15,$ can be applied in each grid point while global positive dependence can be verified by the test rejecting it for small values of $L_*$, see (12) and (11). For the given data, the simulated, on the basis of 10 000 observations, $p$-value of such global test on positive quadrant dependence is equal to 1.\
, $n = 392$\
[*Fig. 5.*]{} Left panel: scatter plot of $(R_i/(n+1),S_i/(n+1))$, $i = 1,\ldots,n$, $n = 392$; middle panel: estimator $Q_n(u,v)$ of $q_C(u, v)$ on the grid $\mathbbm {G}_{16}$; right panel: standardized estimator $L_n(u,v) = \sqrt{n}\, Q_n(u,v)$ on the grid $\mathbbm {G}_{16}$. $L_{*} = 0.8$, $L^{*} = 10.2$.\
[*4.3. Example 3: Aircraft data*]{}\
Consider $n=230$ aircraft span and speed data, on log scales, from years 1956-1984, reported and analyzed in Bowman and Azzalini (1997). We summarize the data in Figure 6. Since in this example both negative and positive correlations appear, we added respective signs to the colors in the heat maps. The figure exhibits that small and moderately large values of log speed are positively correlated with log span, while for the remaining cases the relation is reversed. Two, approximately symmetrically located, regions of relatively strong dependence are seen. In general, the strength of dependence is weaker than in previous cases. Similarly as in the previous example, also here our approach provides simpler and more regular picture of the dependence structure than this one presented in Figure 2 of Jones and Koch (2003).\
, $n = 230$\
[*Fig. 6.*]{} Left panel: scatter plot of $(R_i/(n+1),S_i/(n+1))$, $i = 1,\ldots,n$, $n = 230$; middle panel: estimator $Q_n(u,v)$ of $q_C(u, v)$ on the grid $\mathbbm {G}_{16}$; right panel: standardized estimator $L_n(u,v) = \sqrt{n}\, Q_n(u,v)$ on the grid $\mathbbm {G}_{16}$. $L_{*} = -6.5$, $L^{*} = 4.6$.\
Bowman and Azzalini (1997) used these data to discuss some drawbacks of standard correlation measures. Indeed, for these data classical Pearson’s, and Spearman’s and Blomqvist’s rank statistics for assessing an association yield simulated $p$-values 0.81, 0.74, and 0.79, respectively. Kendall’s rank correlation gives simulated $p$-value 0.31, which also seems to be too high, when one is looking at the magnitude of standardized local correlations in Figure 6. Combining the local correlation into global statistic $L^o =\max_{1 \leq i,j \leq 15} |L_n(i/16,j/16)|$, with large values being significant, basing on simulation of size 10 000, we get $p$-value 0.00 for such global independence test. This shows that local correlations are more informative than each of the above single classical global indices of association.\
[**5. Discussion**]{}\
We have introduced the novel function valued measure of dependence of two random variables. Its definition, based on Studentized difference of two cdf’s, is general, simple, and natural. In our considerations, we mainly focus attention on copula-based variant of the measure. It allows for simple estimation and guarantees appealing finite sample properties of the resulting estimate. The estimate is tightly linked to the popular scatter plot and helps to extract explicit dependence structure from it. Both, the measure and the estimate, allow for comparison and visualization of different association structures. The value of the measure in a fixed point has useful interpretation as correlation coefficient of some specific increasing functions of the marginals. Also, the proposed estimate features simple interpretation and easy implementation. Its performance in real data analysis yields relatively simple, in comparison to alternative method, dependence structure. We believe that the proposed approach will be useful in practice.
Also, simple and reliable tests for local and global association, based on estimated dependencies have been proposed. It is worth noticing that, in particular, statistic like $L^*$, cf. (12), can be considered as a usable approximation of empirical isotonic canonical correlation coefficient, introduced in Schriever (1987). Similarly, $L^o$, given in (13), can be serve as an easy to implement approximate exemplification of Rényi’s idea to calculate maximal correlation over large class of functions. In Ledwina and Wy[ł]{}upek (2014) empirical correlations close to $L_n(u,v)$’s were successfully applied to construct highly sensitive test for detection of positive quadrant dependence. Recently, there is much of interest in detecting dependencies in some conditional copulas; see Veraverbeke et al. (2011), and Li et al. (2014) for discussion and further references. It seems that some graphical presentation of dependence structure of two random variables, conditionally upon some fixed values of a covariate, and formal application of pertaining counterparts of $L^o, L^*$, and $L_*$ could be useful in such considerations, as well. It is also worthy noting that the definition of $q$ can be naturally extended to higher dimensions and applied to construct tests for positive orthant dependence, for example. These questions are however beyond the scope of this initial article.\
[**References**]{}\
I. Bairamov, S. Kotz, T.J. Kozubowski (2003), A new measure of linear local dependence, [*Statistics*]{}, [ 37]{}, 243-258.
N. Balakrishnan, Ch.-D. Lai (2009), [*Continuous Bivariate Distributions*]{}, Springer, Dordrecht.
G. Bertensen, B. St[ø]{}ve, D. Tj[ø]{}stheim, T. Nordb[ø]{} (2013) Recognizing and visualizing copulas: an approach using local Gaussian approximation, manuscript, Department of Mathematics, University of Bergen, http://folk.uib.no/gbe062/localgaussian-correlation.
S. Bjerve, K. Doksum (1993), Correlation curves: measures of association as function of covariate values, [*The Annals of Statistics*]{}, [ 21]{}, 890-902.
A.W. Bowman, A. Azzalini (1997), [*Applied Smoothing Techniques for Data Analysis*]{}, Clarendon Press, Oxford.
S. Cambanis, G. Simons, W. Stout (1976), Inequalities for $Ek(X,Y)$when marginals are fixed, [*Zeitschrift für Wahrscheinlichkeitstheorie und vervandte Gebiete*]{}, [ 36]{}, 285-294.
P. Deheuvels, La Fonction de depéndance empirique et ses propriétś (1979), [*Academie Royale Belgique, Bulletin de la Classe des Sciences, 5e Série*]{}, [65]{}, 274-292.
J. Dhaene, M. Denuit, S. Vanduffel (2009), Correlation order, merging and diversification, [*Insurance: Mathematics and Economics*]{}, [ 45]{}, 325-332.
D. Drouet Mari, S. Kotz (2001), [*Correlation and Dependence*]{}, Imperial College Press, London.
P. Embrechts, A. McNeil, D. Straumann (2002), Correlation and dependency in risk management: properties and pitfalls, in: Dempster, M., Moffatt, H. (Eds.), [*Risk Management: Value at Risk and Beyond*]{}, Cambridge: Cambridge University Press, Cambridge, pp. 176-223.
J.-D. Fermanian, D. Radulović, M. Wegkamp (2004), Weak convergence of empirical copula process, [*Bernoulli*]{}, [ 10]{}, 847-860.
N.I. Fisher, P. Switzer (1985), Chi-plots for assessing dependence, [*Biometrika*]{}, 72, 253-65.
D.M. Hawkins, Fitting monotonic polynomials to data (1994), [*Computational Statistics*]{}, 9, 233-247.
P.W. Holland, Y.J. Wang (1987), Dependence function for continuous bivariate densities, [*Communications in Statistics - Theory and Methods*]{}, [ 16]{}, 863-876.
R.J. Hyndman, Y. Fan (1996), Sample quantiles in statistical packages, [*The American Statistician*]{}, 50, 361-365.
P. Janssen, J. Swanepoel, N. Veraverbeke (2012), Large sample behavior of the Bernstein copula estimator, [*Journal of Statistical Planning and Inference*]{}, [142]{}, 1189-1197.
K. Jogdeo (1982). Dependence, concepts of, in: S. Kotz, N.L. Johnson (Eds), [*Encyclopedia of Statistical Sciences*]{}, Vol. 2. Wiley, New York, pp. 324-334.
M.C. Jones (1998), Constant local dependence, [*Journal of Multivariate Analysis*]{}, [ 64]{}, 148-155.
M.C. Jones, I. Koch (2003), Dependence maps: local dependence in practice, [*Statistics and Computing*]{}, [ 13]{}, 241-255.
T. Kowalczyk, E. Pleszczyńska (1997), Monotonic dependence functions of bivariate distributions, [*The Annals of Statistics*]{}, [ 5]{}, 1221-1227.
H.O. Lancaster (1982), Dependence, measures and indices of, in: S. Kotz, N.L. Johnson (Eds), [*Encyclopedia of Statistical Sciences*]{}, Vol. 2. Wiley, New York, pp. 334-339.
T. Ledwina, G. Wy[ł]{}upek (2014), Validation of positive quadrant dependence, [*Insurance: Mathematics and Economics*]{}, 56, 38-47.
R. Li, Y. Cheng, J.P. Fine (2014), Quantile association regression models, [*Journal of the American Statistical Association*]{}, 109, 230-242.
J.-F. Mai, M. Scherer (2011), Bivariate extreme-value copulas with discrete Pikands dependence measure, [*Extremes*]{}, [ 14]{}, 311-324.
R.B. Nelsen (2006), [*An Introduction to Copulas*]{}, Springer, New York.
M. Omelka, I. Gijbels, N. Veraverbeke (2009), Improved kernel estimation of copulas: weak convergence and goodness-of-fit, [*The Annals of Statistics*]{} [37]{}, 3023-3058.
B. Póczos, Z. Ghahramani, J. Schneider (2012), Copula-based kernel dependency measures, in [*Proceedings of the 29th International Conference on Machine Learning*]{}, New York; Omnipress, pp. 775-782.
B. Schweizer, E.F. Wolff (1981), On nonparametric measures of dependence for random variables, [*The Annals of Statistics*]{}, [9]{}, 879-885.
B.F. Schriever (1987), An ordering for positive dependence, [*The Annals of Statistics*]{}, 15, 1208-1214.
J.W.H. Swanepoel, J.S. Allison (2013), Some new results on the empirical copula estimator with applications, [*Statistics and Probability Letters*]{}, [ 83]{}, 1731-1739.
D. Tj[ø]{}stheim, K.O. Hufthammer (2013), Local Gaussian correlation: A new measure of dependence, [*Journal of Econometrics*]{}, [172]{}, 33-48.
N. Veraverbeke, M. Omelka, I. Gijbels (2011), Estimation of a conditional copula and association measures, [*Scandinavian Journal of Statistics*]{}, 38, 766-780.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Named entity recognition (/) is the very first step in the linguistic processing of any new domain. It is currently a common process in BioNLP on English clinical text. However, it is still in its infancy in other major languages, as it is the case for Spanish. Presented under the umbrella of the / shared task, this paper describes a very simple method for the annotation and normalization of pharmacological, chemical and, ultimately, biomedical named entities in clinical cases. The system developed for the shared task is based on limited knowledge, collected, structured and munged in a way that clearly outperforms scores obtained by similar dictionary-based systems for English in the past. Along with this recovering of the knowledge-based methods for / in subdomains, the paper also highlights the key contribution of *resource-based* systems in the validation and consolidation of both the annotation guidelines and the human annotation practices. In this sense, some of the authors discoverings on the overall quality of human annotated datasets question the above-mentioned ‘official’ results obtained by this system, that ranked second (0.91 F1-score) and first (0.916 F1-score), respectively, in the two / subtasks.'
author:
- |
Fernando Sánchez León\
*unaffiliated*\
[[email protected]]{}\
Ana González Ledesma\
*unaffiliated*\
[[email protected]]{}
bibliography:
- 'PharmaCoNER-FSL-AGL\_final.bib'
title: 'Annotating and normalizing biomedical with limited knowledge[^1]'
---
Introduction
============
Named Entity Recognition (/) is considered a necessary first step in the linguistic processing of any new domain, as it facilitates the development of applications showing co-occurrences of domain entities, cause-effect relations among them, and, eventually, it opens the (still to be reached) possibility of understanding full text content. On the other hand, Biomedical literature and, more specifically, clinical texts, show a number of features as regards / that pose a challenge to NLP researchers [@Cohen_and_Demner-Fushman:2014]: (1) the clinical discourse is characterized by being conceptually very dense; (2) the number of different classes for /s is greater than traditional classes used with, for instance, newswire text; (3) they show a high formal variability for /s (actually, it is rare to find entities in their “canonical form”); and, (4) this text type contains a great number of ortho-typographic errors, due mainly to time constraints when drafted.
Many ways to approach / for biomedical literature have been proposed, but they roughly fall into three main categories: rule-based, dictionary-based (sometimes called knowledge-based) and machine-learning based solutions. Traditionally, the first two approaches have been the choice before the availability of Human Annotated Datasets (/), albeit rule-based approaches require (usually hand-crafted) rules to identify terms in the text, while dictionary-based approaches tend to miss medical terms not mentioned in the system dictionary [@Rebholz-Schumann_et_al:2011]. Nonetheless, with the creation and distribution of / as well as the development and success of supervised machine learning methods, a plethora of data-driven approaches have emerged —from Hidden Markov Models (HMMs) [@Ephraim:2002], Support Vector Machines (SVMs) [@Habib_and_Kalita:2010] and Conditional Random Fields (CRFs) [@He_and_Kayaalp:2008], to, more recently, those founded on neural networks [@Armengol-Estapeetal:2019]. This fact has had an impact on knowledge-based methods, demoting them to a second plane. Besides, this situation has been favoured by claims on the uselessness of gazetteers for / in, for example, Genomic Medicine (GM), as it was suggested by @Cohen_and_Demner-Fushman:2014 :
> One of the findings of the first BioCreative shared task was the demonstration of the long-suspected fact that gazetteers are typically of little use in GM.
Although one might think that this view could strictly refer to the subdomain of GM and to the past —BioCreative I was a shared task held back in 2004—, we can still find similar claims today, not only referred to rule-based and dictionary-based methods, but also to stochastic ones [@Armengol-Estapeetal:2019].\
In this paper, in spite of previous statements, we present a system that uses rule-based and dictionary-based methods combined (in a way we prefer to call *resource-based*). Our final goals in the paper are two-fold: on the one hand, to describe our system, developed for the / shared task[^2], dealing with the annotation of some of the /s in health records (namely, pharmacological, chemical and biomedical entities) using a revisited version of rule- and dictionary-based approaches; and, on the other hand, to give pause for thought about the quality of datasets (and, thus, the fairness) with which systems of this type are evaluated, and to highlight the key role of resource-based systems in the validation and consolidation of both the annotation guidelines and the human annotation practices.\
In section \[sect:resources\], we describe our initial resources and explain how they were built, and try to address the issues posed by features (1) and (2) above. Section \[sect:devel\] depicts the core of our system and the methods we have devised to deal with text features (3) and (4). Results obtained in / by our system are presented in section \[sect:results\]. Section \[sect:discussion\] details some of our errors, but, most importantly, focusses on the errors and inconsistencies found in the evaluation dataset, given that they may shed doubts on the scores obtained by any system in the competition. Finally, we present some concluding remarks in section \[sect:conclusions\].
Resource building {#sect:resources}
=================
As it is common in resource-based system development, special effort has been devoted to the creation of the set of resources used by the system. These are mainly two —a flat subset of the / medical ontology[^3], and the library and a part of the contextual regexp grammars developed by @FSL:2018 for a previous competition on abbreviation resolution in clinical texts written in Spanish. The process of creation and/or adaptation of these resources is described in this section.
SNOMED CT {#subsect:snomed}
---------
Although the competition proposes two different scenarios, in fact, both are guided by the / ontology —for subtask 1, entities must be identified with offsets and mapped to a predefined set of four classes (`PROTEINAS`, `NORMALIZABLES`, `NO_NORMALIZABLES` and `UNCLEAR`); for subtask 2, a list of all / <span style="font-variant:small-caps;">id</span>s (/) for entities occurring in the text must be given, which has been called *concept indexing* by the shared task organizers[^4]. Moreover, / organizers decided to promote / substance <span style="font-variant:small-caps;">id</span>s over product, procedure or other possible interpretations also available in this medical ontology for a given entity. This selection must be done even if the context clearly refers to a different concept, according to the annotation guidelines[^5] (henceforth, /) and the praxis. Finally, `PROTEINAS` is ranked as the first choice for substances in this category.
These previous decisions alone on the part of the organizers greatly simplify the task at hand, making it possible to build (carefully compiled) subsets of the entities to be annotated. This is a great advantage over open domain /, where (like in GM) the texts may contain an infinite (and very creative indeed) number of /s. For clinical cases, although the / density is greater, there exist highly structured terminological resources for the domain. Moreover, the set of classes to use in the annotation exercise for subtask 1 has been dramatically cut down by the organizers.
With the above-mentioned initial constraints in mind, we have painstakingly collected, from the whole set of / terms, instances of entities as classified by the human annotators in the datasets released by the organizers and, when browsing the / web version, we have tried to use the ontological hierarchical relations to pull a complete class down from /. This way, we have gathered 80 classes —from lipids to proteins to peptides or peptide hormones, from plasminogen activators to dyes to drugs or medicaments—, that have been arranged in a ranked way so as to mimic human annotators choices[^6]. The number of entities so collected (henceforth, ‘primary entities’) is 51,309.
Contextual regexp grammars {#subsect:regexp}
--------------------------
Some of the entities to be annotated, specially those in abbreviated form, are ambiguous without a context. This is the case, for instance, of *PCR*, whose expanded forms are (among other meanings; we use only English expanded forms) ‘reactive protein c’, ‘polymerase chain reaction’, ‘cardiorespiratory arrest’. In order to deal with these cases, we use a contextual regexp rule system with a lean and simple rule formalism previously developed [@FSL:2018]. As an exemplification, we include one rule to deal with one of the cases of the preceeding ambiguity:
b:[il::bioquímica|en sangre|hemoglobina|
hemograma|leucocit|parásito|plaqueta|
prote.na|recuento|urea] - [PCR] - >
[m=proteína]
A rule has a left hand side (LHS) and a right hand side (RHS). There is a focus in the LHS (`PCR`, within dashes) and a left and right context (that may be empty). When the left context includes a `b:` (like in this case), it indicates either left or right context. The words in the context can take other qualifiers —in this case, the matching will be case insensitive (`i` to the left of `bioquímica`) and local (`l`), which means the disjunction of words and/or stems can be found in a distance of 40 characters (this can be modifified by the user). Hence, the rule applies, selecting the `proteína` expansion (in RHS) of `PCR` if any of the words/stems specified as local context (40 chars maximum) is matched either to the left or right of the focus term (which is usually an abbreviation).
With no tweaking at all for the datasets in / competition, the system annotates correctly 18 out of 20 occurrences of *PCR* in the test dataset (a precision of 0.9)[^7].
This component of the system is important because, only when the previous abbreviation is expanded as the first string (that of a protein name), it must be annotated, according to the /. The same ambiguity happens with *Cr*, which may mean ‘creatinine’ or ‘chrome’[^8]. These expansions are both `NORMALIZABLES`, but, obviously, their / is different.
The system currently uses 104 context rules, only for abbreviations and acronyms in the clinical cases. These rules, contrary to what is commonly referred in the biomedical processing literature [@Armengol-Estapeetal:2019], do not require a special domain knowledge (none of the authors do have it) and can be written, most of the times, in a very straightforward way in the formalism briefly described above.
Development {#sect:devel}
===========
In general, dictionary-based methods rely on strict string matching over a fixed set of lexical entries from the domain. This is clearly insufficient to deal with non-canonical linguistic forms of /s as used in clinical texts. For this reason, we have devised two different solutions to this shortcoming.
In the first place, we have munged a great number of our primary entities, in a way similar to that described in @FSL:2019a for gazetteers used for protected information anonymization in clinical texts. We basically transform canonical forms in other possible textual forms observed when working with biomedical texts. With such transformations, a system module converts a salt compound like *clorhidrato de ciclopentolato* into *ciclopentolato clorhidrato*, or simply the `PP` *de potasio* into its corresponding adjective *potásico*. Other, more complex conversions include the treatment of antibodies —for instance *anticuerpo contra especie de Leishmania* becomes *ac. Leishmania*, among other variants—, or pairs of antibiotics normally prescribed together —which have a unique / and whose order we handle just as the ‘glueing’ characters. Note, incidentally, that, while the input to this pre-processing step is always a string, the output can be a regular expression, that is linked to a /. Plural forms are also generated through this module, that uses 45 transformations (not all equally productive). Using these transformation rules, we produce 139,150 ‘secondary entities’, many of them regexps. As a final (simple) example of this, consider the entity *antígeno CD13*: after applying one of the previous string-to-regexp transformations, it is converted to:
(?:antígeno )?CD[- ]?13
With the previous regexp, the system is able to identify (and string-normalize) six different textual realizations of the same unique / term. There are more complex rules that, thus, produce many more potential strings. The important thing with this strategy is that through the generative power of these predictably-created regexps from / entities the system is able to improve its recall and overcome the limitations of traditional dictionary-based approaches.\
Secondly, to tackle with careless drafting of clinical reports, a Levenshtein edit distance library[^9] is used on the whole background dataset. The process is run once, using our secondary entities as lexicon[^10] and a general vocabulary lexicon to rule out common words in the candidate search process. We have used distances in the range 1-3 (depending on string length) for sequences up to 3 words long[^11]. The output of this process, which links forms with spelling errors with canonical ones and, thus, to /s, can be inspected prior to its inclusion in the system lexicon, if so desired.\
Annotation process
------------------
As such, the annotation process is very simple. The program reads the input byte stream trying to identify known entities by means of a huge regexp built through the pre-processing of the available resources. If the candidate entity is ambiguous and (at least) one contextual rule exists for it, it is applied. For the rest of the /s, the system assigns them the class and / found in our ranked in-memory lexicon. As already mentioned in passing, the system does not tokenize text prior to /, a processing order that we consider the right choice for highly entity-dense texts. The data structures built during pre-processing are efficiently stored on disk for subsequent runs, so the pre-processing is redone only when resources are edited.
Results {#sect:results}
=======
According to the organizers, and taking into account the / of the tiny subset from the background dataset released to the participants[^12], the system obtained the scores presented in table \[table:results\], ranking as second best system for subtask1 and best system for subtask2 [@Agirre:2019].[^13].
/
Precision Recall F1-score
--------------- ----------- --------- ----------
**Subtask 1** 0.90625 0.91314 0.90968
**Subtask 2** 0.91108 0.92083 0.91593
: Results for / test dataset (both subtasks)[]{data-label="table:results"}
Our results are consistent with our poor understanding of the classes for subtask 1. Having a null knowledge of Pharmacology, Biomedicine or even Chemistry, assigning classes (as requested for subtask 1) to entities is very hard, while providing a / (subtask 2) seems an easier goal. We will explain the point with an example entity —*ácido hialurónico* (‘hyaluronic acid’). Using the ontological structure of /, one can find the following parent relations (just in English):\
*hyaluronic acid* <span style="font-variant:small-caps;">`is-a`</span> *mucopolysaccharide* <span style="font-variant:small-caps;">`is-a`</span> *protein*\
The authors have, in this case, promoted the `PROTEINAS` annotation for this entity, disregarding its interpretation as a replacement agent and overlooking a recommendation on polysaccharides in the /. Fortunately, all its interpretations share a unique /. The same may be true for\
*haemosiderin* <span style="font-variant:small-caps;">`is-a`</span> *protein*\
which is considered `NORMALIZABLE` in the test dataset. Similar cases are responsible for the lower performance on subtask 1 with respect to the more complex subtask 2.\
In spite of these human classification errors, our system scores outperform those obtained by PharmacoNER Tagger[^14] [@Armengol-Estapeetal:2019], a simpler system using a binary classification and a very different organization of the dataset with a smaller fragment for test (10% of the data as opposed to 25% for the official competition). In fact, our system improves their F1-score (89.06) by 1.3 points when compared with our results for the more complex / subtask 1.
Discussion {#sect:discussion}
==========
In this section, we perform error analysis for our system run on the test dataset. We will address both recall and precision errors, but mainly concentrate on the latter type, and on a thorough revision of mismatches between system and human annotations.\
In general, error analysis is favoured by knowledge-based methods, since it is through the understanding of the underlying reasons for an error that the system could be improved. Moreover, and differently to what happens with the current wave of artificial neural network methods, the whole annotation process —its guidelines for human annotators, the collection and appropriate structuring of resources, the adequate means to assign tags to certain entities but not to other, similar or even pertaining to the same class— must be clearly understood by the designer/developer/data architect of such systems. As a natural consequence of this attempt to mimick a task defined by humans to be performed, in the first place, also by humans, some inconsistencies, asystematic or missing assignments can be discovered, and this information is a valuable treasure not only for system developers but also for task organizers, guideline editors and future annotation campaigns, not to mention for the exactness of program evaluation results.
Most of the error types made by the system (i.e., by the authors) in class assignment for subtask 1 have already been discussed. In the same vein, as regards subtask 2, a great number of errors come from the selection of the ‘product containing substance’ reading from / rather to the ‘substance’ itself. This is due to inexperience of the authors on the domain and the wrong consideration of context when tagging entities —the latter being clearly obviated in the /.
In the following paragraphs, some of the most relevant inconsistencies found when performing error analysis of our system are highlighted. The list is necessarily incomplete due to space constraints, and it is geared towards the explanation of our possible errors.
Inconsistency in the [****]{} {#subsect:inconsistency:guide}
-----------------------------
Among some of the paradoxical examples in the / it stands out the double explicit consideration of *gen* (‘gene’), when occurs alone in context, as both an entity to be tagged (positive rule P2 of the /) and a noun not to be tagged (negative rule N2). This inconsistency (and a bit of bad luck) has produced that none of the 6 occurrences as an independent noun —not introducing an entity— is tagged in the train[[`+`]{}]{}dev (henceforth, t[[`+`]{}]{}d) while the only 2 in the same context in the test dataset have been tagged. This amounts for 2 true negatives (<span style="font-variant:small-caps;">tn</span>s) for the evaluation script.
Inconsistency in [****]{} as regards [****]{} {#subsect:inconsistency:interpreting-guide}
---------------------------------------------
The / proposal for the treatment of elliptical elements is somewhat confusing. For these cases, a longest match annotation is proposed, which is difficult to replicate automatically and not easy to remember for the human annotator. In many contexts, the annotator has made the right choice —for instance, in *receptores de estrógeno y de progesterona*— whereas in others do not —*$|$anticuerpos anticardiolipina$|$ $|$IgG$|$ e $|$IgM$|$*, with ‘$|$’ marking the edges of the annotations. The last example occurs twice in the test dataset. Hence, the disagreement counts as 6 <span style="font-variant:small-caps;">tn</span>s and 2 false positives (<span style="font-variant:small-caps;">fp</span>s)[^15].
On the other hand, there is a clear reference to food materials and nutrition in the /, where they are included in the class of substances. However, none of the following entities is tagged in the test dataset: *azúcar* (which is mandatory according to / and was tagged in t[[`+`]{}]{}d; 1 <span style="font-variant:small-caps;">fp</span>); *almidón de maíz* (also mandatory in /; 1 <span style="font-variant:small-caps;">fp</span>); and *Loprofín*, *Aglutella*, *Aproten* (hypoproteic nutrition products, 3 <span style="font-variant:small-caps;">fp</span>s in total)[^16].
There is an explicit indication in the / to annotate salts, with the example *iron salts*. However, in the context *sales de litio* (‘lithium salts’), only the chemical element has been tagged (1 <span style="font-variant:small-caps;">fp</span>[^17]).
There exist other differing-span mismatches between human and automatic annotation. These include *anticuerpos anticitoplasma de neutrófilo*, where the / considers the first two words only (in one of the occurrences, 1 <span style="font-variant:small-caps;">fp</span>); in the text fragment *b2 microglobulina, CEA y CA 19,9 normales*, *CA 19,9* is the correct span for the last entity (and not *CA*, 1 <span style="font-variant:small-caps;">fp</span>); *A.S.T* is the span selected (for *A.S.T.*, 1 <span style="font-variant:small-caps;">fp</span>); finally, in the context *lgM anticore* only *lgM* has been tagged (1 <span style="font-variant:small-caps;">fp</span>).
Other prominent mismatch between / and / is that of *DNA*, which is explicitly included in the / (sects. P2 and O1). It accounts for 2 <span style="font-variant:small-caps;">fp</span>s.\
But perhaps one of the most common discrepancies between human and automatic annotation has to do with medicaments normally prescribed together, which have a unique /. Examples include *amiloride/hidroclorotiazida* (1 <span style="font-variant:small-caps;">fp</span>); and *betametasona + calcipotriol* (1 <span style="font-variant:small-caps;">fp</span>) in the test set. This situation was also observed in the t[[`+`]{}]{}d corpus fragment (*tenofovir + emtricitabina*, *carbonato cálcico /colecalciferol*, *lopinavir/ritonavir*).
Inconsistency in [****]{} on the test set as regards t[[`+`]{}]{}d sets {#subsect:inconsistency:train}
-----------------------------------------------------------------------
Some inconsistencies between dataset annotations have turned the authors crazy: *NPT* (acronym for ‘total parenteral nutrition, TPN’) is tagged in the train[[`+`]{}]{}dev dataset 15 out of 21 times it occurs[^18]. The common sense of frequency in the / of texts has led us to tag it in the background set. Unluckily, neither *NPT* nor its expansion have been tagged in the test dataset. This has also been the behaviour in / for ‘parenteral nutrition’ and ‘enteral nutrition’ (and their corresponding acronyms) in test dataset, since these entities have not been tagged. We asked the organizers about this and other entities for which we had doubts, either because the / didn’t cover their cases or because the / didn’t match the recommendations in the /. Woefully, communication with the organizers has not been very fluent on this respect. All in all, this bad decision on the part of the authors amounts for 6 <span style="font-variant:small-caps;">fp</span>s (more than 7.5% of our <span style="font-variant:small-caps;">fp</span>s according to evaluation script).
For other cases, decisions that may be clearly induced from the tagging of train[[`+`]{}]{}dev datasets, have not been applied in the test corpus fragment. These include *cadenas ligeras* (5 times in t[[`+`]{}]{}d, 1 <span style="font-variant:small-caps;">fp</span> in test); *enzimas hepáticas* (tagged systematically in t[[`+`]{}]{}d, 1 <span style="font-variant:small-caps;">fp</span>); *p53* (also tagged in t[[`+`]{}]{}d, 1 <span style="font-variant:small-caps;">fp</span>).
Another entity that stands out is *hidratos de carbono* (‘carbohydrates’). It is tagged twice in the t[[`+`]{}]{}d dataset, occurring 4 times in the set (once as *HC*). However, although the form *carbohidratos* has been annotated twice in the test set, *hidratos de carbono* has been not (1 <span style="font-variant:small-caps;">fp</span>).
Moreover, *suero* (‘Sodium chloride solution’ or ‘serum’) deserves its own comment. Both entity references are tagged in the train[[`+`]{}]{}dev datasets (although with the latter meaning it is tagged only 4 out of 12 occurrences). We decided to tag it due to its relevance. In the test dataset, it occurs 5 times with the blood material meaning, but it has only been tagged twice as such (one of them being an error, since it refers to the former meaning). Our system tagged all occurrences, but tagged also one of the instances with the former meaning as serum (3 <span style="font-variant:small-caps;">fp</span>s).
Finally, there are some inconsistencies within the same dataset. For example, nutricional agent *Kabiven* is tagged as both `NORMALIZABLES` (with /) and `NO_NORMALIZABLES` in the very same text. The same happens with another nutritional complement, *Cernebit*, this time in two different files. The perfusion solution *Isoplasmal G* (with a typo in the datasets —*Isoplasmar G*) is tagged as `NORMALIZABLES` and `UNCLEAR`. These examples reveal a vague understanding (or definition) of criteria as regards fluids and nutrition, as we pointed out at the beginning of this section.
Asystematic/incomplete annotation {#subsect:asystematic}
---------------------------------
Some of the entities occurring in the test dataset have not always been tagged. This is the case for *celulosa* (annotated only once but used twice, 1 <span style="font-variant:small-caps;">fp</span>); *vimentina* (same situation as previous, 1 <span style="font-variant:small-caps;">fp</span>); *LDH* (tagged 20 times in t[[`+`]{}]{}d but not in one of the files, 1 <span style="font-variant:small-caps;">fp</span>); *cimetidina* (1 <span style="font-variant:small-caps;">fp</span>); *reactantes de fase aguda* (2 <span style="font-variant:small-caps;">fp</span>s; 2 other occurrences were tagged); *anticuerpos antinucleares* (human annotators missed 1, considered <span style="font-variant:small-caps;">fp</span>).
Incorrect [s]{} {#subsect:incorrect}
---------------
On our refinement work with the system, some incorrect /s have emerged. These errors impact on subtask 2 (some also on subtask 1). A large sample of them is enumerated below.
*ARP* (‘actividad de renina plasmática’, ‘plasma renin activity’, PRA) cannot be linked to / for *renina*, which happens twice. In the context ‘perfil de antigenos \[sic\] extraíbles del núcleo (ENA)’, *ENA* has been tagged with / of the antibody (1 <span style="font-variant:small-caps;">fp</span>). In one of the files, *tioflavina* is linked to / of *tioflavina T*, but it could be *tioflavina S*. Thus, it should be `NO_NORMALIZABLE`. *Harvoni* is and not (1 <span style="font-variant:small-caps;">fp</span>). *AcIgM contra CMV* has a wrong / (1 <span style="font-variant:small-caps;">fp</span>). *HBsAg* has no / in the test set; it should be (‘Hepatitis B surface antigen’) (1 <span style="font-variant:small-caps;">fp</span>).\
There are other incorrect annotations, due to inadvertent human errors, like *biotina* tagged as `PROTEINAS` or *VEB* (‘Epstein-Barr virus’) being annotated when it is not a substance. Among these mismatches between / and system annotation, the most remarkable is the case of synonyms in active principles. For instance, the brand name drug *Dekapine* has been linked to ‘ácido valproico’ in the former case and to ‘valproato sódico’ in the latter. These terms are synonymous[^19], but sadly they don’t share /. Hence, this case also counts as a <span style="font-variant:small-caps;">fp</span>.\
A gold standard dataset for any task is very hard to develop, so a continuous editing of it is a must[^20]. In this discussion, we have focused on false positives (<span style="font-variant:small-caps;">fp</span>s) according to the script used for system evaluation, with the main purpose of *understanding* the domain knowledge encoded in the linguistic conventions (lexical/terminological items and constructions) used by health professionals, but also the decisions underlying both the / and the / practice.
In this journey to system improvement and authors enlightenment, some inconsistencies, errors, omissions have come up, as it has been reflected in this section, so both the guidelines for and the practice of annotation can also be improved in future use scenarios of the clinical case corpus built and maintained by the shared task organizers.
Our conclusion on this state of affairs is that some of the inconsistencies spotted in this section show that there were not a rational approach to the annotation of certain entities contained in the datasets (apart from other errors and/or oversights), and, hence, the upper bound of any tagging system is far below the ideal 1.0 F1-score. To this respect, in very many cases, the authors have made the wrong choice, but in others they were guided by analogy or common sense. Maybe a selection founded on probability measures estimated on training material could have obtained better results with this specific test dataset. However, in the end, this cannot be considered as an indication of a better system performance, since, as it has been shown, the test dataset used still needs more refinement work to be used as the right dataset for automatic annotation evaluation.
Conclusions {#sect:conclusions}
===========
With this resource-based system developed for the / shared task on / of pharmacological, chemical and biomedical entities, we have demonstrated that, having a very limited knowledge of the domain, and, thus, making wrong choices many times in the creation of resources for the tasks at hand, but being more flexible with the matching mechanisms, a simple-design system can outperform a / tagger for biomedical entities based on state-of-the-art artificial neural network technology. Thus, knowledge-based methods stand on their own merits in task resolution.
But, perhaps most importantly, the other key point brought to light in this contribution is that a resource-based approach also favours a more critical stance on the dataset(s) used to evaluate system performance. With these methods, system development can go hand in hand with dataset refinement in a virtuous circle that let us think that maybe next time we are planning to add a new gazetteer or word embedding to our system in order to try to improve system performance, we should first look at our data and, like King Midas, turn our Human Annotated Dataset into a true Gold Standard Dataset.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank three anonymous reviewers of our manuscript for their careful reading and their many insightful comments and suggestions. We have made our best in providing a revised version of the manuscript that reflects their suggestions. Any remaining errors are our own responsability.
[^1]: This paper should have been published in the *Proceedings of the 5th Workshop on BioNLP Shared Tasks*. Unfortunately, due to their complete lack of funding, the authors could not afford the registration fees, a mandatory expense for a contribution to be published in the aforementioned proceedings.
[^2]: <http://temu.bsc.es/pharmaconer/>
[^3]: From <https://browser.ihtsdotools.org/>.
[^4]: In the train[[`+`]{}]{}dev datasets, only 17 of the `PROTEINAS` (‘proteins’) and `NORMALIZABLES` (‘standardizable’) entities have an <span style="font-variant:small-caps;">id</span> not in the / ontology. Besides, just 40 out of 5,615 annotations —not taking into account the class `UNCLEAR`, which is not considered for the system evaluation— are tagged as `NO_NORMALIZABLES` (‘non standardizable’), many of them due to the fact that they include elliptical constructions.
[^5]: <https://bit.ly/2qxofgd>, p. 4.
[^6]: Note that we have gathered the complete set of medical terms included in /, but, for the purpose of this shared task, we only use a subset of it.
[^7]: Note that 2 of the *PCR* occurrences in the train[[`+`]{}]{}dev datasets have been incorrectly mapped to the protein interpretation (file `S1130-63432014000100012-1`, 2 times).
[^8]: Again, one of the occurrences of *Cr* has been incorrectly mapped to the former extended form (file `S0212-16112012000500042-1`).
[^9]: We use `Text::Levensthtein::Flexible` library, from `Perl` ecosystem. One of the anonymous reviewers has shed doubts about the use of `Perl` as a language for “NLP and text-mining nowadays”. In this respect, we are not committed with a given programming language more than we are with our native language —and we have submitted our paper in English, a foreign language for us. The system could have been implemented in any other programming language more popular “nowadays”, provided that we were as *proficient* in it as we are in `Perl` and the language used were as efficient in string and regexp handling and in I/O operations as `Perl` is. In this regard, the most popular language nowadays —`Python`— is 2 to 10 times slower for these particular features. `Perl` is even faster for regexp processing than `Python PyPy` —see, for instance, <https://github.com/mariomka/regex-benchmark>. Idiomatic `Perl` is even faster. Finally, `Perl` has a long tradition in biology and medicine text processing.
[^10]: With enumeration of strings from non-infinite-loop regexps.
[^11]: These words are not isolated from the byte stream, and the process uses textual anchors to delimit them as word candidates. Consequently, no proper tokenization is performed.
[^12]: When compared with the rest of the tasks in BioNLP-OST 2019, the time given to participants to submit their system runs is 4 times longer than the mean —longer time that is unnecessary if system is mature enough. On the other hand, the dataset released for evaluation purposes is more than 4 times larger than the mean. As a consequence, participating groups have to annotate full domain corpora rather than just test dataset(s). A shorter submission period and a smaller test dataset would be preferable, and besides fairer, in future calls.
[^13]: The authors have been unable to obtain these results with the official script, downloaded from <https://github.com/PlanTL-SANIDAD/PharmaCoNER-CODALAB-Evaluation-Script>. In their execution of the evaluation script, system results are better (?).
[^14]: The tagger authors, some of them also organizers of shared task, have changed the casing of the name for the program.
[^15]: When we indicate this kind of information, mostly using only <span style="font-variant:small-caps;">fp</span>s, it must be understood that the system made the choice(s) that the authors judge as correct, although disagreeing with / and/or /.
[^16]: On nutrition replacements, see also section \[subsect:inconsistency:train\].
[^17]: Note, in passing, that these span errors account for 1 <span style="font-variant:small-caps;">tn</span> also for the evaluation scripts.
[^18]: However, at least one expanded variant of it —*nutrición parenteral*, ‘parenteral nutrition’— is never tagged.
[^19]: Although, ‘valproato sódico’ is the name used in the leaflet, as it can be seen in the Spanish Medicament Agency, AEMPS, web page ([https://cima.aemps.es/cima/dochtml/p/
48828/P\_48828.html](https://cima.aemps.es/cima/dochtml/p/
48828/P_48828.html): last consulted on 16.07.2019).
[^20]: Besides, when the dataset is being used in a shared task, this refinement process should be available to participants while the task is open.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We describe a fast method to eliminate features (variables) in $l_1$-penalized least-square regression (or LASSO) problems. The elimination of features leads to a potentially substantial reduction in running time, especially for large values of the penalty parameter. Our method is not heuristic: it only eliminates features that are guaranteed to be absent after solving the LASSO problem. The feature elimination step is easy to parallelize and can test each feature for elimination independently. Moreover, the computational effort of our method is negligible compared to that of solving the LASSO problem - roughly it is the same as single gradient step. Our method extends the scope of existing LASSO algorithms to treat larger data sets, previously out of their reach. We show how our method can be extended to general $l_1$-penalized convex problems and present preliminary results for the Sparse Support Vector Machine and Logistic Regression problems.'
author:
- |
Laurent El Ghaoui [email protected]\
Vivian Viallon [email protected]\
Tarek Rabbani [email protected]\
Department of EECS\
University of California\
Berkeley, CA 94720-1776, USA
bibliography:
- 'ffe\_jmlr.bib'
title: |
Safe Feature Elimination for the LASSO\
and Sparse Supervised Learning Problems
---
Sparse Regression, LASSO, Feature Elimination, SVM, Logistic Regression
Introduction
============
“Sparse” classification or regression problems, which involve an $\ell_{1}-\text{norm}$ regularization has attracted a lot of interest in the statistics [@tibshirani1996regression], signal processing [@chen2001atomic], and machine learning communities. The $\ell_{1}$ regularization leads to sparse solutions, which is a desirable property to achieve model selection, or data compression. For instance, consider the problem of $\ell_{1}$-regularized least square regression commonly referred to as the LASSO [@tibshirani1996regression]. In this context, we are given a set of $m$ observations $a_{i}\in{\mathbb{R}}^{n},\, i=1,\ldots ,m$ and a response vector $y\in{\mathbb{R}}^{m}$ . Denoting by $X=\left(a_{1},\ldots ,a_{m}\right)^{T}\in{\mathbb{R}}^{m\times n}$ the feature matrix of observations, the LASSO problem is given by $${\cal P}(\lambda)\;:\;\phi(\lambda):=\min_{w}\frac{1}{2}\left\Vert Xw-y\right\Vert _{2}^{2}+\lambda\left\Vert w\right\Vert _{1},\label{eq:LASSO}$$ where $\lambda$ is a regularization parameter and $w\in{\mathbb{R}}^{n}$ is the optimization variable. For large enough values of $\lambda$, any solution $w^{\star}\in{\mathbb{R}}^{n}$ of (\[eq:LASSO\]) is typically sparse, i.e. $w^{\star}$ has few entries that are non-zero, and therefore identifies the features in $X$ (columns of $X$) that are useful to predict $y$.
Several efficient algorithms have been developed for the LASSO problem, including [@lars; @boyd_linear; @ParkHastie; @Homotopy; @Friedman_Pathwise; @becker2010templates; @FHT:10] and references therein. However, the complexity of these algorithms, when it is known, grows fast with the number of variables. While the LASSO problem is particularly appealing in presence of very high-dimensional problems, the available algorithms can be quite slow in such contexts. In some applications, the feature matrix is so big that it can not even be loaded and LASSO solvers cannot be used at all. Hence it is of paramount interest to be able to efficiently eliminate features in a pre-processing step, in order to reduce dimensionality and solve the optimization problem on a reduced matrix.
Assume that a sparse solution exists to (\[eq:LASSO\]) and that we were able to identify $e$ zeros of $w^{\star}$ **a priori** to solving the LASSO problem. Identifying $e$ zeros in $w^{\star}$ a priori to solving (\[eq:LASSO\]) is equivalent to removing $e$ features (columns) from the feature matrix $X$. If $e$ is large, we can obtain $w^{\star}$ by solving (\[eq:LASSO\]) with a “small” feature matrix $X$.
In this paper we propose a “safe” feature elimination (SAFE) method that can identify zeros in the solution $w^{\star}$ a priori to solving the LASSO problem. Once the zeros are identified we can safely remove the corresponding features and then solve the LASSO problem (\[eq:LASSO\]) on the reduced feature matrix.
Feature selection methods are often used to accomplish dimensionality reduction, and are of utmost relevance for data sets of massive dimension, see for example [@Fan_Lv_Review]. These methods, when used as a pre-processing step, have been referred to in the literature as [*screening*]{} procedures [@Fan_Lv_Review; @Fan_Lv]. They typically rely on univariate models to score features, independently of each other, and are usually computationally fast. Classical procedures are based on correlation coefficients, two-sample $t$-statistics or chi-square statistics [@Fan_Lv_Review]; see also [@BNS] and the references therein for an overview in the specific case of text classification. Most screening methods might remove features that could otherwise have been selected by the regression or classification algorithm. However, some of them were recently shown to enjoy the so-called “sure screening” property [@Fan_Lv]: under some technical conditions, no relevant feature is removed, with probability tending to one.
Screening procedures typically ignore the specific classification task to be solved after feature elimination. In this paper, we propose to remove features based on the supervised learning problem considered, that is on both the structure of the loss function and the problem data. While we focus mainly on the LASSO problem here, we provide results for a large class of convex classification or regression problems. The features are eliminated according to a sufficient, in general conservative, condition, which we call SAFE (for SAfe Feature Elimination). With SAFE, we never remove features unless they are [*guaranteed*]{} to be absent if one were to solve the full-fledged classification or regression problem.
An interesting fact is that SAFE becomes extremely aggressive at removing features for large values of the penalty parameter $\lambda$. The specific application we have in mind involves large data sets of text documents, and sparse matrices based on occurrence, or other score, of words or terms in these documents. We seek extremely sparse optimal coefficient vectors, even if that means operating at values of the penalty parameter that are substantially larger than those dictated by a pure concern for predictive accuracy. The fact that we need to operate at high values of this parameter opens the hope that, at least for the application considered, the number of features eliminated by using our fast test is high enough to allow a dramatic reduction in computing time and memory requirements. Our experimental results indicate that for many of these data sets, we do observe a dramatic reduction in the number of variables, typically by an order of magnitude or more. The method has two main advantages: for medium- to large-sized problem, it enables to reduce the computational time. More importantly, SAFE allows to tackle problems that are too huge to be even loaded in memory, thereby expanding the reach of current algorithms
The paper is organized as follows. In section \[sec:Proposed-SAFE-method\], we derive the SAFE method for the LASSO problem. In section \[sec:using\_safe\], we illustrate the use of SAFE and detail some relevant algorithms. In section \[sec:SAFE-for-general\], we extend the results of SAFE to general convex problems and derive preliminary SAFE results for the Sparse Support Vector Machine and Logistic regression problems. In section \[s:num\], we experiment the SAFE for LASSO method on synthetic data and on data derived from text classification sources. Numerical results demonstrate that SAFE provides a substantial reduction in problem size, and, as a result, it enables the LASSO algorithms to run faster and solve huge problems originally out of their reach.
#### Notation.
We use ${\mathbf 1}$ and ${\mathbf 0}$ to denote a vector of ones and zeros, with size inferred from context, respectively. For a scalar $a$, $a_{+}$ denotes the positive part of $a$. For a vector $a$, this operation is component-wise, so that ${\mathbf 1}^{T}a_{+}$ is the sum of the positive elements in $a$. We take the convention that a sum over an empty index sets, such as $\sum_{i=1}^{k}a_{i}$ with $k\le0$, is zero.
The SAFE method for the LASSO\[sec:Proposed-SAFE-method\]
=========================================================
The SAFE method crucially relies on duality and optimality conditions. We begin by reviewing the appropriate facts.
Dual problem and optimality conditions for the LASSO\[sec:Dual-Problem-and\]
----------------------------------------------------------------------------
A dual to the LASSO problem (\[eq:LASSO\]) [@boyd_linear] can be written as $${\cal D}(\lambda)\;:\;\phi(\lambda):=\max_{\theta}\: G(\theta)\::\:\left|\theta^{T}x_{k}\right|\leq\lambda,\, k=1,\ldots ,n,\label{eq:dualLASSO}$$ with $x_{k}\in{\mathbb{R}}^{m},\, k=1,\ldots ,n$, the $k$-th column of $X$ and $G(\theta)=\frac{1}{2}\left\Vert y\right\Vert_{2}^{2}-\frac{1}{2}\left\Vert \theta+y\right\Vert _{2}^{2}$. In this context, we call ${\cal P}(\lambda)$ the primal problem, $w$ the primal variable, and $w^{\star}$ a primal optimal point. The dual problem ${\cal D}(\lambda)$ is a convex optimization problem with dual variable $\theta\in{\mathbb{R}}^{m}$. We call $\theta$ dual feasible when it satisfies the constraints in ${\cal D}(\lambda)$. Figure \[subfig:geo\_interp\_a\] shows the geometry of the feasibility set in the dual space. The quantity $G(\theta)$ gives a lower bound on the optimal value $\phi(\lambda)$ for any dual feasible point $\theta$, i.e. $G(\theta)\leq\phi(\lambda),\;\left|\theta^{T}x_{k}\right|\leq\lambda,\, k=1,\ldots ,n$. For the LASSO problem (\[eq:LASSO\]) strong duality holds and the optimal value of ${\cal D}(\lambda)$ achieves $\phi(\lambda)$ at $\theta^{\star}$ the solution of (\[eq:dualLASSO\]) or the dual optimal point. Furthermore, the following relation holds at optimum: $\theta^{\star}=Xw^{\star}-y$.
We consider the dual problem ${\cal D}(\lambda)$ because of an important property that helps us derive our SAFE method. Assuming $w^{\star}$ is sparse, knowledge of $\theta^{\star}$ allows us to identify the zeros in $w^{\star}$ by checking the optimality condition [@BV:04]: $$\left|\theta^{\star T}x_{k}\right|<\lambda\Rightarrow\left(w^{\star}\right)_{k}=0.\label{eq:opt_sec2}$$
Figure \[subfig:geo\_interp\_b\] illustrates the geometric interpretation of the inequality test $\left|\theta^{\star T}x_{k}\right|<\lambda$ in (\[eq:opt\_sec2\]).
Basic idea
----------
The basic idea behind SAFE is to use the optimality condition (\[eq:opt\_sec2\]) with $\theta^{\star}$ in the inequality test replaced by a set $\Theta$ that contains the dual optimal point, i.e. $\left|\theta^{T}x_{k}\right|<\lambda,\;\forall\theta\in\Theta$ and $\theta^{\star}\in\Theta$. If the inequality test holds for the whole set $\Theta$, then the $k$-th entry of $w^{\star}$ is zero, $(w^{\star})_{k}=0$.
In the following sections, we show how to construct the set $\Theta$ using optimality conditions of the dual problem, and derive the corresponding SAFE test.
In our derivation, we assume that we have knowledge of a solution $w_{0}^{\star}$ of ${\cal P}(\lambda_{0})$ for some $\lambda_{0}$, and we seek to apply SAFE for ${\cal P}(\lambda)$ with $\lambda\leq\lambda_{0}$. By default, we can choose $\lambda_{0}$ to be large enough for $w_{0}^{\star}$ to be identically zero. To find such a $\lambda_{0}$, we substitute $w_{0}^{\star}=0$ in (\[eq:LASSO\]) to obtain $\phi(\lambda_{0})=\frac{1}{2}\left\Vert y\right\Vert _{2}^{2}$. By strong duality, ${\cal D}(\lambda_{0})$ achieves a value of $\phi(\lambda_{0})=\frac{1}{2}\left\Vert y\right\Vert _{2}^{2}$ at the unique solution $\theta_{0}^{\star}=-y$. The point $\theta_{0}^{\star}$ is a dual feasible point and satisfies the constraints $\lambda_{0}\geq\left|(-y)^{T}x_{k}\right|,\: k=1,\ldots ,n.$ Note that $\lambda_{0}$ is not uniquely defined but we choose the smallest value above which $w_{0}^{\star}=0$, that is $\lambda_{0}=\max_{1\le j\le n}\:|y^{T}x_{j}|=\|X^{T}y\|_{\infty}$.
Constructing $\Theta$\[sub:Constructing\]
-----------------------------------------
We start by finding a set $\Theta$ that contains the dual optimal point $\theta^{\star}$ of ${\cal D}(\lambda)$. We express $\Theta$ as the intersection of two sets $\Theta_{1}$ and $\Theta_{2}$, where each set corresponds to different optimality conditions.
We construct $\Theta_{1}$ using the optimality condition of ${\cal D}(\lambda)$: $\theta^{\star}$ is a dual optimal point if $G(\theta^{\star})\geq G(\theta)$ for all dual feasible points $\theta$. Let $\theta_{s}$ be a dual feasible point to ${\cal D}(\lambda)$, and $\gamma:=G(\theta_{s})$. Obviously $G(\theta^{\star})\geq\gamma$ and the set $\Theta_{1}:=\left\{ \theta\:\mid\: G(\theta)\geq\gamma\right\} $ contains $\theta^{\star}$, i.e. $\theta^{\star}\in\Theta_{1}$.
One way to obtain a lower bound $\gamma$ is by dual scaling. We set $\theta_{s}$ to be a scaled feasible dual point in terms of $\theta_{0}^{\star}$, $\theta_{s}:=s\theta_{0}^{\star}$ with $s\in{\mathbb{R}}$ constrained so that $\theta_{s}$ is a dual feasible point for ${\cal D}(\lambda)$, that is, $\|X^{T}\theta_{s}\|_{\infty}\le\lambda$ or $|s|\le\lambda/\lambda_{0}$. We then set $\gamma$ according to the convex optimization problem: $$\gamma=\max_{s}\:\left\{ G(s\theta_{0}^{\star})~:~|s|\le\frac{\lambda}{\lambda_{0}}\right\} =\max_{s}\:\left\{ \beta_{0}s-\frac{1}{2}s^{2}\alpha_{0}~:~|s|\le\frac{\lambda}{\lambda_{0}}\right\} ,$$ with $\alpha_{0}:=\theta_{0}^{\star T}\theta_{0}^{\star}>0$, $\beta_{0}:=|y^{T}\theta_{0}^{\star}|$. We obtain $$\gamma=\frac{\beta_{0}^{2}}{2\alpha_{0}}\left(1-\left(1-\frac{\alpha_{0}}{\beta_{0}}\frac{\lambda}{\lambda_{0}}\right)_{+}^{2}\right).\label{eq:gamma}$$
We construct $\Theta_{2}$ by applying a first order optimality condition on ${\cal D}(\lambda_{0})$: $\theta_{0}^{\star}$ is a dual optimal point if $g^{T}(\theta_{0}-\theta_{0}^{\star})\leq0$ for every dual point $\theta_{0}$ that is feasible for ${\cal D}(\lambda_{0})$, where $g:=\nabla G(\theta_{0}^{\star})=\theta_{0}^{\star}+y$. For $\lambda\leq\lambda_{0}$, any dual point $\theta$ feasible for ${\cal D}(\lambda)$ is also dual feasible for ${\cal D}(\lambda_{0})$ ($|\theta^{T}x_{k}|\le\lambda\leq\lambda_{0}\;\; k=1,\ldots,n$). Since $\theta^{\star}$ is dual feasible for ${\cal D}(\lambda_{0})$, we conclude $\theta^{\star}\in\Theta_{2}:=\left\{ \theta\mid g^{T}(\theta-\theta_{0}^{\star})\leq0\right\}$.
Figure \[fig:Theta\] shows the geometry of $\Theta_{1}$, $\Theta_{2}$ and $\Theta$ in the dual space; Figure \[subfig:inequality\_test\] shows the geometric interpretation of the inequality test when it is applied to the set $\Theta$.
SAFE-LASSO theorem\[sub:SAFE-method\]
-------------------------------------
Our criterion to identify the $k$-th zero in $w^{\star}$ and thus remove the $k$-th feature (column) from the feature matrix $X$ in problem ${\cal P}(\lambda)$ becomes $$\lambda>\left|\theta^{T}x_{k}\right|=\text{max}(\theta^{T}x_{k},-\theta^{T}x_{k})\;:\;\theta\in\Theta.\label{eq:test1}$$
An equivalent formulation of condition (\[eq:test1\]) is $$\lambda>\text{max}(P(\gamma,x_{k}),P(\gamma,-x_{k})),$$ where $P(\gamma,x_{k})$ is the optimal value of a convex optimization problem with constraints $\theta\in\Theta_{1}$ and $\theta\in\Theta_{2}$:
$$P(\gamma,x_{k}):=\;\max_{\theta}\, x_{k}^{T}\theta\;:\; G(\theta)\geq\gamma,\; g^{T}\left(\theta-\theta_{0}^{\star}\right)\geq0.\label{eq:testproblem}$$
It turns out that the above problem is simple enough to admit a closed-form solution (see Appendix \[app:p-gamma\]). The resulting test can be summarized as follows.
[**Theorem (SAFE-LASSO)**]{} [ *Consider the LASSO problem ${\cal P}(\lambda)$ in (\[eq:LASSO\]). Let $\lambda_{0}\ge\lambda$ be a value for which an optimal solution $w_{0}^{\star}\in{\mathbb{R}}^{n}$ is known. Denote by $x_{k}$ the $k$-th feature (column) of the matrix $X$. Define $${\cal E}=\left\{ k\:|\:\lambda>\max(P(\gamma,x_{k}),P(\gamma,-x_{k})\right\} ,\label{eq:test-LASSO-lambda0-recursive}$$ where $$P(\gamma,x_{k})=\begin{cases}
\theta_{0}^{\star T}x_{k}+\Psi_{k}\tilde{D}(\gamma)\qquad & \left\Vert g\right\Vert _{2}^{2}\left\Vert x_{k}\right\Vert _{2}\geq D(\gamma)x_{k}^{T}g,\\
-y^{T}x_{k}+\left\Vert x_{k}\right\Vert _{2}D(\gamma)\qquad & \left\Vert g\right\Vert _{2}^{2}\left\Vert x_{k}\right\Vert _{2}\leq D(\gamma)x_{k}^{T}g,\end{cases}\label{eq:thmP}$$ with $$\begin{array}{l}
\theta_{0}^{\star}=Xw_{0}^{\star}-y, \;\; g:=\theta_{0}^{\star}+y, \;\;
\alpha_{0}:=\theta_{0}^{\star T}\theta_{0}^{\star}, \;\;
\beta_{0}:=|y^{T}\theta_{0}^{\star}|, \;\;
\gamma:=\frac{\beta_{0}^{2}}{2\alpha_{0}}\left(1-\left(1-\frac{\alpha_{0}}{\beta_{0}}\frac{\lambda}{\lambda_{0}}\right)_{+}^{2}\right), \\
D(\gamma)=\left(\left\Vert y\right\Vert _{2}^{2}-2\gamma\right)^{1/2}, \;\;
\tilde{D}(\gamma)=\left(D(\gamma)^{2}-\left\Vert g\right\Vert _{2}^{2}\right)^{1/2}, \;\;
\Psi_{k}:=\left(\left\Vert x_{k}\right\Vert _{2}^{2}-\frac{\left(x_{k}^{T}g\right)^{2}}{\left\Vert g\right\Vert _{2}^{2}}\right)^{1/2}.
\end{array}$$ Then, for every index $e\in{\cal E}$, the $e$-th entry of $w^{\star}$ is zero, i.e. $\left(w^{\star}\right)_{e}=0$, and feature $x_{e}$ can be safely eliminated from $X$ a priori to solving the LASSO problem (\[eq:LASSO\]).*]{} When we don’t have access to a solution $w_{0}^{\star}$ of ${\cal P}(\lambda_{0})$, we can set $w_{0}^{\star}=0$ and $\lambda_{0}=\lambda_{\rm max}:=\|X^{T}y\|_{\infty}$. In this case, the inequality test $\lambda>\text{max}(P(\gamma,x_{k}),P(\gamma,-x_{k})$ in the SAFE-LASSO theorem takes the form $\lambda>\rho_{k}\lambda_{\rm
max}$, with $$\rho_{k}=\frac{\left\Vert y\right\Vert _{2}\left\Vert x_{k}\right\Vert _{2}+|y^{T}x_{k}|}{\left\Vert y\right\Vert _{2}\left\Vert x_{k}\right\Vert _{2}+\lambda_{\rm max}}.$$
In the case of scaled data sets, for which $\left\Vert y\right\Vert _{2}=1$ and $\left\Vert x_{k}\right\Vert _{2}=1$ for every $k$, $\rho_{k}$ has a convenient geometrical interpretation:$$\rho_{k}=\frac{1+\left|\cos\alpha_{k}\right|}{1+\underset{1\le j\le n}{\text{max}}\:|\cos\alpha_{j}|},$$ where $\alpha_{k}$ is the angle between the $k$-th feature and the response vector $y$. Our test then consists in eliminating features based on how closely they are aligned with the response, *relative* to the most closely aligned feature. For scaled data sets, our test is very similar to standard correlation-based feature selection [@Fan_Lv]; in fact, for scaled data sets, the ranking of features it produces is then exactly the same. The big difference here is that our test is not heuristic, as it only eliminates features that are *guaranteed* to be absent when solving the full-fledged sparse supervised learning problem.
SAFE for LASSO with intercept problem {#ss:intercept-lasso}
-------------------------------------
The SAFE-LASSO theorem can be applied to the LASSO with intercept problem$${\cal P}_{\text{int}}(\lambda)\;:\;\phi(\lambda):=\min_{w,\nu}\frac{1}{2}\left\Vert Xw+\nu-y\right\Vert _{2}^{2}+\lambda\left\Vert w\right\Vert _{1},$$ with $\nu\in{\mathbb{R}}^{m}$ the intercept term, by using a simple transformation. Taking the derivative of the objective function of ${\cal P}_{\text{int}}(\lambda)$ w.r.t $\nu$ and setting it to zero, we obtain $\nu=\bar{y}-\bar{X}^{T}w$ with $\bar{y}=(1/m){\mathbf 1}^{T}y$, $\bar{X}=(1/m)X{\mathbf 1}$ and ${\mathbf 1}\in{\mathbb{R}}^{m}$ the vector of ones . Using the expression of $\nu$, ${\cal P}_{\text{int}}(\lambda)$ can be expressed as$${\cal P}_{\text{int}}(\lambda)\;:\;\phi(\lambda):=\min_{w}\frac{1}{2}\left\Vert X_{\text{cent}}w-y_{\text{cent}}\right\Vert _{2}^{2}+\lambda\left\Vert w\right\Vert _{1},$$ with $X_{\text{cent}}:=X-\bar{X}{\mathbf 1}^{T}$ and $y_{\text{cent}}=y-\bar{y}{\mathbf 1}$. Thus the SAFE-LASSO theorem can be applied to ${\cal P}_{\text{int}}$ and eliminate features (columns) from $X_{\text{cent}}$ .
SAFE for elastic net
--------------------
The elastic net problem $${\cal P}_{\text{elastic}}(\lambda)\;:\;\phi(\lambda):=\min_{w}\frac{1}{2}\left\Vert Xw-y\right\Vert _{2}^{2}+\lambda\left\Vert w\right\Vert _{1}+\frac{1}{2}\epsilon\left\Vert w\right\Vert _{2}^{2},$$ can be expressed in the form of ${\cal P}(\lambda)$ by replacing $X$ and $y$ of (\[eq:LASSO\]) with $X_{\text{elastic}}=\left(X^{T},\sqrt{\epsilon}I\right)^{T}$ and $y_{\text{elastic}}=\left(y^{T},\mathbf{0}^{T}\right)^{T}$. This transformation allows us to apply the SAFE-LASSO theorem on ${\cal
P}_{\text{elastic}}(\lambda)$ and eliminate features from $X_{\text{elastic}}$.
Using SAFE {#sec:using_safe}
==========
In this section we illustrate the use of SAFE and detail the relevant algorithms.
SAFE for reducing memory limit problems
---------------------------------------
SAFE can extend the reach of LASSO solvers to larger size problems than what they could originally handle. In this section, we are interested in solving for $w_{d}^{\star}$ the solution of ${\cal P}(\lambda_{d})$ under a memory constraint of loading only $M$ features. We can compute $w_{d}^{\star}$ by solving a sequence of problems, where each problem has a number of features less than our memory limit $M$. We start by finding an appropriate $\lambda$ where our SAFE method can eliminate at least $n-M$ features, we then solve a reduced size problem with $L_{F}\leq M$ features, where $L_{F}=\left|{\cal E}^{c}\right|$ is the number of features left after SAFE and ${\cal E}^{c}=\left\{ 1,\ldots ,n\right\} \backslash{\cal E}$ is the complement of the set ${\cal E}$ in the SAFE-LASSO theorem. We proceed to the next stage as outlined in algorithm \[alg:SAFE-for-memory\].
**given** a feature matrix $X\in{\mathbb{R}}^{m\times n}$, response $y\in{\mathbb{R}}^{m}$, penalty parameter $\lambda_{d}$ , memory limit $M$ and LASSO solver: `LASSO`, i.e. $w^\star=\verb"LASSO"(X,y,\lambda)$.
**initialize** $\lambda_{0}=\|X^{T}y\|_{\infty}$, $w_{0}^{\star}={\mathbf 0}\in{\mathbb{R}}^{n}$,
**repeat**
1. *Use SAFE to search for a $\lambda$ with $LF\leq M$ .* Obtain $\lambda$ and ${\cal E}$. % $L_{F}$ is the number of features left after SAFE and ${\cal E}$ is the set defined in the SAFE-LASSO theorem.
2. **if** $\lambda<\lambda_{d}$ **then** $\lambda=\lambda_{d}$, apply SAFE to obtain ${\cal E}$ **end if.**
3. Compute the solution $w^{\star}$. $w^{\star}({\cal E}^{c})=\verb"LASSO"(X({\cal E}^{c},:),y,\lambda)$, $w^{*}({\cal E})=0$; % $w^{\star}({\cal E}^{c})$ and $X({\cal E}^{c},:)$ are the elements and columns of $w^{\star}$ and $X$ defined by the set ${\cal E}^{c}$, respectively. ${\cal E}^{c}=\left\{ 1,\ldots ,n\right\} \backslash{\cal E}$ is the complement of the set ${\cal E}$ .
4. $\lambda_{0}:=\lambda$, $w_{0}^{\star}=w^{*}$.
**until $\lambda_{0}=\lambda_{d}$**
We use a bisection method to find an appropriate value of $\lambda$ for which SAFE leaves $L_{F}\in\left[M-\epsilon_{F},\: M\right]$ features, where $\epsilon_{F}$ is a number of feature tolerance. The bisection method on $\lambda$ is outlined in algorithm \[alg:Bisection-method-for\].
**given** a feature matrix $X\in{\mathbb{R}}^{m\times n}$, response $y\in{\mathbb{R}}^{m}$, penalty parameter $\lambda_{0}$ with LASSO solution $w_{0}^{\star}$, tolerance $\epsilon_{F}>0$ and memory limit $M$.
**initialize** $l=0$, and $u=\lambda_{0}$.
**repeat**
1. Set $\lambda:=\left(l+u\right)/2$.
2. Use the SAFE-LASSO theorem to obtain ${\cal E}$.
3. Set $L_{F}=\left|{\cal E}^{c}\right|$.
4. **if** $L_{F}>M$ **then** set $l:=\lambda$ **else** set $u:=\lambda$ **end if**
**until** $M-L_{F}\leq\epsilon_{F}$ and $L_{F}\leq M$.
SAFE for LASSO run-time reduction
---------------------------------
In some applications like [@GJMEYC:10], it is of interest to solve a sequence of problems ${\cal P}(\lambda_{1}),\:\ldots {\cal P}(\lambda_{s})$ for decreasing values of the penalty parameters, i.e. $\lambda_{1}\geq\ldots \geq\lambda_{s}$. The computational complexities of LASSO solvers depend on the number of features and using SAFE might result in run-time improvements. For each problem in the sequence, we can use SAFE to reduce the number of features a priori to using our LASSO solver as shown in algorithm \[alg:Recursive-SAFE-for-1\].
**given** a feature matrix $X\in{\mathbb{R}}^{m\times n}$, response $y\in{\mathbb{R}}^{m}$, a sequence of penalty parameters $\lambda_{s}\leq\ldots \leq\lambda_{1}\leq\|X^{T}y\|_{\infty},$ and LASSO solver: `LASSO`.
**initialize** $\lambda_{0}=\|X^{T}y\|_{\infty}$, $w_{0}^{\star}={\mathbf 0}\in{\mathbb{R}}^{n}$.
**for** $i=1$ **until** $i=s$ **do**
1. Set $\lambda_{0}=\lambda_{i-1}$, and $\lambda=\lambda_{i}$.
2. Use the SAFE-LASSO theorem to obtain ${\cal E}$.
3. Compute the solution $w^{\star}$. $w^{\star}({\cal E}^{c})=\verb"LASSO"(X({\cal E}^{c},:),y,\lambda)$, $w^{*}({\cal E})=0$. % $w^{\star}({\cal E}^{c})$ and $X({\cal E}^{c},:)$ are the elements and columns of $w^{\star}$ and $X$ defined by the set ${\cal E}^{c}$, respectively. ${\cal E}^{c}=\left\{ 1,\ldots ,n\right\} \backslash{\cal E}$ is the complement of the set ${\cal E}$ .
4. Set $w_{0}^{\star}=w^{*}$.
**end for**
SAFE applied to general $\ell_{1}$-regularized convex problems\[sec:SAFE-for-general\]
======================================================================================
The SAFE-LASSO result presented in section \[sub:SAFE-method\] for the LASSO problem (\[eq:LASSO\]) can be adapted to a more general class of $l_{1}-$ regularized convex problems. We consider the family of problems$${\cal P}(\lambda)\;:\;\phi(\lambda):=\min_{w,\,\nu}\sum_{i=1}^{m}f(a_{i}^{T}w+b_{i}v+c_{i})+\lambda\left\Vert w\right\Vert _{1},\label{eq:general-p}$$ where $f$ is a closed convex function, and non-negative everywhere, $a_{i}\in{\mathbb{R}}^{n}$, $i=1,\ldots,m$, $b,c\in{\mathbb{R}}^{m}$ are given. The LASSO problem (\[eq:LASSO\]) is a special case of (\[eq:general-p\]) with $f(\zeta)=(1/2)\zeta^{2}$, $a_{i}\in{\mathbb{R}}^{n}$, $i=1,\ldots,m$ the observations, $c=-y$ is the (negative) response vector, and $b=0$. Hereafter, we refer to the LASSO problem as ${\cal P}_{\text{LASSO}}(\lambda)$ and to the general class of $l_{1}$-regularized problems as ${\cal P}(\lambda)$. In this section, we outline the steps necessary to derive a SAFE method for the general problem ${\cal P}(\lambda)$. We show some preliminary results for deriving SAFE methods when $f(\zeta)$ is the hing loss function, $f_{\text{hi}}(\zeta)=\left(1-\zeta\right)_{+}$, and the logistic loss function $f_{{\rm log}}(\xi)=\log(1+e^{-\xi})$.
Dual Problem
------------
The first step is to devise the dual of problem (\[eq:general-p\]), which is $${\cal D}(\lambda)~:~~\phi(\lambda)=\max_{\theta}\: G(\theta)~:~\theta^{T}b=0,\;\;|\theta^{T}x_{k}|\le\lambda,\;\; k=1,\ldots,n,\label{eq:pb-generic-dual}$$ where $$G(\theta):=c^{T}\theta-\sum_{i=1}^{m}f^{\ast}(\theta_{i})\label{eq:pb-generic-dual-fcn}$$ with $f^{\ast}(\vartheta)=\max_{\xi}\:\xi\vartheta-f(\xi)$ the conjugate of the loss function $f(\zeta)$, and $x_{k}$ the $k$-th column or feature of the feature matrix $X=\left(a_{1},\ldots ,a_{m}\right)^{T}\in{\mathbb{R}}^{m\times n}$. $G(\theta)$ is the dual function, which is, by construction, concave. We assume that strong duality holds and primal and dual optimal points are attained. Due to the optimality conditions for the problem (see [@BV:04]), constraints for which $|\theta^{T}x_{k}|<\lambda$ at optimum correspond to a zero element in the primal variable: $\left(w^{\star}\right)_{k}=0$, i.e.$$\left|\theta^{\star T}x_{k}\right|<\lambda\Rightarrow\left(w^{\star}\right)_{k}=0.\label{eq:opt_sec6}$$
Optimality set $\Theta$ {#ss:dual-scaling-generic}
-----------------------
For simplicity, we consider only the set $\Theta:=\left\{ \theta\:|\: G(\theta)\geq\gamma\right\} $ which contains $\theta^{\star}$ the dual optimal point of ${\cal D}(\lambda)$. One way to get a lower bound $\gamma$ is to find a dual point $\theta_{s}$ that is feasible for the dual problem ${\cal D}(\lambda)$, and then set $\gamma=G(\theta_{s})$.
To obtain a dual feasible point, we can solve the problem for a higher value $\lambda_{0}\ge\lambda$ of the penalty parameter. (In the specific case examined below, we will see how to set $\lambda_{0}$ so that the vector $w_{0}^{\star}=0$ at optimum.) This provides a dual point $\theta_{0}^{\star}$ that is feasible for ${\cal D}(\lambda_{0})$, which satisfies $\lambda_{0}=\|X\theta_{0}\|_{\infty}$. In turn, $\theta_{0}^{\star}$ can be scaled so as to become feasible for ${\cal D}(\lambda)$. Precisely, we set $\theta_{s}=s\theta_{0}$, with $\|X\theta_{s}\|_{\infty}\le\lambda$ equivalent to $|s|\le\lambda/\lambda_{0}$. In order to find the best possible scaling factor $s$, we solve the one-dimensional, convex problem $$\gamma(\lambda):=\max_{s}\: G(s\theta_{0})~:|s|\le\frac{\lambda}{\lambda_{0}}.\label{eq:gamma-pb}$$ Under mild conditions on the loss function $f$, the above problem can be solved by bisection in $O(m)$ time. By construction, $\gamma(\lambda)$ is a lower bound on $\phi(\lambda)$. We can generate an initial point $\theta_{0}^{\star}$ by solving ${\cal P}(\lambda_{0})$ with $w_{0}=0$. We get $$\min_{v_{0}}\:\sum_{i=1}^{m}f(b_{i}v_{0}+c_{i})=\min_{v_{0}}\:\max_{\theta_{0}}\:\theta_{0}^{T}(bv_{0}+c)-\sum_{i=1}^{m}f^{\ast}\left(\left(\theta_{0}\right)_{i}\right)=\max_{\theta_{0}\::\: b^{T}\theta_{0}=0}\: G(\theta_{0}).$$ Solving the one-dimensional problem above can be often done in closed-form, or by bisection, in $O(m)$. Choosing $\theta_{0}^{\star}$ to be any optimal for the corresponding dual problem (the one on the right-hand side) generates a point that is dual feasible for it, that is, $G(\theta_{0}^{\star})$ is finite, and $b^{T}\theta_{0}=0$.
The point $\theta_{0}^{\star}$ satisfies all the constraints of problem ${\cal D}(\lambda)$, except perhaps for the constraint $\|X\theta\|_{\infty}\le\lambda$, i.e. $\|X\theta_{0}^{\star}\|_{\infty}>\lambda$. Hence, if $\lambda\ge\lambda_{{\rm 0}}:=\|X\theta_{0}^{\star}\|_{\infty}$, then $\theta_{0}^{\star}$ is dual optimal for ${\cal D}(\lambda)$ and by the optimality condition (\[eq:opt\_sec6\]) we have $w^{\star}=0$ . Note that, since $\theta_{0}^{\star}$ may not be uniquely defined, $\lambda_{0}$ may not necessarily be the smallest value for which $w^{\star}=0$ is optimal for the primal problem.
SAFE method
-----------
Assume that a lower bound $\gamma$ on the optimal value of the learning problem $\phi(\lambda)$ is known: $\gamma\le\phi(\lambda)$. (Without loss of generality, we can assume that $0\le\gamma\le\sum_{i=1}^{m}f(c_{i})$). The test $$\lambda>\text{max}(P(\gamma,x_{k}),P(\gamma,-x_{k})),$$ allows to eliminate the $k$-th feature from the feature matrix $X$, where $P(\gamma,x_{k})$ is the optimal value of a convex optimization problem with two constraints: $$P(\gamma,x_{k}):=\max_{\theta}\:\theta^{T}x_{k}~:~G(\theta)\ge\gamma,\;\;\theta^{T}b=0.\label{eq:P-def}$$ Since $P(\gamma,x_{k})$ decreases when $\gamma$ increases, the closer $\phi(\lambda)$ is to its lower bound $\gamma$, the more aggressive (accurate) our test is.
By construction, the dual function $G$ is decomposable as a sum of functions of one variable only. This particular structure allows to solve problem (\[eq:P-def\]) very efficiently, using for example interior-point methods, for a large class of loss functions $f$. Alternatively, we can express the problem in dual form as a convex optimization problem with two scalar variables: $$P(\gamma,x_{k})=\min_{\mu>0,\:\nu}\:-\gamma\mu+\mu\sum_{i=1}^{m}f\left(\frac{\left(x_{k}\right)_{i}+\mu c_{i}+\nu b_{i}}{\mu}\right).\label{eq:P-def-dual}$$ Note that the expression above involves the perspective of the function $f$, which is convex (see [@BV:04]). For many loss functions $f$, the above problem can be efficiently solved using a variety of methods for convex optimization, in (close to) $O(m)$ time. We can also set the variable $\nu=0$, leading to a simple bisection problem over $\mu$. This amounts to ignore the constraint $\theta^{T}b=0$ in the definition of $P(\gamma,x)$, resulting in a more conservative test. More generally, any pair $(\mu,\nu)$ with $\mu>0$ generates an upper bound on $P(\gamma,x)$, which in turn corresponds to a valid, perhaps conservative, test.
SAFE for Sparse Support Vector Machine {#s:hi}
--------------------------------------
We turn to the sparse support vector machine classification problem: $${\cal P}_{{\rm hi}}(\lambda)~:~~\phi(\lambda):=\min_{w,v}\:\sum_{i=1}^{m}(1-y_{i}(z_{i}^{T}w+v))_{+}+\lambda\|w\|_{1},\label{eq:pb-svm-primal}$$ where $z_{i}\in{\mathbb{R}}^{n}$, $i=1,\ldots,m$ are the data points, and $y\in\{-1,1\}^{m}$ is the label vector. The above is a special case of the generic problem (\[eq:general-p\]), where $f(\zeta):=(1-\xi)_{+}$ is the hinge loss, $b=y$, $c=0$, and the feature matrix $X$ is given by $X=[y_{1}z_{1},\ldots,y_{m}z_{m}]^{T}$, so that $x_{k}=[y_{1}z_{1}(k),\ldots,y_{m}z_{m}(k)]^{T}$.
We denote by ${\cal I}_{+},{\cal I}_{-}$ the set of indicies corresponding to the positive and negative classes, respectively, and denote by $m_{\pm}=|{\cal I}_{\pm}|$ the associated cardinalities. We define $\underline{m}:=\min(m_{+},m_{-})$. Finally, for a generic data vector $x$, we set $x^{\pm}=(x_{i})_{i\in{\cal I}_{\pm}}\in{\mathbb{R}}^{m_{\pm}}$, $k=1,\ldots,n$, the vectors corresponding to each one of the classes.
The dual problem takes the form $${\cal D}_{hi}(\lambda)~:~~\phi(\lambda):=\max_{\theta}\: G_{\text{hi}}(\theta)~:~-{\mathbf 1}\le\theta\le0,\;\;\theta^{T}y=0,\;\;|\theta^{T}x_{k}|\le\lambda,\;\; k=1,\ldots,n.\label{eq:pb-svm-dual}$$ with $G_{\text{hi}}(\theta)={\mathbf 1}^{T}\theta$.
### Test, $\gamma$ given {#ss:test-gamma-svm}
Let $\gamma$ be a lower bound on $\phi(\lambda)$. The optimal value obtained upon setting $w=0$ in (\[eq:pb-svm-primal\]) is given by $$\min_{v}\:\sum_{i=1}^{m}(1-y_{i}v)_{+}=2\min(m_{+},m_{-}):=\gamma_{{\rm max}}.\label{eq:pb-svm-primal-w0}$$ Hence, without loss of generality, we may assume $0\le\gamma\le\gamma_{{\rm max}}$.
The feature elimination test hinges on the quantity $$\begin{array}{rcl}
P_{{\rm hi}}(\gamma,x) & = & {\displaystyle}\max_{\theta}\:\theta^{T}x~:~{\mathbf 1}^{T}\theta\ge\gamma,\;\;\theta^{T}y=0,\;\;-{\mathbf 1}\le\theta\le0\\
& = & {\displaystyle}\min_{\mu>0,\:\nu}\:-\gamma\mu+\mu\sum_{i=1}^{m}f_{{\rm hi}}\left({\displaystyle}\frac{x_{i}-\nu y_{i}}{\mu}\right)\\
& = & {\displaystyle}\min_{\mu>0,\:\nu}\:-\gamma\mu+\sum_{i=1}^{m}(\mu+\nu y_{i}-x_{i})_{+}.\end{array}\label{eq:P-hi-pb}$$ In appendix \[app:P-gamma-svm\], we show that for any $x$, the quantity $P(\gamma,x)$ is finite if and only if $0\le\gamma\le\gamma_{{\rm max}}$, and can be computed in $O(m\log m)$, or less with sparse data, via a closed-form expression. That expression is simpler to state for $P_{{\rm hi}}(\gamma,-x)$: $$\begin{array}{rcl}
P_{{\rm hi}}(\gamma,-x) & = & {\displaystyle}\sum_{j=1}^{\lfloor\gamma/2\rfloor}\bar{x}_{j}-(\frac{\gamma}{2}-\lfloor\frac{\gamma}{2}\rfloor)(\bar{x}_{\lfloor\gamma/2\rfloor+1})_{+}+\sum_{j=\lfloor\gamma/2\rfloor+1}^{\underline{m}}(\bar{x}_{j})_{+},\;\;0\le\gamma\le\gamma_{{\rm max}}=2\underline{m},\\
& & \bar{x}_{j}:=x_{[j]}^{+}+x_{[j]}^{-},\;\; j=1,\ldots,\underline{m},\end{array}$$ with $x_{[j]}$ the $j$-th largest element in a vector $x$, and with the convention that a sum over an empty index set is zero. Note that in particular, since $\gamma_{{\rm max}}=2\underline{m}$: $$P_{{\rm hi}}(\gamma_{{\rm max}},-x)=\sum_{i=1}^{\underline{m}}(x_{[j]}^{+}+x_{[j]}^{-}).$$
### SAFE-SVM theorem {#ss:safe-thm-svm}
Following the construction proposed in section \[ss:dual-scaling-generic\] for the generic case, we select $\gamma=G_{{\rm hi}}(\theta)$, where the point $\theta$ is feasible for (\[eq:pb-svm-dual\]), and can found by the scaling method outlined in section \[ss:dual-scaling-generic\], as follows. The method starts with the assumption that there is a value $\lambda_{0}\ge\lambda$ for which we know the optimal value $\gamma_{0}$ of ${\cal P}_{{\rm hi}}(\lambda_{0})$.
#### Specific choices for $\lambda_{0},\gamma_{0}$.
Let us first detail how we can find such values $\lambda_{0}$, $\gamma_{0}$.
We can set a value $\lambda_{0}$ such that $\lambda>\lambda_{0}$ ensures that $w=0$ is optimal for the primal problem (\[eq:pb-svm-primal\]). The value that results in the least conservative test is $\lambda_{0}=\lambda_{{\rm max}}$, where $\lambda_{{\rm max}}$ is the smallest value of $\lambda$ above which $w=0$ is optimal: $$\lambda_{{\rm max}}:=\min_{\theta}\:\|X\theta\|_{\infty}~:~-\theta^{T}{\mathbf 1}\ge\gamma_{{\rm max}},\;\;\theta^{T}y=0,\;\;-{\mathbf 1}\le\theta\le0.\label{eq:lambdamax-def-svm}$$ Since $\lambda_{{\rm max}}$ may be relatively expensive to compute, we can settle for an upper bound $\overline{\lambda}_{{\rm max}}$ on $\lambda_{{\rm max}}$. One choice for $\overline{\lambda}_{{\rm max}}$ is based on the test derived in the previous section: we ask that it passes for all the features when $\lambda=\overline{\lambda}_{{\rm max}}$ and $\gamma=\gamma_{{\rm max}}$. That is, we set $$\begin{array}{rcl}
\overline{\lambda}_{{\rm max}} & = & {\displaystyle}\max_{1\le k\le n}\:\max\left(P_{{\rm hi}}(\gamma_{{\rm max}},x_{k}),P_{{\rm hi}}(\gamma_{{\rm max}},-x_{k})\right)\\
& = & {\displaystyle}\max_{1\le k\le n}\:\max\left({\displaystyle}\sum_{i=1}^{\underline{m}}(x_{k}^{+})_{[j]}+(x_{k}^{-})_{[j]},\sum_{i=1}^{\underline{m}}(-x_{k}^{+})_{[j]}+(-x_{k}^{-})_{[j]}\right).\end{array}\label{eq:lambda0-def-svm}$$ By construction, we have $\overline{\lambda}_{{\rm max}}\ge\lambda_{{\rm max}}$, in fact: $$\begin{array}{rcl}
\overline{\lambda}_{{\rm max}} & = & {\displaystyle}\max_{1\le k\le n}\:\max_{\theta}\:|x_{k}^{T}\theta|~:~-\theta^{T}{\mathbf 1}\ge\gamma_{{\rm max}},\;\;\theta^{T}y=0,\;\;-{\mathbf 1}\le\theta\le0\\
& = & {\displaystyle}\max_{\theta}\:\|X\theta\|_{\infty}~:~-\theta^{T}{\mathbf 1}\ge\gamma_{{\rm max}},\;\;\theta^{T}y=0,\;\;-{\mathbf 1}\le\theta\le0,\end{array}$$ The two values $\lambda_{{\rm max}},\overline{\lambda}_{{\rm max}}$ coincide if the feasible set is a singleton, that is, when $m_{+}=m_{-}$. On the whole interval $\lambda_{0}\in[\lambda_{{\rm max}},\overline{\lambda}_{{\rm max}}]$, the optimal value of problem ${\cal P}_{{\rm hi}}(\lambda_{0})$ is $\gamma_{{\rm max}}$.
#### Dual scaling.
The remainder of our analysis applies to any value $\lambda_{0}$ for which we know the optimal value $\gamma_{0}\in[0,\gamma_{{\rm max}}]$ of the problem ${\cal P}_{{\rm hi}}(\lambda_{0})$.
Let $\theta_{0}$ be a corresponding optimal dual point (as seen shortly, the value of $\theta_{0}$ is irrelevant, as we will only need to know $\gamma_{0}={\mathbf 1}^{T}\theta_{0}$). We now scale the point $\theta_{0}$ to make it feasible for ${\cal P}_{{\rm hi}}(\lambda)$, where $\lambda$ ($0\le\lambda\le\lambda_{0}$) is given. The scaled dual point is obtained as $\theta=s\theta_{0}$, with $s$ solution to (\[eq:gamma-pb\]). We obtain the optimal scaling $s=\lambda/\lambda_{0}$, and since $\gamma_{0}=-{\mathbf 1}^{T}\theta_{0}$, the corresponding bound is $$\gamma(\lambda)={\mathbf 1}^{T}(s\theta_{0})=s\gamma_{0}=\gamma_{0}\frac{\lambda}{\lambda_{0}}.$$ Our test takes the form $$\lambda>\max\left(P_{{\rm hi}}(\gamma(\lambda),x),P_{{\rm hi}}(\gamma(\lambda),-x)\right).$$
Let us look at the condition $\lambda>P_{{\rm hi}}(\gamma(\lambda),-x)$: $$\exists\:\mu\ge0,\:\nu~:~\lambda>-\gamma(\lambda)\mu+\sum_{i=1}^{m}(\mu+\nu y_{i}+x_{i})_{+},$$ which is equivalent to: $$\lambda>\min_{\mu\ge0,\nu}\:\frac{{\displaystyle}\sum_{i=1}^{m}(\mu+\nu y_{i}+x_{i})_{+}}{1+(\gamma_{0}/\lambda_{0})\mu}.$$ The problem of minimizing the above objective function over variable $\nu$ has a closed-form solution. In appendix \[app:Phi-svm\], we show that for any vectors $x^{\pm}\in{\mathbb{R}}^{m_{\pm}}$, we have $$\Phi(x^{+},x^{-}):=\min_{\nu}\:\sum_{i=1}^{m_{+}}(x_{i}^{+}+\nu)_{+}+\sum_{i=1}^{m_{-}}(x_{i}^{-}-\nu)_{+}=\sum_{i=1}^{\underline{m}}(x_{[i]}^{+}+x_{[i]}^{-})_{+},$$ with $x_{[j]}$ the $j$-th largest element in a vector $x$. Thus, the test becomes $$\lambda>\min_{\mu\ge0}\:\frac{{\displaystyle}\sum_{i=1}^{\underline{m}}(2\mu+x_{[i]}^{+}+x_{[i]}^{-})_{+}}{1+(\gamma_{0}/\lambda_{0})\mu}.$$
Setting $\kappa=\lambda_{0}/(\lambda_{0}+\gamma_{0}\mu)$, we obtain the following formulation for our test: $${\lambda}>\min_{0\le\kappa\le1}\:\sum_{i=1}^{\underline{m}}((1-\kappa)\frac{2\lambda_{0}}{\gamma_{0}}+\kappa(x_{[i]}^{+}+x_{[i]}^{-}))_{+}=\frac{2\lambda_{0}}{\gamma_{0}}G(\frac{\gamma_{0}}{2\lambda_{0}}\overline{x}),\label{eq:test-svm-kappa}$$ where $\overline{x}_{i}:=x_{[i]}^{+}+x_{[i]}^{-}$, $i=1,\ldots,\underline{m}$, and for $z\in{\mathbb{R}}^{m}$, we define $$G(z):=\min_{0\le\kappa\le1}\:\sum_{i=1}^{m}(1-\kappa+\kappa z_{i})_{+}.$$ We show in appendix \[app:G-svm\] that $G(z)$ admits a closed-form expression, which can be computed in $O(d\log d)$, where $d$ is the number of non-zero elements in vector $z$. By construction, the test removes all the features if we set $\lambda_{0}=\lambda_{{\rm max}}$, $\gamma_{0}=\gamma_{{\rm max}}$, and when $\lambda>\lambda_{{\rm max}}$.
[**Theorem (SAFE-SVM)**]{} [*Consider the SVM problem ${\cal P}_{{\rm hi}}(\lambda)$ in (\[eq:pb-svm-primal\]). Denote by $x_{k}$ the $k$-th row of the matrix $[y_{1}z_{1},\ldots,y_{m}z_{m}]$, and let ${\cal I}_{\pm}:=\{i\::\: y_{i}=\pm1\}$, $m_{\pm}:=|{\cal I}_{\pm}|$, $\underline{m}:=\min(m_{+},m_{-})$, and $\gamma_{{\rm max}}:=2\underline{m}$. Let $\lambda_{0}\ge\lambda$ be a value for which the optimal value $\gamma_{0}\in[0,\gamma_{{\rm max}}]$ of ${\cal P}_{{\rm sq}}(\lambda_{0})$ is known. The following condition allows to remove the $k$-th feature vector $x_{k}$: $${\lambda}>\frac{2\lambda_{0}}{\gamma_{0}}\max\left(G(\frac{\gamma_{0}}{2\lambda_{0}}\overline{x}_{k}),G(\frac{\gamma_{0}}{2\lambda_{0}}\underline{x}_{k})\right),\label{eq:test-svm-lambda0}$$ where $(\overline{x}_{k})_{i}:=(x_{k})_{[i]}^{+}+(x_{k})_{[i]}^{-}$, $(\underline{x}_{k})_{i}:=(-x_{k})_{[i]}^{+}+(-x_{k})_{[i]}^{-}$, $i=1,\ldots,\underline{m}$, and for $z\in{\mathbb{R}}^{m}$: $$G(z)=\min_{z}\:{\displaystyle}\frac{1}{1-z}\sum_{i=1}^{p}(z_{i}-z)_{+}~:~z\in\{-\infty,0,(z_{j})_{j\::\: z_{j}<0}\}$$ A specific choice for $\lambda_{0}$ is $\overline{\lambda}_{{\rm max}}$ given by (\[eq:lambda0-def-svm\]), with corresponding optimal value $\gamma_{0}=\gamma_{{\rm max}}$.*]{}
SAFE for Sparse Logistic Regression {#s:lo}
-----------------------------------
We now consider the sparse logistic regression problem: $${\cal P}_{{\rm lo}}(\lambda)~:~~\phi(\lambda):=\min_{w,v}\:\sum_{i=1}^{m}\log\left(1+\exp(-y_{i}(z_{i}^{T}w+v))\right)+\lambda\|w\|_{1},\label{eq:pb-lo-primal}$$ with the same notation as in section \[s:hi\]. here $x_{i}\in{\mathbb{R}}^{n}$, $i=1,\ldots,m$ are the data points, and $y\in\{-1,1\}^{m}$ is the label vector. The above is a special case of the generic problem (\[eq:pb-generic-primal\]), where $f=f_{{\rm lo}}$ is the logistic loss, $b=y$, $c=0$, and the feature matrix $X$ is given by $X=[y_{1}z_{1},\ldots,y_{m}z_{m}]^{T}$, so that $x_{k}=[y_{1}z_{1}(k),\ldots,y_{m}z_{m}(k)]^{T}$. As before, we denote by ${\cal I}_{+},{\cal I}_{-}$ the set of indices corresponding to the positive and negative classes, respectively, and denote by $m_{\pm}=|{\cal I}_{\pm}|$ the associated cardinalities. Without loss of generality, we may assume $m_{+}\le m_{-}$. The dual problem takes the form $${\cal D}_{{\rm lo}}(\lambda)~:~~\phi(\lambda):=\max_{\theta}\:\sum_{i=1}^{m}\left(\theta_{i}\log(-\theta_{i})-(1+\theta_{i})^{T}\log(1+\theta_{i})\right)~:~\begin{array}[t]{l}
-{\mathbf 1}\le\theta\le0,\;\;\theta^{T}y=0,\\
|\theta^{T}x_{k}|\le\lambda,\;\; k=1,\ldots,n.\end{array}\label{eq:pb-lo-dual}$$
### Test, $\gamma$ given {#test-gamma-given}
Assume that we know a lower bound on the problem, $\gamma\le\phi(\lambda)$. Since $0\le\phi(\lambda)\le m\log2$, we may assume that $\gamma\in[0,m\log2]$ without loss of generality. We proceed to formulate problem (\[eq:P-def-dual\]). For given $x\in{\mathbb{R}}^{m}$, and $\gamma\in{\mathbb{R}}$, we have $$\begin{array}{rcl}
P_{{\rm log}}(\gamma,x) & = & {\displaystyle}\min_{\mu>0,\:\nu}\:-\gamma\mu+\mu\sum_{i=1}^{m}f_{{\rm log}}\left({\displaystyle}\frac{x_{i}+y_{i}\nu}{\mu}\right),\end{array}\label{eq:P-lo-pb}$$ which can be computed in $O(m)$ by two-dimensional search, or by the dual interior-point method described in appendix. (As mentioned before, an alternative, resulting in a more conservative test, is to fix $\nu$, for example $\nu=0$.) Our test to eliminate the $k$-th feature takes the form $$\lambda>T_{{\rm log}}(\gamma,x_{k}):=\max(P_{{\rm log}}(\gamma,x_{k}),P_{{\rm log}}(\gamma,-x_{k})).$$ If $\gamma$ is known, the complexity of running this test through all the features is $O(nm)$. (In fact, the terms in the objective function that correspond to zero elements of $x$ are of two types, involving $f_{{\rm log}}(\pm\nu/\mu)$. This means that the effective dimension of problem (\[eq:P-lo-pb\]) is the cardinality $d$ of vector $x$, which in many applications is much smaller than $m$.)
### Obtaining a dual feasible point
We can construct dual feasible points based on scaling one obtained by choice of a primal point (classifier weight) $w_{0}$. This in turn leads to other possible choices for the bound $\gamma$.
For $w_{0}\in{\mathbb{R}}^{n}$ given, we solve the one-dimensional, convex problem $$v_{0}:=\arg\min_{b}\:{\displaystyle}\sum_{i=1}^{m}f_{{\rm log}}(y_{i}x_{i}^{T}w_{0}+y_{i}b).$$ This problem can be solved by bisection in $O(m)$ time [@boyd_linear]. At optimum, the derivative of the objective is zero, hence $y^{T}\theta_{0}=0$, where $$\theta_{0}(i):=-\frac{1}{1+\exp(y_{i}x_{i}^{T}w_{0}+y_{i}v_{0})},\;\; i=1,\ldots,m.$$ Now apply the scaling method seen before, and set $\gamma$ by solving problem (\[eq:gamma-pb\]).
### A specific example of a dual point
A convenient, specific choice in the above construction is to set $w_{0}=0$. Then, the intercept $v_{0}$ can be explicitly computed, as $v_{0}=\log(m_{+}/m_{-})$, where $m_{\pm}=|\{i\::\: y_{i}=\pm1\}|$ are the class cardinalities. The corresponding dual point $\theta_{0}$ is $$\theta_{0}(i)=\left\{ \begin{array}{ll}
-{\displaystyle}\frac{m_{-}}{m} & (y_{i}=+1)\\[0.1in]
-{\displaystyle}\frac{m_{+}}{m} & (y_{i}=-1),\end{array}\right.\;\; i=1,\ldots,m.\label{eq:theta0-logreg}$$ The corresponding value of $\lambda_{0}$ is (see [@boyd_linear]): $$\lambda_{0}:=\|X^{T}\theta_{0}\|_{\infty}=\max_{1\le k\le n}|\theta_{0}^{T}x_{k}|.$$
We now compute $\gamma(\lambda)$ by solving problem (\[eq:gamma-pb\]), which expresses as $$\gamma(\lambda)=\max_{|s|\le\lambda/\lambda_{0}}\: G_{{\rm log}}(s\theta_{0})=\max_{|s|\le\lambda/\lambda_{0}}\:-m_{+}f_{{\rm log}}^{\ast}(-s\frac{m_{-}}{m})-m_{-}f_{{\rm log}}^{\ast}(-s\frac{m_{+}}{m}).\label{eq:theta-logreg}$$ The above can be solved analytically: it can be shown that $s=\lambda/\lambda_{0}$ is optimal.
### Solving the bisection problem
In this section, we are given $c\in{\mathbb{R}}^{m}$, $\gamma\in(0,m\log2)$, and we consider the problem $$\begin{array}{rcl}
F^{\ast}:={\displaystyle}\min_{\mu>0}\: F(\mu) & := & -\gamma\mu+\mu{\displaystyle}\sum_{i=1}^{m}f_{{\rm log}}(c(i)/\mu).\end{array}\label{eq:F-logreg}$$ Problem (\[eq:F-logreg\]) corresponds to the problem (\[eq:P-lo-pb\]), with $\nu$ set to a fixed value, and $c(i)=y_{i}x_{i}$, $i=1,\ldots,m$. We assume that $c(i)\ne0$ for every $i$, and that $\kappa:=m\log2-\gamma>0$. Observe that $F^{\ast}\le F_{0}:=\lim_{\mu\rightarrow0^{+}}\: F(\mu)={\mathbf 1}^{T}c_{+}$, where $c_{+}$ is the positive part of vector $c$.
To solve this problem via bisection, we initialize the interval of confidence to be $[0,\mu_{u}]$, with $\mu_{u}$ set as follows. Using the inequality $\log(1+e^{-x})\ge\log2-(1/2)x_{+}$, which is valid for every $x$, we obtain that for every $\mu>0$: $$F(\mu)\ge-\gamma\mu+\mu\sum_{i=1}^{m}\left(\log2-\frac{(c(i))_{+}}{2\mu}\right)=\kappa\mu-\frac{1}{2}{\mathbf 1}^{T}c_{+}.$$ We can now identify a value $\mu_{u}$ such that for every $\mu\ge\mu_{u}$, we have $F(\mu)\ge F_{0}$: it suffices to ensure $\kappa\mu-(1/2){\mathbf 1}^{T}c_{+}\ge F_{0}$, that is, $$\mu\ge\mu_{u}:=\frac{(1/2){\mathbf 1}^{T}c_{+}+F_{0}}{\kappa}=\frac{3}{2}\frac{{\mathbf 1}^{T}c_{+}}{m\log2-\gamma}.$$
### Algorithm summary
An algorithm to check if a given feature can be removed from a sparse logistic regression problem works as follows.
[*Given:*]{} $\lambda$, $k$ ($1\le k\le n$), $f_{{\rm log}}(x)=\log(1+e^{-x})$, $f_{{\rm log}}^{\ast}(\vartheta)=(-\vartheta)\log(-\vartheta)+(\vartheta+1)\log(\vartheta+1)$.
1. Set $\lambda_{0}={\displaystyle}\max_{1\le k\le n}|\theta_{0}^{T}x_{k}|$, where $\theta_{0}(i)=-m_{-}/m$ ($y_{i}=+1$), $\theta_{0}(i)=-m_{+}/m$ ($y_{i}=-1$), $i=1,\ldots,m$.
2. Set $$\gamma(\lambda):=-m_{+}f_{{\rm log}}^{\ast}(-\frac{\lambda}{\lambda_{0}}\frac{m_{-}}{m})-m_{-}f_{{\rm log}}^{\ast}(-\frac{\lambda}{\lambda_{0}}\frac{m_{+}}{m}).$$
3. Solve via bisection a pair of one-dimensional convex optimization problems $$P_{\epsilon}={\displaystyle}\min_{\mu>0}\:-\gamma(\lambda)\mu+\mu{\displaystyle}\sum_{i=1}^{m}f_{{\rm log}}(\epsilon y_{i}(x_{k})_{i}/\mu)\;\;(\epsilon=\pm1),$$ each with initial interval $[0,\mu_{u}]$, with $$\mu_{u}=\frac{3}{2}\frac{{\displaystyle}\sum_{i=1}^{m}(\epsilon y_{i}(x_{k})_{i})_{+}}{m\log2-\gamma}.$$
4. If $\lambda>\max(P_{+},P_{-})$, the $k$-th feature can be safely removed.
Numerical results {#s:num}
=================
In this section we explore the benefits of SAFE by running numerical experiments[^1] with different LASSO solvers. We present two kinds of experiments to highlight the two main benefits of SAFE. One kind, in our opinion the most important, shows how memory limitations can be reduced, by allowing to treat larger data sets. The other focuses on measuring computational time reduction when using SAFE a priori to the LASSO solver.
We have used a variety of available algorithms for solving the LASSO problem. We use acronyms to refer to the following methods: IPM stands for the Interior-Point Method for LASSO described in [@boyd_linear]; GLMNET corresponds to the Generalized Linear Model algorithm described in [@FHT:10]; TFOCS corresponds to Templates for First-Order Conic Solvers described in [@becker2010templates]; FISTA and Homotopy stand for the Fast Iterative Shrinkage-Thresholding Algorithm and homotopy algorithm, described and implemented in [@yang2010fast], respectively. Some methods (like IPM, TFOCS) do not return exact zeros in the final solution of the LASSO problem and the issue arises in evaluating the its cardinality. In appendix \[app:thresholding\], we discuss some issue related to the thresholding of the LASSO solution.
In our experiments, we use data sets derived from text classification sources in [@Frank+Asuncion:2010]. We use medical journal abstracts from PubMed represented in a bag-of-words format, where stop words have been eliminated and capitalization removed. The dimensions of the feature matrix $X$ we use from PubMed is $m=1,000,000$ abstracts and $n=127,025$ features (words). There is a total of $82,209,586$ non-zeros in the feature matrix, with an average of about $645$ non-zeros per feature (word). We also use data-sets derived from the headlines of [*The New York Times,*]{} (NYT) spanning a period of about $20$ years (from 1985 to 2007). The number of headlines in the entire NYT data-set is $m=3,241,260$ and the number of features (words) is $n=159,943$. There is a total of $14,083,676$ non-zeros in the feature matrix, with an average of about $90$ non-zeros per feature.
In some applications such as [@GJMEYC:10], the goal is to learn a short list of words that are predictive of the appearance of a given query term (say, “lung” or “china”) in the abstracts of medical journals or NYT news. The LASSO problem can be used to produce a summarization of the query term across the many abstracts or headlines considered. To be manageable by a human reader, the list of predictive terms should be very short (say at most $100$ terms) with respect to the size of the dictionary $n$. To produce such a short list, we solve the LASSO problem (\[eq:LASSO\]) with different penalty parameters $\lambda$, and choose the appropriate penalty $\lambda$ that would generate enough non-zeros in the LASSO solution (around $100$ non-zeros in our case).
SAFE for reducing memory limit problems
---------------------------------------
We experiment with PubMed data-set which is too large to be loaded into memory, and thus not amenable to current LASSO solvers. As described before, we are interested in solving the LASSO problem for a regularization parameter that would result in about $100$ non-zeros in the solution. We implement algorithm \[alg:SAFE-for-memory\] with a memory limit $M=1,000$ features, where we have observed that for the PubMed data loading more than $1,000$ features causes memory problems in the machine and platform we are using. The memory limit is approximately two orders of magnitudes less than the original number of features $n$, i.e. $M\approx0.01n$. Using algorithm \[alg:SAFE-for-memory\], we were able to solved the LASSO problem for $\lambda=0.04\lambda_{max}$ using a sequence of $25$ LASSO problem with each problem having a number of features less than $M=1,000$. Figure \[fig:mem\_lim\_pubmed\] shows the simulation result for the PubMed data-set.
![\[fig:mem\_lim\_pubmed\]A LASSO problem solved for the PubMed data-set and $\lambda=0.04\lambda_{max}$ using a sequence of $25$ smaller size problems. Each LASSO problem in the sequence has a number of features $L_{F}$ that satisfies the memory limit $M=1,000$, i.e $L_{F}\leq1,000$. ](Figures/mem_pubmed_f1){width="12cm"}
SAFE for LASSO run-time reduction
---------------------------------
We have used a portion of the NYT data-set corresponding to all headlines in year $1985,$ the corresponding feature matrix has dimensions $n=38,377$ features and $m=192,182$ headlines, with an average of $21$ non-zero per feature. We solved the plain LASSO problem and the LASSO problem with SAFE as outlined in algoirthm \[alg:Recursive-SAFE-for-1\] for a sequence of $\lambda$ logarithmically distributed between $0.03\lambda_{max}$ and $\lambda_{max}$. We have used four LASSO solvers, IPM, TFOCS, FISTA and Homotopy to solve the LASSO problem. Figure \[subfig:perc\_savings\]shows the computational time saving when using SAFE. Figure \[subfig:nyt\_soln\] shows the number of features we used to solve the LASSO problem when using SAFE, and the number of non-zeros in the solution. We realize that when using algorithm \[alg:Recursive-SAFE-for-1\] we solve problems with a number of features at most $10,000$ instead of $n=38,377$ features, this reduction has a direct impact on the solving time of the LASSO problem as demonstrated in figure \[subfig:perc\_savings\].
SAFE for LASSO with intercept problem {#safe-for-lasso-with-intercept-problem}
-------------------------------------
We return to the LASSO with intercept problem discussed in section \[ss:intercept-lasso\]. We generate a feature matrix $X\in{\mathbb{R}}^{m\times n}$ with $m=500$, $n=10^{6}$. The entries of $X$ has a ${\cal N}(0,1)$ normal distributed and sparsity density $d=0.1$. We also generate a vector of coefficients $\omega\in{\mathbb{R}}^{n}$ with $50$ non-zero entries. The response $y$ is generated by setting $y=X\omega+0.01\eta$, where $\eta$ is a vector in ${\mathbb{R}}^{m}$ with ${\cal N}(0,1)$ distribution. We use GLMNET implemented in `R` to solve the LASSO problem with intercept. The generated data, $X$ and $y$ can be loaded into `R` , yet memory problems occur when we try to solve the LASSO problem. We use algorithm \[alg:SAFE-for-memory\] with memory limit $M=10,000$ features and $\lambda=0.33\lambda_{max}$. Figure \[fig:mem\_lim\_glmnet\] shows the number of non-zeros in the solution of the $352$ sequence of problems used to obtain the solution at $\lambda=0.33\lambda_{max}$.
![\[fig:mem\_lim\_glmnet\]A LASSO problem with intercept solved for randomly generated data-set and $\lambda=0.33\lambda_{max}$ using a sequence of $352$ smaller size problems. Each LASSO problem in the sequence has a number of features $L_{F}$ that satisfies the memory limit $M=10,000$, i.e $L_{F}\leq1000$. ](Figures/glmnetR_fig1){width="12cm"}
Expression of $P(\gamma,x_k)$ (LASSO) {#app:p-gamma}
=====================================
We can express problem (\[eq:testproblem\]) in dual form as a convex optimization problem with two scalar variables, $\mu_{1}$ and $\mu_{2}$:
$$\begin{aligned}
P(\gamma,x_{k}) & = & \;\min_{\mu_{1},\mu_{2}\geq0}\max_{\theta}\, x_{k}^{T}\theta+\mu_{1}\left(G(\theta)-\gamma\right)+\mu_{2}g^{T}\left(\theta-\theta_{0}^{\star}\right)\\
& = & \;\min_{\mu_{1},\mu_{2}\geq0}-\mu_{1}\gamma-\mu_{2}g^{T}\theta_{0}^{\star}+\max_{\theta}\, x_{k}^{T}\theta+\mu_{1}G(\theta)+\mu_{2}g^{T}\theta\\
& = & \;\min_{\mu_{1},\mu_{2}\geq0}-\mu_{1}\gamma-\mu_{2}g^{T}\theta_{0}^{\star}+\mu_{1}\max_{\theta}\,\left(\frac{x_{k}^{T}-\mu_{1}y^{T}+\mu_{2}g^{T}}{\mu_{1}}\theta-\frac{1}{2}\left\Vert \theta\right\Vert _{2}^{2}\right)\end{aligned}$$
We obtain:$$\begin{aligned}
P(\gamma,x_{k}) & = & \min_{\mu_{1},\mu_{2}\geq0}L(\mu_{1},\mu_{2})\label{eq:safe3test_1}\end{aligned}$$ with $$L(\mu_{1},\mu_{2})=-x_{k}^{T}y+\frac{\mu_{1}}{2}D^{2}+\frac{1}{2\mu_{1}}\left\Vert x_{k}\right\Vert _{2}^{2}+\frac{\mu_{2}^{2}}{2\mu_{1}}\left\Vert g\right\Vert _{2}^{2}+\frac{\mu_{2}}{\mu_{1}}x_{k}^{T}g-\mu_{2}\left\Vert g\right\Vert _{2}^{2},\label{eq:objec_of_test}$$ and $D:=\left(\left\Vert y\right\Vert _{2}^{2}-2\gamma\right)^{1/2}$.
To solve (\[eq:safe3test\_1\]), we take the derivative of (\[eq:objec\_of\_test\]) w.r.t $\mu_{2}$ and set it to zero:
$$\mu_{2}\left\Vert g\right\Vert _{2}^{2}+x_{k}^{T}g-\mu_{1}\left\Vert g\right\Vert _{2}^{2}=0.$$
This implies that $\mu_{2}=\text{max}(0,\mu_{1}-\frac{x_{k}^{T}g}{\left\Vert g\right\Vert _{2}^{2}})$. When $\mu_{1}\leq\frac{x_{k}^{T}g}{\left\Vert g\right\Vert _{2}^{2}}$, we have $\mu_{2}=0$, $\mu_{1}=\frac{\left\Vert x_{k}\right\Vert _{2}}{D}$ and $P(\gamma,x_{k})$ takes the value:
$$P(\gamma,x_{k})=-y^{T}x_{k}+\left\Vert x_{k}\right\Vert _{2}D.$$
On the other hand, when $\mu_{1}\geq\frac{x_{k}^{T}g}{\left\Vert g\right\Vert _{2}^{2}}$, we take the derivative of (\[eq:objec\_of\_test\]) w.r.t $\mu_{1}$ and set it to zero:
$$\tilde{D}^{2}\mu_{1}^{2}=\Psi_{k}^{2},$$ with $\Psi_{k}=\left(\left\Vert x_{k}\right\Vert _{2}^{2}-\frac{\left(x_{k}^{T}g\right)^{2}}{\left\Vert g\right\Vert _{2}^{2}}\right)^{1/2}$ and $\tilde{D}=\left(D^{2}-\left\Vert g\right\Vert _{2}^{2}\right)^{1/2}$. Substituting $\mu_{1}$ and $\mu_{2}$ in (\[eq:safe3test\_1\]), $P(\gamma,x_{k})$ takes the value:
$$P(\gamma,x_{k})=\theta_{0}^{\star T}x_{k}+\Psi_{k}\tilde{D}.$$
Expression of $P(\gamma,x)$, general case
=========================================
We show that the quantity $P(\gamma,x)$ defined in (\[eq:P-def\]) can be expressed in dual form (\[eq:P-def-dual\]). This is a simple consequence of duality: $$\begin{array}{rcl}
P(\gamma,x) & = & {\displaystyle}\max_{\theta}\:\theta^{T}x~:~G(\theta)\ge\gamma,\;\;\theta^{T}b=0\\
& = & {\displaystyle}\max_{\theta}\:\min_{\mu>0,\:\nu}\:\theta^{T}x+\mu(G(\theta)-\gamma)-\nu\theta^{T}b\\
& = & {\displaystyle}\min_{\mu>0,\:\nu}\:{\displaystyle}\max_{\theta}\:\theta^{T}x+\mu(-y^{T}\theta-\sum_{i=1}^{m}f^{\ast}(\theta(i))-\gamma)-\nu\theta^{T}b\\
& = & {\displaystyle}\min_{\mu>0,\:\nu}\:-\gamma\mu+{\displaystyle}\max_{\theta}\:\theta^{T}(x-\mu y-\nu z)-\mu\sum_{i=1}^{m}f^{\ast}(\theta(i))\\
& = & {\displaystyle}\min_{\mu>0,\:\nu}\:-\gamma\mu+\mu\left({\displaystyle}\max_{\theta}\:\frac{1}{\mu}\theta^{T}(x-\mu y-\nu z)-\sum_{i=1}^{m}f^{\ast}(\theta(i))\right)\\
& = & {\displaystyle}\min_{\mu>0,\:\nu}\:-\gamma\mu+\mu\sum_{i=1}^{m}f\left(\frac{x_{i}-\mu y(i)-\nu b_{i}}{\mu}\right).\end{array}$$
SAFE test for SVM {#app:min-poly-2d}
=================
In this section, we examine various optimization problems involving polyhedral functions in one or two variables, which arise in section \[ss:test-gamma-svm\] for the computation of $P_{{\rm hi}}(\gamma,x)$ as well as in the SAFE-SVM theorem of section \[ss:safe-thm-svm\].
Computing $P_{{\rm hi}}(\gamma,x)$ {#app:P-gamma-svm}
----------------------------------
We first focus on the specific problem of computing the quantity defined in (\[eq:P-hi-pb\]). To simplify notation, we will consider the problem of computing $P_{{\rm hi}}(\gamma,-x)$, that is: $$P_{{\rm hi}}(\gamma,-x)={\displaystyle}\min_{\mu\ge0,\:\nu}\:-\gamma\mu+\sum_{i=1}^{m}(\mu+\nu y_{i}+x_{i})_{+},\label{eq:app-P-hi}$$ where $y\in\{-1,1\}^{m}$, $x\in{\mathbb{R}}^{m}$ and $\gamma$ are given, with $0\le\gamma\le\gamma_{0}:=2\min(m_{+},m_{-})$. Here, ${\cal I}_{\pm}:=\{i\::\: y_{i}=\pm1\}$, and $x^{+}=(x_{i})_{i\in{\cal I}_{+}}$, $x^{-}=(x_{i})_{i\in{\cal I}_{-}}$, $m_{\pm}=|{\cal I}_{\pm}|$, and $\underline{m}=\min(m_{+},m_{-})$. Without loss of generality, we assume that both $x^{+},x^{-}$ are both sorted in descending order: $x_{1}^{\pm}\ge\ldots\ge x_{m_{\pm}}^{\pm}$.
Using $\alpha=\mu+\nu$, $\beta=\mu-\nu$, we have $$\begin{array}{rcl}
P_{{\rm hi}}(\gamma,-x) & = & {\displaystyle}\min_{\alpha+\beta\ge0}\:-{\displaystyle}\frac{\gamma}{2}(\alpha+\beta)+\sum_{i=1}^{m_{+}}(x_{i}^{+}+\alpha)_{+}+\sum_{i=1}^{m_{-}}(x_{i}^{-}+\beta)_{+}\\
& = & {\displaystyle}\min_{\alpha,\:\beta}\:\max_{t\ge0}\:-{\displaystyle}\frac{\gamma}{2}(\alpha+\beta)+\sum_{i=1}^{m_{+}}(x_{i}^{+}+\alpha)_{+}+\sum_{i=1}^{m_{-}}(x_{i}^{-}+\beta)_{+}-t(\alpha+\beta)\\
& = & {\displaystyle}\max_{t\ge0}\:{\displaystyle}\min_{\alpha,\:\beta}\:-({\displaystyle}\frac{\gamma}{2}+t)(\alpha+\beta)+\sum_{i=1}^{m_{+}}(x_{i}^{+}+\alpha)_{+}+\sum_{i=1}^{m_{-}}(x_{i}^{-}+\beta)_{+}\\
& = & {\displaystyle}\max_{t\ge0}\: F({\displaystyle}\frac{\gamma}{2}+t,x^{+})+F({\displaystyle}\frac{\gamma}{2}+t,x^{-}),\end{array}\label{eq:P-hi-via-F}$$ where, for $h\in{\mathbb{R}}$ and $x\in{\mathbb{R}}^{p}$, $x_{1}\ge\ldots\ge x_{p}$, we set $$F(h,x):=\min_{z}\:-hz+\sum_{i=1}^{p}(z+x_{i})_{+},\label{eq:P-hi-F-xh}$$
#### Expression of the function $F$.
If $h>p$, then with $z\rightarrow+\infty$ we obtain $F(h,x)=-\infty$. Similarly, if $h<0$, then $z\rightarrow-\infty$ yields $F(h,x)=-\infty$. When $0\le h\le p$, we proceed by expressing $F$ in dual form: $$F(h,x)=\max_{u}\: u^{T}x~:~0\le u\le{\mathbf 1},\;\; u^{T}{\mathbf 1}=h.$$
If $h=p$, then the only feasible point is $u={\mathbf 1}$, so that $F(p,x)={\mathbf 1}^{T}x$. If $0\le h<1$, choosing $u_{1}=h$, $u_{2}=\ldots=u_{p}=0$, we obtain the lower bound $F(h,x)\ge hx_{1}$, which is attained with $z=-x_{1}$.
Assume now that $1\le h<p$. Let $h=q+r$, with $q=\lfloor h\rfloor$ the integer part of $h$, and $0\le r<1$. Choosing $u_{1}=\ldots=u_{q}=1$, $u_{q+1}=r$, we obtain the lower bound $$F(h,x)\ge\sum_{j=1}^{q}x_{j}+rx_{q+1},$$ which is attained by choosing $z=-x_{q+1}$ in the expression (\[eq:P-hi-F-xh\]).
To summarize: $$F(h,x)=\left\{ \begin{array}{ll}
hx_{1} & \mbox{if }0\le h<1,\\
{\displaystyle}\sum_{j=1}^{\lfloor h\rfloor}x_{j}+(h-\lfloor h\rfloor)x_{\lfloor h\rfloor+1} & \mbox{if }1\le h<p,\\
{\displaystyle}\sum_{j=1}^{p}x_{j} & \mbox{if }h=p,\\
-\infty & \mbox{otherwise.}\end{array}\right.\label{eq:P-hi-F-xh-soln}$$ A more compact expression, valid for $0\le h\le p$ if we set $x_{p+1}=x_{p}$ and assume that a sum over an empty index sets is zero, is $$F(h,x)={\displaystyle}\sum_{j=1}^{\lfloor h\rfloor}x_{j}+(h-\lfloor h\rfloor)x_{\lfloor h\rfloor+1},\;\;0\le h\le p.$$ Note that $F(\cdot,x)$ is the piece-wise linear function that interpolates the sum of the $h$ largest elements of $x$ at the integer break points $h=0,\ldots,p$.
#### Expression of $P_{{\rm hi}}(\gamma,-x)$.
We start with the expression found in (\[eq:P-hi-via-F\]): $$P_{{\rm hi}}(\gamma,-x)={\displaystyle}\max_{t\ge0}\: F({\displaystyle}\frac{\gamma}{2}+t,x^{+})+F({\displaystyle}\frac{\gamma}{2}+t,x^{-}).$$ Since the domain of $F(\cdot,x^{+})+F(\cdot,x^{-})$ is $[0,\underline{m}]$, and with $0\le\gamma/2\le\gamma_{0}/2=\underline{m}$, we get $$P_{{\rm hi}}(\gamma,-x)={\displaystyle}\max_{\gamma/2\le h\le\underline{m}}\: G(h,x^{+},x^{-}):=F(h,x^{+})+F(h,x^{-}).$$ Since $F(\cdot,x)$ with $x\in{\mathbb{R}}^{p}$ is a piece-wise linear function with break points at $0,\ldots,p$, a maximizer of $G(\cdot,x^{+},x^{-})$ over $[\gamma/2,\underline{m}]$ lies in $\{\gamma/2,\lfloor\gamma/2\rfloor+1,\ldots,\underline{m}\}$. Thus, $$P_{{\rm hi}}(\gamma,-x)={\displaystyle}\max\left(G(\frac{\gamma}{2},x^{+},x^{-}),\max_{h\in\{\lfloor\gamma/2\rfloor+1,\ldots,\underline{m}\}}\: G(h,x^{+},x^{-})\right).$$
Let us examine the second term, and introduce the notation $\bar{x}_{j}:=x_{j}^{+}+x_{j}^{-}$, $j=1,\ldots,\underline{m}$: $$\begin{aligned}
\max_{h\in\{\lfloor\gamma/2\rfloor+1,\ldots,\underline{m}\}}\: G(h,x^{+},x^{-}) & = & \max_{h\in\{\lfloor\gamma/2\rfloor+1,\ldots,\underline{m}\}}\:\sum_{j=1}^{h}(x_{j}^{+}+x_{j}^{-})\\
& = & \sum_{j=1}^{\lfloor\gamma/2\rfloor+1}\bar{x}_{j}+\sum_{j=\lfloor\gamma/2\rfloor+2}^{\underline{m}}(\bar{x}_{j})_{+},\end{aligned}$$ with the convention that sums over empty index sets are zero. Since $$G(\frac{\gamma}{2},x^{+},x^{-})={\displaystyle}\sum_{j=1}^{\lfloor\gamma/2\rfloor}\bar{x}_{j}+(\frac{\gamma}{2}-\lfloor\frac{\gamma}{2}\rfloor)\bar{x}_{\lfloor\gamma/2\rfloor+1},$$ we obtain $$P_{{\rm hi}}(\gamma,-x)={\displaystyle}\sum_{j=1}^{\lfloor\gamma/2\rfloor}\bar{x}_{j}+{\displaystyle}\max\left((\frac{\gamma}{2}-\lfloor\frac{\gamma}{2}\rfloor)\bar{x}_{\lfloor\gamma/2\rfloor+1},\bar{x}_{\lfloor\gamma/2\rfloor+1}+\sum_{j=\lfloor\gamma/2\rfloor+2}^{\underline{m}}(\bar{x}_{j})_{+}\right).$$ An equivalent expression is: $$\begin{array}{rcl}
P_{{\rm hi}}(\gamma,-x) & = & {\displaystyle}\sum_{j=1}^{\lfloor\gamma/2\rfloor}\bar{x}_{j}-(\frac{\gamma}{2}-\lfloor\frac{\gamma}{2}\rfloor)(-\bar{x}_{\lfloor\gamma/2\rfloor+1})_{+}+\sum_{j=\lfloor\gamma/2\rfloor+1}^{\underline{m}}(\bar{x}_{j})_{+},\;\;0\le\gamma\le2\underline{m},\\
& & \bar{x}_{j}:=x_{j}^{+}+x_{j}^{-},\;\; j=1,\ldots,\underline{m}.\end{array}$$ The function $P_{{\rm hi}}(\cdot,-x)$ linearly interpolates the values obtained for $\gamma=2q$ with $q$ integer in $\{0,\ldots,\underline{m}\}$: $$P_{{\rm hi}}(2q,-x)={\displaystyle}\sum_{j=1}^{q}\bar{x}_{j}+\sum_{j=q+1}^{\underline{m}}(\bar{x}_{j})_{+}.$$
Computing $\Phi(x^{+},x^{-})$ {#app:Phi-svm}
-----------------------------
Let us consider the problem of computing $$\Phi(x^{+},x^{-}):=\min_{\nu}\:\sum_{i=1}^{m_{+}}(x_{i}^{+}+\nu)_{+}+\sum_{i=1}^{m_{-}}(x_{i}^{-}-\nu)_{+},$$ with $x^{\pm}\in{\mathbb{R}}^{m_{\pm}}$, $x_{1}^{\pm}\ge\ldots\ge x_{m_{\pm}}^{\pm}$, given. We can express $\Phi(x^{+},x^{-})$ in terms of the function $F$ defined in (\[eq:P-hi-F-xh\]): $$\begin{array}{rcl}
\Phi(x^{+},x^{-}) & = & {\displaystyle}\min_{\nu_{+},\nu_{-}}\:\sum_{i\in{\cal I}_{+}}(x_{i}^{+}+\nu^{+})_{+}+\sum_{i\in{\cal I}_{-}}(x_{i}^{-}-\nu^{-})_{+}~:~\nu^{+}=\nu^{-}\\
& = & {\displaystyle}\max_{h}\:{\displaystyle}\min_{\nu^{+},\nu^{-}}\:-h(\nu^{+}-\nu^{-})+\sum_{i\in{\cal I}_{+}}(x_{i}^{+}+\nu^{+})_{+}+\sum_{i\in{\cal I}_{-}}(x_{i}^{-}-\nu^{-})_{+}\\
& = & {\displaystyle}\max_{h}\:{\displaystyle}\min_{\nu^{+},\nu^{-}}\:-h\nu^{+}+\sum_{i\in{\cal I}_{+}}(x_{i}^{+}+\nu^{+})_{+}+h\nu^{-}+\sum_{i\in{\cal I}_{-}}(x_{i}^{-}-\nu^{-})_{+}\\
& = & {\displaystyle}\max_{h}\:\left({\displaystyle}\min_{\nu}\:-h\nu+\sum_{i\in{\cal I}_{+}}(x_{i}^{+}+\nu)_{+}\right)+\left({\displaystyle}\min_{\nu}\:-h\nu+\sum_{i\in{\cal I}_{-}}(x_{i}^{-}+\nu)_{+}\right)\;\;(\nu_{+}=-\nu_{-}=\nu)\\
& = & {\displaystyle}\max_{h}\: F(h,x^{+})+F(h,x^{-})\\
& = & {\displaystyle}\max_{0\le h\le\underline{m}}\: F(h,x^{+})+F(h,x^{-})\\
& = & \max(A,B,C),\end{array}$$ where $F$ is defined in (\[eq:P-hi-F-xh\]), and $$A={\displaystyle}\max_{0\le h<1}\: F(h,x^{+})+F(h,x^{-}),\;\; B:={\displaystyle}\max_{1\le h<\underline{m}}\: F(h,x^{+})+F(h,x^{-})),\;\; C=F(\underline{m},x^{+})+F(\underline{m},x^{-}).$$
We have $$A:={\displaystyle}\max_{0\le h<1}\: F(h,x^{+})+F(h,x^{-})={\displaystyle}\max_{0\le h<1}\: h(x_{1}^{+}+x_{1}^{-})=(x_{1}^{+}+x_{1}^{-})_{+}.$$ Next: $$\begin{aligned}
B & = & \max_{1\leq h<\underline{m}}\: F(h,x^{+})+F(h,x^{-})\\
& = & \max_{q\in\{1,\ldots ,\underline{m}-1\},r\in[0,1[}\:\sum_{i=1}^{q}(x_{i}^{+}+x_{i}^{-})+r(x_{q+1}^{+}+x_{q+1}^{-})\\
& = & \max_{q\in\{1,\ldots,\underline{m}-1\}}\sum_{i=1}^{q}(x_{i}^{+}+x_{i}^{-})+(x_{q+1}^{+}+x_{q+1}^{-})_{+}\\
& = & (x_{1}^{+}+x_{1}^{-})+\sum_{i=2}^{\underline{m}}(x_{i}^{+}+x_{i}^{-})_{+}.\end{aligned}$$ Observe that $$B\geq C=\sum_{i=1}^{\underline{m}}(x_{i}^{+}+x_{i}^{-}).$$ Moreover, if $(x_{1}^{+}+x_{1}^{-})\geq0$, then $B=\sum_{i=1}^{\underline{m}}(x_{i}^{+}+x_{i}^{-})_{+}\geq A$. On the other hand, if $x_{1}^{+}+x_{1}^{-}\leq0$, then $x_{i}^{+}+x_{i}^{-}\le0$ for $2\le j\le\underline{m}$, and $A=\sum_{i=1}^{\underline{m}}(x_{i}^{+}+x_{i}^{-})_{+}\geq x_{1}^{+}+x_{1}^{-}=B$. In all cases, $$\Phi(x^{+},x^{-})=\max(A,B,C)=\sum_{i=1}^{\underline{m}}(x_{i}^{+}+x_{i}^{-})_{+}.$$
SAFE-SVM test {#app:G-svm}
-------------
Now we consider the problem that arises in the SAFE-SVM test (\[eq:test-svm-kappa\]): $$G(z):=\min_{0\le\kappa\le1}\:\sum_{i=1}^{p}(1-\kappa+\kappa z_{i})_{+},$$ where $z\in{\mathbb{R}}^{p}$ is given. (The SAFE-SVM condition (\[eq:test-svm-kappa\]) involves $z_{i}=\gamma_{0}/(2\lambda_{0})(x_{[i]}^{+}+x_{[i]}^{-})$, $i=1,\ldots,p:=\underline{m}$.) We develop an algorithm to compute the quantity $G(z)$, the complexity of which grows as $O(d\log d)$, where $d$ is (less than) the number of non-zero elements in $z$.
Define ${\cal I}_{\pm}=\{i\::\:\pm z_{i}>0\}$, $k:=|{\cal I}_{+}|$, $h:=|{\cal I}_{-}|$, $l={\cal I}_{0}$, $l:=|{\cal I}_{0}|$.
If $k=0$, ${\cal I}_{+}$ is empty, and $\kappa=1$ achieves the lower bound of $0$ for $G(z)$. If $k>0$ and $h=0$, that is, $k+l=p$, then ${\cal I}_{-}$ is empty, and an optimal $\kappa$ is attained in $\{0,1\}$. In both cases (${\cal I}_{+}$ or ${\cal I}_{-}$ empty), we can write $$G(z)=\min_{\kappa\in\{0,1\}}\:\sum_{i=1}^{p}(1-\kappa+\kappa z_{i})_{+}=\min\left(p,S_{+}\right),\;\; S_{+}:=\sum_{i\in{\cal I}_{+}}z_{i},$$ with the convention that a sum over an empty index set is zero.
Next we proceed with the assumption that $k\ne0$ and $h\ne0$. Let us re-order the elements of ${\cal I}_{-}$ in decreasing fashion, so that $z_{i}>0=z_{k+1}=\ldots=z_{k+l}>z_{k+l+1}\ge\ldots\ge z_{p}$, for every $i\in{\cal I}_{+}$. (The case when ${\cal I}_{0}$ is empty is handled simply by setting $l=0$ in our formula.) We have $$G(z)=k+l+\min_{0\le\kappa\le1}\:\left\{ \kappa\alpha+\sum_{i=k+l+1}^{p}(1-\kappa+\kappa z_{i})_{+}\right\} ,$$ where, $\alpha:=S_{+}-k-l$. The minimum in the above is attained at $\kappa=0,1$ or one of the break points $1/(1-z_{j})\in(0,1)$, where $j\in\{k+l+1,\ldots,p\}$. At $\kappa=0,1$, the objective function of the original problem takes the values $S_{+},p$, respectively. The value of the same objective function at the break point $\kappa=1/(1-z_{j})$, $j=k+l+1,\ldots,p$, is $k+l+G_{j}(z)$, where $$\begin{aligned}
G_{j}(z) & := & \frac{\alpha}{1-z_{j}}+{\displaystyle}\sum_{i=k+l+1}^{p}\left(\frac{z_{i}-z_{j}}{1-z_{j}}\right)_{+}\\
& = & \frac{\alpha}{1-z_{j}}+{\displaystyle}\frac{1}{1-z_{j}}\sum_{i=k+l+1}^{j-1}(z_{i}-z_{j})\\
& = & {\displaystyle}\frac{1}{1-z_{j}}\left(\alpha-(j-k-l-1)z_{j}+{\displaystyle}\sum_{i=k+l+1}^{j-1}z_{i}\right)\\
& = & {\displaystyle}\frac{1}{1-z_{j}}\left(S_{+}-(j-1)z_{j}-(k+l)(1-z_{j})+{\displaystyle}\sum_{i=k+l+1}^{j-1}z_{i}\right)\\
& = & -(k+l)+{\displaystyle}\frac{1}{1-z_{j}}\left({\displaystyle}\sum_{i=1}^{j-1}z_{i}-(j-1)z_{j}\right).\end{aligned}$$ This allows us to write $$G(z)=\min\left(p,\sum_{i=1}^{k}z_{i},\min_{j\in\{k+l+1,\ldots,p\}}\:{\displaystyle}\frac{1}{1-z_{j}}\left({\displaystyle}\sum_{i=1}^{j-1}z_{i}-(j-1)z_{j}\right)\right).$$ The expression is valid when $k+l=p$ ($h=0$, ${\cal I}_{-}$ is empty), $l=0$ (${\cal I}_{0}$ is empty), or $k=0$ (${\cal I}_{+}$ is empty) with the convention that the sum (resp. minimum) over an empty index set is $0$ (resp. $+\infty$).
We can summarize the result with the compact formula: $$G(z)=\min_{z}\:{\displaystyle}\frac{1}{1-z}\sum_{i=1}^{p}(z_{i}-z)_{+}~:~z\in\{-\infty,0,(z_{j})_{j\::\: z_{j}<0}\}.$$
Let us detail an algorithm for computing $G(z)$. Assume $h>0$. The quantity $$\underline{G}(z):=\min_{k+l+1\le j\le p}\:(G_{j}(z))$$ can be evaluated in less than $O(h)$, via the following recursion: $$\begin{array}{rcl}
G_{j+1}(z) & = & {\displaystyle}\frac{1-z_{j}}{1-z_{j+1}}G_{j}(z)-j\frac{z_{j+1}-z_{j}}{1-z_{j+1}}\\
\underline{G}_{j+1}(z) & = & \min(\underline{G}_{j}(z),G_{j+1}(z))\end{array},\;\; j=k+l+1,\ldots,p,\label{eq:rec-underlineG-svm-app}$$ with initial values $$G_{k+l+1}(z)=\underline{G}_{k+l+1}(z)={\displaystyle}\frac{1}{1-z_{k+l+1}}\left({\displaystyle}\sum_{i=1}^{k+l}z_{i}-(k+l)z_{k+l+1}\right).$$ On exit, $\underline{G}(z)=\underline{G}_{p}$.
Our algorithm is as follows.
**Algorithm for the evaluation of $G(z)$.**
1. Find the index sets ${\cal I}_{+}$, ${\cal I}_{-}$, ${\cal I}_{0}$, and their respective cardinalities $k,h,l$.
2. If $k=0$, set $G(z)=0$ and exit.
3. Set $S_{+}=\sum_{i=1}^{k}z_{i}$.
4. If $h=0$, set $G(z)=\min(p,S_{+})$, and exit.
5. If $h>0$, order the negative elements of $z$, and evaluate $\underline{G}(z)$ by the recursion (\[eq:rec-underlineG-svm-app\]). Set $G(z)=\min(p,S_{+},\underline{G}(z))$ and exit.
The complexity of evaluating $G(z)$ thus grows in $O(k+h\log h)$, which is less than $O(d\log d)$, where $d=k+h$ is the number of non-zero elements in $z$.
Computing $P_{{\rm log}}(\gamma,x)$ via an interior-point method
================================================================
We consider the problem (\[eq:P-lo-pb\]) which arises with the logistic loss. We can use a generic interior-point method [@BV:04], and exploit the decomposable structure of the dual function $G_{{\rm log}}$. The algorithm is based on solving, via a variant of Newton’s method, a sequence of linearly constrained problems of the form $$\min_{\theta}\:\tau x^{T}\theta+\log(G_{{\rm log}}(\theta)-\gamma)+\sum_{i=1}^{m}\log(-\theta-\theta^{2})~:~z^{T}\theta=0,$$ where $\tau>0$ is a parameter that is increased as the algorithm progresses, and the last terms correspond to domain constraints $\theta\in[-1,0]^{m}$. As an initial point, we can take the point $\theta$ generated by scaling, as explained in section \[ss:dual-scaling-generic\]. Each iteration of the algorithm involves solving a linear system in variable $\delta$, of the form $H\delta=h$, with $H$ is a rank-two modification to the Hessian of the objective function in the problem above. It is easily verified that the matrix $H$ has a “diagonal plus rank-two” structure, that is, it can be written as $H=D-gg^{T}-vv^{T}$, where the $m\times m$ matrix $D$ is diagonal and $g,v\in{\mathbb{R}}^{m}$ are computed in $O(m)$. The matrix $H$ can be formed, as the associated linear system solved, in $O(m)$ time. Since the number of iterations for this problem with two constraints grows as $\log(1/\epsilon)O(1)$, the total complexity of the algorithm is $\log(1/\epsilon)O(m)$ ($\epsilon$ is the absolute accuracy at which the interior-point method computes the objective). We note that memory requirements for this method also grow as $O(m)$.
On thresholding methods for LASSO {#app:thresholding}
=================================
Sparse classification algorithms may return a classifier vector $w$ with many small, but not exactly zero, elements. This implies that we need to choose a thresholding rule to decide which elements to set to zero. In this section, we discuss an issue related to the thresholding rule originally proposed for the interior point method for Logistic algorithm in [@boyd_logistic], and propose a new thresholding rule.
#### The KKT thresholding rule.
Recall that the primal problem for LASSO is $$\phi(\lambda)=\min_{w}\frac{1}{2}\|X^{T}w-y\|_{2}^{2}+\lambda\|w\|_{1}.$$ Observing that the KKT conditions imply that, at optimum, $(X(X^{T}w-y))_{k}=\lambda\mbox{sign}(w_{k})$, with the convention $\mbox{sign}(0)\in[-1,1]$, and following the ideas of [@boyd_logistic], the following thresholding rule can be proposed: at optimum, set component $w_{k}$ to $0$ whenever $$|(X(X^{T}w-y))_{k}|\leq0.9999\lambda.\label{thresholding_Boyd_rule}$$ We refer to this rule as the “KKT” rule.
The IPM-LASSO algorithm takes as input a “duality gap” parameter $\epsilon$, which controls the relative accuracy on the objective. When comparing the IPM code results with other algorithms such as GLMNET, we observed chaotic behaviors when applying the KKT rule, especially when the duality gap parameter $\epsilon$ was not small enough. More surprisingly, when this parameter is not small enough, some components $w_{k}$ with absolute values not close to $0$ can be thresholded. This suggests that the KKT rule should only be used for problems solved with a small enough duality gap $\epsilon$. However, setting the duality gap to a small value can dramatically slow down computations. In our experiments, changing the duality gap from $\epsilon=10^{-4}$ to $10^{-6}$ (resp.$10^{-8}$) increased the computational time by $30\%$ to $40\%$ (resp. $50$ to $100\%$).
#### An alternative method.
We propose an alternative thresholding rule, which is based on controlling the perturbation of the objective function that is induced by thresholding.
Assume that we have solved the LASSO problem above, with a given duality gap parameter $\epsilon$. If we denote by $w^{\ast}$ the classifier vector delivered by the IPM algorithm, $w^{\ast}$ is $\epsilon$-sub-optimal, that is, achieves a value $$\phi^{\ast}=\frac{1}{2}\|Xw^{\ast}-y\|_{2}^{2}+\lambda\|w^{\ast}\|_{1},$$ with $0\le\phi^{\ast}-\phi(\lambda)\le\epsilon\phi(\lambda)$.
For a given threshold $\tau>0$, consider the thresholded vector $\tilde{w}(\tau)$ defined as $$\begin{aligned}
\tilde{w}_{k}(\tau) & = & \left\{ \begin{array}{ll}
0 & \mbox{if }|w_{k}^{\ast}|\le\tau,\\
w_{k}^{\ast} & \mbox{otherwise,}\end{array}\right.\;\; k=1,\ldots,n.\end{aligned}$$ We have $\tilde{w}(\tau)=w^{\ast}+\delta(\tau)$ where the vector of perturbation $\delta(\tau)$ is such that $$\begin{aligned}
\delta_{k}(\tau) & = & \left\{ \begin{array}{ll}
-w_{k}^{\ast} & \mbox{if }|w_{k}^{\ast}|\le\tau,\\
0 & \mbox{otherwise,}\end{array}\right.\;\; k=1,\ldots,n.\end{aligned}$$ Note that, by construction, we have $\|w^{\ast}\|_{1}=\|w^{\ast}+\delta\|_{1}+\|\delta\|_{1}$. Also note that if $w^{\ast}$ is sparse, so is $\delta$.
Let us now denote by $\phi_{\tau}$ the LASSO objective that we obtain upon replacing the optimum classifier $w^{\ast}$ with its thresholded version $\tilde{w}(\tau)=w^{\ast}+\delta(\tau)$: $$\begin{aligned}
\phi_{\tau} & := & \frac{1}{2}\|X(w^{\ast}+\delta(\tau))-y\|_{2}^{2}+\lambda\|w^{\ast}+\delta(\tau)\|_{1}.\end{aligned}$$
Since $w(\tau)$ is (trivially) feasible for the primal problem, we have $\phi_{\tau}\ge\phi(\lambda)$. On the other hand, $$\begin{aligned}
\phi_{\tau} & = & \frac{1}{2}\|Xw^{\ast}-y\|_{2}^{2}+\lambda\|w^{\ast}+\delta(\tau)\|_{1}+\frac{1}{2}\|X\delta(\tau)\|_{2}^{2}+\delta(\tau)^{T}X^{T}(Xw^{\ast}-y)\\
& \le & \frac{1}{2}\|Xw^{\ast}-y\|_{2}^{2}+\lambda\|w^{\ast}\|_{1}+\frac{1}{2}\|X\delta(\tau)\|_{2}^{2}+\delta(\tau)^{T}X^{T}(Xw^{\ast}-y).\end{aligned}$$ For a given $\alpha>1$, the condition $${\cal C}(\tau):=\frac{1}{2}\|X\delta(\tau)\|_{2}+\delta(\tau)^{T}X^{T}(Xw^{\ast}-y)\leq\kappa\phi^{\ast},\;\;\kappa:=\frac{1+\alpha\epsilon}{1+\epsilon}-1\ge0,\label{eq:new-thresholding-rule}$$ allows to write $$\phi(\lambda)\le\phi_{\tau}\le\ (1+\alpha\epsilon)\phi(\lambda).$$ The condition (\[eq:new-thresholding-rule\]) then implies that the thresholded classifier is sub-optimal, with relative accuracy $\alpha\epsilon$.
Our proposed thresholding rule is based on the condition (\[eq:new-thresholding-rule\]). Precisely, we choose the parameter $\alpha>0$, then we set the threshold level $\tau$ by solving, via line search, the largest threshold $\tau$ allowed by condition (\[eq:new-thresholding-rule\]): $$\tau_{\alpha}=\arg\max_{\tau\geq0}\:\left\{ \tau~:~\|X\delta(\tau)\|_{2}\le\left(\sqrt{\frac{1+\alpha\epsilon}{1+\epsilon}}-1\right)\|Xw^{\ast}-y\|_{2}\right\} .$$ The larger $\alpha$ is, the more elements the rule allows to set to zero; at the same time, the more degradation in the objective will be observed: precisely, the new relative accuracy is bounded by $\alpha\epsilon$. The rule also depends on the duality gap parameter $\epsilon$. We refer to the thresholding rule as TR($\alpha$) in the sequel. In practice, we observe that the value $\alpha=2$ works well, in a sense made more precise below.
The complexity of the rule is $O(mn)$. More precisely, the optimal dual variable $\theta^{\ast}=Xw^{\ast}-y$ is returned by IPM-LASSO. The matrix $X\theta^{\ast}=X(X^{T}w^{\ast}-y)$ is computed once for all in $O(mn)$. We then sort the optimal vector $w^{\ast}$ so that $|w_{(1)}^{\ast}|\le\ldots\le|w_{(n)}^{\ast}|$, and set $\tau=\tau_{0}=|w_{(n)}^{\ast}|$, so that $\delta_{k}(\tau_{0})=-w_{k}^{\ast}$ and $\tilde{w}_{k}(\tau_{0})=0$ for all $k=1,\ldots,n$. The product $X\delta(\tau_{0})$ is computed in $O(mn)$, while the product $\delta(\tau_{0})^{T}(X^{T}\theta^{\ast})$ is computed in $O(n)$. If the quantity ${\cal C}(\tau_{0})=\frac{1}{2}\|X\delta(\tau_{0})\|_{2}+\delta(\tau_{0})^{T}(X^{T}\theta^{\ast})$ is greater than $\kappa\phi^{\star}$, then we set $\tau=\tau_{1}=|w_{(n-1)}^{\ast}|$. We have $\delta_{k}(\tau_{1})=\delta_{k}(\tau_{0})$ for any $k\neq(n)$ and $\delta_{(n)}(\tau_{1})=0$. Therefore, ${\cal C}(\tau_{1})$ can be deduced from ${\cal C}(\tau_{0})$ in $O(n)$. We proceed by successively setting $\tau_{k}=|w_{(n-k)}^{\ast}|$ until we reach a threshold $\tau_{k}$ such that ${\cal C}(\tau_{k})\le\kappa\phi^{\ast}$.
#### Simulation study.
We conducted a simple simulation study to evaluate our proposal and compare it to the KKT thresholding rule. Both methods were further compared to the results returned by the `glmnet` `R` package. The latter algorithm returns hard zeros in the classifier coefficients, and we have chosen the corresponding sparsity pattern as the “ground truth”, which the IPM should recover.
We first experimented with synthetic data. We generated samples of the pair $(X,y)$ for various values of $(m,n)$. We present the results for $(m,n)=(5000,2500)$ and $(m,n)=(100,500)$. The number $s$ of relevant features was set to $\min(m,n/2)$. Features were drawn from independent ${\mathcal{N}}(0,1)$ distributions and $y$ was computed as $y=X^{T}w+\xi$, where $\xi\sim{\mathcal{N}}(0,0.2)$ and $w$ is a vector of $\mathbb{R}^{n}$ with first $s$ components equal to $0.1+1/s$ and remaining $n-s$ components set to $0$. Because `glmnet` includes an unpenalized intercept while IPM method does not, both $y$ and $X$ were centered before applying either methods to make their results comparable.
Results are presented on Figures \[fig:thresholding\_Sim\_1\]. First, the KKT thresholding rule was observed to be very chaotic when the duality gap was set to $\epsilon=10^{-4}$ (we recall here that the default value for the duality gap in IPM `MATLAB` implementation is $\epsilon=10^{-3}$), while it was way better when duality gap was set to $\epsilon=10^{-8}$ (somehow justifying our choice of considering the sparsity pattern returned by `glmnet` as the ground truth). Therefore, for applications where computational time is not critical, running IPM method and applying KKT thresholding rule should yield appropriate results. However, when computational time matters, passing the duality gap from, say, $10^{-4}$ to $10^{-8}$, is not a viable option. Next, regarding our proposal, we observed that it was significantly better than KKT thresholding rule when the duality gap was set to $10^{-4}$ and equivalent to KKT thresholding rule for a duality gap of $10^{-8}$. Interestingly, setting $\alpha=1.5$ in (\[eq:new-thresholding-rule\]) generally enabled to achieved very good results for low values of $\lambda$, but lead to irregular results for higher values of $\lambda$ (in the case $m=100$, results were unstable for the whole range of $\lambda$ values we considered). Overall, the choices $\alpha=2$, $3$ and $4$ lead to acceptable results. A little irregularity remained with $\alpha=2$ for high values of $\lambda$, but this choice of $\alpha$ performed the best for lower values of $\lambda$. As for choices $\alpha=3$ and $\alpha=4$, it is noteworthy that the results were all the better as the dimension $n$ was low.
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --
![\[fig:thresholding\_Sim\_1\] Comparison of several thresholding rules on synthetic data: the case $m=5000$, $n=100$ ([*top panel*]{}) and $m=100$, $n=500$ ([*bottom panel*]{}) with duality gap in IPM method set to $(i)$ $10^{-4}$ ([*left panel*]{}) and $(iii)$ $10^{-8}$ ([*right panel*]{}). The curves represent the differences between the number of active features returned after each thresholding method and the one returned by `glmnet` (this difference is further divided by the total number of features $n$). The graphs present the results attached to six thresholding rules: the one proposed by [@boyd_logistic] and five versions of our proposal, corresponding to setting $\alpha$ in (\[eq:new-thresholding-rule\]) to $1.5$, $2$, $3$, $4$ and $5$ respectively. Overall, these results suggest that by setting $\alpha\in(2,5)$, our rule is less sensitive to the value of the duality gap parameter in IPM-LASSO than is the rule proposed by [@boyd_logistic].](Figures/Meth3pNbe_feat_kept_Comp_Glmnet_m5000_n2500_Tolb1e-004 "fig:"){width="48.00000%"} ![\[fig:thresholding\_Sim\_1\] Comparison of several thresholding rules on synthetic data: the case $m=5000$, $n=100$ ([*top panel*]{}) and $m=100$, $n=500$ ([*bottom panel*]{}) with duality gap in IPM method set to $(i)$ $10^{-4}$ ([*left panel*]{}) and $(iii)$ $10^{-8}$ ([*right panel*]{}). The curves represent the differences between the number of active features returned after each thresholding method and the one returned by `glmnet` (this difference is further divided by the total number of features $n$). The graphs present the results attached to six thresholding rules: the one proposed by [@boyd_logistic] and five versions of our proposal, corresponding to setting $\alpha$ in (\[eq:new-thresholding-rule\]) to $1.5$, $2$, $3$, $4$ and $5$ respectively. Overall, these results suggest that by setting $\alpha\in(2,5)$, our rule is less sensitive to the value of the duality gap parameter in IPM-LASSO than is the rule proposed by [@boyd_logistic].](Figures/Meth3pNbe_feat_kept_Comp_Glmnet_m5000_n2500_Tolb1e-008 "fig:"){width="48.00000%"}
![\[fig:thresholding\_Sim\_1\] Comparison of several thresholding rules on synthetic data: the case $m=5000$, $n=100$ ([*top panel*]{}) and $m=100$, $n=500$ ([*bottom panel*]{}) with duality gap in IPM method set to $(i)$ $10^{-4}$ ([*left panel*]{}) and $(iii)$ $10^{-8}$ ([*right panel*]{}). The curves represent the differences between the number of active features returned after each thresholding method and the one returned by `glmnet` (this difference is further divided by the total number of features $n$). The graphs present the results attached to six thresholding rules: the one proposed by [@boyd_logistic] and five versions of our proposal, corresponding to setting $\alpha$ in (\[eq:new-thresholding-rule\]) to $1.5$, $2$, $3$, $4$ and $5$ respectively. Overall, these results suggest that by setting $\alpha\in(2,5)$, our rule is less sensitive to the value of the duality gap parameter in IPM-LASSO than is the rule proposed by [@boyd_logistic].](Figures/Meth3pNbe_feat_kept_Comp_Glmnet_m100_n500_Tolb1e-004 "fig:"){width="48.00000%"} ![\[fig:thresholding\_Sim\_1\] Comparison of several thresholding rules on synthetic data: the case $m=5000$, $n=100$ ([*top panel*]{}) and $m=100$, $n=500$ ([*bottom panel*]{}) with duality gap in IPM method set to $(i)$ $10^{-4}$ ([*left panel*]{}) and $(iii)$ $10^{-8}$ ([*right panel*]{}). The curves represent the differences between the number of active features returned after each thresholding method and the one returned by `glmnet` (this difference is further divided by the total number of features $n$). The graphs present the results attached to six thresholding rules: the one proposed by [@boyd_logistic] and five versions of our proposal, corresponding to setting $\alpha$ in (\[eq:new-thresholding-rule\]) to $1.5$, $2$, $3$, $4$ and $5$ respectively. Overall, these results suggest that by setting $\alpha\in(2,5)$, our rule is less sensitive to the value of the duality gap parameter in IPM-LASSO than is the rule proposed by [@boyd_logistic].](Figures/Meth3pNbe_feat_kept_Comp_Glmnet_m100_n500_Tolb1e-008 "fig:"){width="48.00000%"}
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --
Real data examples
------------------
We also applied our proposal and compared it to KKT rule (\[thresholding\_Boyd\_rule\]) on real data sets arising in text classification. More precisely, we used the New York Times headlines data set presented in the Numerical results Section. For illustration, we present here results we obtained for the topic China and the year $1985$. We successively ran IPM-LASSO method with duality gap set to $10^{-4}$ and $10^{-8}$ and compare the number of active features returned after applying KKT thresholding rule (\[thresholding\_Boyd\_rule\]) and TR $(1.5)$, TR $(2)$, TR $(3)$ and TR $(4)$. Results are presented on Figure \[fig:thresholding\_NYT\_China\]. Because we could not applied `glmnet` on this data set, the ground truth was considered as the result of KKT rule, when applied to the model returned by IPM-LASSO ran with duality gap set to $10^{-10}$. Applying KKT rule on the model built with a duality gap of $10^{-4}$ lead to very misleading results again, especially for low values of $\lambda$. In this very high-dimensional setting ($n=38377$ here), our rule generally resulted in a slight underestimation of the true number of active features for the lowest values of $\lambda$ when the duality gap was set to $10^{-4}$. This suggests that the “optimal” $\alpha$ for our rule might depend on both $n$ and $\lambda$ when the duality gap is not small enough. However, we still observed that our proposal significantly improved upon KKT rule when the duality gap was set to $10^{-4}$.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --
![\[fig:thresholding\_NYT\_China\] Comparison of several thresholding rules on the NYT headlines data set for the topic China and year $1985$. Duality gap in IPM-LASSO was successively set to $10^{-4}$ ([*left panel*]{}) and $10^{-8}$ ([*right panel*]{}). The curves represent the differences between the number of active features returned after each thresholding method and the one returned by the KKT rule when duality gap was set to $10^{-10}$. The graphs present the results attached to five thresholding rules: the KKT rule and four versions of our rule, corresponding to setting $\alpha$ in (\[eq:new-thresholding-rule\]) to $1.5$, $2$, $3$ and $4$ respectively. Results obtained following our proposal appear to be less sensitive to the value of the duality gap used in IPM-LASSO. For instance, for the value $\lambda=\lambda_{{\rm max}}/1000$, the KKT rule returns $1758$ active feature when the duality gap is set to $10^{-4}$ while it returns $2357$ features for a duality gap of $10^{-8}$.](Figures/M3pNbe_feat_kept_172_1985_Tolb1e-004 "fig:"){width="48.00000%"} ![\[fig:thresholding\_NYT\_China\] Comparison of several thresholding rules on the NYT headlines data set for the topic China and year $1985$. Duality gap in IPM-LASSO was successively set to $10^{-4}$ ([*left panel*]{}) and $10^{-8}$ ([*right panel*]{}). The curves represent the differences between the number of active features returned after each thresholding method and the one returned by the KKT rule when duality gap was set to $10^{-10}$. The graphs present the results attached to five thresholding rules: the KKT rule and four versions of our rule, corresponding to setting $\alpha$ in (\[eq:new-thresholding-rule\]) to $1.5$, $2$, $3$ and $4$ respectively. Results obtained following our proposal appear to be less sensitive to the value of the duality gap used in IPM-LASSO. For instance, for the value $\lambda=\lambda_{{\rm max}}/1000$, the KKT rule returns $1758$ active feature when the duality gap is set to $10^{-4}$ while it returns $2357$ features for a duality gap of $10^{-8}$.](Figures/M3pNbe_feat_kept_172_1985_Tolb1e-008 "fig:"){width="48.00000%"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --
[^1]: In our experiments, we have used an Apple Mac Pro 64-bit workstation, with two $2.26$ GHz Quad-Core Intel Xeon processors, $8$ MB on-chip shared L3 cache per processor, with $6$ GB SDRAM, operating at $1066$ MHz.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The Siegel upper half space, $\mathcal{S}_n$, the space of complex symmetric matrices, $Z$ with positive definite imaginary part, is the generalization of the complex upper half plane in higher dimensions. In this paper, we study a generalization of linear fractional transformations, $\Phi_S$, where $S$ is a complex symplectic matrix, on the Siegel upper half space. We partially classify the complex symplectic matrices for which $\Phi_S(Z)$ is well defined. We also consider $\mathcal S_n$ and $\overline{\mathcal S}_n$ as metric spaces and discuss distance properties of the map $\Phi_S$ from $\mathcal S_n$ to $\mathcal{S}_n$ and $\overline{\mathcal S}_n$ respectively.'
address:
- 'Department of Mathematics, Embry-Riddle Aeronautical University, 600 S. Clyde Moris Blvd., Daytona Beach, FL 32114, U.S.A.'
- 'Department of Mathematics and Statistics, Mississippi State University, 175 President’s Cir., Mississippi State, MS 39762, U.S.A.'
author:
- Keshav Raj Acharya
- Matt McBride
title: Action of Complex Symplectic Matrices on the Siegel Upper Half Space
---
[^1]
Introduction
============
The Siegel upper half-space which we denote by $\mathcal{S}_n$ is the set of all $n\times n$ symmetric matrices $Z$ of the form: $Z=X+iY$ where $X$ and $Y$ are symmetric matrices and $Y$ is positive definite. When $n =1$, then $$\mathcal{S}_1 = {\mathbb{C}}^{+} =\{z\in {\mathbb{C}}: z= x+iy, \,\, y>0 \}\ .$$ Therefore the space $\mathcal{S}_n$ is a generalization of the complex upper half plane. For more on details on the Siegel space see [@Sie]. In this paper we study the action of symplectic group $SP_{2n}({\mathbb{C}})$ on $\mathcal S_n$.\
The symplectic group $SP_{2n}({\mathbb{C}})$ is the group of all $2n\times 2n$ complex matrices $S$ satisfying $S^tJS =J$ with
$$J=
\begin{pmatrix}
O & I_n \\ -I_n & O
\end{pmatrix}$$
where $I_n$ is the $n\times n$ identity matrix and $O$ is the $n\times n$ zero matrix.
When restriced to ${\mathbb{R}}$, the symplectic group $ SP_{2n}({\mathbb{R}})$ is a generalization of the group $SL_2({\mathbb{R}})$ to higher dimension. The action that we consider here has similarities with the action of $SL_2({\mathbb{R}})$ on the hyperbolic plane ${\mathbb{C}}^+$. A study of this action was done by Carl Ludwig Siegel in 1943 and published his work in the book “Symplectic Geometry" in which not only the analytical and geometrical aspects of the action are considered but also some applications to number theory. In 1999, Freitas gave more insight into this theory from the point of view of topological spaces, see [@F] for details. In particular he compactified the space and extended the action continuously to a compact domain. There are several studies on the Siegel upper half space motivated from other areas, for example [@Oh; @Jose]. However, the action of a complex symplectic matrix $SP_{2n}({\mathbb{C}})$ has yet to be studied. Here we consider complex symplectic matrix with non zero imaginary part.\
To relate the symplectic group to the Siegel upper half space, we write a symplectic matrix $S \in SP_{2n}({\mathbb{C}})$ in the form of block matrices $$S = \begin{pmatrix}
A & B \\ C & D
\end{pmatrix}$$ where $A, B, C, D $ are $ n\times n $ complex matrices. It can be easily seen by using the equation $S^tJS =J$ that $S$ is symplectic if and only if $A^tC$ and $B^tD$ are symmetric and $A^tD-C^tB= I_n$.\
For any $2n\times 2n$ complex block matrix $$S = \begin{pmatrix}
A & B \\ C & D
\end{pmatrix}$$ define the following matrix $\Phi_S$ by $$\Phi_S(Z) = S(Z)= (AZ+B)(CZ+D)^{-1}$$ where $Z$ is a $n\times n$ complex matrix. This map is a generalization of a fractional linear transformation on complex plane. This map is well defined only if $CZ+D$ is invertible.\
In [@F; @Sie], it is shown that if $S \in SP_{2n}({\mathbb{R}})$ and $Z \in \mathcal S_n$ then $\Phi_S $ is well defined and it maps the Siegel upper half space to itself. Since $SP_{2n}({\mathbb{R}})$ is a group, the map $ \Phi_S: \mathcal S_n \rightarrow \mathcal S_n $ has been considered as a group action of $SP_{2n}({\mathbb{R}})$ on $\mathcal S_n$. However, this is not true if $ S\in SP_{2n}({\mathbb{C}})$.
Let $S = \begin{pmatrix} I_n & iI_n\\ iI_n & O \end{pmatrix}$. Notice that $S\in SP_{2n}({\mathbb{C}})$ and $iI_n \in \mathcal S_n $. Therefore $$\Phi_S(iI_n)= (iI_n +iI_n)(iI_niI_n)^{-1} = -2iI_n\ .$$ Thus $\Phi_S(iI_n) \notin \mathcal S_n$ .
Let $A(2n, {\mathbb{C}}) $ be the space of all matrices $S$ that satisfy $S^tJS = -J$, that is $S$ is antisymplectic. Define the following space: $$\overline{\mathcal S}_n = \{ Z\in {\mathbb{C}}^{d\times d}: Z = X +iY, \,\, X^t=X, Y^t =Y, \,\, Y< 0 \}\ .$$
Following the methods of [@F], it is easy to see that if $S\in A(2n, {\mathbb{R}})$, $\Phi_S$ maps the Siegel half space to the space $\overline{\mathcal S}_n$. We will show this in Section 2. Since the product of antisymplectic matrices is a symplectic matrix, the set $A(2n, {\mathbb{C}})$ is not a group. Thus $\Phi_S $ is not a group action for such $S$.\
Again, if we consider $S$ to be antisymplectic complex matrix, $ \Phi_S $ does not map $\mathcal{S}_n$ to $\overline{\mathcal{S}}_n$.
Let $S = \begin{pmatrix} iI_n & O\\ O & iI_n \end{pmatrix} $ then $S\in A(2n, {\mathbb{C}})$ and $iI_n \in \mathcal S_n$. However $\Phi_S(iI_n) \notin \overline{\mathcal S}_n $ since $$\Phi_S(iI_n)= (iI_niI_n)(iI_n)^{-1} = iI_n \ .$$
One of the main goals in this paper is to partially classify all complex symplectic matrices for which the map $ \Phi_S: \mathcal S_n \rightarrow \mathcal S_n $ is well defined on $\mathcal S_n.$
We also want to consider a metric on $\mathcal S_n$ as a generalization of the hyperbolic metric
$$\label{hm}
ds = \frac{\sqrt{dx^2 +dy^2}}{y}$$
on the complex hyperbolic plane ${\mathbb{C}}^+.$ For this we consider the Finsler metric as in [@Froese] as follows.
Let $Z= X+iY \in \mathcal S_n $ and let the complex symmetric matrix $W$ be an element of the tangent space at $Z$. Consider the Finsler norm $F_Z$ defined by $$F_Z(W) = \| Y^{-\frac{1}{2}} W Y^{-\frac{1}{2}}\|$$ where $ \|\cdot\|$ is the usual matrix (operator) norm. This Finsler norm defines a metric on $\mathcal S_n$ as $$d_{\infty}(Z_1, Z_2) = \inf_{Z(t)} \int _0^1 F_{Z(t)}(\dot{Z}(t))\ dt$$ where the infimum is taken over all differentiable paths $Z(t)$ joining $Z_1$ to $Z_2.$
In the case when $n=1$, the Siegel upper half space $\mathcal S_1 = {\mathbb{C}}^+$ and the metric $d_{\infty}$ is same as the the hyperbolic metric defined in equation (\[hm\]) on ${\mathbb{C}}^+$. The length of a curve in the Euclidean plane is measured by $\sqrt{dx^2 +dy^2}$. Similarly, the length of a curve $\gamma(t) = x(t) +i y(t)$, $t\in [0,1]$ on the hyperbolic plane ${\mathbb{C}}^+$ is defined by
$$h(\gamma) = \int _0^1 \frac{1}{y(t)}\sqrt{\left(\frac{dx}{dt}\right)^2 + \left(\frac{dy}{dt}\right)^2}\ dt\ .$$
The hyperbolic distance between two points $z, w \in C^+$ is defined by
$$\rho(z,w) = \inf h(\gamma)$$
where the infimum is taken over all piecewise differentiable curves joining $z$ and $w$. So for $n=1$, we have $d_{\infty}(z, w)= \rho (z, w)$. Thus the Finsler metric $d_{\infty}$ on $\mathcal S_n$ is a generalization of the hyperbolic metric $\rho$ on ${\mathbb{C}}^+$.
We also want to define a metric on $\mathcal \overline{\mathcal S}_n$. First note that the complex conjugate of a given matrix $Z= X+iY \in\overline{\mathcal S}_n$ is a matrix $\overline{Z} = X-iY \in \mathcal S_n$. A metric $d_-$ on $ \overline{\mathcal S}_n $ is defined by $$d_-(Z_1, Z_2) = d_{\infty}(\overline{Z}_1, \overline{Z}_2 )\ .$$ Thus $(\mathcal S_n,d_{\infty})$ and $(\overline{\mathcal S}_n, d_-)$ are metric spaces.
Preliminaries and Results
=========================
As mentioned above, it is shown in [@F] that for each real symplectic matrix $S$, the map $\Phi_S$ maps the Siegel upper half space to itself. However, it is not true when $S$ is real antisymplectic matrix. In fact, it maps $\mathcal{S}_n$ onto $\overline{\mathcal{S}}_n$. We have the following proposition.
\[p\] For $S\in A(2n, {\mathbb{R}}) $ the map $\Phi_S $ maps $\mathcal S_n$ onto $\overline{\mathcal S}_n$.
We show that $CZ+D$ is invertible for any $Z \in \mathcal S_n $ and $\Phi_S(Z)=(AZ+B)(CZ+D)^{-1} \in \overline{\mathcal S}_n .$\
Set $E=AZ+B$ and $F=CZ+D$. First we show that $F$ is invertible. We can write $\Phi_S(Z)$ as
$$S \begin{pmatrix} Z\\I\end{pmatrix} = \begin{pmatrix} E\\F\end{pmatrix}\ .$$
From which we have,
$$\label{lab2}
\begin{aligned}
\frac{1}{2i}(E^*\,\, F^*) J \begin{pmatrix} E\\F\end{pmatrix} & = \frac{1}{2i}(Z^*\,\, I_n)S^tJS \begin{pmatrix} Z\\I_n\end{pmatrix} = \frac{1}{2i}(Z^*\,\, I_n) (-J)\begin{pmatrix} Z\\I_n\end{pmatrix} \\
&= \frac{1}{2i}(Z^*\,\, I_n) \begin{pmatrix} O & -I_n\\ I_n & O\end{pmatrix} \begin{pmatrix} Z\\I_n\end{pmatrix} \\
&=\frac{1}{2i}(Z-Z^*) > 0\ .
\end{aligned}$$
It follows that,
$$\label{lab4}
\frac{1}{2i}(E^*F-F^*E)>0 \ .$$
To see $F$ invertible, suppose $v$ is a solution of $Fv=0$ then we have $v^*F^*= 0$ and
$$v^*(E^*F-F^*E)v=0\ .$$
From equation (\[lab4\]), we have $v=0$ which implies $F$ invertible.
Next we show that $\Phi_S (Z)$ is symmetric and $\operatorname{Im} \Phi_S (Z) <0$. We have
$$\begin{aligned}
(E^t\,\, F^t) J \begin{pmatrix} E\\F\end{pmatrix} & = (Z^t\,\, I_n)S^tJS \begin{pmatrix} Z\\I_n\end{pmatrix} = (Z^t\,\, I_n) (-J)\begin{pmatrix} Z\\I_n\end{pmatrix} \\
&= Z-Z^t = O\ .
\end{aligned}$$
This implies that $F^tE= E^tF$. From this we get,
$$(EF^{-1})^t = (F^{-1})^tE^t =(F^{-1})^t F^t E F^{-1} = EF^{-1}$$
which means $\Phi_S (Z) $ is symmetric. Moreover we have,
$$\begin{aligned}
\operatorname{Im} \Phi_S (Z)= \operatorname{Im} (EF^{-1}) & = \frac{1}{2i}(EF^{-1}- (EF^{-1})^*) \\ &= \frac{1}{2i} (F^{-1})^*F^*(EF^{-1}- (EF^{-1})^*)F F^{-1} \\ &= \frac{1}{2i} (F^{-1})^*(F^*E- E^*F) F^{-1} \ .
\end{aligned}$$
Thus $\frac{1}{2i}(F^*E- E^*F) < 0$. Now for any $v \in {\mathbb{C}}^n$ and $F^{-1}v\in {\mathbb{C}}^n$ we have
$$\begin{aligned}
&v^*\frac{1}{2i} (F^{-1})^*F^*(EF^{-1}- (EF^{-1})^*)F F^{-1}v \\
&=(F^{-1}v)^* \frac{1}{2i} (F^*(EF^{-1}- (EF^{-1})^*)F) F^{-1}v \\
&=(F^{-1}v)^* \frac{1}{2i} (F^*E- E^*F) (F^{-1}v) < 0 \ .
\end{aligned}$$
It follows that $ \operatorname{Im} (EF^{-1}) < 0$. Hence $\Phi_S$ maps $\mathcal S_n$ to $\overline{\mathcal S}_n$.
Finally we show the map $\Phi_S $ is onto. Suppose $Z= X+iY \in \overline{\mathcal S}_n.$ We have
$${S_Z}_- = \begin{pmatrix} - \sqrt{-Y} & X \sqrt{-Y^{-1}}\\ O & \sqrt{-Y^{-1}} \end{pmatrix} \in A(2n, {\mathbb{R}})$$
and $\Phi_{{S_Z}_-}(iI_n) = Z$. $-Y$ is a symmetric positive definite matrix, therefore the Spectral Theorem guarantees that $\sqrt{-Y}$ and $\sqrt{-Y^{-1}}$ exist. This completes the proof.
As a consequence of this proposition we have the following theorem.
If $S\in SP_{2n}({\mathbb{C}})$ is purely imaginary then $\Phi_S$ maps Siegel half space to the space $\overline{\mathcal S}_n$.
Let $S = iQ$, where
$$Q = \begin{pmatrix} A & B \\ C & D \end{pmatrix}\ .$$
Since $S^tJS = J$, we have $Q^tJQ = -J$, hence $Q \in A(2n,{\mathbb{R}})$. Moreover we have
$$\begin{aligned}
\Phi_{iQ}(Z) & = (iAZ+iB)(iCZ+iD)^{-1}\\
& =(AZ+B)(CZ+D)^{-1} \\
& =\Phi_{Q}(Z)\ .
\end{aligned}$$
Since $Q \in A(2n,{\mathbb{R}})$, by Proposition \[p\] $\Phi_{Q}$ maps $\mathcal S_n$ to $\overline{\mathcal S}_n$ and hence $\Phi_{iQ}$ maps $\mathcal{S}_n$ to $\overline{\mathcal{S}}_n$. Thus completing the proof.
In the following theorem we partially classify the matrices from $ SP_{2n}({\mathbb{C}})$ so that the action is well defined.
Let $$S = \begin{pmatrix} A & B\\ C & D \end{pmatrix} \in SP_{2n} ({\mathbb{C}})$$ and $ Z \in \mathcal S_n$. If $i(S^*JS-J)\ge 0$ then $\Phi_S(Z) = S(Z)= (AZ+B)(CZ+D)^{-1} \in \mathcal S_n$.
For convenience set $E = AZ+B$ and $F=CZ+D$. Suppose $i(S^*JS -J) \ge 0 $. For any $ Z \in \mathcal S_n$ we have
$$(Z^*\,\, I_n) \left(i(S^*JS-J) \right) \begin{pmatrix} Z \\ I_n \end{pmatrix} \ge 0\ .$$
On the other hand
$$\begin{aligned}
&(Z^*\,\, I_n) \left(i(S^*JS-J) \right) \begin{pmatrix} Z \\ I_n \end{pmatrix}\\
&=i\left(\left(S \begin{pmatrix} Z \\ I_n \end{pmatrix} \right)^*\left(S\begin{pmatrix} Z \\ I_n \end{pmatrix}\right) - (Z^*\,\, I_n)J\begin{pmatrix} Z \\ I_n \end{pmatrix}\right) \\
&= i(E^*\,\, F^*)J\begin{pmatrix} F \\ -E\end{pmatrix} - i(Z^*\,\, I_n)\begin{pmatrix} I_n \\ -Z\end{pmatrix} \\
&=i(E^*F - F^*E) - \frac{1}{i}(Z - Z^*)\ .
\end{aligned}$$
Therefore we have $i(E^*F - F^*E) +i(Z - Z^*) \ge 0$ and thus
$$\frac{1}{i}(F^*E - E^*F) \ge \frac{1}{i}(Z-Z^*) = 2\textrm{Im}(Z)>0$$
since $Z\in\mathcal{S}_n$. To show that $S(Z)\in\mathcal{S}_n$ we need to show that $F$ is invertible, $S(Z)$ is symmetric, and $\textrm{Im}(S(Z))>0$. First we show that $F$ is invertible. Suppose $Fv = 0$ for $v\in {\mathbb{C}}^n$, then it follows that $v^*(F^*E-E^*F)v/{2i} = 0$. Since $-i(F^*E - E^*F)>0$, we have $v=0$. Hence $F^{-1}$ exists and $EF^{-1}$ is well-defined.
Next we show symmetry. Notice that since $S$ is a symplectic matrix and $Z\in\mathcal{S}_n$, we have
$$\begin{aligned}
E^tF-F^tE &= \begin{pmatrix} Z \\ I_n \end{pmatrix}^tS^tJS\begin{pmatrix} Z \\ I_n \end{pmatrix} = \begin{pmatrix} Z \\ I_n \end{pmatrix}^tJ\begin{pmatrix} Z \\ I_n \end{pmatrix} = Z^t - Z = O\ .
\end{aligned}$$
Therefore $E^tF = F^tE$ and thus by the invertibility of $F$ we have
$$(EF^{-1})^t = (F^{-1})^tE^t = (F^{-1})^tF^tEF^{-1} = EF^{-1}\ ,$$
thus showing the symmetry of $S(Z)$. Finally a similar calculation to the one at the beginning of the proof shows that
$$\textrm{Im}(S(Z)) = (F^{-1})^*\frac{1}{2i}\left(F^*E - E^*F\right)F^{-1}$$
and since $-i(F^*E - E^*F)>0$, it follows that $\textrm{Im}(S(Z))>0$. This completes the proof.
It should be noted the other direction may or may not be true. One of the many issues in showing the other direction is going from $n\times n$ complex matrices to $2n\times 2n$ complex matrices. The direction proven is simpler since the dimensions are going down. In fact, one can give necessary and sufficient conditions on a self-adjoint block matrix when it is positive definite. In fact [@BV] stated without proof, and [@Gallier] proved in unpublished work, conditions on when a symmetric block matrix is positive definite. This proof can be easily modified to allow for generic self-adjoint block matrices. We state it without proof.
Let $M$ be a $2n\times 2n$ self-adjoint block matrix of the form: $$M = \begin{pmatrix} \alpha & \beta \\ \beta^* & \gamma \end{pmatrix}\ .$$ The following are equivalent:
1. $M\ge 0$
2. $\alpha\ge 0$,$(I_n - \alpha\alpha^\dag)\beta = O$,and $\gamma - \beta^*\alpha^\dag\beta \ge 0$
3. $\gamma \ge 0 $,$(I_n - \gamma\gamma^\dag)\beta^* = O$,and $\alpha - \beta\gamma^\dag\beta^* \ge 0$
where $P^\dag$ is the pseudo-inverse of $P$.
For more details on pseudo-inverses of square matrices see [@Gallier1]. Let $M = i(S^*JS-J)$, then it’s clear that $M$ is self-adjoint. A computation shows that
$$M = \begin{pmatrix}
i(A^*C-C^*A) & i(A^*D-C^*B-I_n) \\ i(B^*C-D^*A+I_n) & i(B^*D-D^*B)
\end{pmatrix} :=\begin{pmatrix} \alpha & \beta \\ \beta^* & \gamma\end{pmatrix}$$
Based on the discussion from above, we know since $S$ is symplectic, then $A^tC$ and $B^tD$ are symmetric and $A^tD-C^tB= I_n$. Moreover in this case, the entries of $S$ can not be all real, otherwise, we are the case already proven by [@F]. Based on this proposition, for $M\ge 0$, we must have
$$\begin{aligned}
&(1)\quad i(A^*C-C^*A)\ge 0 \\
&(2)\quad [I_n + (A^*C-C^*A)(A^*C-C^*A)^\dag](A^*D-C^*B-I_n) = O \\
&(3)\quad i(B^*D-D^*B) + \\
&\qquad +i(B^*C-D^*A+I_n)(A^*C-C^*A)^\dag(A^*D-C^*B-I_n) \ge 0\ .
\end{aligned}$$
Whether or not these conditions are equivalent to $i(S^*JS- J)\ge 0$ remains unknown.
The set of all matrices $S\in SP_{2n}({\mathbb{C}})$ for which the action is well defined is a subgroup of $SP_{2n}({\mathbb{C}})$. This was stated in [@Froese] without proof, therefore we provide a proof.
\[r1\] If $S$ and $R$ are any $2n\times 2n$ complex matrices then $\Phi_S \circ \Phi_{R} = \Phi_{SR}$
Let
$$S = \begin{pmatrix} A & B \\ C & D \end{pmatrix} \textrm{ and } R = \begin{pmatrix} E & F \\ G & H \end{pmatrix}\ .$$
For any $Z \in \mathcal S_n,$ for which $\Phi_S(Z)$ and $\Phi_R (Z) $ are well defined, we have
$$\begin{aligned}
\Phi_S \circ\Phi_R (Z) & = \Phi_S (\Phi_R (Z)) = \Phi_S ((EZ+F)(GZ+H)^{-1}) \\
& = \left(A(EZ+F)(GZ+H)^{-1} + B\right)\\
&\qquad\times \left(C(EZ+F)(GZ+H)^{-1} +D \right)^{-1} \\
&= \left(A(EZ+F)(GZ+H)^{-1} +B\right)(GZ+H) \\
&\qquad\times (GZ+H)^{-1}\left(C(EZ+F)(GZ+H)^{-1} +D \right)^{-1} \\
& = \left(A(EZ+F)+B(GZ+H)\right)\\
&\qquad\times\left(C(EZ+F) +D(GZ+H)\right)^{-1}\\
&= \left((AE+BG)Z+(AF+BH)\right)\\
&\qquad\times\left((CE+DG)Z+ (CF+DH))\right)^{-1} \\
& = \Phi_{SR}(Z)\ .
\end{aligned}$$
This works for an arbitrary $Z$, thus completing the proof.
This next proposition was due to [@F].
The action of $\Phi_S$ on $\mathcal S_n$ is transitive.
We show that $\mathcal S_n$ has a single orbit. In other words, every $Z \in \mathcal S_n$ is an image of some $S\in SP_{2n}({\mathbb{C}})$. Let $Z= X+iY \in \mathcal S_n$. Define $S_Z$ by
$$S_Z = \begin{pmatrix} \sqrt{Y} & X \sqrt{Y^{-1}}\\ O & \sqrt{Y^{-1}} \end{pmatrix}\ .$$
Recall $Y$ is a symmetric positive definite matrix, therefore the Spectral Theorem guarantees that $\sqrt{Y}$ and $\sqrt{Y^{-1}}$ exist. It is easy to verify that $S_Z^tJS_Z = J$. Therefore, $S_Z\in SP_{2n}({\mathbb{C}})$. Now consider $S_Z(iI_n)$. Indeed we have
$$S_Z(iI_n) = (\sqrt{Y} iI_n + X \sqrt{Y^{-1}} ) (\sqrt{Y^{-1}})^{-1} = X+iY =Z\,$$
thus showing $\mathcal{S}_n$ has a single orbit completing the proof.
Since the action of $$I_{2n} := \begin{pmatrix} I_n & O\\ O & I_n \end{pmatrix}$$ is same as the action of $$-I_{2n} = \begin{pmatrix} -I_n & O\\ O & -I_n \end{pmatrix}$$ we can identify them so that we can form the quotient space
$$SP_{2n}({\mathbb{R}})/\{I_{2n}, -I_{2n}\}$$
again denoted by $SP_{2n}({\mathbb{R}}).$ This makes the action of $SP_{2n}({\mathbb{R}})$ on $\mathcal S_n$ bijective. The following proposition was stated in [@Froese] again without proof. Like before we present a proof for completeness of this paper.
The set of matrices $$K= \{U \in SP_{2n}({\mathbb{R}}): \Phi_U(iI_n) = iI_n\}$$ is precisely the subgroup of orthogonal matrices in $SP_{2n}({\mathbb{R}})$.
Suppose $$U = \begin{pmatrix} A & B \\ C & D \end{pmatrix} \in SP_{2n}({\mathbb{R}})$$ such that $\Phi_U(iI_n) = iI_n$ then $(A iI_n +B)(C iI_n +D)^{-1} = iI_n$. This implies that $iA+B = iI_n (iC+D)$, therefore it follows that $B= -C$ and $A=D$. Hence $ U$ and $U^t$ are of the forms
$$U= \begin{pmatrix} A & B \\ -B & A \end{pmatrix} \textrm{ and }U^t = \begin{pmatrix} A^t & -B^t \\ B^t & A^t \end{pmatrix}\ .$$
Thus we have
$$\begin{aligned}
UU^t & = \begin{pmatrix} A & B \\ -B & A \end{pmatrix} \begin{pmatrix} A^t & -B^t \\ B^t & A^t \end{pmatrix} \\
&= \begin{pmatrix} AA^t +BB^t & -AB^t+BA^t \\ -BA^t+AB^t & BB^t +AA^t \end{pmatrix} = I_{2n}
\end{aligned}$$
since $AA^t +BB^t = I_n$ and $-AB^t$ is symmetric. This shows that $ U^t= U^{-1}$, hence $U$ is orthogonal. Moreover if
$$U = \begin{pmatrix} A & B \\ -B & A \end{pmatrix}\in K\ ,$$
then $U^{-1} \in K$. This is true since
$$\begin{aligned}
\Phi_{U^{-1}}(iI_n) & = \begin{pmatrix} A^t & -B^t \\ B^t & A^t \end{pmatrix}(iI_n) = \left( A^t (iI_n) -B^t \right) \left( B^t (iI_n) + A^t \right)^{-1} \\
&= ( Ai -B )^t \left[( B(iI_n) + A)^t \right]^{-1} = \left[(Bi + A)^{-1}(Ai -B)\right]^t \\
&= \left[i(Bi + A)^{-1}(A + Bi) \right]^t = iI_n\ .
\end{aligned}$$
In addition, if $U$ and $V$ are both in $K$ then by Proposition \[r1\] we have $\Phi_{UV^{-1}}(iI_n) =\Phi_U(\Phi_{V^{-1}}(iI_n)) = \Phi_U(iI_n)) = iI_n$. This shows that $K$ is a subgroup. Thus the result follows.
Since $K$ is the subgroup of orthogonal matrices in $SP_{2n}({\mathbb{R}})$ general theory tells us that $K$ is a normal subgroup. Therefore we can consider the quotient space $$SP_{2n}({\mathbb{R}})/K = \{SK : S\in SP_{2n}({\mathbb{R}})\}$$ where $SK= \{SU : U \in K\}$ is a coset of $K$. As shown in [@F], this space is metrizable with a metric given by
$$d^{S}(S_1K, S_2K) = 2 \ln \| S_1^{-1}S_2\|\ .$$
Notice that the metric is independent of the choice of representative in the equivalence class $SK$.
Define a map $\Psi$ in the following way
$$\label{psi_isometry}
\Psi: SP_{2n}({\mathbb{R}})/K\rightarrow \mathcal{S}_n \textrm{ via } SK\mapsto \Phi_S(iI_n)\ .$$
Since $\Phi_S$, when acting $SP_{2n}({\mathbb{R}})$, is a bijective map, the map $\Psi$ is also bijective. The following theorem was proved in [@Froese].
The map $\Psi$, defined in equation (\[psi\_isometry\]), is an isometry for the metrics $d^{S}$ and $d_{\infty}$, that is
$$d^{S}(S_1 K, S_2 K) = d_{\infty}(S_1(iI_n), S_2(iI_n))\ .$$
In particular,
$$d_{\infty}(Z_1, Z_2) = 2 \ln \|{S_{Z_1}} ^{-1} S_{Z_2}\|\ .$$
We skip the proof for brevity. Based on this result we have the following theorem.
If $S\in SP_{2n}({\mathbb{C}}) $ is purely imaginary then the map $\Phi_S$ is an isometry between the metric spaces $(\mathcal S_n,d_{\infty})$ and $(\overline{\mathcal S}_n, d_-)$.
Let $S = iQ, $ where
$$Q = \begin{pmatrix} A & B \\ C & D \end{pmatrix}\ .$$
Since $S^tJS = J$ we have $Q^tJQ = -J$, hence $Q \in A(2n, {\mathbb{R}})$. For any $Z= X+iY \in \mathcal S_n$ we have $\Phi_{S_Z}(iI_n)= S_Z(iI_n)= Z$. Let
$${S_Z}_- = \begin{pmatrix} - \sqrt{Y} & X \sqrt{Y^{-1}}\\ O & \sqrt{Y^{-1}} \end{pmatrix} \textrm{ and } I_- = \begin{pmatrix} -I_n & O\\ O & I_n \end{pmatrix}\ .$$
Then ${S_Z}_- = S_Z I_-$ and ${S_Z}_-^tJ{S_Z}_- = -J$ implying that $ {S_Z}_- \in A(2n,{\mathbb{R}})$. Since both $Q$ and ${S_Z}_-$ are in $A(2n, {\mathbb{R}})$ this implies that $Q{S_Z}_-$ is in $SP_{2n}({\mathbb{R}})$. Notice that $ {S_Z}_-(iI_n)= \overline{Z}$. Therefore we have
$$\begin{aligned}
d_-(\Phi_S(Z), \Phi_S(W)) & = d_-(\Phi_{iQ}(Z),\Phi_{iQ}(W) = d_-(\Phi_Q(Z),\Phi_Q (W)) \\
& = d_{\infty}(\overline{\Phi_Q (Z)}, \overline{\Phi_Q (W)}) = d_{\infty}(\Phi_Q(\overline{Z}),\Phi_Q (\overline{W}))\\
& = d_{\infty}(Q{S_Z}_-(iI_n), Q{S_W}_-(iI_n))\\
&= 2\ln\|(Q {S_Z}_-)^{-1} Q{S_W}_- \| = 2\ln\|{S_Z}_-^{-1}{S_W}_- \|\\
&= 2\ln\|I_- {S_Z} ^{-1}{S_W} I_- \| = 2\ln\|{S_Z} ^{-1} {S_W}\| \\
&=d_{\infty}(Z, W)
\end{aligned}$$
since $I_-$ is a unitary. Thus this completes the proof.
Unfortunately in general, the map $\Phi_S$ is not an isometry. In fact we have the following theorem.
If the matrix $S$ is simplectically similar to
$$R = \begin{pmatrix} I_n & Z\\ I_n & 0 \end{pmatrix} \in SP_{2n} ({\mathbb{C}})$$
where $ Z\in \mathcal S_n$, that is $S = PRP^{-1} $ for some $P\in SP_{2n} ({\mathbb{R}}) $ then there exists a $\delta$ with $0<\delta<1$ such that for all $Z_1,Z_2\in\mathcal{S}_n$,
$$d_{\infty}(S(Z_1), S(Z_2)) \leq \delta d_{\infty}(Z_1, Z_2)\ .$$
Claim: for $Z= X+iY, W=W_1 +i W_2 \in \mathcal S_n$
$$F_{W+Z}(\gamma ) < \delta F_W(\gamma)$$
for some $0 < \delta < 1$. Indeed we have
$$\begin{aligned}
F_{W+Z}(\gamma) & = F_{(W_1 + X )+i(W_2+Y)}(\gamma ) = \| (W_2+Y)^{-\frac{1}{2}} \gamma (W_2+Y)^{-\frac{1}{2}}\| \\
& = \| (W_2+Y)^{-\frac{1}{2}} W_2^{\frac{1}{2}}W_2^{-\frac{1}{2}}\gamma W_2^{-\frac{1}{2}}W_2^{\frac{1}{2}}(W_2+Y)^{-\frac{1}{2}} \| \\
&\le \|(W_2+Y)^{-\frac{1}{2}} W_2^{\frac{1}{2}} \|^2 \| (W_2^{-\frac{1}{2}} \gamma W_2^{-\frac{1}{2}}\|\ .
\end{aligned}$$
Since $Y>0$, we have $W_2^{\frac{1}{2}} < (W_2+Y)^{\frac{1}{2}}$ and $ (W_2+Y)^{-\frac{1}{2}}>0$. It follows
$$(W_2+Y)^{-\frac{1}{2}} W_2^{\frac{1}{2}} < (W_2+Y)^{-\frac{1}{2}}(W_2+Y)^{\frac{1}{2}} = I_n\ .$$
Therefore we have $\|(W_2+Y)^{-\frac{1}{2}} W_2^{\frac{1}{2}}\| < 1$. Thus we can choose
$$\delta = \| (W_2+Y)^{-\frac{1}{2}} W_2^{\frac{1}{2}}\|$$
proving the claim. Now let $\gamma(t)$ be a path joining $Z_1, Z_2$ so that $\gamma(t)+Z$ is a path joining $Z_1+Z, Z_2+Z$ and
$$\int_0^1 F_{{\gamma(t)}+Z}(\dot{\gamma}(t))\ dt<\delta\int_0^1 F_{\gamma(t)}(\dot{\gamma}(t))\ dt\ .$$
Taking the infimum over all the paths $\gamma(t)$ joining $Z_1$ and $Z_2$, we get
$$\int_0^1 F_{\gamma(t)+Z}(\dot{\gamma}(t))\ dt< \delta d_{\infty}(Z_1, Z_2)\ .$$
Again taking the infimum over all the paths $\gamma(t)+Z$ joining $Z_1 + Z$ and $Z_2 +Z$, we get
$$d_{\infty}(R(Z_1), R(Z_2)) \le \delta d_{\infty}(Z_1, Z_2)\ .$$
Finally we get
$$\begin{aligned}
d_{\infty}(S(Z_1), S(Z_2)) &= d_{\infty}(PRP^{-1}(Z_1), PRP^{-1}(Z_2))\\
&= d_{\infty}(RP^{-1}(Z_1), RP^{-1}(Z_2)) \\
&\le \delta d_{\infty}(P^{-1}(Z_1), P^{-1}(Z_2)) \\
&= \delta d_{\infty}(Z_1, Z_2 )\ ,
\end{aligned}$$
completing the proof.
Recall that the Finsler metric $d_\infty(\cdot,\cdot)$ becomes the hyperbolic metric when $n=1$. Therefore we can compare distances of numbers in the upper half plane and corresponding counterparts in the Siegel space. We end the paper with the following result.
Let $\mathbb{B}(0)$ be the unit ball in ${\mathbb{C}}^n$. For any $Z_1, Z_2 \in \mathcal S_n$ and $v\in \mathbb B(0)$,
$$d_{\infty}(v^*Z_1v, v^*Z_2v) \leq d_{\infty}(Z_1, Z_2)\ .$$
If $Z(t)$ is a path joining $Z_1$ and $Z_2$ then $v^*Z(t)v$ is a path joining $v^*Z_1v$ and $v^*Z_2v$ with $v\in \mathbb B(0).$ For $Z(t) = X(t)+iY(t)$ and $v \in \mathbb B(0)$ we have
$$\begin{aligned}
\operatorname{Im}(v^*Z(t)v) & = \frac{1}{2i} (v^*Z(t)v - (v^*Z(t)v)^*)
\\ & =\frac{1}{2i} v^*(Z(t) - Z(t)^*) v^* \\ & = v^*Y(t) v^*
\end{aligned}$$
with $v^*Y(t)v >0$. Recall that for a $n\times n$ matrix $A$ and a vector $x$ we have the following submultiplicative estimate: $\|Ax\|\le \|A\|\|x\|$. Using this, the Cauchy-Schwarz inequality, and the fact that the Finsler metric becomes the hyperbolic metric in one dimension we get
$$\begin{aligned}
F_{v^* Z(t)v}(v^*\dot{Z}(t) v) & = \| (v^*Yv)^{-\frac{1}{2}} (v^* \dot{Z}(t)v) (v^*Yv)^{-\frac{1}{2}} \| \\
& =\frac{1}{v^*Yv} \| v^*Y^{\frac{1}{2}} Y^{-\frac{1}{2}} \dot{Z}(t)Y^{-\frac{1}{2}} Y^{\frac{1}{2}}v\|\\
& =\frac{1}{v^*Yv} \langle Y^{\frac{1}{2}}v, Y^{-\frac{1}{2}}\dot{Z}(t)Y^{-\frac{1}{2}} Y^{\frac{1}{2}}v\rangle \\
& \leq \frac{1}{v^*Yv} \|Y^{\frac{1}{2}}v \|^2 \|Y^{-\frac{1}{2}}\dot{Z}(t)Y^{-\frac{1}{2}}\| \\
&= \|Y^{-\frac{1}{2}}\dot{Z}(t)Y^{-\frac{1}{2}}\| = F_{Z(t)}(\dot{Z}(t))\ .
\end{aligned}$$
Upon integration we get,
$$\int_0^1 F_{v^*Z(t)v}(v^*\dot{Z}(t)v)\ dt \le \int_0^1 F_{Z(t)}(\dot{Z}(t))\ dt\ .$$
As before, taking the infimum over all the paths $Z(t)$ joining $Z_1$ and $Z_2$, we get
$$\int_0^1 F_{v^*Z(t)v}(v^*\dot{Z}(t)v)\ dt \le d_{\infty}(Z_1, Z_2)\ .$$
Finally taking the infimum over all the paths $v^*Z(t)v$ joining $v^*Z_1v$ and $v^*Z_2v$, we get
$$d_{\infty}(v^*Z_1v, v^*Z_2v) \leq d_{\infty}(Z_1, Z_2)\ .$$
This completes the proof.
[99]{}
Boyd, S. and Vandenberghe, L., [*Convex Optimization*]{}, Cambridge University Press, 2004.
Froese, R., Hasler, D. and Spitzer, W., Transfer Matrices, Hyperbolic Geometry and Absolutely Continuous Spectrum for Some Discrete Schrödinger Operators on Graphs, [*Jour. Func. Anal.*]{}, 230(1), 184-221, 2006.
Freitas, P. J., On the Action of the Symplectic Group on the Siegel Upper Half Plane, PhD thesis, University of Illinois at Chicago, 1999.
Gallier, J., The Schur Complement and Symmetric Positive Semidefinite (and Definite) Matrices, unpublished work.
Gallier, J. [*Geometric Methods and Applications, For Computer Science and Engineering*]{}, TAM, Vol. 38, Springer, 2000.
Ohsawa, T., The Siegel Upper Half Space is a Marsden-Weinstein Quotient: Symplectic Reduction and Gaussian Wave Packets, [*Lett. Math. Phys.*]{}, 105(9), 1301-1320, 2015.
Pontoja, J., Andrade, J. S. and Vargas, J. A., On the Construction of a Finite Siegel Space, [*Jour. of Lie Theo.*]{}, 25, 1045-1071, 2015.
Siegel, C. L., [*Symplectic Geometry*]{}, Academic Press. New York and London, 1964.
[^1]:
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The reduced dynamics of two interacting qubits coupled to two independent bosonic baths is investigated. The one-excitation dynamics is derived and compared with that based on the resolution of appropriate non-Markovian master equations. The Nakajima-Zwanzig and the time-convolutionless projection operator techniques are exploited to provide a description of the non-Markovian features of the dynamics of the two-qubits system. The validity of such approximate methods and their range of validity in correspondence to different choices of the parameters describing the system are brought to light.'
author:
- 'E. Ferraro'
- 'M. Scala'
- 'R. Migliore'
- 'A. Napoli'
title: |
Non-Markovian dissipative dynamics of two coupled qubits in independent reservoirs:\
a comparison between exact solutions and master equation approaches
---
Introduction
============
Despite its simplicity a two-state system is of great significance being it exploitable to effectively describe many real situations. The theoretical analysis as well as the practical implementation of interacting or not two-level systems thus represents a central topic in several branches of modern physics ranging from high energy to nuclear and condensed matter physics [@polyakov; @belitsky]. During the last decade the interest towards two-level systems has further been stimulated by the fact that a qubit represents the basic element in the context of the new applicative area of quantum information and communication. It has been for example shown that a fundamental quantum gate like the C-NOT gate can be implemented using dipole-dipole interacting quantum dots modeled as two qubits [@baranco]. Moreover due to the development of new technologies, today there are several possible routes to the creation of what might be termed quantum bit, each based on a different physical system. These include quantum optics, microscopic quantum objects (electrons, ions, atoms) in traps, quantum dots and quantum circuits [@amico; @6; @7; @8; @9; @10; @bellomo; @garraway1; @garraway2; @garraway4]. In describing real systems however it is mandatory to take into account the effects stemming from the presence of the surroundings. Thus the dissipative dynamics of two-level systems has been the subject of numerous papers appeared in literature in the last decades [@agarwal; @sinaysky; @quiroga; @maniscalco; @Mazzola; @yu; @scala]. Generally speaking the research has been developed assuming a Markovian environment [@agarwal; @sinaysky]. But memory effects are in general present and could affect quantitatively and qualitatively the dynamics of the small system. Unfortunately, there are no fully systematic investigations of non-Markovian environments. Projection operator techniques, such as the time-convolutionless (TCL) [@Chaturvedi] and the Nakajima-Zwanzig (NZ) [@Nakajima; @Zwanzig] approaches, are in general exploited in order to perform a description of the non-Markovian features of the dynamics of open systems. On the one hand, the NZ provides a generalized master equation in which the time derivative of the density operator is connected to the past history of the state through the convolution of the density operator and an appropriate integral kernel. On the other hand the TCL approach provides a generalized master equation which is local in time. Intuitively, one might argue that the NZ should work better than the TCL approach in describing the memory effect, since it explicitly takes into account the past history of the open system. Anyway there are examples in which the exact dynamics of the open system can be described by means of a master equation which is local in time, as in the well known case of the Hu-Paz-Zhang generalized master equation for the non-Markovian theory of quantum Brownian motion [@hu; @petruccionebook].
In general it is not easy to establish whether one method is better than the other one. In fact, the performance of these perturbation schemes strongly depends on the details of the system under investigation. In this paper we consider two interacting two-level systems, each coupled to its own bosonic bath, exactly solving their dynamics in a one excitation subspace. The knowledge of the exact dynamics is exploited to test perturbative approaches based on TCL and NZ techniques. We show that, counterintuitively, the TCL approach works better than the NZ one, since the latter approach does not guarantee the positivity of the density matrix when the correlations inside the reservoir become moderately strong. On the contrary the TCL approach describes all the qualitative features of non-Markovian dynamics for a wider range of values of reservoir memory time.
The paper is organized as follows. In Sec. II we derive the exact equations governing the evolution of the two-qubit system coupled to two independent reservoirs, in the case of one initial excitation, and find their exact solution. In Sec. III the derivation of the second order NZ equation is presented and the features of the second order TCL, derived in Ref.[@Ferraro], are recalled. An extensive comparison among the exact, the NZ and the TCL is presented in Sec. IV, while in Sec. V some conclusive remarks are given.
Exact dynamics
==============
The model
---------
The physical system on which we focus our attention is composed by two interacting two-level systems. Each qubit is moreover coupled to an external environment modeled as a bosonic bath [@quiroga]. Assuming $\hbar=1$, the Hamiltonian model describing the total system can be written in the following form $$\label{Htot}
H=H_0+H_I.$$ Here $$\begin{aligned}
\label{H0}
H_0=\frac{\omega_0}{2}\sigma_z^{(1)}+\frac{\omega_0}{2}\sigma_z^{(2)}+
\sum_{j=1,2}\sum_k\omega_k^{(j)}b_k^{(j)\dagger}b_k^{(j)}\end{aligned}$$ is the unperturbed part containing the free Hamiltonian of the two qubits as well as that of the two independent environments. The transition frequency of the two two-level systems, supposed coincident for simplicity, is indicated by $\omega_0$ whereas $\sigma_z^{(j)}$ ($j = 1,2$) denotes the Pauli operator describing the $j$-th subsystem. The two independent bosonic baths are characterized by proper frequencies $\omega_k^{(j)}$, $b_k^{(j)\dagger}$ and $b_k^{(j)}$ being correspondingly the creation and annihilation bosonic operators.
The interaction term $$\label{Hint}
\begin{split}
H_I=&\Omega(\sigma_+^{(1)}\sigma_-^{(2)}+\sigma_-^{(1)}\sigma_+^{(2)})+\\
&+\sum_{j=1,2}\left(\sigma_+^{(j)}\sum_kg_k^{(j)}b_k^{(j)}+\sigma_-^{(j)}\sum_kg_k^{(j)\ast}b_k^{(j)\dagger}\right)
\end{split}$$ includes both the direct interaction between the two qubits, characterized by the coupling constant $\Omega$, and the interaction between each qubit and its respective bosonic bath, with coupling constants $g_k^{(j)}$. In eq.(\[Hint\]) $\sigma_{\pm}^{(j)}\equiv
\frac{1}{2}(\sigma_{x}^{(j)} \pm i \sigma_{y}^{(j)})$ are, as usual, the lowering and raising Pauli operators.
It is worth underlining that the Hamiltonian model is quite versatile in the sense that it can be successfully adopted to describe many different physical systems. In the framework of cavity QED [@bellomo] or circuit QED [@6; @7; @8; @9; @10] this model can be indeed exploited for the description of two atoms in spatially separated cavities as well as of two far enough Josephson charge, flux or phase qubits so that it is reasonable to assume that they interact with two different electromagnetic environments. Model in addition allows to study the influence of spurious microwave resonators within Josephson tunnel junctions on the coherent dynamics of a phase qubit [@martinis]. In this case the environment coupled to the spurious resonator (modeled as a two state system) is a phononic bath.
Very recently the Markovian dynamics stemming from Hamiltonian has been analyzed [@sinaysky]. In Ref.[@Ferraro] instead the TCL approach has been exploited in order to investigate on the non-Markovian regime. In this paper we exactly solve the time-dependent Schrödinger equation confining ourselves to the one excitation subspace. At the same time we adopt the NZ technique to derive a non-Markovian master equation for the reduced density operator of the two coupled qubits. Having at our disposal both the exact dynamics and the approximate master equations, a comparison may be done in order to test the effectiveness of the perturbative approaches. From now on, we work in the interaction picture defined by $H_0$ in which the interaction Hamiltonian reads $$\label{Hint2}
H_I(t)=H_I^{(s)}+H_I^{(D_1)}+H_I^{(D_2)}$$ with $$H_I^{(s)}=\Omega(\sigma_+^{(1)}\sigma_-^{(2)}+\sigma_-^{(1)}\sigma_+^{(2)})$$ and $$\begin{split}
H_I^{(D_j)}&=\sigma_+^{(j)}\sum_kg_k^{(j)}b_k^{(j)}e^{i(\omega_0-\omega_k^{(j)})t}+\\
&+\sigma_-^{(j)}\sum_kg_k^{(j)\ast}b_k^{(j)\dagger}e^{-i(\omega_0-\omega_k^{(j)})t}.
\end{split}$$
One excitation time evolution
-----------------------------
Let us begin by looking at the exact solution of the time-dependent Schrödinger equation. It is easy to verify that the number operator $$\begin{aligned}
\hat{N}=\sum_{j=1,2}\sigma_+^{(j)}\sigma_-^{(j)}+\sum_{j=1,2}\sum_{k}b_k^{(j)\dagger}b_k^{(j)}\end{aligned}$$ is a constant of motion. This in particular means that, starting from an eigenstate of $\hat{N}$, the system evolves remaining in the subspace correspondent to the same eigenvalue $n$ of $\hat{N}$. In what follows we consider the dynamics of the system in the subspace with one excitation, that means $n=1$. To this end let us suppose to prepare at $t=0$ the two qubits in a linear superposition of states with one excitation and both the baths in the vacuum state denoted by $|0_k^{(j)}\rangle$ ($j=1,2$) $$|\psi(0)\rangle=\left(a(0)|10\rangle+b(0)|01\rangle\right)|0_k^{(1)}0_k^{(2)}\rangle,$$ with $|a(0)|^2+|b(0)|^2=1$. Since $[H,\hat{N}]=0$, at a generic time instant $t$ we may write $$\label{psi}\begin{split}
&|\psi(t)\rangle=\left(a(t)|10\rangle+b(t)|01\rangle\right)|0_k^{(1)}0_k^{(2)}\rangle+\\
&+|00\rangle\left(\sum_kc_k^{(1)}(t)|1_k^{(1)}0_k^{(2)}\rangle+\sum_kc_k^{(2)}(t)|0_k^{(1)}1_k^{(2)}\rangle\right),
\end{split}$$ where $|1_k^{(j)}\rangle$ denotes a state of the $j-th$ bath ($j=1,2$) with one excitation in the mode $k$ and the probability amplitudes $a(t)$, $b(t)$ and $c_k^{(j)}$ ($j=1,2$) are solutions of the following system of differential equations $$\label{diff_eq_ampl}
\left\{\begin{array}{c} \dot{a}(t)=-i\left(\Omega\,
b(t)+\sum_kc_k^{(1)}(t)g_k^{(1)}e^{i(\omega_0-\omega_k^{(1)})t}\right)\\
\dot{b}(t)=-i\left(\Omega\,
a(t)+\sum_kc_k^{(2)}(t)g_k^{(2)}e^{i(\omega_0-\omega_k^{(2)})t}\right)\\
\dot{c}_k^{(1)}(t)=-i\,a(t)g_k^{(1)\ast}e^{-i(\omega_0-\omega_k^{(1)})t}\\
\dot{c}_k^{(2)}(t)=-i\,b(t)g_k^{(2)\ast}e^{-i(\omega_0-\omega_k^{(2)})t}.\\
\end{array}\right.$$
From eqs.(\[diff\_eq\_ampl\]) it is easy to verify that the amplitudes $c_k^{(j)}(t)$ formally evolve as follows: $$c_k^{(1)}(t)=-ig_k^{(1)\ast}\int_0^ta(t')e^{-i(\omega_0-\omega_k^{(1)})t'}\,dt'$$ $$c_k^{(2)}(t)=-ig_k^{(2)\ast}\int_0^tb(t')e^{-i(\omega_0-\omega_k^{(2)})t'}\,dt'.$$ Inserting these formal solutions in the equations for $a(t)$ and $b(t)$ we achieve $$\label{system}
\left\{\begin{array}{c} \dot{a}(t)=-i\Omega\,
b(t)-i\int_0^ta(t')f_1(t-t')\,dt'\\
\dot{b}(t)=-i\Omega\,
a(t)-i\int_0^tb(t')f_2(t-t')\,dt'\\
\end{array}\right.$$ where the kernel $f_j(t-t')$ is given by the correlation function defined as $$\label{fj}
f_j(t-t')=\sum_k|g_k^{(j)}|^2e^{i(\omega_0-\omega_k^{(j)})(t-t')}$$ that in the continuum limit becomes $$\label{fjcont}
f_j(t-t')=\int_0^{+\infty}d\omega\,J_j(\omega)e^{i(\omega_0-\omega)(t-t')},$$ $J_j(\omega)$ being the spectral density of the $j-th$ bath.
Making use of the Laplace transform, the system becomes $$\label{systemLaplace}
\left\{\begin{array}{c} s\,\tilde{a}(s)-a(0)=-i\Omega\,\tilde{b}(s)-\tilde{a}(s)\tilde{f}_1(s)\\
s\,\tilde{b}(s)-b(0)=-i\Omega\,\tilde{a}(s)-\tilde{b}(s)\tilde{f}_2(s),
\end{array}\right.$$ where $\tilde{a}(s)$, $\tilde{b}(s)$ and $\tilde{f}_j(s)$ denote the Laplace transforms of $a(t)$, $b(t)$ and $f_j(t-t')$ respectively. It is thus immediate to obtain $$\label{a}
\tilde{a}(s)=\frac{a(0)(s+\tilde{f}_2(s))-i\Omega
b(0)}{(s+\tilde{f}_1(s))(s+\tilde{f}_2(s))+\Omega^2}$$ $$\label{b}
\tilde{b}(s)=\frac{b(0)(s+\tilde{f}_1(s))-i\Omega
a(0)}{(s+\tilde{f}_1(s))(s+\tilde{f}_2(s))+\Omega^2}.$$ Once fixed the spectral densities for both baths $J_1(\omega)$ and $J_2(\omega)$, it is quite easy to obtain the time behavior of $a(t)$, $b(t)$ and $c_k^{(j)}(t)$, simply antitransforming the amplitudes $\tilde{a}(s)$ and $\tilde{b}(s)$ given by eqs. and . The results thus obtained will be discussed in Sec. IV.
Nakajima-Zwanzig master equation
================================
In this section we will apply the projection operator techniques in order to derive a non-Markovian master equation for the reduced density matrix $\rho_S(t)$ of the two qubits [@petruccionebook]. To this end, it is convenient to introduce a super-operator according to $$\rho\rightarrow\mathcal{P}\rho=\operatorname{Tr}_B{\rho}\otimes\rho_B\equiv\rho_S\otimes\rho_B,$$ where $\rho_B$ is the density matrix of the environment. The super-operator $\mathcal{P}$ projects any state of the total system $\rho$ onto its relevant part $\mathcal{P}\rho$, expressing formally the elimination of the irrelevant degrees of freedom from the full dynamical description of the model under scrutiny. Following the NZ approach we get an integro-differential equation $$\label{maste}
\frac{d}{dt}\mathcal{P}\rho(t)=\int_0^tdt'\,\mathcal{K}(t,t')\mathcal{P}\rho(t')$$ describing the reduced dynamics of the system. Here the memory kernel $\mathcal{K}(t,t')$ is a super-operator in the relevant subspace. In order to discuss the reduced dynamics we perform a perturbation expansions of $\mathcal{K}(t,t')$ with respect to the strength of the interaction Hamiltonian [@petruccionebook]. If we restrict ourselves to the second order, the two relevant terms of $\mathcal{K}(t,t')$ can be written down as $$\mathcal{K}_1(t,t')=\mathcal{P}\mathcal{L}(t)\mathcal{P}$$ and $$\mathcal{K}_2(t,t')=\mathcal{P}\mathcal{L}(t)\mathcal{Q}\mathcal{L}(t')\mathcal{P},$$ where $\mathcal{Q}\equiv I-\mathcal{P}$ and $\mathcal{L}$ is the Liouville super-operator defined by $\mathcal{L}(t)\rho(t)\equiv-i[H_I(t),\rho(t)]$. Starting from eq. it is possible to demonstrate that the second order non-Markovian master equation assumes the following form $$\label{master0}
\begin{split}
\frac{d}{dt}\rho_S(t)=&-i[H_I^{(s)},\rho_S(0)]+\\
&-\int_0^tdt'\,\operatorname{Tr}_B\left\{[H_I(t),[H_I(t'),\rho_S(t')\otimes\rho_B]]\right\},
\end{split}$$ that, contrarily to the TCL master equation, is a non-local evolution equation. After some manipulations it is possible to recast eq. in the more compact form $$\label{master}
\begin{split}
\frac{d}{dt}\rho_S(t)=&-i[H_I^{(s)},\rho_S(0)]+\mathcal{L}^{(s)}\rho_S(t)+\\
&+\mathcal{L}^{(D_1)}\rho_S(t)+\mathcal{L}^{(D_2)}\rho_S(t),
\end{split}$$ where the dissipators are expressed by $$\mathcal{L}^{(s)}\rho_S(t)=-\int_0^tdt'\,[H_I^{(s)}(t),[H_I^{(s)}(t'),\rho_S(t')]]$$ and $$\label{dj}
\begin{split}
\mathcal{L}^{(D_j)}\rho_S(t)=\int_0^t&dt'\,\left\{g_j(t-t')[\sigma_-^{(j)}\rho(t'),\sigma_+^{(j)}]+\right.\\
&\left.+g_j^{\ast}(t-t')[\sigma_-^{(j)},\rho_S(t')\sigma_+^{(j)}]\right\}.
\end{split}$$ In eq. $g_j(t-t')$ is the correlation function, that in correspondence to a thermal bath with $T=0$ coincides with $f_j(t-t')$ defined in eq.. Once again, in order to solve the master equation we can exploit the Laplace transform. The results we have obtained are reported in the next section where they are compared with the exact dynamics as well as with the results obtained exploiting the TCL approach [@Ferraro]. The master equation solved in Ref.[@Ferraro] is a time-local differential equation and it can be obtained from eq. by replacing $\rho(t')$ with $\rho(t)$.
Comparison between exact and approximate solutions
==================================================
Exploiting the results obtained in the previous sections we now analyze the time behavior of the two qubits comparing in particular the exact dynamics with the ones stemming from the NZ and TCL approaches. Our aim is to highlight the performances of two approximate approaches and to point out their range of validity. As underlined in the introduction there is not indeed a general theory that predict which one of the two methods is to be preferred and generally speaking their range of validity strongly depends on the specific features of the system, namely the interaction hamiltonian, the interaction time, the environmental state and the spectral density.
In what follows we assume that each qubit interacts resonantly with a reservoir with Lorentzian spectral density $$J_1(\omega)=J_2(\omega)\equiv
J(\omega)=\frac{1}{2\pi}\frac{\gamma\lambda^2}{(\omega_0-\omega)^2+\lambda^2},$$ where $\gamma$ is a parameter which in the Markovian limit coincides with the system decay rate, and $\lambda$ is the reservoir bandwidth. This is for instance the case of two atoms interacting each of them with their own cavity field in presence of cavity losses [@garraway1]. Thus the two correlation functions $f_1(t-t')$ and $f_2(t-t')$ defined by eq. coincide and are given by $$\label{corr_lorentz}
f_1(t-t')=f_2(t-t')\equiv f(t-t')=\frac{1}{2}\gamma\lambda
e^{-\lambda|t-t'|}.$$ From the latter equation, it is clear that the bandwidth $\lambda$ plays the role of the inverse of the reservoir memory time.
The system dynamics will be analyzed considering the two baths in a thermal state at $T=0$ whereas the two qubits are supposed at $t=0$ in the Bell state $|\psi(0)\rangle=\frac{1}{\sqrt{2}}\left(|10\rangle-|01\rangle\right)$. Let us concentrate on the population $P_{10}(t)$, that is on the probability of finding the two qubits in the state $|10\rangle$. To test the validity of the approximate approaches we explore three different regimes varying the width of the Lorentzian spectral density $\lambda$. This investigation will allow us to assess in which cases the solutions of the master equations are efficient in the description of the true dynamics of the system.
![Time evolution of the population of the state ${|{10}\rangle}$ for a system initially prepared in the Bell state $\frac{1}{\sqrt{2}}\left({|{10}\rangle}-{|{01}\rangle}\right)$. The width of the Lorentzian spectral density is $\lambda=10\gamma$, the strength of the coupling constant between the two qubits is $\Omega=0.001\gamma$.[]{data-label="figura1"}](Fig1inset.eps){width="40.00000%"}
Figure shows a comparison among the exact, the TCL and the NZ solutions in the case of large reservoir bandwidth $\lambda=10\gamma$. The plot is done against the dimensionless variable $\gamma t$ and the coupling constant between the qubits is fixed to $\Omega=0.001\gamma$. We can clearly appreciate the perfect agreement between the exact analytical solution and the approximate ones for the short time behavior but also for long interaction times. In this case the two approaches TCL and NZ both provide a very good description of the dynamics and we may conclude that there is no way to establish if one method is to be preferable with respect to the other, they indeed give the same results. However in such cases the TCL master equation might be preferred since it involves a time local first order differential equation and therefore it is easier to solve. In the inset of fig. the short time behavior is shown. It is interesting to underline the initially quadratic behavior that witnesses the non-Markovian features of the dynamics.
![Time evolution of the population of the state ${|{10}\rangle}$ for a system initially prepared in the Bell state $\frac{1}{\sqrt{2}}\left({|{10}\rangle}-{|{01}\rangle}\right)$. The width of the Lorentzian spectral density is $\lambda=\gamma$, the strength of the coupling constant between the two qubits is $\Omega=0.001\gamma$.[]{data-label="figura2"}](Fig2.eps){width="40.00000%"}
In fig. the same quantity is reported choosing $\lambda=\gamma$. Despite the good agreement for the short time dynamics, we observe significant deviations when time increases. In particular concerning the long time behavior, the NZ equation leads to a very bad approximation. For times longer than some critical values the solution for the population $P_{10}(t)$ cannot represent a true diagonal element of a density matrix anymore, since it indeed assumes negative values. We may conclude that for this range of parameters the TCL solution gives a better description of the dynamics since it reproduces all the qualitative features of the exact solution.
![Time evolution of the population of the state ${|{10}\rangle}$ for a system initially prepared in the Bell state $\frac{1}{\sqrt{2}}\left({|{10}\rangle}-{|{01}\rangle}\right)$. The width of the Lorentzian spectral density is $\lambda=0.01\gamma$, the strength of the coupling constant between the two qubits is $\Omega=0.001\gamma$.[]{data-label="figura3"}](Fig3.eps){width="40.00000%"}
Finally in fig. we examine the regime $\lambda=0.01\gamma$ which, according to eq., corresponds to very strong reservoir correlations and very long memory time. We observe once again a perfect agreement among all the three approaches in the short time behavior but in this case we assist at a failure of the TCL approach too. The solution of the TCL master equation (dashed line) doesn’t succeed to follow the Rabi’s oscillations witnessed by the exact dynamics (solid line). The NZ approaches presents the same problem of not conserving the positivity of the density matrix as in the previous case. Thus, in this case, both the perturbative approaches are not suitable to describe the dynamics of the system. It is interesting to observe that the oscillations appearing in the exact evolution of the populations $P_{10}(t)$ are not due to the spin-spin interaction constant $\Omega$, which we have taken small on purpose. The oscillations are instead due to the fact that, in this case, the Lorentzian peak is so narrow to make the environment equivalent to a cavity with losses, as one might verify for example by using the pseudomode approach in Refs.[@Mazzola; @garraway1; @garraway2; @garraway4]. Therefore, our exact solution in this regime, could be exploitable to describe the dynamics of two interacting qubits put inside two different optical cavities.
Discussion and conclusive remarks
=================================
The problem of the proper description of open quantum systems is still far from having a complete and general solution. In particular, while the features of the Markovian dissipative dynamics are all well established and accepted, a lot of work has still to be done to claim general statements on the validity of the different possible non-Markovian approaches available in the literature.
In this paper we concentrated on the study of two popular methods aimed at describing non-Markovian dynamics: the Nakajima-Zwanzig and the time-convolutionless master equation approaches. Exploiting an exact solution for the dissipative dynamics of two coupled qubits interacting with independent reservoirs, we have shown that the TCL approach reproduces all the features of the non-Markovian dynamics for a range of parameters much wider than the one in which the NZ equation gives results which are physically reasonable, since the latter approach may violate the positivity condition on the density matrix already for reservoir correlations which are not very strong.
The discrimination of the best master equation approach for the problem under study is a very important issue, because the problem of the dissipative dynamics of two interacting qubits has been given a lot of attention in recent years, especially from the point of view of the entanglement dynamics [@quiroga; @sinaysky; @Tanas; @tanas; @2008; @grifoni; @storcz]. So, once we have shown that for some parameters the TCL master equations provides a very good description of the non-Markovian dynamics, one may use the same approach for the description of temperature effects on the system dynamics, since we do not have at our disposal an exact solution of the dynamics at general reservoir temperatures.
We finally note that our exact solution can moreover be exploited in order to study the zero temperature entanglement dynamics in a range of parameters in which none of the two master equation approaches can be used: in this way one could extend the studies given in Refs. [@quiroga; @scala; @sinaysky], so that one could get the most general features of the quantum correlations between two qubits coupled to two independent reservoirs. These points will be the subject of our future research.
Acknowledgements
================
The authors acknowledge stimulating discussions with Prof. A. Messina. The authors also acknowledge partial support by MIUR Project N. II04C0E3F3. M.S. acknowledges financial support by the European Commission project EMALI.
[10]{}
A. Polyakov, Phys. Lett. **72B**, 224 (1977).
A. Belitsky, V. Braun, A. Gorsky and G. Korchemsky, Int. J. Mod. Phys. A **19**, 4715 (2004).
A. Baranco et al., Phys. Rev. Lett. **74**, 4083 (1995).
L. Amico, R. Fazio, A. Ostorloh and V. Vedral, Rev.Mod. Phys. **80**, 517 (2008).
R. McDermott, R. W. Simmonds, M. Steffen, K. B. Cooper, K. Cicak, K. D. Osborn, S. Oh, D. P. Pappas, and J. M. Martinis, Science **307**, 1299 (2005).
J. B. Majer, F. G. Paauw, A. C. J. ter Haar, C. J. P. M. Harmans, and J. E. Mooij, Phys. Rev. Lett. **94**, 090501 (2005).
T. Yamamoto, Y. A. Pashkin, O. Astafiev, Y. Nakamura, and J. Tsai, Nature **425**, 941 (2003).
A. J. Berkley, H. Xu, R. C. Ramos, M. A. Gubrud, F. W. Strauch, P. R. Johnson, J. R. Anderson, A. J. Dragt, C. J. Lobb, and F. C. Wellstood, Science **300**, 1548 (2003).
Y. A. Pashkin, T. Yamamoto, O. Astafiev, Y. Nakamura, D. Averin, and J. Tsai, Nature **421**, 823 (2003).
B. Bellomo, R. Lo Franco, G. Compagno, Phys. Rev. Lett. **99**, 160502 (2007).
B.M. Garraway, Phys. Rev. A [**55**]{}, 2290 (1997).
B.M. Garraway, Phys. Rev. A [**55**]{}, 4636 (1997).
B.J. Dalton, S.M. Barnett and B.M. Garraway, Phys. Rev. A [**64**]{}, 053813 (2001).
Sumanta Das, G.S. Agarwal, arXiv:0905.3399 [\[]{}quant-ph\].
I. Sinaysky, F. Petruccione, D. Burgarth, Phys. Rev. A **78**, 062301 (2008).
L. Quiroga, F.J. Rodríguez, M.E. Ramírez, R. París, Phys. Rev. A **75**, 032308 (2007).
S. Maniscalco, F. Francica, R. L. Zaffino, N. Lo Gullo, F. Plastina, Phys. Rev. Lett. **100**, 090503 (2008).
L. Mazzola, S. Maniscalco, J. Piilo, K.-A. Suominen, and B. M. Garraway, Phys. Rev. A **80**, 012104 (2009)
T. Yu, J.H. Eberly, arXiv:0906.5378 [\[]{}quant-ph\].
M. Scala, R. Migliore, A. Messina, J. Phys. A: Math. Theor. **41**, 435304 (2008).
S. Chaturvedi and F. Shibata, Z. Phys. B **35**, 297 (1979).
S. Nakajima, Prog. Theor. Phys. **70**, 948 (1958).
R. Zwanzig, J. Chem. Phys. **33**, 1338 (1960).
B.L. Hu, J.P. Paz, Y. Zhang, Phys. Rev. D **45**, 2843 (1992).
H.-P. Breuer and F. Petruccione, [*The Theory of Open Quantum Systems*]{}, Oxford University Press (2002).
I. Sinayskiy, E. Ferraro, A. Napoli, A. Messina, F. Petruccione, arXiv:0906.1796 [\[]{}quant-ph\].
R. W. Simmonds, K. M. Lang, D. A. Hite, S. Nam, D. P. Pappas, and John M. Martinis, Phys. Rev. Lett. [**93**]{}, 077003 (2004).
R. Tanas, Z. Ficek, J. Opt. B Quantum Semiclass. Opt. [**6**]{}, S90 (2004); Phys. Rev. A, [**74**]{}, 024304 (2008).
R. Tanas, Z. Ficek, Phys. Rev. A, [**77**]{}, 054301 (2008).
M. Governale, M. Grifoni and G. Schön, Chem. Phys. [**268**]{}, 273 (2001).
M. J. Storcz, F. Hellmann, C. Hrelescu, and F. K. Wilhelm, Phys. Rev. A [**72**]{}, 052314 (2005).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Queuing models provide insight into the temporal inhomogeneity of human dynamics, characterized by the broad distribution of waiting times of individuals performing tasks. We study the queuing model of an agent trying to execute a task of interest, the priority of which may vary with time due to the agent’s “state of mind.” However, its execution is disrupted by other tasks of random priorities. By considering the priority of the task of interest either decreasing or increasing algebraically in time, we analytically obtain and numerically confirm the bimodal and unimodal waiting time distributions with power-law decaying tails, respectively. These results are also compared to the updating time distribution of papers in the arXiv.org and the processing time distribution of papers in Physical Review journals. Our analysis helps to understand human task execution in a more realistic scenario.'
author:
- 'Hang-Hyun Jo'
- Raj Kumar Pan
- Kimmo Kaski
title: 'Time-Varying Priority Queuing Models for Human Dynamics'
---
Introduction
============
Understanding human behavior is a fascinating and challenging subject that has been explored extensively in the fields of sociology, economics, and psychology for centuries [@Zipf1949; @Becker1978; @Ajzen1980]. Recently, with the need to gain deeper insight into complex social systems, the human behavior has also been studied from the perspective of other scientific disciplines including physical and computational sciences [@Castellano2009]. This is because the concepts and methods developed for investigating physical processes have turned out to be applicable to understand the human and social phenomena. Moreover, the huge amounts of digital data have enabled us to explore the human behavior at the very high resolution. The empirical studies based on such dataset have shown that the timing of human actions is not simply random but correlated or bursty [@Barabasi2005; @Wu2010; @Karsai2011]. Such correlations are often characterized by the heavy tailed or power-law distribution of waiting or response time $\tau$ as $P(\tau)\sim\tau^{-\alpha}$, where $\alpha\approx 1$ for e-mail communication and $1.5$ for letter correspondence, respectively [@Barabasi2005; @Oliveira2005]. Here $\tau$ is defined as the time taken by the user to respond to a received message.
The origin of bursts in human dynamics has been explored for years. The long inactive periods at nighttime as well as on weekends are known to result in the bursts of e-mail communication [@Malmgren2008; @Malmgren2009a], where the circadian and weekly cyclic patterns of humans were modeled by the non-homogeneous Poisson process. In addition to these human cyclic patterns, the human dynamics can be also affected by other human factors, such as the task execution behavior. To figure out the role of task execution behavior we have adopted the framework of queuing theory [@Barabasi2005; @Vazquez2005; @Vazquez2006; @Anteneodo2009], where the to-do list of an agent is modeled as a finite length queue. The agent assigns priorities to tasks in the queue and then executes them according to some protocol. As in the Barabási’s model [@Barabasi2005], if the task with highest priority is executed first, then the one with low priority must wait for a relatively long time in the queue, leading to the heavy tailed distribution of waiting times. Barabási’s priority queuing model is based on a few key assumptions: (i) It is a single agent model, where the agent has no explicit interaction with other agents, (ii) all the tasks in the queue are of the same type, such as replying to the received message, and (iii) the priority assigned to each task is kept fixed over time. Several variants of this model have been studied by means of relaxing at least one of these assumptions [@Blanchard2007; @Oliveira2009; @Min2009; @Jo2011]. Here we perform the similar study by relaxing the assumptions (ii) and (iii) and consider that only one type of the task has the time-varying priority. In the similar context, Blanchard and Hongler studied the queuing systems on the basis of population dynamics, where the individuals in a city (tasks in a queue) are assigned with the priorities increasing with time due to aging or the deadline effect, and finally dying (executed) according to “the highest priority first” protocol [@Blanchard2007]. However, instead of the population dynamics based models, the microscopic models with simpler setups might help us to better understand the effect of time-varying priorities on the waiting time distributions.
In this paper, we adopt the microscopic approach and study a single agent queuing model with two tasks: one is the task of interest with time-varying priority and the other is the unexpected and random task with random priority. As an example let us consider a situation in which a researcher wants to complete the research project of current interest, but he or she gets continuously distracted by meetings, coffee breaks and so on. The priority of the task of interest recognized by the researcher can vary due to his or her state of mind, such as mood and happiness. The priority may increase in some cases, such as when the deadline assigned to the task approaches [@Alfi2007], or decrease in other cases, for example, when the researcher gradually loses interest in that task. For convenience, we call the task of interest and the random task as type-A and type-B tasks, respectively. This model with one type-A and one type-B tasks is minimal and can be generalized to more complex scenarios, such as the case of an agent with many tasks of various types or the case with interacting agents.
The paper is organized as follows. In Section \[sec:model\], we study the models with different cases of time-varying priorities. For each case we obtain the analytic solutions for the waiting time distribution that are confirmed by numerical simulations. In Section \[sec:empirical\], we show how the extended versions of our models to many agents can be used to understand the paper updating mechanism in arXiv.org and the paper reviewing process in Physical Review journals. Finally we summarize the results and make conclusions in Section \[sec:conclusion\].
![Schematic diagram of time-varying priority queuing model. (a) The task of interest (type-A) has the time-varying priority $x(t)$, which is compared with the random task (type-B) with random number $r_t$ at each time step and executed if $x(t)\geq r_t$. (b) The case with power-law decreasing priority. (c) The case with power-law increasing priority: bounded and unbounded cases. See the text for details.[]{data-label="fig:model"}](fig1.pdf){width="\columnwidth"}
Model {#sec:model}
=====
We consider an agent with a queue of size $2$ as depicted in Fig. \[fig:model\](a). The first site of the queue is occupied by the task of interest with time-varying priority (type-A task). When the type-A task is introduced to the queue at the time step $t=0$, its priority is given by the random number $x_0$ drawn from the uniform distribution of $(0,1)$ and then it varies with time. The second site of the queue is for the random task (type-B task), which is replaced by a new random task at every time step. The priority of the type-B task introduced at time step $t$ is given by the random number $r_t$ drawn from the uniform distribution of $(0,1)$. At each time step $t$, the priority $x(t)$ is compared with the random number $r_t$ and executed when $x(t)\geq r_t$. Here the execution time defines the waiting time $\tau$ for which the type-A task waits for the execution since it is introduced to the queue. This is the deterministic case where only the highest priority task is executed.
We also consider a stochastic version of the model, where the lower priority task is allowed to be executed with small probability. At each time step, the highest priority task is executed with the probability $p$, while with probability $1-p$ one of two tasks is selected at random for the execution, which we call random selection. This implies that the priority is not the only determinant for the execution. This can also be interpreted as the “trembling hand” effect in the game theory [@Weibull1997]. In general, the waiting time distribution $P(\tau)$ is obtained by $$\begin{aligned}
\label{eq:P_tau}
P(\tau)&=&\int_0^1 dx_0 \left[
\prod_{t=0}^{\tau-1}\left\{p[1-x(t)]+\tfrac{1-p}{2}\right\} \right] \left[p x(\tau)+\tfrac{1-p}{2}\right] \nonumber\\
&=&p^{\tau+1}\int_0^1 dx_0
\left[ \prod_{t=0}^{\tau-1}\left[1-x(t)+\epsilon\right] \right] \left[x(\tau)+\epsilon\right],\end{aligned}$$ where $\epsilon\equiv \frac{1-p}{2p}$ is assumed to be very small in this paper.
Fixed priority
--------------
We first consider the case of fixed priority. Here the priority of the type-A task is fixed, i.e. $x(t)=x_0~\forall t$. The waiting time distribution is given by $$\begin{aligned}
\label{eq:fixed_sol}
P(\tau)&=&p^{\tau+1}\left[\tfrac{(1+\epsilon)^{\tau+2}-\epsilon^{\tau+2}}{(\tau+1)(\tau+2)}
+ \tfrac{\epsilon(1+\epsilon)[(1+\epsilon)^\tau-\epsilon^\tau]}{\tau+1}\right].\end{aligned}$$ In the limit of $\tau\gg 1$, it can be approximated by $$\begin{aligned}
\label{eq:fixed_sol1}
P(\tau) &\approx & (\tau^{-2}+2\epsilon\tau^{-1})e^{-\tau/\tau_c(p)},\end{aligned}$$ where we define the cutoff of distribution as $\tau_c(p)\equiv [\ln(1/p)]^{-1}\approx (2\epsilon)^{-1}$. In the scaling regime, i.e. $1\ll \tau\ll \tau_c\approx (2\epsilon)^{-1}$, the first term $\tau^{-2}$ dominates the distribution function, so that $P(\tau) \approx \tau^{-\alpha}$ with $\alpha=2$. In the deterministic case with $p=1$, $P(\tau)=\frac{1}{(\tau+1)(\tau+2)}\approx\tau^{-2}$ in the limit of $\tau\gg 1$. Note that the small probability of random selection does not change the scaling exponent but leads to the finite cutoff of the distribution. Figure \[fig:fixed\] shows that the analytic solution is confirmed by the numerical simulations for both deterministic ($p=1$) and stochastic ($p=0.999$) cases.
![The case with fixed priority as $x(t)=x_0$ for deterministic ($p=1$) and stochastic ($p=0.999$) versions. For the numerical results we used $10^9$ samples. The solid line for the stochastic version represents the analytic solution of Eq. (\[eq:fixed\_sol\]).[]{data-label="fig:fixed"}](fig2.pdf){width=".9\columnwidth"}
Power-law decreasing priority {#subsec:decrease}
-----------------------------
Now we consider the case when the priority of the type-A task changes over time. This change might be caused by the internal factors related to the agent as well as to the external factors. We consider the simple case with the algebraically decreasing priority as $$x(t)=\frac{x_0}{(t+1)^{\gamma}}.$$ The decreasing speed of the priority is controlled by the exponent $\gamma$. If $\gamma=0$, the model reduces to the fixed priority case, where the waiting time distribution decays as the power-law with exponent $\alpha=2$. For the larger value of $\gamma$, the priority decays faster so that the type-A task becomes less likely to be executed. Hence one may expect that this will lead to a heavier tail of waiting time distribution and consequently to a smaller value of $\alpha$.
At first we analyze the case with $\gamma\ll 1$. Since the priority $x(t)$ decreases very slowly, we can use the approximation of $x(t)\approx x(\tau)=\frac{x_0}{(\tau+1)^{\gamma}}$. Putting this into Eq. (\[eq:P\_tau\]), we get $$\begin{aligned}
P(\tau)&\approx& p^{\tau+1}\bigg\{\tfrac{(\tau+1)^{\gamma-1}}{\tau+2} \left[ (1+\epsilon)^{\tau+2}-\left(1+\epsilon-\tfrac{1}{(\tau+1)^\gamma}\right)^{\tau+2}\right]\nonumber\\
&& - (\tau+1)^{\gamma-1}\left[\epsilon+\tfrac{1}{(\tau+1)^\gamma}\right] \left[1+\epsilon-\tfrac{1}{(\tau+1)^\gamma}\right]^{\tau+1}\nonumber\\
&& + (\tau+1)^{\gamma-1}\epsilon(1+\epsilon)^{\tau+1} \bigg\}.\label{eq:smallGamma}\end{aligned}$$ For large values of $\tau$ and with the fact that $\tau^\gamma$ is close to $1$, the above equation reduces to $$\begin{aligned}
\label{eq:smallGammaApprox}
P(\tau)&\approx&(\tau^{-(2-\gamma)}+2\epsilon\tau^{-(1-\gamma)})e^{-\tau/\tau_c(p)}.\end{aligned}$$ In the scaling regime of $1\ll \tau\ll \tau_c\approx (2\epsilon)^{-1}$, the first term $\tau^{-(2-\gamma)}$ dominates the distribution function, so that $P(\tau) \approx \tau^{-\alpha}$ with $\alpha=2-\gamma$. Thus the exponent $\alpha$ is smaller for larger value of $\gamma$, as expected.
![The case with power-law decreasing priority as $x(t)=\frac{x_0} {(t+1)^\gamma}$ for deterministic version with $p=1$ (a) and for stochastic version with $p=0.999$ (b). For the numerical results we used up to $10^8$ samples. In the inset of (a) we plot the numerical results of $\alpha(\gamma)$ by the red filled circles with the analytic expectation by the black lines. In (b) the solid lines represent the analytic solutions of Eq. (\[eq:smallGamma\]) for $\gamma=0.2$ and Eq. (\[eq:largeGamma\]) for $\gamma=2$, respectively.[]{data-label="fig:decrease"}](fig3a-b.pdf){width=".9\columnwidth"}
For the deterministic version of the model, i.e. $p=1$, we perform the numerical simulations with various values of $\gamma$ as shown in Fig. \[fig:decrease\](a). We find that for any value of $\gamma$ the waiting time distribution follows the power-law behavior with the power-law exponent $\alpha$ as a function of $\gamma$, see the inset of Fig. \[fig:decrease\](a). Our analytic result of $\alpha=2-\gamma$ is numerically confirmed up to $\gamma\approx 0.6$. The numerical results, however, suggest that the value of $\alpha$ has the minimum at $1.22$ for $\gamma=1$, while it increases when $\gamma>1$. In the stochastic case with $p<1$, the second term $\tau^{-(1-\gamma)}$ in Eq. (\[eq:smallGammaApprox\]) begins to affect the dominant power-law behavior and the exponential cutoff also becomes finite. This approximate solution is again numerically confirmed for various values of $\gamma$, as depicted in Fig. \[fig:decrease\](b).
Next, the other limiting case when $\gamma\gg 1$ is considered. Since the value of $x(t)$ decreases very fast, we use the approximation of $x(t)\ll 1$ for all $t$. Thus the product term in Eq. (\[eq:P\_tau\]) can be expanded in the limit of $\epsilon\ll 1$ as follows $$\begin{aligned}
\prod_{t=0}^{\tau-1}[1-x(t)+\epsilon] &\approx& 1-\sum_{t=0}^{\tau-1} [x(t)-\epsilon]\label{eq:expand}\\
&\approx& 1+\epsilon\tau-x_0\tfrac{1-\tau^{1-\gamma}}{\gamma-1}.\end{aligned}$$ Then the waiting time distribution is obtained by $$\begin{aligned}
P(\tau)&\approx&p^{\tau+1}\bigg\{\tfrac{1}{2(\tau+1)^\gamma} -\tfrac{1-\tau^{1-\gamma}}{3(\gamma-1)(\tau+1)^\gamma}\nonumber\\
&& + \epsilon\left[1-\tfrac{1-\tau^{1-\gamma}}{2(\gamma-1)}+\tfrac{\tau}{2(\tau+1)^\gamma}\right]+\epsilon^2\tau\bigg\}\label{eq:largeGamma}.\end{aligned}$$ Using the assumption of $\gamma\gg 1$ and the condition that $\epsilon\tau\ll 1$, the above equation can be approximated as $$\begin{aligned}
P(\tau)&\approx& \left[\left(\tfrac{1}{2}-\tfrac{1}{3\gamma}\right)\tau^{-\gamma}+\epsilon \left(1-\tfrac{1}{2\gamma}\right)\right]e^{-\tau/\tau_c(p)}.\label{eq:largeGammaFinal}\end{aligned}$$ For the scaling regime, the waiting time distribution is dominated by the first term, leading to $P(\tau)\approx\tau^{-\alpha}$ with $\alpha=\gamma$. This is numerically confirmed up to $\gamma>1.4$ in the deterministic case as shown in the inset of Fig. \[fig:decrease\](a). For the stochastic case with $p=0.999$, we find that the numerical simulations confirm the analytic solutions but with some deviation due to the approximation we have used, see Fig. \[fig:decrease\](b).
The result of Eq. (\[eq:largeGammaFinal\]) can be understood more intuitively as following: For large $\tau$, the probability that the type-A task is not executed becomes of the order of $1$, i.e. $\prod_{t=0}^{\tau-1}[1-x(t)+\epsilon]\approx \mathcal{O}(1)$. Therefore, the only factor relevant to determine $P(\tau)$ is the probability of the task being executed at time step $\tau$. Thus, $$P(\tau)\sim [x(\tau)+\epsilon]e^{-\tau/\tau_c} \approx [\tau^{-\gamma}+\epsilon]e^{-\tau/\tau_c},$$ which implies $\alpha=\gamma$. For the stochastic case we find that the second term in the result is only a function of $\epsilon$ not coupled to $\tau$. That is, the waiting time distribution does not follow a typical power-law with an exponential cutoff as $\tau^{-\alpha}e^{-\tau/\tau_c}$, but shows a bimodal combination of the power-law and exponential distributions. Similar bimodal distribution has been empirically observed for the inter-event time distributions of Short Messages in mobile phone communication [@Wu2010].
Power-law increasing priority {#subsec:increase}
-----------------------------
In this Subsection, we consider the case of the priority of the task increasing as a power-law with time, starting from $x(0)=0$. The priority can increase up to some bounded value, which we call the *bounded* case. Here the priority of type-A task is given by $$x(t)= \begin{cases}
x_0\left(\frac{t}{d}\right)^\beta & \textrm{if $t<d$,}\\
x_0 & \textrm{if $t \geq d$,}
\end{cases}
\label{eq:increase}$$ where the scale factor $d$ plays the role of a deadline given to the task. $x_0$ can be interpreted as an intrinsic priority of the task and its value is drawn from the uniform distribution of $(0,1)$.
On the other hand, the priority of the type-A task can increase up to the maximum possible value of $1$, which we call the *unbounded* case. Once the priority reaches $1$, the task is inevitably executed. In this case, the priority $x(t)$ is given by $$x(t)= \begin{cases}
x_0\left(\frac{t}{d}\right)^\beta & \textrm{if $t<t_c$,}\\
1 & \textrm{if $t\geq t_c$,}
\end{cases}
\label{eq:increase1}$$ where $t_c\equiv x_0^{-1/\beta}d$ is defined by the condition $x(t_c)=1$.
![The case with power-law increasing priority as $x(t)=x_0 (\tfrac{t}{d})^\beta$ for deterministic version with $p=1$ (a) and for stochastic version with $p=0.999$ (b). The value of $x(t)$ can increase to $x_0$ in the bounded case, where $x(t)=x_0$ for $t\geq d$. On the other hand the value of $x(t)$ can increase to $1$ in the unbounded case. For the numerical results we used up to $10^9$ samples. In (b) the solid lines for $\beta=2$ represent the analytic solutions of Eq. (\[eq:smallTau\]) for $\tau<30$ and Eq. (\[eq:largeTau\]) for $\tau>150$, respectively.[]{data-label="fig:increase"}](fig4a-b.pdf){width=".9\columnwidth"}
For both the bounded and unbounded cases, we can adopt the approximation used in Eq. (\[eq:expand\]) in the limit of $\tau\ll d$, resulting in $$\begin{aligned}
P(\tau) &\approx& p^{\tau+1}\int_0^1 dx_0 \left[1+\epsilon\tau-x_0 \sum_{t=0}^{\tau-1} \left(\tfrac{t}{d}\right)^\beta \right] \left[x_0 \left(\tfrac{\tau}{d}\right)^\beta+\epsilon\right] \nonumber\\
&\approx& p^{\tau+1}\big[ \tfrac{1}{2}(\tfrac{\tau}{d})^\beta +
\tfrac{\epsilon d}{2}(\tfrac{\tau}{d})^{\beta+1}
-\tfrac{\epsilon d}{2(\beta+1)}(\tfrac{\tau-1}{d})^{\beta+1} \nonumber\\
&& -\tfrac{d}{3(\beta+1)} (\tfrac{\tau}{d})^{\beta} (\tfrac{\tau-1}{d})^{\beta+1}
+\epsilon(1+\epsilon\tau)\big].\label{eq:smallTau}\end{aligned}$$ In the limit of $\epsilon\tau\ll 1$, the above equation can be approximated as $$\begin{aligned}
P(\tau) &\approx&\left[\tfrac{1}{2}(\tfrac{\tau}{d})^{\beta}+\epsilon\right]e^{-\tau/\tau_c(p)}.\end{aligned}$$ For the range of $\epsilon^{1/\beta}\ll \tfrac{\tau}{d}\ll 1$, the waiting time distribution is dominated by the first term, indicating a power-law increasing behavior as $P(\tau)\sim \tau^{\alpha'}$ with $\alpha'=\beta$. This solution is compared to the numerical results for both the deterministic and stochastic versions of the model, where we set $d=100$, see Fig. \[fig:increase\].
On the other hand, in the limit of $\tau \gg d$, the waiting time distributions decay as a power-law, but with different tail exponents for bounded and unbounded cases. In the bounded case, when $t\geq d$, the priority is fixed as $x(t)=x_0$, corresponding to the fixed priority model. Therefore, we get $\alpha=2$, which is independent of the values of $\beta$ and $d$. To obtain the distribution for the unbounded case, we use the approximations that $x(t)\ll 1$ if $t<d$ and that $x(t)\approx x(\tau)$ if $t\geq d$. In addition, the type-A task always gets executed as the priority reaches $1$, implying that $x(\tau)+\epsilon\leq 1$. This means that the value of $x_0$ has an upper bound of $x_c$ given by $x_c (\tfrac{\tau}{d})^\beta+\epsilon=1$. Thus the upper limit of the integral range in Eq. (\[eq:P\_tau\]) is also limited to $x_c$, leading to $$\begin{aligned}
P(\tau) &\approx& p^{\tau+1}\int_0^{x_c} dx_0 \left[1+d\epsilon-x_0
\sum_{t=0}^{d-1} \left(\tfrac{t}{d}\right)^\beta \right] \nonumber \\
&& \times \left[ 1-x_0(\tfrac{\tau}{d})^\beta+\epsilon \right]^{\tau-d} \left[ x_0 \left(\tfrac{\tau}{d}\right)^\beta+\epsilon \right] \label{eq:largeTau}\nonumber\\
&=&
\tfrac{(1+\epsilon)^{\tau-d+1}\epsilon(1+d\epsilon)-(2\epsilon)^{\tau-d+1}\left[1+d\epsilon
-A(1-\epsilon)/a_\tau\right]}{a_\tau(\tau-d+1)} \nonumber \\
&& + \tfrac{(1+\epsilon)^{\tau-d+2}\left[(1+d\epsilon)a_\tau-A\epsilon\right]-(2\epsilon)^{\tau-d+2}\left[(1+d\epsilon)a_\tau-A(2-\epsilon)\right]}{a^2_\tau (\tau-d+1)(\tau-d+2)}\nonumber \\
&& - \tfrac{2A\left[(1+\epsilon)^{\tau-d+3}-(2\epsilon)^{\tau-d+3}\right]}{a^2_\tau(\tau-d+1)(\tau-d+2)(\tau-d+3)},\end{aligned}$$ where $a_\tau\equiv (\tfrac{\tau}{d})^\beta$ and $A\equiv\tfrac{d}{\beta+1}(\tfrac{d-1}{d})^{\beta+1}$. The above equation can be approximated in the limit of $\tau\gg 1$ as $$\begin{aligned}
P(\tau) &\approx& d^\beta(\tau^{-(\beta+2)}+2\epsilon \tau^{-(\beta+1)})e^{-\tau/\tau_c(p)}.\end{aligned}$$ In the scaling regime, i.e. $\epsilon\tau\ll 1$, we find that $P(\tau) \sim\tau^{-\alpha}$ with $\alpha=\beta+2$. We performed the numerical simulations with various values of $\beta$ for both deterministic and stochastic cases. In Fig. \[fig:increase\] we find that the simulations confirm the analytic solutions for both deterministic and stochastic cases for the range of $\tau<d$. The range of $\tau>d$ in the stochastic case is only qualitatively matched, which is again due to the approximations we have used.
Empirical Analysis {#sec:empirical}
==================
The results of the models for different cases of time-varying priorities in Section \[sec:model\] can be applied to understand the empirical distributions of the updating times of papers that have appeared in arXiv.org (arXiv in short) and of the processing times of papers published in Physical Review journals (PR in short). The essential ingredients of the empirical setup include the following: (i) An agent has a given task of interest, the execution of which gets postponed due to the other tasks and (ii) the priority of the task varies with time.
Description of the datasets
---------------------------
The first dataset consists of all the papers that appeared on arXiv between January 1995 and March 2010. For each paper we consider its *updating time*, defined as the time interval in days between submission and the final update. The papers that were not updated have been discarded in our analysis. This dataset contains 185151 papers. We divide this dataset into three groups depending on the number of authors in each paper, denoted by $n$. The groups of $n=1$, $n=2$, and $n\geq 3$ consist of 55382, 56572, and 73197 papers, respectively. The papers with more authors have the smaller average updating time: 220.8, 188.5, and 160.0 days for groups of $n=1$, $n=2$, and $n\geq 3$.
The second dataset consists of papers that appeared in PR between January 2000 and December 2009. For each paper we consider its *processing time*, defined as the time interval in days between submission and publication. We have retained only the research articles and have removed articles of other types such as errata and publisher’s notes, which often have different processing times. The exact publication dates of papers printed before 2000 were not available and hence were ignored. This dataset consists of 157484 papers. We divide it into two 5-year periods, i.e. sets of papers appearing for 2000-2004 and for 2005-2009, respectively. The average updating time decreased from $200.3$ days for 2000-2004 to $163.9$ days for 2005-2009. We would like to note that many researchers submit their papers to arXiv at the same time while submitting them to journals and update the arXiv versions on acceptance of the article by the journal. Both the updating time and the processing time are denoted by $\tau$.
![(a) Probability distributions of updating times for papers by single authors ($n=1$), two authors ($n=2$), and more than two authors ($n\geq 3$) in arXiv.org. The solid line indicates the bimodal curve $(\tau^{-\alpha}+c)e^{-\tau/\tau_c}$ fitting to the group of $n=2$. The estimated values of parameters are provided in the text. (b) Waiting time distributions of the stochastic model with power-law decreasing priority in Subsection \[subsec:decrease\] and its extended versions to $n$ agents. Here $\tau^{(n)}$ is defined as $\min\{\tau_i\}_{i=1,\cdots,n}$ with each $\tau_i$ drawn from the distribution of the single agent model.[]{data-label="fig:empiricalArxiv"}](fig5a-b.pdf){width=".9\columnwidth"}
![(a) Probability distributions of processing times in Physical Review journals for different periods. We fit the increasing part and the tail part of the distributions with the curves $(\tau^{\alpha'}+c')e^{-\tau/\tau_c}$ and $\tau^{-\alpha}e^{-\tau/\tau_c}$, respectively. The fitting curve for the tail of 2000-2004 is not shown for clear presentation. The estimated values of parameters are provided in the text. (b) Waiting time distributions of the stochastic model with power-law increasing priority in Subsection \[subsec:increase\] and its extended versions to $n$ agents. Here $\tau^{(n)}$ is defined as $\max\{\tau_i\}_{i=1,\cdots,n}$ with each $\tau_i$ drawn from the distribution of the single agent model.[]{data-label="fig:empiricalPR"}](fig6a-b.pdf){width=".9\columnwidth"}
Updating time distributions in arXiv.org
----------------------------------------
We plot the updating time distributions of arXiv papers for groups of $n=1$, $n=2$, and $n\geq 3$, as depicted in Fig. \[fig:empiricalArxiv\](a). In all cases, the initial part of the distribution ($\tau<50$) shows a power-law decaying behavior, which is followed by an exponentially decaying tail. The distribution also shows a hump around $\tau=100$ indicating its bimodal nature. These distributions can be fitted by a simplified bimodal curve of the form $P(\tau) \propto (\tau^{-\alpha}+c)e^{-\tau/\tau_c}$. We ignored very large updating times ($\tau>2000$) due to the inconclusive fitting results. The parameter values are estimated as $\alpha \approx 0.76(7)$ and $\tau_c \approx 234(29)$ for $n=1$ group, $\alpha \approx 0.99(13)$ and $\tau_c \approx 189(17)$ for $n=2$ group, and $\alpha \approx 1.16(15)$ and $\tau_c \approx 166(7)$ for $n\geq 3$ group, respectively. The fitted curve for $n=2$ group is plotted in Fig. \[fig:empiricalArxiv\](a). It is found that the value of $\alpha$ is slightly larger for the papers with more authors. The cutoff updating times $\tau_c$ are consistent with the corresponding average updating times within error bars.
Let us consider the updating mechanism of the papers submitted to arXiv. Immediately after submission the authors are most alert and are likely to find some anomalies in their paper and subsequently update it. As time passes, they get more involved with other projects and the probability to find irregularities decreases, leading to a decreasing priority of the updating task. In addition, there is a finite probability that the authors might receive a referee report, and hence will update the paper with a revised version even when the priority of the updating task is smaller than the other tasks. Thus, the decreasing priority and the stochastic selection of the updating task can qualitatively describe the bimodal updating time distributions in arXiv. However, for the papers by more than one author the interaction among authors is also an important factor for determining the updating time. In this case, we consider the generalized model with $n$ agents. One of the authors who first finds the anomalies in the paper is assumed to update the paper. Then, the resultant updating time is determined by the shortest updating time as $\tau^{(n)}\equiv\min\{\tau_i\}_{i=1,\cdots,n}$, where each $\tau_i$ is drawn from the distribution of the single agent model in Subsection \[subsec:decrease\]. The distribution of $\tau^{(n)}$ reads $$P(\tau^{(n)})=c_nP(\tau_1=\tau^{(n)})\prod_{i=2}^n P(\tau_i\geq \tau^{(n)}),
%P(\tau^{(n)})=\sum_{\tau_1,\cdots,\tau_n}\prod_{i=1}^n P(\tau_i)\cdot \delta \Big(\tau^{(n)}-\sum_{i=1}^n\tau_i\Big).$$ where $c_n$ is the normalization constant. We numerically obtain $P(\tau^{(n)})$ from the numerical results of $P(\tau_i)$ of the single agent model for small values of $n$, as shown in Fig. \[fig:empiricalArxiv\](b). The decreasing average updating time according to $n$ is observed as in the case of empirical results. The power-law exponent for the range of small updating times, denoted by $\alpha_{\rm min}$, tends to increase with $n$. This tendency is also consistent with the empirical observation apart from the exact values, while the discrepancy in the values of power-law exponent would require some additional or different approach beyond the current version of our models.
Processing time distributions in Physical Review journals
---------------------------------------------------------
Next, we plot the processing time distributions of PR papers for periods 2000-2004 and 2005-2009, as depicted in Fig. \[fig:empiricalPR\](a). In both cases, when compared to the bimodality found in the arXiv case, the distributions are unimodal with peaks at $\tau\approx 150$ and $120$ for 2000-2004 and 2005-2009, respectively. We fit the tail of distribution, by using the functional form of $P(\tau)\propto \tau^{-\alpha}e^{-\tau/\tau_c}$. We estimate $\alpha\approx 2.7(3)$ and $\tau_c \approx 234(27)$ for 2000-2004 and $\alpha\approx 3.2(3)$ and $\tau_c \approx 337(38)$ for 2005-2009, respectively. While the values of $\alpha$ are the same within error bars, the value of $\tau_c$ increases from 2000-2004 to 2005-2009. Then we fit the increasing part of distribution with the form $P(\tau)\propto (\tau^{\alpha'}+c')e^{-\tau/\tau_c}$, where we have assumed the same value of $\tau_c$ as for the tail. The estimated values are $\alpha'\approx 3.2(2)$ for 2000-2004 and $\alpha'\approx 1.6(1)$ for 2005-2009, respectively. The fitted curves are plotted in Fig. \[fig:empiricalPR\](a).
The publication process of PR can be understood as a simple queuing process from the point of view of the referee. Among many factors affecting the processing of the paper, we consider only the time taken by the referee as the main factor. The other factors, such as the time taken by editors, are considered to be fixed and hence they are expected not to change the shape of the distribution but only to affect the location of the peak. The referee assigns some initial priority to the received paper, which may depend upon the quality of the paper and the referee as well. The priority of the reviewing task increases with time, as the journal editor hurries up the referee in case when he or she does not submit the review report within a stipulated period. The referees may also select the reviewing task randomly, independent of the priority. Hence the simple mechanism of the power-law increasing priority of the referee can describe the unimodal processing time distributions in the case of PR. As the number of referees per paper varies, we can consider the generalized model with $n$ independent referees. The processing time of a paper is determined by the longest reviewing time, i.e. $\tau^{(n)}\equiv \max\{\tau_i\}_{i=1,\cdots,n}$. Here each $\tau_i$ is drawn from the distribution of the single agent model in Subsection \[subsec:increase\]. The distribution of $P(\tau^{(n)})$ reads $$P(\tau^{(n)})=c_nP(\tau_1=\tau^{(n)})\prod_{i=2}^n P(\tau_i\leq \tau^{(n)}),$$ where $c_n$ is the normalization constant. When $\tau^{(n)}\ll d$, since $P(\tau_i)\sim \tau_i^{\alpha'}$ for each $i$, the leading term of $P(\tau^{(n)})$ becomes $\tau^{(n)\alpha'_{\rm max}}$ with $\alpha'_{\rm max}=n\alpha'+n-1$. On the other hand, if $\tau^{(n)}\gg d$, $P(\tau_i\leq \tau^{(n)})$ is of the order of $1$ due to the unimodality of the distribution. Thus, the tail of the distribution is described by the same exponent as in the single agent model, i.e. $\alpha_{\rm max}=\alpha$, which is independent of $n$. Then, by means of $\alpha=\alpha'+2$ obtained in the unbounded case of the single agent model, we get the following equation: $$n=\frac{\alpha'_{\rm max}+1}{\alpha_{\rm max}-1}.$$ Since the information about the numbers of referees in PR is not accessible, we infer the *effective* number of referees by plugging the empirical values of power-law exponents $\alpha$ and $\alpha'$ into $\alpha_{\rm max}$ and $\alpha'_{\rm max}$, respectively. We obtain $n\approx 2.5$ for 2000-2004 and $n\approx 1.2$ for 2005-2009, respectively. The decreasing value of $n$ is also consistent with the decreasing average processing time, in a sense that the longest reviewing time will be shorter in case of the fewer referees. The additional factors, such as the number of revisions needed and the time taken by the authors for revision, can be considered in an extended version of the model.
Conclusion {#sec:conclusion}
==========
Various forms of waiting time distributions have been observed in the human dynamics, including the bimodal and the unimodal distributions with power-law or exponentially decaying tails. To figure out the origin of various types of bursty correlations, we have studied a single agent queuing model with two tasks: one is the task of interest whose priority algebraically increases or decreases with time, and the other is the random task with the random priority. If we choose the model to be deterministic, only the highest priority task will be executed. In the stochastic version, however, one of the tasks in the queue is randomly selected for execution with small probability. In all the cases we obtain the approximate analytic solutions for the waiting time distributions of the task of interest, which are confirmed by the extensive numerical simulations.
For the case of power-law decreasing priority with exponent $\gamma$, we find that the power-law exponent of the waiting time distribution $P(\tau)\sim\tau^{-\alpha}$ shows the non-monotonous behavior as $\gamma$ increases, i.e. $\alpha=2-\gamma$ for $\gamma\ll 1$ and $\alpha=\gamma$ for $\gamma\gg 1$. The stochastic case of the model leads to both the bimodality and the finite cutoff of distributions. For the case of power-law increasing priority with exponent $\beta$, the resulting waiting time distributions increase as power-law $P(\tau)\sim\tau^\beta$ up to some scale factor. The power-law exponent $\alpha$ of tails turns out to be $2$ and $\beta+2$ for the bounded and unbounded case, respectively. The extended versions of our models to many agents are also compared to the empirical distributions of updating times for the papers appeared in arXiv.org (arXiv) and of processing times for the papers published in the Physical Review journals (PR). The updating process in arXiv can be understood as a model where each author’s priority of updating the paper decreases with time as a power-law. The processing time of a paper in PR can be understood by means of the reviewing process, where each referee’s priority of reviewing the paper increases with time.
Compared to the Barabási’s queuing model [@Barabasi2005], in our models the tasks in the queue are assumed to be of different types. The task of interest among them has the time-varying priority, while its execution is affected by the random tasks. Our simplified models for the human dynamics can be extended or generalized to the more complicated and realistic cases. As an example, a researcher working on two ongoing projects with different time-varying priorities can be studied by means of the model with two tasks of interest and one random task in his or her to-do list.
Financial support by Aalto University postdoctoral program (HJ), from EU’s FP7 FET-Open to ICTeCollective Project No. 238597 (KK), and by the Academy of Finland, the Finnish Center of Excellence program 2006-2011, Project No. 129670 (RKP, KK) are gratefully acknowledged.
[19]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [**]{} (, , ) @noop [**]{} (, ) @noop [**]{}, ed. (, ) [****, ()](\doibase 10.1103/RevModPhys.81.591) [****, ()](\doibase 10.1038/nature03459) [****, ()](\doibase
10.1073/pnas.1013140107) [****, ()](\doibase
10.1103/PhysRevE.83.025102) [****, ()](\doibase 10.1038/4371251a) [****, ()](\doibase 10.1073/pnas.0800332105) [****, ()](\doibase 10.1126/science.1174562) [****, ()](\doibase 10.1103/PhysRevLett.95.248701) [****, ()](\doibase
10.1103/PhysRevE.73.036127) [****, ()](\doibase 10.1103/PhysRevE.80.041131) [****, ()](\doibase 10.1103/PhysRevE.75.026102) [****, ()](\doibase 10.1016/j.physa.2008.08.022) [****, ()](\doibase 10.1103/PhysRevE.79.056110) [****, ()](\doibase 10.1371/journal.pone.0022687) [****, ()](\doibase 10.1038/nphys761) @noop [**]{} (, )
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The Josephson ring modulator (JRM) is a device, based on Josephson tunnel junctions, capable of performing non-degenerate mixing in the microwave regime without losses. The generic scattering matrix of the device is calculated by solving coupled quantum Langevin equations. Its form shows that the device can achieve quantum-limited noise performance both as an amplifier and a mixer. Fundamental limitations on simultaneous optimization of performance metrics like gain, bandwidth and dynamic range (including the effect of pump depletion) are discussed. We also present three possible integrations of the JRM as the active medium in a different electromagnetic environment. The resulting circuits, named Josephson parametric converters (JPC), are discussed in detail, and experimental data on their dynamic range are found to be in good agreement with theoretical predictions. We also discuss future prospects and requisite optimization of JPC as a preamplifier for qubit readout applications.'
author:
- Baleegh Abdo
- Archana Kamal
- Michel Devoret
title: 'Non-degenerate, three-wave mixing with the Josephson ring modulator'
---
The photon energy of microwave radiation in the band from $4-8$ GHz ($\sim$ $8-4$ cm wavelength) is approximately $10^{5}$ smaller than that of the visible light. Yet, at a temperature $10^{4}$ smaller than room temperature, now routinely achievable with a dilution refrigerator, it is now possible to resolve the energy of single microwave photons [@DSnature]. There are three advantages of single photon microwave electronics when compared with quantum optics. First, signal shapes at carrier frequencies of a few GHz with a relative bandwidth of few percent can be controlled with much greater relative precision than their equivalent at a few hundreds of THz. This is partly due to the fact that microwave generators have more short term stability than lasers, but also because microwave components are mechanically very stable, particularly when cooled, compared with traditional optical components. Second, in single photon microwave electronics, the on-chip circuitry can be well in the lumped element regime, and spatial mode structure can be controlled more thoroughly and more reliably than in the optical domain. Finally, there exists a simple, robust non-dissipative component, the Josephson tunnel junction (JJ), whose non-linearity can be ultra-strong even at the single photon level [@MDphysik]. Many quantum signal processing functions have been realized using JJs, both digital and analog, and this short review will not attempt to describe all of them. We will focus on analog Josephson devices pumped with a microwave tone. They recently led to microwave amplifiers working at the single photon level [@CastellanosNat; @JPCnature]. These novel devices have taken the work pioneered by B. Yurke at Bell labs 25 years ago [@YurkePRL; @ParamYurkePRA; @MovshovichPRL] to the point where actual experiments can be performed using Josephson amplifiers as the first link in the chain of measurement [@TeufelNatTech; @QuantumJumps; @QubitJPC].
In this paper, we address one particular subclass of analog signal processing devices based on Josephson tunnel junction, namely those performing non-degenerate three-wave mixing. Examples are Josephson circuits based on the Josephson ring modulator [@JPCnaturePhys; @Jamp] which we will describe below. The Hamiltonian of such a device is of the form $$\begin{aligned}
H_{0} & = & \frac{1}{2}\left( \frac{P_{X}^{2}}{\mathcal{M}_{X}}
+\frac{P_{Y}^{2}}{\mathcal{M}_{Y}}
+\frac{P_{Z}^{2}}{\mathcal{M}_{Z}}\right)
\nonumber\\
& & + \frac{1}{2}\left( \mathcal{K}_{X}X^{2}+\mathcal{K}_{Y}Y^{2} +\mathcal{K}_{Z}Z^{2}\right) +KXYZ, \label{Hamiltonian1}$$ where ($X$, $Y$, $Z$) and ($P_{X}$, $P_{Y}$, $P_{Z}$) are the generalized position and momentum variables for the three independent oscillators, $\mathcal{M}_{X,Y,Z}$ and $\mathcal{K}_{X,Y,Z}$ represent the “mass" and “spring constant" of the relevant oscillator (see table I), and $K$ is the three-wave mixing constant which governs the non-linearity of the system. We will discuss later how such simple minimal non-linear term can arise. The classical equation of motions for the standing waves in such a device are symmetric and are given by: $$\begin{aligned}
\overset{\cdot\cdot}{X}+\gamma_{a}\overset{\cdot}{X}+\omega_{a}^{2} X+K^{\prime}YZ & =x\left( t\right) \cos\omega_{a}t,
\label{non-degenerate1}\\
\overset{\cdot\cdot}{Y}+\gamma_{b}\overset{\cdot}{Y}+\omega_{b}^{2} Y+K^{\prime}XZ & =y\left( t\right) \cos\omega_{b}t,
\label{non-degenerate2}\\
\overset{\cdot\cdot}{Z}+\gamma_{c}\overset{\cdot}{Z}+\omega_{c}^{2}Z+K^{\prime}XY & =z\left( t\right) \cos\omega_{c}t,
\label{non-degenerate3}$$ where $K^{\prime}=K/\mathcal{M}$ (we assume, for simplicity, equal masses $\mathcal{M}_{X,Y,Z}=\mathcal{M}$) and $\omega_{a,b,c}$ $=$ $\sqrt
{\mathcal{K}_{X,Y,Z}/\mathcal{M}}$ are the angular resonant frequencies of the three coordinates satisfying $$\omega_{a}<\omega_{b}\,<\omega_{c}=\omega_{a}+\omega_{b}.$$
-------------------------------------------------
$X $ $P$ $\mathcal{M}$ $\mathcal{K}$
-------- -------- --------------- ---------------
$\Phi$ $Q$ $C$ $L^{-1}
$
$Q$ $\Phi$ $L$ $C^{-1}
$
-------------------------------------------------
\[Table1\]
We also suppose the oscillators are well in the underdamped regime $$\begin{aligned}
\gamma_{a} & \ll\omega_{a},\\
\gamma_{b} & \ll\omega_{b},\\
\gamma_{c} & \ll\omega_{c},\end{aligned}$$ a sufficient but not strictly necessary hypothesis, which has the principal merit of keeping the problem analytically soluble under the conditions of interest. It is worth noting that the system is non-degenerate both spatially and temporally. On the other hand, it is important to suppose that the envelope functions $x(t)$, $y\left( t\right) $ and $z\left( t\right) $ of the drive signals are supposed to be slow compared to the respective drive frequencies $\omega_{b}-\omega_{a}\gg\gamma_{a}+\gamma_{b}$.
The equations (\[non-degenerate1\]-\[non-degenerate3\]) must be contrasted with that of a degenerate three-wave mixing device for which two cases are possible. In the first case, where the $Y$ and $Z$ degrees of freedom have merged into a single oscillator, the Hamiltonian has a non-linear term of the form $KXZ^{2}$ and the equations read: $$\begin{aligned}
\overset{\cdot\cdot}{X}+\gamma_{a}\overset{\cdot}{X}+\omega_{a}^{2}X+K^{\prime}Z^{2} & =x\left( t\right) \cos\omega_{a}t,\\
\overset{\cdot\cdot}{Z}+\gamma_{c}\overset{\cdot}{Z}+\omega_{c}^{2}Z+2K^{\prime}ZX & =z\left( t\right) \cos\omega_{c}t.\end{aligned}$$ This is the case of electromechanical resonators [@TeufelNature] in which one of the capacitance plates of a microwave oscillator ($Z$) is itself the mass of a mechanical resonator ($X$). There $\omega_{c}\gg\omega_{a}$, and pumping the microwave oscillator in the vicinity of $\omega_{c}-\omega_{a}$ leads to cooling of the mechanical oscillator provided $\gamma_{c}\gg
\gamma_{a}$. In the second case, it is the $X$ and the $Y$ degrees of freedom that merge into a single oscillator, leading to a non-linear term in the Hamiltonian of the form $KX^{2}Z$. The equations then read $$\begin{aligned}
\overset{\cdot\cdot}{X}+\gamma_{a}\overset{\cdot}{X}+\omega_{a}^{2}X+2K^{\prime}XZ & =x\left( t\right) \cos\omega_{a}t,
\label{degenerate 2}\\
\overset{\cdot\cdot}{Z}+\gamma_{c}\overset{\cdot}{Z}+\omega_{c}^{2}Z+K^{\prime}X^{2} & =z\left( t\right) \cos\omega_{c}t
\label{degenerate 3}$$ and we have now $$\omega_{c}=2\omega_{a}.
\label{degenerate 1}$$ This case is implemented in Josephson circuits as a dcSQUID whose flux is driven by a microwave oscillating signal at twice the plasma frequency of the SQUID [@Yamamoto]. When $z\left( t\right) =z_{d}\gg K^{\prime}X^{2}$ (so-called stiff“ or non-depleted” pump condition), the system of equations (\[degenerate 2\],\[degenerate 3\]) reduces to the parametrically driven oscillator equation $$\overset{\cdot\cdot}{X}+\gamma_{a}\overset{\cdot}{X}+\omega_{a}^{2}
\left[1+\frac{K^{\prime}z_{d}}{\gamma_{c}\omega_{c}}\sin\left( \omega_{c}t\right)\right] X=x\left( t\right) \cos\omega_{a}t. \label{parametric oscillator}$$ Note that there is, in addition to the parametric drive on the left hand side, a small perturbing drive signal $x\left( t\right) \cos\omega_{a}t$ on the right hand side. The theory of the degenerate parametric amplifier starts with this latter equation, the term $\frac{K^{\prime}z_{d}}{\gamma_{c}\omega_{c}}\sin\left( \omega_{c}t\right) $ corresponding to the pump and $x\left(
t\right) \cos\omega_{a}t$ corresponding to the input signal. The output signal is obtained from a combination of the loss term $\gamma_{a}\dot{X}$ and the input signal.
In the context of Josephson devices, another route to the effective parametric oscillator of equation (\[parametric oscillator\]) can be obtained by a driven, Duffing-type oscillator [@VijayJBAreview; @JBA]. This system (Josephson bifurcation amplifier) has only one spatial mode and quartic non-linearity, $$\overset{\cdot\cdot}{X}+\gamma_{a}\overset{\cdot}{X}+\omega_{a}^{2}X-\lambda
X^{3}=\left[ z_{d}+x\left( t\right) \right] \cos\omega_{d}t.
\label{parametric oscillator2}$$ Driven by a strong tone $z_{d}\cos\omega_{d}t$ in the vicinity of the bifurcation occurring at $$\begin{aligned}
\omega_{d} & =\omega_{a}-\frac{\sqrt{3}}{2}\gamma_{a},\\
z_{d} & =\frac{128}{27}\sqrt{\frac{\gamma_{a}^{3}\omega_{a}}{3\lambda}},\end{aligned}$$ it will lead to an equation of the form (\[parametric oscillator\]) for small deviations around the steady-state solution. It will, therefore, amplify the small drive modulation signal $x\left( t\right) $ of equation (\[parametric oscillator2\]) \[\]. Similar amplifying effects can be found in pumped superconducting microwave resonators without Josephson junctions [@ImBaleegh; @TholenAPL; @Zmuidzinas].
In the following section, we will treat Eqs. (\[non-degenerate1\]-\[non-degenerate3\]) using input-output theory [@LinearScatteringNote] and obtain the quantum-mechanical scattering matrix of the signal and idler amplitudes in the stiff-pump approximation. This allows us to find the photon gain of the device in its photon amplifier mode as a function of the pump amplitude, and the corresponding reduction of bandwidth. We then discuss the implementation of the device using a ring of four Josephson junctions flux-biased at half-quantum in Sec. II. It is the non-dissipative analogue of the semiconductor diode ring modulator [@Pozar]. In Sec. III, we treat the finite amplitude of signals and establish useful relations between the dynamic range, gain and bandwidth. In Sec. IV we introduce the Josephson parametric converter (JPC) as an example of a non-degenerate, three-wave mixing device operating at the quantum limit. We present three different realizations schemes for the JPC and point out their practical advantages and limitations. In Sec. V we present experimental results for different JPC devices and compare the data with the maximum bounds predicted by theory. We follow this with a discussion, in Sec. VI, of general requirements for an amplifier to meet the needs of qubit readout and how the maximum input power of the device can be increased by two orders of magnitude beyond typical values achieved nowadays. We conclude with a brief summary of our results in Sec. VII.
Input-output treatment of a generic non-degenerate, three-wave mixing device
============================================================================
The three oscillators of Eqs. (\[non-degenerate1\]-\[non-degenerate3\]) correspond to three quantum LC oscillators coupled by a non-linear, trilinear mutual inductance, whose mechanism we will discuss in the next section. They are fed by transmission lines which carry excitations both into and out of the oscillators, as shown on Fig. \[three\_osc\_fig\]. The Hamiltonian of the system is (leaving out the transmission lines for the moment),
![General non-degenerate three-wave mixing device consisting of three LC oscillators coupled by a non-linear medium, giving a trilinear term in the Hamiltonian of the form $K\Phi_{a}\Phi_{b}\Phi_{c}$ where the fluxes $\Phi_{a,b,c}$ are those of the inductors. Each oscillator is fed by a transmission line with characteristic impedance $R_{a,b,c}$.[]{data-label="three_osc_fig"}](three_oscillators.pdf "fig:"){width="\columnwidth"}\
$$\begin{aligned}
\frac{H_{0}}{\hbar} & =\omega_{a}a^{\dag}a+\omega_{b}b^{\dag}b+\omega_{c}c^{\dag}c\nonumber\\
& +g_{3}\left( a+a^{\dag}\right) \left( b+b^{\dag}\right) \left(c+c^{\dag}\right) ,\end{aligned}$$
where $a$, $b$ and $c$ are the annihilation operators associated with each of the three degrees of freedom. Their associated angular frequencies are given in terms of the inductances and capacitances as $$\omega_{a,b,c}=\frac{1}{\sqrt{L_{a,b,c}C_{a,b,c}}}.$$ The bosonic operators of different modes (a, b, c) commute with each other and those associated with the same mode satisfy the usual commutation relations of the form $$\left[ a,a^{\dag}\right] =1.$$ The link between the mode amplitude such as $X$, which represents the flux through the inductance of the oscillator, and a quantum operator such as $a$ can be written as, $$X=X^{ZPF}\left( a+a^{\dag}\right),$$ where “ZPF" stands for “zero-point fluctuations" and $$\begin{aligned}
X^{ZPF} & =\sqrt{\frac{\hbar Z_{a}}{2}},\\
Z_{a} & =\sqrt{\frac{L_{a}}{C_{a}}},\end{aligned}$$ the last equation defining the impedance of the oscillator, equal to the modulus of the impedance on resonance of either the inductance or the capacitance. The link between $K$ and $g_{3}$ is therefore $$\hbar g_{3}=KX^{ZPF}Y^{ZPF}Z^{ZPF}.$$ We now work in the framework of Rotating Wave Approximation (RWA), in which we only keep terms commuting with the total photon number $$\frac{H_{0}^{\mathrm{RWA}}}{\hbar}
=\omega_{a}a^{\dag}a+\omega_{b}b^{\dag}b+\omega_{c}c^{\dag}c+g_{3}\left( a^{\dag}b^{\dag}c+abc^{\dag}\right).$$ Treating in RWA the coupling of each oscillator with a transmission line carrying waves in and out of the oscillator (see Appendix for complements of the next 6 equations), one arrives at three coupled quantum Langevin equations for $a\left(t\right)$, $b\left(t\right)$ and $c\left( t\right)$: $$\begin{aligned}
\frac{\mathrm{d}}{\mathrm{d}t}a
& = -i\omega_{a}a-ig_{3}b^{\dag}c-\frac{\gamma_{a}}{2}a+\sqrt{\gamma_{a}}\tilde{a}^{\mathrm{in}}\left(t\right),
\nonumber\\
\frac{\mathrm{d}}{\mathrm{d}t}b
& = -i\omega_{b}b-ig_{3}a^{\dag}c-\frac{\gamma_{b}}{2}b+\sqrt{\gamma_{b}}\tilde{b}^{\mathrm{in}}\left(t\right),
\nonumber\\
\frac{\mathrm{d}}{\mathrm{d}t}c & =-i\omega_{c}c-ig_{3}ab-\frac{\gamma_{c}}{2}c+\sqrt{\gamma_{c}}\tilde{c}^{\mathrm{in}}\left(t\right),
\label{threeAmpEqs}\end{aligned}$$ In these equations, the second term in the right hand side corresponds to the non-linear term producing photon conversion. The third term says that photons introduced in one resonator leave with a rate $$\gamma_{a,b,c}=\omega_{a,b,c}\frac{Z_{a,b,c}}{R_{a,b,c}},$$ with the resistances $R_{a,b,c}$ denoting the characteristic impedances of the transmission lines. Finally, in the fourth term of the Langevin equations, the input fields such as $\tilde{a}^{\mathrm{in}}\left(t\right)$ correspond to the negative frequency component of the drive terms in the classical equations. They obey the relation $$\tilde{a}^{\mathrm{in}}(t)=\frac{1}{\sqrt{2\pi}}\int_{0}^{+\infty}a^{\mathrm{in}}\left[ \omega\right] e^{-i\omega t}\mathrm{d}\omega,$$ where $a^{\mathrm{in}}[\omega]$ are the usual field operators obeying the commutation relations $$\left[a^{\mathrm{in}}\left[ \omega\right] ,a^{\mathrm{in}}\left[\omega^{\prime}\right] \right]
=\mathrm{sgn}\left(\frac{\omega-\omega^{\prime}}{2}\right) \delta\left(\omega+\omega^{\prime}\right)$$ in which $\omega$ denotes a frequency that can be either positive or negative. The transmission lines thus both damp and drive the oscillators. The incoming field operator treats the drive signals and the Nyquist equilibrium noise of the reservoir on the same footing. Photon spectral densities $\mathcal{N}^{\mathrm{in}}[\omega]$ of the incoming fields, introduced by relations of the form $$\left\langle \left\{a^{\mathrm{in}}\left[\omega\right], a^{\mathrm{in}}\left[\omega^{\prime}\right]\right\}\right\rangle
=2\mathcal{N}_{a}^{\mathrm{in}}\left[ \frac{\omega-\omega^{\prime}}{2}\right]\delta\left(\omega+\omega^{\prime}\right),
\label{Na_in_first}$$ have the value $$\begin{aligned}
\mathcal{N}_{a}^{\mathrm{in}}\left[ \omega\right]
& = \frac{\mathrm{sgn}\left( \omega\right) }{2}\coth\left( \frac{\hbar\omega}{2k_{B}T}\right)
\nonumber\\
& + 2\pi P_{a}^{\mathrm{in}}\left[ \delta\left( \omega-\omega_{1}\right)+\delta\left( \omega+\omega_{1}\right) \right],
\label{Na_in_sec}$$ where $P_{a}^{\mathrm{in}}$ is the photon flux of the incoming drive signal at angular frequency $\omega_{1}$ (in units of photons per unit time) and $T$ is the temperature of the electromagnetic excitations of the line. Note that the dimensionless function $\mathcal{N}_{a}^{\mathrm{in}}\left[ \omega\right] $ is defined for both positive and negative frequencies. It is symmetric $\mathcal{N}_{a}^{\mathrm{in}}\left[ \omega\right] =\mathcal{N}_{a}^{\mathrm{in}}\left[ -\omega\right] $ and its value at frequency $\left\vert \omega\right\vert $ represents the average number of photons per unit time per unit bandwidth in the incoming signal, which in the high temperature limit is $k_{B}T/\left( \hbar\left\vert \omega\right\vert
\right) $. It includes the $\frac{1}{2}$ contribution of zero-point quantum noise.
It is worth insisting that we treat the non-linear coupling strength as a perturbation compared with the influence of the reservoirs, treated themselves as a perturbation compared with the Hamiltonian of the oscillators: $$g_{3}\ll\gamma_{a},\gamma_{b}<\gamma_{c}\ll\omega_{a},\omega_{b}<\omega_{c}=\omega_{a}+\omega_{b}.$$ In general, only one strong drive tone is applied to one of the resonators and is called the “pump". Two cases must then be distinguished at this stage, as shown in Fig. \[frequencies\]:
Case 1 (amplification and frequency conversion with photon gain): the pump tone is applied to the $c$ resonator. The device is usually used as an amplifier [@JPCnature; @Jamp]. It can also be used as a two-mode squeezer [@JPCSqueezer].
Case 2 (noiseless frequency conversion without photon gain): the pump tone is applied to either the $a$ or $b$ resonator [@BSconv]. The device is useful as a noiseless up- and down-converter and can perform dynamical cooling of the lowest energy oscillator, transferring its spurious excitations to the highest frequency one, which is more easily void of any excitations and plays the role of a cold source.
\[h\]
![Characteristic frequency landscape of non-degenerate three-wave mixing devices. Three separate oscillators have resonant frequencies $\omega_{a}<\omega_{b}<\omega_{c}=\omega_{a}+\omega_{b}$. They are fed by transmission lines, giving them a full linewidth at half-maximum $\gamma_{a}$, $\gamma_{b}$ and $\gamma_{c}$ respectively. The non-linear coupling strength, expressed in photon amplitude language, is much smaller than these linewidths. The device can be pumped at $\omega_{c}$ and operates then as a phase-preserving amplifier with photon gain for frequencies $\omega_{a}$ and $\omega_{b}$ (top), or it can be pumped at one of the two lower frequencies $\omega_{a}$ or $\omega_{b}$ and operates then as a noiseless frequency converter or dynamical cooler, upconverting signals into oscillator at $\omega_{c}$ (bottom). In this figure, the spectral density of weak signal corresponds to thin arrows whereas the spectral densities of pump signals corresponds to thick arrows.[]{data-label="frequencies"}](converter-frequencies-positive.pdf "fig:"){width="\columnwidth"}\
Photon gain (case 1)
--------------------
We will first suppose that the pump is “stiff", namely $$\begin{aligned}
\left\vert \left\langle \tilde{c}^{\mathrm{in}}\right\rangle \right\vert ^{2}
& \gg1\\
\gamma_{c} & \gg\gamma_{a},\gamma_{b}$$ This means that the pump tone will not be easily depleted despite the fact that its photons are converted into the signal and idler photons at $\omega_{a}$ and $\omega_{b}$. For solving the quantum Langevin equations, we replace the pumped oscillator annihilation operator $c$ by its average value in the coherent state produced by the pump as $$c\left( t\right) \rightarrow\left\langle c\left( t\right) \right\rangle
=\sqrt{\bar{n}_{c}}e^{-i\left( \omega_{c}t+\phi\right) }.$$ The Langevin equations can then be transformed into the linear equations (see equation (\[IOT\]) of Appendix) $$\begin{gathered}
\left[
\begin{array}
[c]{cc}O_{a}^{+} & ig_{b}^{a}e^{-i\omega_{c}t}\\
-ig_{a}^{b\ast}e^{+i\omega_{c}t} & O_{b}^{+\ast}\end{array}
\right] \left[
\begin{array}
[c]{c}\tilde{a}^{\mathrm{out}}\\
\tilde{b}^{\mathrm{out}\dagger}\end{array}
\right] =\\
-\left[
\begin{array}
[c]{cc}O_{a}^{-} & ig_{b}^{a}e^{-i\omega_{c}t}\\
-ig_{a}^{b\ast}e^{+i\omega_{c}t} & O_{b}^{-\ast}\end{array}
\right] \left[
\begin{array}
[c]{c}\tilde{a}^{\mathrm{in}}\\
\tilde{b}^{\mathrm{in}\dagger}\end{array}
\right] ,\end{gathered}$$ where $$\begin{aligned}
O_{a,b}^{\pm} & =\frac{\mathrm{d}}{\mathrm{d}t}+i(\omega_{a,b}\mp
i\Gamma_{a,b}),\\
\Gamma_{a,b} & =\frac{\gamma_{a,b}}{2},\\
g_{b,a}^{a,b} & =g_{3}\sqrt{\bar{n}_{c}}e^{-i\phi}\sqrt{\frac{\Gamma_{a,b}}{\Gamma_{b,a}}}.\end{aligned}$$ After a Fourier transform, we obtain in the frequency domain, a simpler relation $$\begin{gathered}
\left[
\begin{array}
[c]{cc}h_{a}\left[ \omega_{1}\right] & +ig_{b}^{a}\\
-ig_{a}^{b\ast} & h_{b}^{\ast}\left[ \omega_{2}\right]
\end{array}
\right] \left[
\begin{array}
[c]{c}a^{\mathrm{out}}\left[ +\omega_{1}\right] \\
b^{\mathrm{out}}\left[ -\omega_{2}\right]
\end{array}
\right] =\label{in-out-amp5}\\
\left[
\begin{array}
[c]{cc}h_{a}^{\ast}\left[ \omega_{1}\right] & -ig_{b}^{a}\\
+ig_{a}^{b\ast} & h_{b}\left[ \omega_{2}\right]
\end{array}
\right] \left[
\begin{array}
[c]{c}a^{\mathrm{in}}\left[ +\omega_{1}\right] \\
b^{\mathrm{in}}\left[ -\omega_{2}\right]
\end{array}
\right] ,\end{gathered}$$ where $$h_{a,b}\left[ \omega\right] =-i\omega+i(\omega_{a,b}-i\Gamma_{a,b})$$ and the signal and idler angular frequencies $\omega_{1}$ and $\omega_{2}$ are both positive, satisfying the relationship $$\omega_{1}+\omega_{2}=\omega_{c}.$$ The scattering matrix of the device for small signals is defined by $$\left[
\begin{array}
[c]{c}a^{\mathrm{out}}\left[ +\omega_{1}\right] \\
b^{\mathrm{out}}\left[ -\omega_{2}\right]
\end{array}
\right] =\left[
\begin{array}
[c]{cc}r_{aa} & s_{ab}\\
s_{ba} & r_{bb}\end{array}
\right] \left[
\begin{array}
[c]{c}a^{\mathrm{in}}\left[ +\omega_{1}\right] \\
b^{\mathrm{in}}\left[ -\omega_{2}\right]
\end{array}
\right] . \label{RedS}$$ It can be computed from Eq. (\[in-out-amp5\]) and one finds $$\begin{aligned}
r_{aa} & =\frac{\chi_{a}^{-1\ast}\chi_{b}^{-1\ast}+\left\vert \rho
\right\vert ^{2}}{\chi_{a}^{-1}\chi_{b}^{-1\ast}-\left\vert \rho\right\vert
^{2}},\\
r_{bb} & =\frac{\chi_{a}^{-1}\chi_{b}^{-1}+\left\vert \rho\right\vert ^{2}}{\chi_{a}^{-1}\chi_{b}^{-1\ast}-\left\vert \rho\right\vert ^{2}},\\
s_{ab} & =\frac{-2i\rho}{\chi_{a}^{-1}\chi_{b}^{-1\ast}-\left\vert
\rho\right\vert ^{2}},\\
s_{ba} & =\frac{2i\rho^{\ast}}{\chi_{a}^{-1}\chi_{b}^{-1\ast}-\left\vert
\rho\right\vert ^{2}},\end{aligned}$$ where the $\chi$’s are the bare response functions of modes *a* and *b* (whose inverses depend linearly on the signal frequency) $$\begin{aligned}
\chi_{a}^{-1} & =1-i\frac{\omega_{1}-\omega_{a}}{\Gamma_{a}},\\
\chi_{b}^{-1} & =1-i\frac{\omega_{2}-\omega_{b}}{\Gamma_{b}},\end{aligned}$$ and $\rho$ is the dimensionless pump amplitude $$\rho=\frac{g_{3}\sqrt{\bar{n}_{c}}e^{-i\phi}}{\sqrt{\Gamma_{a}\Gamma_{b}}}.
\label{rho}$$ Note that the matrix in Eq. (\[RedS\]) has unity determinant and the property $$\begin{aligned}
\left\vert r_{aa}\right\vert ^{2}-\left\vert s_{ab}\right\vert ^{2} & =1,\\
\left\vert r_{bb}\right\vert ^{2}-\left\vert s_{ba}\right\vert ^{2} & =1.\end{aligned}$$ For zero frequency detuning, i.e. $\chi_{a}^{-1}=\chi_{b}^{-1}=1$, the scattering matrix displays a very simple form $$\left[
\begin{array}
[c]{cc}\cosh\tau_{0} & -ie^{-i\phi}\sinh\tau_{0}\\
+ie^{+i\phi}\sinh\tau_{0} & \cosh\tau_{0}\end{array}
\right] , \label{hyperbolic}$$ where $\tanh(\tau_{0}/2)=|\rho|$. The zero frequency detuning power gain $G_{0}$ is given by $$G_{0}=\left( \cosh\tau_{0}\right) ^{2}=\left( \frac{1+\left\vert
\rho\right\vert ^{2}}{1-\left\vert \rho\right\vert ^{2}}\right) ^{2}.
\label{gain-formula}$$ For non-zero detuning, the scattering matrix acquires extra phase factors but the minimal scattering matrix for a quantum-limited phase-preserving amplifier represented in Fig. \[minimal-amp\] still describes the device.
\[h\]
![An amplifier reaching the quantum limit must have a minimal scattering matrix, with the signal in port $a$ being reflected with amplitude gain $G^{1/2}$ while the signal in port $b$ is phase-conjugated and transmitted to port $a$ with amplitude gain $(G-1)^{1/2}$. This can be realized in case 1 of Fig. \[frequencies\].[]{data-label="minimal-amp"}](phase_preserving_amplifier.pdf "fig:"){width="\columnwidth"}\
The gain $G_{0}$ diverges as $\left\vert \rho\right\vert \rightarrow1^{-}$, i.e. when the photon number $\bar{n}_{c}$ in the pump resonator reaches the critical number given by $$\bar{n}_{c}^{po}=\frac{\Gamma_{a}\Gamma_{b}}{\left\vert g_{3}\right\vert ^{2}}, \label{onset-param-osc}$$ a result that is common to all forms of parametric amplification. Increasing the pump power beyond the critical power yielding $\bar{n}_{c}^{po}$ leads to the parametric oscillation regime. This phenomenon is beyond the scope of our simple analysis and cannot be described by our starting equations, since higher order non-linearities of the system need to be precisely modelled if the saturation of the oscillation is to be accounted for.
Introducing the detuning $$\Delta\omega=\omega_{1}-\omega_{a}=\omega_{b}-\omega_{2},$$ we can give a useful expression for the gain as a function of frequency as $$G\left( \Delta\omega\right) \underset{\left\vert \rho\right\vert
\rightarrow1^{-}}{=}\frac{G_{0}}{1+\left( \frac{\Delta\omega}{\gamma
G_{0}^{-1/2}}\right) ^{2}},$$ which shows that in the limit of large gain, the response of the amplifier for both the signal and idler port is Lorentzian with a bandwidth given by $$B=2\gamma G_{0}^{-1/2}=\frac{2\gamma_{a}\gamma_{b}G_{0}^{-1/2}}{\gamma
_{a}+\gamma_{b}}. \label{B}$$ The product of the maximal amplitude gain times the bandwidth is thus constant and is given by the harmonic average of the oscillator bandwidths. Another interesting prediction of the scattering matrix is the two-mode squeezing function of the device demonstrated in Ref. .
Conversion without photon gain (case 2)
---------------------------------------
The case of conversion without photon gain can be treated along the same line as in the previous subsection, where scattering takes place between *c* and *a* or *c* and *b* modes. Without loss of generality we assume that the pump is applied to the intermediate frequency resonance. In this case the scattering matrix reads $$\left[
\begin{array}
[c]{c}a^{\mathrm{out}}\left[ +\omega_{1}\right] \\
c^{\mathrm{out}}\left[ +\omega_{3}\right]
\end{array}
\right] =\left[
\begin{array}
[c]{cc}r_{aa} & t_{ac}\\
t_{ca} & r_{cc}\end{array}
\right] \left[
\begin{array}
[c]{c}a^{\mathrm{in}}\left[ +\omega_{1}\right] \\
c^{\mathrm{in}}\left[ +\omega_{3}\right]
\end{array}
\right] ,$$ where $$\begin{aligned}
r_{aa} & =\frac{\chi_{a}^{-1\ast}\chi_{c}^{-1}-\left\vert \rho^{\prime
}\right\vert ^{2}}{\chi_{a}^{-1}\chi_{c}^{-1}+\left\vert \rho^{\prime
}\right\vert ^{2}},\nonumber\\
r_{cc} & =\frac{\chi_{a}^{-1}\chi_{c}^{-1\ast}-\left\vert \rho^{\prime
}\right\vert ^{2}}{\chi_{a}^{-1}\chi_{c}^{-1}+\left\vert \rho^{\prime
}\right\vert ^{2}},\nonumber\\
t_{ac} & =\frac{2i\rho^{\prime}}{\chi_{a}^{-1}\chi_{c}^{-1}+\left\vert
\rho^{\prime}\right\vert ^{2}},\nonumber\\
t_{ca} & =\frac{2i\rho^{\prime\ast}}{\chi_{a}^{-1}\chi_{c}^{-1}+\left\vert
\rho^{\prime}\right\vert ^{2}},\nonumber\\
& \label{SparamsConvYSeries}$$ and $$\begin{aligned}
\chi_{c}^{-1} & =1-i\frac{\omega_{3}-\omega_{c}}{\Gamma_{c}},\\
\rho^{\prime} & =\frac{g_{3}\sqrt{\bar{n}_{b}}e^{-i\phi}}{\sqrt{\Gamma
_{a}\Gamma_{c}}}.\end{aligned}$$ The reduced pump strength $\rho^{\prime}$ plays the same role here as $\rho$ in the photon amplification case. Note that the scattering matrix is now unitary (conservation of total number of photons) and satisfies the following relations: $$\begin{aligned}
\left\vert r_{aa}\right\vert ^{2}+\left\vert t_{ac}\right\vert ^{2} & =1,\\
\left\vert r_{cc}\right\vert ^{2}+\left\vert t_{ca}\right\vert ^{2} & =1.\end{aligned}$$ For zero frequency detuning, i.e. $\chi_{a}^{-1}=\chi_{c}^{-1}=1$, the scattering matrix can be written as $$\left[
\begin{array}
[c]{cc}\cos\tau_{0} & e^{-i\phi}\sin\tau_{0}\\
e^{i\phi}\sin\tau_{0} & \cos\tau_{0}\end{array}
\right] ,$$ which corresponds to replacing the parameter $\tau_{0}$ by $i\tau_{0}$ or $\left\vert \rho\right\vert $ by $i\left\vert \rho\right\vert $ in the scattering matrix (\[hyperbolic\]). A scattering representation of the two-port device in conversion mode is shown in Fig. \[convflow\]. In this mode the device operates as a beam splitter, the only difference being that the photons in different arms have different frequencies [@BSconv]. Full conversion $\left( \sin\tau_{0}=1\right) $ is obtained on resonance when the pump power reaches the critical value. However, here, the critical value can be traversed without violating the validity of the equations. Full photon conversion is desirable in dynamical cooling: in that case, the higher frequency resonator will be emptied of photons, and the lower frequency resonator can be cooled to its ground state by pumping the intermediate frequency resonator (see lower panel of Fig. \[frequencies\]).
\[h\]
![Signal flow graph for a three-wave mixing device operating in conversion without photon gain, realized in case 2 of Fig. \[frequencies\]. The incoming signal in port $a$ ($b$) is reflected with amplitude $r$ and transmitted with up-conversion (down-conversion) to port $b$ (a) with amplitude $(1-r^{2})^{1/2}$. []{data-label="convflow"}](ConvFlow.pdf "fig:"){width="\columnwidth"}\
Added Noise
-----------
The number of output photons generated per mode in the amplification (case 1) is given by $$\mathcal{N}_{a,b}^{\mathrm{out}}=\left\vert r\right\vert ^{2}\mathcal{N}_{a,b}^{\mathrm{in}}+\left\vert s\right\vert ^{2}\mathcal{N}_{b,a}^{\mathrm{in}}, \label{Nab_out_ampl}$$ where $\mathcal{N}^{\mathrm{in}}$ is the input photon spectral density given by Eq. (\[Na\_in\_sec\]) and we assume that there is no cross-correlations between the input fields $a^{\mathrm{in}}$ and $b^{\mathrm{in}}$.
Assuming that the three-wave mixing device is in thermal equilibrium at temperature $T\ll\hbar\omega_{1,2}/k_{B}$ and that the dominant noise entering the system at each port is zero-point fluctuations $\hbar\omega_{1,2}/2$ ($\mathcal{N}^{\mathrm{in}}=1/2$), then in the limit of high gain $\left\vert
r\right\vert \gg1$, the number of noise equivalent photons effectively feeding the system is $$\mathcal{N}_{eq}^{\mathrm{in}}=\mathcal{N}^{\mathrm{out}}/\left\vert
r\right\vert ^{2}\simeq1. \label{N_in_eff_refl}$$ This means that the number of noise equivalent photons added by the device to the input is given by $\mathcal{N}^{\mathrm{add}}=\mathcal{N}_{eq}^{\mathrm{in}}-\mathcal{N}^{\mathrm{in}}=1/2$. Hence, when operated as a non-degenerate amplifier with $G_{0}\gg1$, the device adds noise which is equivalent to at least half a photon at the signal frequency to the input, in agreement with Caves theorem [@Caves].
In contrast, in the conversion mode of operation, assuming that there is no correlation between the input fields, the number of generated output photons per mode reads $$\mathcal{N}_{a,b}^{\mathrm{out}}=\left\vert r\right\vert ^{2}\mathcal{N}_{a,b}^{\mathrm{in}}+\left\vert t\right\vert ^{2}\mathcal{N}_{b,a}^{\mathrm{in}}.$$ Therefore, in pure conversion where $\left\vert r\right\vert =0$ and $\left\vert t\right\vert =1$, when referring the noise back to the input, one gets noise equivalent photons $$\mathcal{N}_{eq}^{\mathrm{in}}=\mathcal{N}^{\mathrm{out}}/\left\vert
t\right\vert ^{2}=1/2. \label{N_in_eff_pure}$$ This means that, as a converter, the device is not required to add noise to the input since $\mathcal{N}_{eq}^{\mathrm{in}}=\mathcal{N}^{\mathrm{in}}$.
Three-wave mixing using JRM
===========================
The Josephson ring modulator is a device consisting of four Josephson junctions, each with critical current $I_{0}=\frac{\hbar}{2eL_{J}}$ forming a ring threaded by a flux $\Phi=\Phi_{0}/2$ where $\Phi_{0}$ is the flux quantum (see Fig. \[Josephson\_ring\_modulator\]). The device has the symmetry of a Wheatstone bridge.
\[h\]
![Three-wave mixing element (see ellipse marked $K$ in Fig. \[three\_osc\_fig\]) consisting of a loop of four nominally identical Josephson junctions threaded by a flux in the vicinity of half a flux quantum. Mutual inductances, not shown here, couple this circuit to inductances $L_{a}$, $L_{b}$ and $L_{c}$ of Fig. \[three\_osc\_fig\] via the inductances $L_{X}$, $L_{Y}$ and $L_{Z}$ respectively, which are much larger than the junction inductance $L_{J}$. The three currents $I_{X}$, $I_{Y}$ and $I_{Z}$ correspond to the three orthogonal modes of the structure.[]{data-label="Josephson_ring_modulator"}](ring-modulator-currents.pdf "fig:"){width="\columnwidth"}\
There are thus three orthogonal electrical modes coupled to the junctions, corresponding to the currents $I_{X}$, $I_{Y}$ and $I_{Z}$ flowing in three external inductances $L_{X}$, $L_{Y}$ and $L_{Z}$ that are much larger than the junction inductance $L_{J}=\varphi_{0}^{2}E_{J}^{-1}$, where $\varphi
_{0}=\hbar/2e$ is the reduced flux quantum. Each junction $j\in\left\{
\alpha,\beta,\gamma,\delta\right\} $ is traversed by a current $I_{j}$ and at the working point (i.e. $\Phi=\Phi_{0}/2$) its energy is, keeping terms up to order four in $I_{j}$, given by $$E_{j}=\frac{1}{2}L_{J}^{eff}I_{j}^{2}-\frac{1}{24}\frac{L_{J}^{eff}}{I_{0}^{\prime2}}I_{j}^{4},$$ where $L_{J}^{eff}=\sqrt{2}L_{J}$ and $I_{0}^{\prime}=I_{0}/\sqrt{2}$. The currents in the junctions are expressed by $$\begin{aligned}
I_{\alpha} & =\frac{-I_{X}-I_{Y}}{2}+\frac{I_{Z}}{4}+I_{\Phi},\\
I_{\beta} & =\frac{+I_{X}-I_{Y}}{2}-\frac{I_{Z}}{4}+I_{\Phi},\\
I_{\gamma} & =\frac{+I_{X}+I_{Y}}{2}+\frac{I_{Z}}{4}+I_{\Phi},\\
I_{\delta} & =\frac{-I_{X}+I_{Y}}{2}-\frac{I_{Z}}{4}+I_{\Phi},\end{aligned}$$ where $I_{\Phi}$ is the supercurrent induced in the ring by the externally applied flux $\Phi$. The total energy of the ring is, keeping terms up to third order in the currents [@HuardProc], $$E_{ring}=\frac{1}{2}L_{J}^{eff}\left( I_{X}^{2}+I_{Y}^{2}+\frac{1}{4}I_{Z}^{2}\right) -\frac{1}{4}\frac{L_{J}^{eff}I_{\Phi}}{I_{0}^{\prime2}}I_{X}I_{Y}I_{Z}.$$ We can express the currents as $$I_{X,Y,Z}=\frac{\Phi_{a,b,c}}{L_{a,b,c}}\frac{M_{a,b,c}}{L_{X,Y,Z}}=\frac
{\Phi_{a,b,c}}{L_{a,b,c}^{eff}},$$ where $M_{a,b,c}$ are the mutual inductances between $L_{X,Y,Z}$ and the oscillator inductances $L_{a,b,c}$. The non-linear coefficient in the energy is, therefore, $$K=\frac{\left( L_{J}^{eff}\right) ^{2}}{4\varphi_{0}}\frac{1}{L_{a}^{eff}L_{b}^{eff}L_{c}^{eff}},$$ and we finally arrive at the result $$g_{3}^{2}=\frac{p_{a}p_{b}p_{c}\omega_{a}\omega_{b}\omega_{c}}{\omega
_{J}^{eff}}. \label{eqng3}$$ Here the participation ratios are defined as $$p_{a,b,c}=\frac{L_{J}^{eff}}{L_{a,b,c}^{eff}},$$ and, at $\Phi=\Phi_{0}/2$, $$\omega_{J}^{eff}=\frac{128}{\sqrt{2}}\frac{E_{J}}{\hbar}.$$ The participation ratios are linked to the maximal number of photons in each resonator, defined as those corresponding to an oscillation amplitude reaching a current of $I_{0}$ in each junction of the ring modulator, $$p_{a,b,c}\bar{n}_{a,b,c}^{\max}=\frac{E_{J}^{a,b,c}}{\hbar\omega_{a,b,c}},
\label{p-min}$$ where the $E_{J}^{a,b,c}$ are of order $E_{J}$ with factors accounting for the different participation of modes $X,$ $Y$ and $Z$ in the current of each junction. Equations (\[eqng3\]) and (\[p-min\]) are valid for all types of coupling between the Josephson ring modulator and signal/pump oscillators, which can be realized in practice by inductance sharing rather than by the mutual inductances discussed here.
Equation (\[p-min\]) can also be rewritten in terms of the maximum circulating power in cavities *a* and *b* as $$P_{\mathrm{cav}}^{\mathrm{\max}}=\frac{\gamma_{a,b}}{p_{a,b}}\frac{E_{J}}{\sqrt{2}} \label{Pmaxcav}$$ where we substituted $E_{J}/\sqrt{2}$ as an upper bound for $E_{J}^{a,b}$. The maximum number of photons in equation (\[p-min\]) determine the maximum signal input power handled by the device $$P_{a,b}^{\max}=\frac{1}{G}\gamma_{a,b}\hbar\omega_{a,b}\bar{n}_{a,b}^{\max}.
\label{P-Max}$$ We can now combine the notion of maximum power in resonator $c$ compatible with weak non-linearity with that of a critical power for the onset of parametric oscillation given by Eq. (\[onset-param-osc\]): $$\bar{n}_{c}^{\max}=\frac{E_{J}^{c}}{p_{c}\hbar\omega_{c}}>\bar{n}_{c}^{\mathrm{po}}=\frac{\Gamma_{1}\Gamma_{2}}{g_{3}^{2}},$$ arriving at the important relation $$p_{a}p_{b}Q_{a}Q_{b}>\Xi, \label{PQ-product}$$ where $\Xi$ is a number of order unity depending on the exact implementation of the coupling between the ring modulator and the oscillators. The quality factors of the resonators obey the well-known relation $$Q_{a,b}=\frac{\omega_{a,b}}{\gamma_{a,b}}.$$
Another maximum limit on the gain of the amplifier is set by the saturation of the device due to amplified zero-point fluctuations present at the input given by $$G_{\mathrm{ZPF}}^{\max}=\frac{E_{J}}{\sqrt{2}p_{a,b}}\frac{2}{\hbar
\omega_{a,b}}. \label{Gmaxzpf}$$ Eqs. (\[p-min\]), (\[P-Max\]) and (\[PQ-product\]) show that it is not possible to maximize simultaneously gain, bandwidth and dynamic range.
\[c\][|c|c|]{} &\
$\omega_{a,b}/2\pi$ & 1 - 16\
$Q_{a,b}$ & 50 - 500\
$Z_{a,b} $ & 10 - 150 $\Omega$\
$\gamma_{c} $ & 0.5 - 10\
$I_{0} $ & 0.5 - 10 $\mu$\
$E_{J} $ & 10 - 230\
$p_{a,b,c} $ & 0.01 - 0.5\
$g_{3}/2\pi$ & 0.1 - 15\
$\bar{n}^{\;\mathrm{max}}_{a,b,c} $ & $20 - 10^{4}$\
\[Table2\]
In table II we enlist general bounds on the characteristic parameters of the three-wave mixing device, which are feasible with superconducting microwave circuits and standard Al-AlOx-Al junction fabrication technology. A few comments regarding the values listed in the table are in order. The frequency ranges of resonators *a* and *b* is mainly set by the center frequency of the system whose signal one needs to amplify or process. It is also important that these frequencies are very small compared to the plasma frequency of the Josephson junction. The total quality factor range listed in the table $\left(50-500\right)$ is suitable for practical devices. Quality factors in excess of $500$ can be easily achieved with superconducting resonators but, as seen from Eq. (\[B\]), higher the quality factor, smaller the dynamical bandwidth of the device. Quality factors lower than $50$ on the other hand are not recommended either for a variety of reasons. For example, in the limit of very low $Q$ the pump softens (becomes less stiff), and the dynamic range decreases as more quantum noise will be admitted by the device bandwidth and amplified “unintentionally" by the junctions. The characteristic impedance of the resonators $Z_{a,b}$ is set by microwave engineering considerations as discussed in Sec. IV but, in general, this value varies around 50 $\Omega$. The rate $\gamma_{c}$ at which pump photons leave the circuit varies from one circuit design to the other as discussed in Sec. IV and is limited by $\omega_{c}$. This parameter also affects the maximum input power performance of the device as explained in Sec. III. As to the values of $I_{0}$, on the one hand it is beneficial to work with large Josephson junctions in order to increase the processing capability of the device; on the other hand a critical current larger than 10 $\mu$A adds complexity to the microwave design of the resonators and makes the fabrication process of the Josephson junction more involved. This might even require switching to a different fabrication process such as Nb-AlOx-Nb trilayer junctions [@Trilayer] or nanobridges [@Nanobridges]. The other parameters listed in the table, namely $p_{a,b,c}$, $E_{J}$, $g_{3}$, $\bar{n}_{a,b,c}^{\max}$, their values depend, to a large extent, on the device parameters already discussed.
Limitation of dynamic range due to pump depletion
=================================================
In the last two sections, we were using results obtained by solving only the first two of the equations of motion Eqs. (\[threeAmpEqs\]) under the restriction of the stiff pump approximation. In this section, we extend our analysis and include the third equation describing the dynamics of the pump to calculate the pump depletion and its effect on the dynamic range of the device. For this purpose, we consider the average value of the third equation of motion for field $c$ $$\frac{\mathrm{d}}{\mathrm{d}t}\left\langle c\right\rangle =-i\omega
_{c}\left\langle c\right\rangle -ig_{3}\left\langle ab\right\rangle
-\frac{\gamma_{c}}{2}\left\langle c\right\rangle +\sqrt{\gamma_{c}}\left\langle \tilde{c}^{\mathrm{in}}\left( t\right) \right\rangle .$$ In steady state and using RWA we obtain $$ig_{3}\left\langle ab\right\rangle +\frac{\gamma_{c}}{2}\left\langle c\left(
t\right) \right\rangle =\sqrt{\gamma_{c}}\left\langle \tilde{c}^{\mathrm{in}}\left( t\right) \right\rangle . \label{cin_c_relation}$$ In the limit of vanishing input, the cross-correlation term $\left\langle
ab\right\rangle $ is negligible and, therefore, $$\left\langle c\left( t\right) \right\rangle =\frac{2}{\sqrt{\gamma_{c}}}\left\langle \tilde{c}^{\mathrm{in}}\left( t\right) \right\rangle .$$ The average number of photons in the $c$ resonator in this case is, thus, $$\underset{\left\langle ab\right\rangle \rightarrow0}{\lim}\bar{n}_{c}=\frac
{4}{\gamma_{c}}\left\vert \left\langle \tilde{c}^{\mathrm{in}}\left(
t\right) \right\rangle \right\vert ^{2}. \label{c_sq_c_in_sq}$$ We now establish a self-consistent equation for $\bar{n}_{c}$, taking into account input signals of finite amplitude. We first evaluate the value of $\left\langle a\left( t\right) b\left( t\right) \right\rangle $ in the frame rotating with the pump phase, $$\begin{aligned}
& \left\langle a\left( t\right) b\left( t\right) \right\rangle
\nonumber\\
& =\frac{1}{2\pi}\int_{-\infty}^{+\infty}\int_{-\infty}^{+\infty}\left\langle
a\left[ \omega\right] b\left[ \omega^{\prime}\right] \right\rangle
e^{-i\left( \omega+\omega^{\prime}\right) t}\mathrm{d}\omega\mathrm{d}\omega^{\prime}.\end{aligned}$$ Using the field relations (see Appendix) $$\begin{aligned}
\sqrt{\gamma_{a}}a\left[ \omega\right] & =\tilde{a}^{\mathrm{in}}\left[
\omega\right] +\tilde{a}^{\mathrm{out}}\left[ \omega\right] ,\\
\sqrt{\gamma_{b}}b\left[ \omega\right] & =\tilde{b}^{\mathrm{in}}\left[
\omega\right] +\tilde{b}^{\mathrm{out}}\left[ \omega\right]\end{aligned}$$ and the input-output relations given by Eq. (\[RedS\]), we obtain (transforming back into the time domain) $$-ig_{3}\left\langle a\left( t\right) b\left( t\right) \right\rangle
=-\frac{\gamma_{eff}\left( G\right) }{2}\left\langle c\left( t\right)
\right\rangle ,$$ where, in the limit of large gains $G\gg1$, $$\gamma_{eff}\left( G\right) =\frac{1}{2\pi}\frac{\gamma_{c}}{4\overline
{n}_{c}^{\mathrm{in}}}\int_{0}^{+\infty}\mathrm{d}\omega\left( \mathcal{N}_{a}^{\mathrm{in}}\left[ \omega\right] +\mathcal{N}_{b}^{\mathrm{in}}\left[
\omega\right] \right) G\left( \Delta\omega\right) \label{gamma_eff}$$ denotes an effective decay rate of pump photons due to generation of entangled signal and idler photons. This last relation expresses, in another form, the Manley-Rowe relations [@ManleyProc] that establish the equality between the number of created signal photons by the amplifier to the number of destroyed pump photons. It shows that even in the absence of any deterministic signal applied to the oscillator $a$ or $b$, pump photons are used to amplify zero-point fluctuations. Therefore, the pump tone always encounters a dissipative load even when no signals are injected into the device.
For a continuous wave (CW) input power sent at the center frequency of the $a$ or $b$ oscillator, or both, we have $$\gamma_{eff}\left( G_{0},P^{\mathrm{in}}\right) =\frac{\gamma_{c}}{4\overline{n}_{c}^{\mathrm{in}}}G_{0}P^{\mathrm{in}}, \label{gamma_eff1}$$ where $P^{\mathrm{in}}=P_{a}^{\mathrm{in}}+P_{b}^{\mathrm{in}}$ is given in units of photon number per unit time and, in steady state, $$\overline{n}_{c}\left( G_{0},P^{\mathrm{in}}\right) =\frac{4\gamma_{c}}{\left( \gamma_{c}+\gamma_{eff}\left( G_{0},P^{\mathrm{in}}\right)
\right) ^{2}}\overline{n}_{c}^{\mathrm{in}}. \label{nc_ncin}$$ As a finite input power is applied to the signal oscillators, oscillator $c$ depopulates and, keeping the pump power constant, we get $$\begin{aligned}
\frac{\overline{n}_{c}\left( G_{0},P^{\mathrm{in}}\right) }{\overline{n}_{c}\left( G_{0},P^{\mathrm{in}}=0\right) } & =\frac{1}{\left(
1+\frac{G_{0}P^{\mathrm{in}}}{4\overline{n}_{c}^{\mathrm{in}}}\right) ^{2}}\\
& \simeq1-\frac{G_{0}P^{\mathrm{in}}}{2\overline{n}_{c}^{\mathrm{in}}}.\end{aligned}$$ On the other hand, from Eqs. (\[rho\]) and (\[gain-formula\]), the left hand side is given by $$\frac{\overline{n}_{c}\left( G_{0},P^{\mathrm{in}}\right) }{\overline{n}_{c}\left( G_{0},P^{\mathrm{in}}=0\right) }=\frac{\frac{\sqrt{G}-1}{\sqrt
{G}+1}}{\frac{\sqrt{G_{0}}-1}{\sqrt{G_{0}}+1}},$$ where $G$ denotes the gain in the presence of $P_{\mathrm{in}}$. In the large gain limit, if we fix the maximum decrease of gain due to pump depletion to be $$\frac{G}{G_{0}}>1-\varepsilon$$ with $\varepsilon\ll1$, then we obtain $$\frac{P^{\mathrm{in}}}{2\overline{n}_{c}^{\mathrm{in}}}<\varepsilon
G_{0}^{-3/2}, \label{pumpDepletion1}$$ which can also be rewritten as $$\frac{2\overline{n}_{c}^{\mathrm{in}}}{G_{0}P^{\mathrm{in}}}>\varepsilon
^{-1}\sqrt{G_{0}}.$$ This relation shows that the ratio of the power of the pump tone to that of the signal at the output of the amplifier must always be much larger than the amplitude gain, in order for the linearity of the amplifier not to be compromised by pump depletion effects.
In Fig. \[drthyfig\] we plot a calculated response of the signal output power $P_{\mathrm{out}}$ versus the signal input power $P_{\mathrm{in}}$ for a typical three-wave mixing device. The device parameters employed in the calculation and listed in the figure caption are practical values yielding a maximum input power, which is limited by the effect of pump depletion. The different blue curves are obtained by solving Eq. (\[nc\_ncin\]) for $G$ and using the input-output relation $P_{\mathrm{out}}=GP_{\mathrm{in}}$, where $P_{\mathrm{in}}$ expressed in units of power is taken as the independent variable and $G_{0}$ is treated as a parameter. Note that in solving Eq. (\[nc\_ncin\]), equations (\[gamma\_eff1\]), (\[c\_sq\_c\_in\_sq\]), (\[rho\]) and (\[gain-formula\]) are used. When drawn on logarithmic scale, the device gain translates into a vertical offset (arrow indicating $G_{0}$) off the $P_{\mathrm{out}}=P_{\mathrm{in}}$ line, indicated in red. The dashed black vertical line corresponds to a signal input power of 1 photon at the signal frequency per inverse dynamical bandwidth of the device at $G_{0}=20$ dB. The dashed green line corresponds to the maximum gain set by the amplified zero-point fluctuations given by Eq. (\[Gmaxzpf\]), while the cyan line corresponds to the maximum circulating power in the cavity given by Eq. (\[Pmaxcav\]).
\[h\]
![(Color online). A calculated response of the signal output power $P_{\mathrm{out}}$ versus the signal input power $P_{\mathrm{in}}$ of a typical three-wave mixing device which exhibits a pump depletion effect. The different blue curves correspond to different $G_{0}$ setpoints. The definition of the other lines in the figure is given in the text. The parameters used in the calculation are: $\omega_{a}/2\pi=7\operatorname{GHz}$, $\omega_{b}/2\pi=8\operatorname{GHz}$, $\omega_{c}/2\pi=15\operatorname{GHz}$, $\gamma_{a}/2\pi=\gamma_{b}/2\pi=50\operatorname{MHz}$, $\gamma_{c}/2\pi=0.6\operatorname{GHz}$, $Q_{a}=140$, $Q_{b}=160,$ $p_{a}=p_{b}=0.03$, $p_{c}=0.02,$ $I_{0}=1\operatorname{\mu A}$, $E_{J}^{a,b}=E_{J}/\sqrt
{2}=16.3\operatorname{K}$, $P_{\mathrm{1ph}}=-128$ dBm, $G_{\mathrm{ZPF}}^{\max}=35$ dB, $P_{\mathrm{cav}}^{\mathrm{\max}}=P_{b}^{\mathrm{\max}}=-86$ dBm and $g_{3}/2\pi=0.7\operatorname{MHz}$.[]{data-label="drthyfig"}](DRthyfig.pdf "fig:"){width="\columnwidth"}\
Furthermore, the maximum bound $P^{\max}$ indicated by the solid magenta line corresponds to $P_{\mathrm{out}}^{\max}=G_{0}P_{\mathrm{in}}^{\max}$, where $P_{\mathrm{in}}^{\max}=P_{b}^{\max}/G_{0}^{3/2}$ and $P_{b}^{\max
}=P_{\mathrm{cav}}^{\mathrm{\max}}$. As can be seen in the figure the predicted power scaling due to pump depletion effect, expressed in relation (\[pumpDepletion1\]), follows the calculated response quite well. Finally, it is straightforward to see that the usable region in the parameter space of the device with respect to gain, bandwidth and maximum input power lies within the boundaries of the fictitious triangle ABC indicated in the figure which is formed by the intersection of the magenta, black and red lines.
The Josephson parametric converter
==================================
We discuss here three different realizations of the Josephson parametric converter (JPC), which constitutes a fully non-degenerate three-wave mixing device capable of amplification and conversion as discussed in the previous sections. The three schemes differ in the resonator circuit design and the coupling between the feedline and the resonator.
Microstrip Resonator JPC (MRJ)
------------------------------
\[h\]
![Circuit model of the Microstrip Resonator JPC (MRJ).[]{data-label="MJcirc"}](MJCirc.pdf "fig:"){width="\columnwidth"}\
The MRJ comprises two superconducting microstrip resonators which intersect at a JRM at the center as shown in the circuit model of the device in Fig. \[MJcirc\]. The resonance frequencies of the MRJ are determined by the lengths of the microstrips $l_{a}\simeq\lambda_{a}/2$ and $l_{b}\simeq
\lambda_{b}/2$ and the Josephson inductance of the JRM, where $\lambda_{a}$, $\lambda_{b}$ are the wavelengths of the fundamental resonances at $\omega
_{a}$ and $\omega_{b}$. It is worth mentioning that in addition to the differential modes *a* and *b*, this configuration of two coupled resonators also supports a common (even) mode. The angular frequency $\omega_{e}$ at which this even mode resonates lies between $\left(
\omega_{b}+\omega_{a}\right) /2$ and $\omega_{b}$ (where $\omega_{b}>\omega_{a}$). The characteristic impedance of the resonators in the MRJ model is designed to be 50 $\Omega$ to ensure optimal coupling to the feedlines. Figure \[MJphoto\] exhibits an optical image of a typical MRJ device. The resonators are usually made of Al or Nb over sapphire or high-resistivity silicon and are coupled to the (transmission-line) feedlines using gap capacitors. The main role of these coupling capacitors is to set the external quality factor of the resonators. For a large bandwidth device operating in the $6-10$ GHz band, the external $Q$ of the resonators is typically in the range $60-100$. In all JPC designs discussed here the total $Q$ essentially coincides with the external $Q$, since the internal losses of the resonators are less than $10^{-4}$. Signals at $\omega_{1}$ and $\omega_{2}$, which lie within the bandwidths of resonators *a* and *b*, are fed into the JPC through the delta port of a $180$ degree hybrid, whereas the pump drive applied at $\omega_{3}=\omega_{1}+\omega_{2}$, for amplification, is a non-resonant tone and is injected into the device through the sigma port of the hybrid (Fig. \[MJcirc\]). The main advantage of the MRJ is that it is easy to design and fabricate. On the other hand, the main disadvantages are: (1) the area of the device can be relatively large depending on the frequencies of interest, (2) the characteristic impedance of the device is limited to around 50 $\Omega$, (3) the pump can be less stiff than the designs discussed below. The latter is due to the fact that the transmission-line resonators support higher resonance modes such as $2\omega_{a}$ and $2\omega_{b}$ with finite $Q$, which can be relatively close to the pump angular frequency $\omega_{3}$.
\[h\]
![(Color online). Optical microscope image of a microstrip resonator JPC (MRJ). The resonators denoted *a* and *b* are half-wave microstrip resonators which intersect at a JRM. A zoomed-in view of the JRM, which consists of four Josephson junctions arranged in Wheatstone bridge configuration, is shown on the right. The MRJ is coupled to 50 $\Omega$ feedlines via gap capacitors.[]{data-label="MJphoto"}](MJfig.pdf "fig:"){width="\columnwidth"}\
Compact Resonator JPC (CRJ)
---------------------------
\[h\]
![Circuit model of the Compact Resonator JPC (CRJ).[]{data-label="CJcirc"}](CJCirc.pdf "fig:"){width="\columnwidth"}\
In order to mitigate some of the drawbacks of the MRJ, we developed a new JPC design based on compact resonators known as CRJ. The circuit model of the CRJ, shown in Fig. \[CJcirc\], consists of four equal capacitors denoted as $C$ and two pairs of linear inductors connected in series with the JRM whose total inductance is $L_{a}$ and $L_{b}$ respectively. Using symmetry considerations one can verify that this circuit has three eigenmodes. Two differential eigenmodes which resonate at bare angular frequencies $\omega_{a}=1/\sqrt{\left( L_{a}+L_{J}^{eff}\right) C}$, $\omega_{b}=1/\sqrt{\left(
L_{b}+L_{J}^{eff}\right) C}$, where $L_{J}^{eff}$ is the equivalent Josephson inductance of the JRM biased at half a flux quantum, and an even eigenmode which resonates at a lower bare angular frequency $\omega_{e}=1/\sqrt{\left(
L_{a}+L_{b}+L_{J}^{eff}\right) C}$. Figure \[cjphoto\] shows an optical image of a typical compact JPC. The resonators of the device are made of Nb deposited over sapphire substrate. They are fabricated using a standard photolithography step and RIE etching. The JRM at the center of the device is made of Aluminum. It is fabricated using e-beam lithography, and angle shadow evaporation. As can be seen in the figure, the capacitance elements (including the coupling capacitors) of the device are implemented using interdigitated capacitors, whereas the inductive elements are realized using long narrow superconducting lines. Unlike the microstrip resonator JPC, the compact resonator JPC does not have higher harmonic resonances. The next closest resonance of this structure resides above $4\omega_{a}$, therefore the pump applied at $\omega_{a}+\omega_{b}$ can be considered stiff to a very good approximation. Other advantages of this realization are: (1) small size, with dimensions much smaller than the wavelengths corresponding to the resonance frequencies, (2) no requirement of a definite ground plane, unlike the MRJ, (3) greater flexibility in engineering the characteristic impedance of the resonators higher or lower than 50 $\Omega$, (4) higher internal quality factor resonators than the microstrip design. On the other hand, the main disadvantages of this design are: (1) the narrow lines and the interdigitated capacitors (as well as the lines connecting them) have parasitic capacitances and parasitic inductances associated with them, therefore scaling these devices to match a certain frequency or certain characteristic impedance requires using a microwave simulation tool, (2) there is a limit to how big the capacitance can be using the interdigitated configuration (values above 0.5 pF is difficult to achieve), therefore engineering characteristic impedances below 30 $\Omega$ is not quite feasible with this design.
![(Color online). Optical microscope image of a compact resonator JPC (CRJ). The device consists of four equal interdigitated capacitors denoted $C$ and two inductive elements denoted $L_{a}$ and $L_{b}$ which are realized using narrow superconducting lines of different lengths. The JRM of the device resides at the intersection of the two lines. An optical image of the JRM is shown in the inset. The CRJ is coupled to $50\operatorname{\Omega }$ microstrip feedlines via interdigitated capacitors denoted $C_{c}$.[]{data-label="cjphoto"}](CJ3fig.pdf "fig:"){width="\columnwidth"}\
Shunted JPC (SJ)
----------------
![Circuit model of the Shunted JPC (SJ).[]{data-label="LJcirc"}](LJCirc.pdf "fig:"){width="\columnwidth"}\
In this subsection we discuss a third promising design called the capacitively and inductively shunted JPC (SJ) which is still a work in progress in our lab. In this version of the JPC, the capacitive elements are parallel plate capacitors and the inductive elements are mainly Josephson junctions. A schematic circuit model of the SJ is drawn in Fig. \[LJcirc\]. It is straightforward to show that the SJ model has two differential eigenmodes with angular resonance frequencies $\omega_{a}=1/\sqrt{L_{J}^{^{\prime}}C_{a}}$, $\omega_{b}=1/\sqrt{L_{J}^{^{\prime}}C_{b}}$, where $L_{J}^{^{\prime}}$ corresponds to the equivalent inductance of the JRM shunted by linear inductors [@Roch], as shown in Fig. \[LJcirc\]. The main purpose of these shunting inductors is to eliminate the hysteretic flux response of the JRM and extend the frequency tunability of the device beyond the bandwidth limit of the resonators. Such frequency tunability is achieved by varying the flux threading the loop which, in turn, varies $L_{J}^{^{\prime}}$. Note that the addition of these shunting inductors can be employed in other realizations of the JPC also, such as the MRJ, as shown in Ref. [@Roch] and the CRJ. It is important to emphasize, however, that the main difference between the SJ and the CRJ or MRJ schemes is that the shunted JRM in the SJ design is the only inductive element in the circuit that forms an integral part of the resonators *a* and *b*. Thus, the larger lumped capacitors employed in the SJ design play a crucial role in keeping the resonance frequencies of the device below $10$ GHz.
Similar to the Josephson bifurcation amplifier (JBA) implementation [@VijayJBAreview], the plate capacitors in the SJ design can be made of Nb electrodes separated by a thin SiN dielectric layer. Using plate capacitors in this realization has two advantages: (1) the plate capacitors can be made very large, i.e. their capacitance can vary in the range $1-40$ pF, (2) they are easy to design as their capacitance scales linearly with the electrode area. Furthermore, due to the lumped nature of the capacitive and inductive elements in the SJ design and the fact that the capacitors can be large, the SJ has three important advantages over the previous designs: (1) the characteristic impedance of the resonators can be of the order of a few ohms, which yields an improved coupling between the resonators and the JRM, (2) due to the impedance mismatch between the characteristic impedance of the resonators and the 50 $\Omega$ feedlines, the coupling capacitors are unnecessary to achieve low external Q and the feedlines can be connected directly to the resonators, (3) the maximum input power of the amplifier can be increased by increasing the critical current of the JRM junctions while keeping the resonance frequencies fixed by enlarging the capacitors.
Experimental results
====================
The set of JPC parameters which can be directly measured in an experiment are: the angular resonance frequencies of the resonators *a* and *b* $\omega_{a}$, $\omega_{b}$, the inverse of residence times of photons at resonance $\gamma_{a}$, $\gamma_{b}$, the participation ratios $p_{a}$, $p_{b}$, the maximum input power which the device can handle with no applied pump tone $P_{a}^{\max}$, $P_{b}^{\max}$, and the maximum measured gain at vanishing input power $G_{0}^{\max}$.
One way to find $p_{a}$, $p_{b}$ is by measuring $\omega_{a}$, $\omega_{b}$ as a function of applied magnetic flux threading the JRM loop. To establish this relation, we model the resonators near resonance as an LC oscillator with effective inductance $L_{a,b}$ and effective capacitance $C_{a,b}$. In this model, the bare angular resonance frequencies of the device (with the junctions) $\omega_{a}$, $\omega_{b}$, can be written as $$\omega_{a,b}\left( \varphi\right) =\frac{1}{\sqrt{C_{a,b}\left(
L_{a,b}+L_{J}\left( \varphi\right) \right) }},$$ where $L_{J}\left( \varphi\right) $ is the effective Josephson inductance of the JRM given by $$L_{J}\left( \varphi\right) =\frac{L_{J}}{\cos\left( \frac{\varphi}{4}\right) }$$ with $\varphi=2\pi\Phi/\Phi_{0}$. By calculating the derivative of $\omega_{a,b}\left( \varphi\right) $ with respect to the reduced flux $\varphi$, one gets $$\begin{aligned}
\frac{1}{\omega_{a,b}}\frac{\mathrm{d}\omega_{a,b}}{\mathrm{d}\varphi} &
=-\frac{1}{8}\tan\left( \frac{\varphi}{4}\right) \frac{L_{J}\left(
\varphi\right) }{\left( L_{a,b}+L_{J}\left( \varphi\right) \right) },\\
& =-\frac{1}{8}\tan\left( \frac{\varphi}{4}\right) p_{a,b}(\varphi).\end{aligned}$$
Hence, at the device working point $\Phi=\Phi_{0}/2$ ($\varphi=\pi$), $p_{a,b}$ reads $$p_{a,b}=-8\left. \left( \frac{1}{\omega_{a,b}}\frac{\mathrm{d}\omega_{a,b}}{\mathrm{d}\varphi}\right) \right\vert _{\varphi=\pi}.$$
Furthermore, using Eq. (\[p-min\]) and the measured values $P_{a}^{\max}$, $P_{b}^{\max}$, one can infer the Josephson energy $E_{J}^{a,b}$ which is available for amplification $$E_{J}^{a,b}=p_{a,b}\frac{P_{a,b}^{\max}}{\gamma_{a,b}}.$$
It is important to mention that, in our experiments, we find that this value is lower by about one order of magnitude than the Josephson energy of the junctions at the working point $E_{J}=I_{0}\varphi_{0}/\sqrt{2}$, where $I_{0}$ is evaluated using dc resistance measurement of the junctions.
Using Eqs. (\[rho\]) and (\[gain-formula\]) for the case of maximum gain $G_{0}^{\max}$ yields $$g_{3}^{2}\overline{n}_{c,\rho\rightarrow1}=\frac{\gamma_{a}\gamma_{b}}{4}\frac{\sqrt{G_{0}^{\max}}-1}{\sqrt{G_{0}^{\max}}+1}, \label{g3_sq_times_n_c}$$ which in the limit of high gains gives an upper bound on the product $g_{3}^{2}\overline{n}_{c,\rho\rightarrow1}$ $$g_{3}^{2}\overline{n}_{c,\rho\rightarrow1}\leq\frac{\gamma_{a}\gamma_{b}}{4}.$$ Here $\overline{n}_{c,\rho\rightarrow1}$ is the number of pump photons in the device at $G_{0}^{\max}$.
\[h\]
![(Color online). Output power $P_{\mathrm{out}}$ measurement of a CRJ amplifier (device A) as a function of input power $P_{\mathrm{in}}$ measured at $\omega_{b}$. The data curves plotted in blue correspond to different $G_{0}$ setpoints obtained for different pump powers. The red line corresponds to $0$ dB (unity gain) where $P_{\mathrm{out}}=P_{\mathrm{in}}$. The dashed black vertical line indicates the input power of $1$ photon at the signal frequency per inverse dynamical bandwidth of the device at $G_{0}=20$ dB. The top horizontal line labelled $P_{\mathrm{cav}}^{\max}$ corresponds to the maximum circulating power in the resonator cavity given by Eq. (\[Pmaxcav\]). The green line corresponds to an upper limit on the device gain set by the saturation of the amplifier due to zero-point fluctuations given by Eq. (\[Gmaxzpf\]). The dashed magenta line is a theoretical prediction for $P_{\mathrm{out}}^{\max}$, which corresponds to the maximum circulating power in the device given by Eq. (\[DRstiff\]). The measured and calculated parameters of this device (A) are listed in table III.[]{data-label="drcjfig"}](DRCJfig.pdf "fig:"){width="\columnwidth"}\
In Figs. (\[drcjfig\]), (\[drpsfig\]), (\[drchfig\]) we plot on logarithmic scale the output power $P_{\mathrm{out}}$ of three different JPCs with different characteristics as a function of input power $P_{\mathrm{in}}$. For simplicity, we refer to the three devices as A, B and C respectively. The parameters of the three devices are listed in table III. The data curves plotted in blue are measured at resonance and satisfy the relation $$P_{\mathrm{out}}=G\left( P_{\mathrm{in}},G_{0}\right) P_{\mathrm{in}},$$ where $G\left( P_{\mathrm{in}},G_{0}\right) $ is the amplifier gain. This depends on $P_{\mathrm{in}}$ and $G_{0}$, the device gain for $P_{\mathrm{in}}=0$ which is set by the applied pump power. In this measurement, we apply a fixed pump power and vary $P_{\mathrm{in}}$ treating $G_{0}$ as a parameter. In log units, the device gain translates into a vertical offset from the $0$ dB baseline (red line) which corresponds to $P_{\mathrm{out}}=P_{\mathrm{in}}$.
As expected, the devices maintain an almost constant gain $G_{0}$ as a function of $P_{\mathrm{in}}$ before they saturate and their gain drops for elevated input powers. However, as can be seen in Figs. (\[drcjfig\]), (\[drpsfig\]), (\[drchfig\]), the three devices exhibit qualitatively different behaviors in the vicinity of their maximum input power, which correspond to different saturation mechanisms taking place in the device as will be discussed shortly. Note that the order in which the different results are presented in this section does not depend on the specific implementation of the device (see Sec. IV) but rather on the saturation mechanism involved in each case.
In Fig. \[drcjfig\], device A exhibits almost a plateau in $P_{\mathrm{out}}$ as it reaches its maximum input power for different $G_{0}$ setpoints. This result can be explained by assuming a stiff pump for which Eq. (\[P-Max\]) applies. By employing $P_{b}^{\max}$, measured with no applied pump tone, we plot the dashed magenta line labelled $P^{\max}$ which corresponds to $$P_{\mathrm{out}}^{\max}=P_{b}^{\max}. \label{DRstiff}$$ The dashed black vertical line indicates the input power of $1$ photon at the signal frequency per inverse dynamical bandwidth of the device at $G_{0}=20$ dB. In practice, as we discuss in Sec. VI, the usable region in the parameter space of the device with respect to gain, bandwidth and maximum input power lies within the boundaries of the fictitious triangle formed by the magenta, red and black lines.
![(Color online). Output power $P_{\mathrm{out}}$ measurement of a MRJ amplifier (device B) as a function of input power $P_{\mathrm{in}}$ measured at $\omega_{a}$. The data curves plotted in blue correspond to different $G_{0}$ setpoints obtained for different pump powers. The red line corresponds to $0$ dB (unity gain) where $P_{\mathrm{out}}=P_{\mathrm{in}}$. The dashed black vertical line indicates the input power of $1$ photon at the signal frequency per inverse dynamical bandwidth of the device at $G_{0}=20$ dB. The top horizontal line labelled $P_{\mathrm{cav}}^{\max}$ corresponds to the maximum circulating power in the resonator cavity given by Eq. (\[Pmaxcav\]). The green line corresponds to an upper limit on the device gain set by the saturation of the amplifier due to zero-point fluctuations given by Eq. (\[Gmaxzpf\]). The solid magenta line is a theoretical prediction for $P_{\mathrm{in}}^{\max}$ of the device and the corresponding $P_{\mathrm{out}}^{\max}$ due to pump depletion effect given by Eq. (\[DRpumpDep\]). The measured and calculated parameters of this device (B) are listed in table III.[]{data-label="drpsfig"}](DRPSfig.pdf "fig:"){width="\columnwidth"}\
Furthermore, in Figs. (\[drcjfig\]), (\[drpsfig\]), (\[drchfig\]) we plot two fundamental limits on the maximum gain $G_{\mathrm{ZPF}}^{\max}$ (green line) which corresponds to saturation of the device due to amplified zero-point fluctuations and the maximum circulating power $P_{\mathrm{cav}}^{\mathrm{\max}}$ (cyan line), given by Eq. (\[Gmaxzpf\]) and Eq. (\[Pmaxcav\]) respectively.
The fact that these lines lie considerably above the experimental data in Figs. (\[drcjfig\]), (\[drpsfig\]), (\[drchfig\]), suggests that the energy threshold, at which nonlinear effects in these devices become significant, is much lower than the Josephson energy of the junctions, i.e. $E_{J}^{a,b}\ll E_{J}$.
In contrast to Fig. \[drcjfig\], the data curves shown in Fig. \[drpsfig\] for device B, exhibit a gradual decrease in the gain in the vicinity of the maximum input power which can be explained in terms of pump depletion effect discussed in Sec. III. The maximum bound $P^{\max}$ indicated by the solid magenta line corresponds to $P_{\mathrm{out}}^{\max}=G_{0}P_{\mathrm{in}}^{\max}$, where in this case $P_{\mathrm{in}}^{\max}$ satisfies the inequality \[pumpDepletion1\] and is given by $$P_{\mathrm{in}}^{\max}=\frac{P_{a}^{\max}}{G_{0}^{3/2}}. \label{DRpumpDep}$$
![(Color online). Output power $P_{\mathrm{out}}$ measurement of another CRJ amplifier (device C, with different parameters from device A) as a function of input power $P_{\mathrm{in}}$ measured at $\omega_{b}$. The data curves plotted in blue corresponds to different $G_{0}$ setpoints obtained for different pump powers. The data curves of this device exhibit abrupt drop in the gain in the vicinity of the maximum input powers which suggests that the device enters an unstable regime at elevated input powers. The red line corresponds to $0$ dB (unity gain) where $P_{\mathrm{out}}=P_{\mathrm{in}}$. The dashed black vertical line indicates the input power of $1$ photon at the signal frequency per inverse dynamical bandwidth of the device at $G_{0}=20$ dB. The top horizontal line labelled $P_{\mathrm{cav}}^{\max}$ corresponds to the maximum circulating power in the resonator cavity given by Eq. (\[Pmaxcav\]). The green line correspond to an upper limit on the device gain set by the saturation of the amplifier due to zero-point fluctuations given by Eq. (\[Gmaxzpf\]). The solid magenta line is a theoretical prediction for $P_{\mathrm{in}}^{\max}$ of the device and the corresponding $P_{\mathrm{out}}^{\max}$ due to pump depletion effect given by Eq. (\[DRpumpDep\]). The measured and calculated parameters of this device (C) are listed in table III.[]{data-label="drchfig"}](DRChfig.pdf "fig:"){width="\columnwidth"}\
On the other hand the data curves shown in Fig. \[drchfig\] for device C exhibit an abrupt drop in the device gain in the vicinity of $P_{\mathrm{in}}^{\max}$ of the device, which indicates that the device enters an unstable regime at elevated input powers. As can be seen in this case the solid magenta line — which satisfies $P_{\mathrm{out}}^{\max}=G_{0}P_{\mathrm{in}}^{\max}$, where $P_{\mathrm{in}}^{\max}$ is given by Eq. (\[DRpumpDep\]) — lies above the experimental data. This suggests that the maximum input power in this sample, which displays a steeper power scaling than Eq. (\[DRpumpDep\]), is mainly limited by nonlinear effects arising from higher order terms in the Hamiltonian of the system and cannot be attributed to a pump depletion effect alone. It is worthwhile noting that a similar power scaling for the maximum input power has been observed as well for an MRJ amplifier in Ref. .
To understand which properties are responsible for the different gain behaviors exhibited by devices A, B and C, we point out a few important distinctions in their design (respective parameters are listed in table III). The data in Fig. \[drcjfig\] (device A) and Fig. \[drchfig\] (device C) is measured on JPC devices realized using the CRJ configuration which yields, in general, a stiff pump response as explained in Sec. IV (B). However, the main two differences between devices A and C are: (1) device A has a narrower bandwidth as compared to C (70 MHz vs. 142 MHz) and (2) the JRM junctions in A have a smaller $I_{0}$ compared to those in C (2 $\mu$A vs. 4 $\mu$A). The relatively large bandwidth of device C leads to a larger dynamical bandwidth 14 MHz at $G_{0}=16$ dB, as opposed to 10 MHz achieved in device A for the same gain, and also yields (with the larger $I_{0}$ of device C) higher $P_{a,b}^{\max}$ values. However, the large bandwidth translates into a lower $pQ$ product for C as compared to A, thus making it more susceptible to parametric oscillation (at high gains or high input powers) as implied by inequality (\[PQ-product\]).
Device B, on the other hand, exhibits a pump depletion effect as shown in Fig. \[drpsfig\]. This can be attributed to its MRJ configuration, which, in general, exhibits a less stiff pump response than the CRJ, due to the presence of high order modes as explained in Sec. IV (A). Furthermore, as opposed to the MRJ amplifier in Ref. with an idler frequency of 6.4 GHz, device B has a higher idler frequency of 15 GHz which leads to a higher $pQ$ product.
$\backslash$
-------------------------------------------------- ---------- ------------- ----------
$\omega_{a}/2\pi$ $6.576$ $8.436$ $7.051$
$\omega_{b}/2\pi$ $6.873$ $15.087$ $7.673$
$\omega_{3}/2\pi$ $13.449$ $23.523$ $14.724$
$\gamma_{a}/2\pi$ $69$ $116$ $79$
$\gamma_{b}/2\pi$ $71$ $250\pm25$ $142$
$Q_{a},Q_{b}$ $94,96$ $73,60$ $89,54$
$p_{a},p_{b}$ $0.02$ $0.03,0.05$ $0.03$
$p_{a}p_{b}Q_{a}Q_{b}$ $8.1$ $6.6$ $4.3$
$I_{0}$ $(\mu A)$ $2$ $3$ $4$
$P_{\;\mathrm{cav}}^{\;\mathrm{max}}$ $-82$ $-77$ $-76$
$P_{a,b}^{\;\mathrm{max}}$ $-97$ $-89$ $-87$
$P_{\;\mathrm{1ph}}$ $-127$ $-125$ $-123$
$G_{\;\mathrm{ZPF}}^{\;\mathrm{max}}$ $38$ $39$ $40$
$G_{0}^{\;\mathrm{max}}$ $22$ $20$ $16$
$g_{3}\bar{n}_{c,\rho\rightarrow1}^{1/2}/2\pi$ $33$ $77$ $45$
: Parameters of JPCs A, B and C. Precision is last significant digit unless indicated otherwise.
\[Table3\]
Requirements for qubit readout
==============================
One of the leading architectures which is used to manipulate and readout the state of superconducting qubits such as transmons and fluxoniums [@TransmonThy; @fluxonium] is circuit Quantum Electrodynamics (cQED). In such a system a quantum non-demolition measurement of the qubit state can be performed using dispersive readout in which the frequencies of the qubit and the cavity are detuned. As a result, the qubit and the cavity interact via exchanging virtual microwave photons [@RSL1] and the qubit state gets encoded in the output microwave field of the cavity. However, since the energy of microwave photons is very small, the detection of single photons is difficult especially considering the fact that state-of-the-art cryogenic amplifiers (i.e. high electron mobility transistor (HEMT) [@HEMT]) following the cQED setup add noise to the input signal, equivalent to about $20-40$ photons at the signal frequency. Therefore, adding a quantum-limited amplifier in series between the cQED sample and the HEMT amplifier can substantially decrease the noise temperature of the system and enable real-time tracking of the qubit state [@QuantumJumps; @QubitJPC]. The desired requirements of a Josephson parametric amplifier for such high-fidelity qubit readout can be summarized as follows:
- A center frequency in the range $5-12$ GHz which is widely used in readout cavities of superconducting qubits.
- A large power gain on the order of $20$ dB in order to beat the noise of the following amplifier, i.e. the HEMT.
- A minimum added noise to the signal, equivalent to a half input photon at the signal frequency $T_{\mathrm{N}}=\hbar\omega_{a,b}/2k_{B}$ when operated in the phase preserving mode [@Caves].
- A large dynamical bandwidth of the order of $10$ MHz, which corresponds to a signal processing time of less than $100$ ns and matches the bandwidth of most readout cavities.
- A maximum input power of a few photons per inverse dynamical bandwidth of the device at the highest gain. Such requirement is essential in quantum non-demolition readout schemes which employ of the order of a photon on average [@fluxonium].
- A tunable bandwidth of more than 100 MHz so that the center frequency of the amplifier can match the frequency of the readout signal. Recently, Roch *et al.* [@Roch] have achieved a tunable bandwidth of more than 500 MHz in a MRJ device by shunting the Josephson junctions of the JRM with linear inductors realized using superconducting wires \[see Fig. \[LJcirc\]\]. Similar results were obtained by our group in a MRJ device by utilizing large Josephson junctions instead of superconducting wires [@QubitJPC].
- Minimal out-of-band back-action to avoid qubit relaxation.
In table IV we enlist the parameters achievable with a JPC which show its viability as a low-noise preamplifier for qubit measurements.
--------------------------- ------------ ---------------------
$\omega_{a,b}/2\pi$ 5 - 12 6.4 & 8.1
$T_{\;\mathrm{N}}$ $@$ 8 0.2 0.4
$G$ $20$ $21$
$B$ 10 11
100 60
$P_{\;\mathrm{max}}$ $1$ photon 3 photons $@$ 20 dB
OB back-action Negligible None measurable
: Preamplifier requirements and JPC merits achieved to date (OB=out-of-band).
\[Table4\]
The last question which we would like to address in this paper is whether there exists a set of technologically feasible parameters for which the JPC can be optimized with respect to dynamic range while maintaining a gain of $20$ dB and a reasonable dynamical bandwidth of more than 2 MHz. In order to provide a quantitative answer we choose a signal frequency of 12 GHz, which is a good choice for readout frequency for qubits as it is higher than most qubit frequencies. We set an ambitious goal for the processing capability of the JPC of about $100$ input photons at the signal frequency with gain $20$ dB.
![(Color online). An optimized JPC response drawn in blue for large maximum input power in excess of $100$ input photons at $12$ $\operatorname{GHz}$ per inverse dynamical bandwidth of the device at $20$ dB. The definition of the other lines shown in the figure is similar to Fig. (\[drcjfig\]). The parameters used in the calculation are: $\omega_{a}/2\pi=11\operatorname{GHz}$, $\omega_{b}/2\pi=12\operatorname{GHz}$, $\omega_{c}/2\pi=23\operatorname{GHz}$, $Z_{a}=36\operatorname{\Omega }$, $Z_{b}=33\operatorname{\Omega }$, $L_{a}=0.51$ n$\operatorname{H}$, $L_{b}=0.42$ n$\operatorname{H}$, $C_{a}=C_{b}=0.4\operatorname{pF}$, $C_{C_{a}}=31$ f$\operatorname{F}$, $C_{C_{b}}=28$ f$\operatorname{F}$, $\gamma_{a}/2\pi=\gamma_{b}/2\pi=44.2\operatorname{MHz}$, $\gamma_{c}/2\pi=3\operatorname{GHz}$, $Q_{a}=249$, $Q_{b}=271$, $Q_{c}=8$, $p_{a}=0.028$, $p_{b}=0.034$, $p_{c}=0.02$, $I_{0}=30\operatorname{\mu A}$, $E_{J}/\sqrt{2}=490\operatorname{K}$, $E_{J}^{a,b}=49\operatorname{K}$, $P_{\mathrm{1ph}}=-126$ dBm, $G_{\mathrm{ZPF}}^{\max}=48$ dB, $P_{\mathrm{cav}}^{\mathrm{\max}}=P_{b}^{\mathrm{\max}}=-86.3$ dBm and $g_{3}/2\pi
=0.6\operatorname{MHz}$. []{data-label="DRprospectsfig"}](DRprospectsThyfig.pdf "fig:"){width="\columnwidth"}\
To that end, we choose to perform the optimization process for the CRJ configuration which inherently yields large $\gamma_{c}$ values and also allows variation of the external quality factor of the resonators more easily than the SJ scheme. We also choose a resonance frequency for mode *a* of $\omega_{a}/2\pi=$ 11 GHz, a relatively high critical current $I_{0}=$ 30 $\mu$A and limit ourselves to capacitance values below or equal to 0.4 pF. The advantage of working with large $I_{0}$ for the purpose of large dynamic range is that it increases $E_{J}$ and lowers $g_{3}$. However, such large $I_{0}$ yields very low $L_{J}^{eff}=\sqrt{2}\varphi_{0}/I_{0}=15$ pH which requires coupling to relatively low impedance resonators while maintaining participation ratios of a few percent. The next challenge in the optimization is to increase the $pQ$ product of the device which promotes high gains by increasing the quality factor of the resonators. Nevertheless, care must be taken not to increase the quality factors beyond what is strictly necessary for two reasons (i) high Q resonators limit the dynamical bandwidth of the device as can be seen in Eq. (\[B\]), (ii) high Q resonators increase the pump depletion effect and, in turn, lower the dynamic range of the device. In Fig. \[DRprospectsfig\] we plot the calculated response of such an optimized JPC which takes into account the above considerations and limitations. As can be seen in the figure the optimized device exhibits, for the chosen set of parameters, a maximum input power of about $100$ photons at the signal frequency per inverse dynamical bandwidth of the device $B/2\pi =$ 4.4 MHz at $20$ dB of gain. The device parameters which are used in the calculation are listed in the figure caption. It is important to note that in the calculation of the expected response, which is indicated by the blue curves for different values of $G_{0}$, we assumed an available Josephson energy $10$ times smaller than $E_{J}/\sqrt{2}$ of the junctions, in agreement with experimental conditions. Finally, we verify that the set of parameters of the optimized device satisfy the inequalities $\bar{n}_{c}^{\max}=3.7\cdot10^{3}>\bar{n}_{c}^{\mathrm{po}}=1.3\cdot10^{3}> \bar{n}_{c}^{20\mathrm{dB}}=10^{3}$.
Conclusion
==========
We have addressed in this paper a new type of quantum signal processing device based on Josephson tunnel junctions. In contrast with the devices based on SQUIDS and driven non-linear Josephson oscillators, it performs a fully non-degenerate three-wave mixing in which the modes of the signal, pump and idler are separate both spatially and temporally. The heart of the device consists of a ring modulator constructed from four Josephson junctions arranged in a loop. Both quantum-limited amplification and noiseless frequency conversion are possible with this device, and the characteristics of these analog signal processing operations are entirely calculable analytically. We have established the limitations preventing the simultaneous maximization of photon number gain, bandwidth and dynamic range. Nevertheless, we have shown that a device satisfying all the requirements of superconducting qubit readout is realizable with present day technology.
Discussions with Flavius Schackert, Michael Hatridge, Nicolas Bergeal, Benjamin Huard and Ananda Roy are gratefully acknowledged. The assistance of Michael Power and Luigi Frunzio in the fabrication process is highly appreciated. This work was supported by Yale University, NSF, IARPA, ARO and College de France.
Appendix: Quantum signals propagating along a transmission line and input-output formalism {#appendix Quantum signals .unnumbered}
==========================================================================================
This appendix treats quantum-mechanically the damping of a circuit by a resistance modelled as a semi-infinite transmission line, as shown in Fig. \[Nyquist-model\]. It borrows heavily from the book by Gardiner and Zoller [@QuantumNoise] but uses slightly different notations that are adapted to the specificities of our Josephson circuits. We first describe an infinite transmission line extending from $x=-\infty$ to $x=+\infty$. Later, we will cut the line at $x=0$ and replace the left portion by two terminals of the circuit.
![The damping of a circuit by a resistance $R$ can take place in a parallel or series way, depending on whether the resistance is placed across a branch or in series with it. The Nyquist model represents the resistance by a transmission line with characteristic impedance $Z_{c}=R$. The noise source associated with the resistance (fluctuation-dissipation theorem) is a parallel current source in the parallel case and a series voltage source in the series case. The noise source is replaced in the Nyquist model by incoming thermal radiation whose amplitude $A^{\mathrm{in}}$ is the square root of the power flux of the radiation ($A^{\mathrm{in}}$ should not be associated to a vector potential and is rather like the square root of the length of the Poynting vector).[]{data-label="Nyquist-model"}](nyquist_model_parallel-series.pdf "fig:"){width="\columnwidth"}\
Infinite transmission line {#infinite-transmission-line .unnumbered}
--------------------------
The capacitance and inductance per unit length of the line are $C_{\ell}$ and $L_{\ell}$, respectively. The equations obeyed by the current $I$ along and the voltage $V$ across the line are $$\begin{aligned}
-\frac{\partial}{\partial x}V\left( x,t\right) & =L_{\ell}\frac{\partial
}{\partial t}I\left( x,t\right) ,\label{propagation_eq_1}\\
-\frac{\partial}{\partial x}I\left( x,t\right) & =C_{\ell}\frac{\partial
}{\partial t}V\left( x,t\right) , \label{propagation_eq_2}$$ in which, for the moment, we treat the fields classically. The characteristic impedance and propagation velocity are given by $$\begin{aligned}
Z_{c} & =\sqrt{\frac{L_{\ell}}{C_{\ell}}},\\
v_{p} & =\sqrt{\frac{1}{L_{\ell}C_{\ell}}}.\end{aligned}$$ In order to solve Eqs. (\[propagation\_eq\_1\]) and (\[propagation\_eq\_2\]), we introduce two new fields: the left-moving and right-moving wave amplitudes, $$\begin{aligned}
A^{\rightarrow}\left( x,t\right) & =\frac{1}{2}\left[ \frac{1}{\sqrt{Z_{c}}}V\left( x,t\right) +\sqrt{Z_{c}}I\left( x,t\right) \right]
,\\
A^{\leftarrow}\left( x,t\right) & =\frac{1}{2}\left[ \frac{1}{\sqrt
{Z_{c}}}V\left( x,t\right) -\sqrt{Z_{c}}I\left( x,t\right) \right] ,\end{aligned}$$ which have the advantage of treating currents and voltage on the same footing (note that these amplitudes are not directly related to the vector potential). The dimension of these fields is \[watt\]$^{1/2}$ and they are normalized such that the total power $P$ traversing, in the forward direction, a section of the line at position $x$ and time $t$ is given by $$P\left( x,t\right) =\left[ A^{\rightarrow}\left( x,t\right) \right]
^{2}-\left[ A^{\leftarrow}\left( x,t\right) \right] ^{2}. \label{poynting}$$ The quantity $P$ here plays the role of the Poynting vector in full 3D electrodynamics. Each of the terms at the right hand side of the last equation is thus the separate contribution of the corresponding wave to the total power flow.
When solving Eqs. (\[propagation\_eq\_1\]-\[propagation\_eq\_2\]), we find $$\frac{\partial}{\partial x}A^{\rightleftarrows}\left( x,t\right) =\mp
\frac{1}{v_{p}}\frac{\partial}{\partial t}A^{\rightleftarrows}\left(
x,t\right) . \label{Eq._of_motion}$$ This relation means that $A^{\rightleftarrows}$ does not depend separately on $x$ or $t$ but a combination of both and thus: $$\begin{aligned}
A^{\rightarrow}\left( x,t\right) & =A^{\rightarrow}\left( x=0,t-\frac
{x}{v_{p}}\right) =A^{\rightarrow}\left( x-v_{p}t,t=0\right) ,\nonumber\\
A^{\leftarrow}\left( x,t\right) & =A^{\leftarrow}\left( x=0,t+\frac
{x}{v_{p}}\right) =A^{\leftarrow}\left( x+v_{p}t,t=0\right) .\nonumber\\
&\end{aligned}$$ The properties of the wave amplitude can be summarized by writing $$\begin{aligned}
A^{\rightleftarrows}\left( x,t\right) & =A_{0}^{\rightleftarrows}\left(
\tau\right) ,\\
\tau & =t+\frac{\varepsilon^{\rightleftarrows}}{v_{p}}x,\\
\varepsilon^{\rightleftarrows} & =\mp1.\end{aligned}$$ Note that the detailed definition of the retardation $\tau$ depends on the wave direction. We now turn to the energy density $U\left( x,t\right) $, related to $P$ by the local energy conservation law $$\frac{\partial U}{\partial t}=-\frac{\partial P}{\partial x}.$$ Combining Eqs. (\[poynting\]) and (\[Eq.\_of\_motion\]), we get $$\begin{aligned}
& \frac{\partial U\left( x,t\right) }{\partial t}\nonumber\\
& =\frac{2}{v_{p}}\left[ A^{\rightarrow}\left( x,t\right) \frac{\partial
}{\partial t}A^{\rightarrow}\left( x,t\right) +A^{\leftarrow}\left(
x,t\right) \frac{\partial}{\partial t}A^{\rightarrow}\left( x,t\right)
\right] ,\nonumber\\
& =\frac{1}{v_{p}}\frac{\partial}{\partial t}\left\{ \left[ A^{\rightarrow
}\left( x,t\right) \right] ^{2}+\left[ A^{\leftarrow}\left( x,t\right)
\right] ^{2}\right\} .\end{aligned}$$ The total energy of the line at time $t$ is, thus [@SumofEMenergy], $$H=\frac{1}{v_{p}}\int_{-\infty}^{+\infty}\left\{ \left[ A^{\rightarrow
}\left( x,t\right) \right] ^{2}+\left[ A^{\leftarrow}\left( x,t\right)
\right] ^{2}\right\} \mathrm{d}x. \label{HinA}$$ When $H$ in Eq. (\[HinA\]) is considered as a functional of dynamical field variables $A^{\rightarrow}$ and $A^{\leftarrow}$, the equation of motion Eq. (\[Eq.\_of\_motion\]) can be recovered from Hamilton’s equation of motion as $$\frac{\partial}{\partial t}A^{\rightleftarrows}\left( x,t\right) =-\left\{
H,A^{\rightleftarrows}\left( x,t\right) \right\} _{P.B.},$$ on imposing the Poisson bracket $$\begin{aligned}
& \left\{ A^{\rightleftarrows}\left( x_{1},t_{1}\right)
,A^{\rightleftarrows}\left( x_{2},t_{2}\right) \right\} _{P.B.}
=\frac{1}{2}\frac{\partial}{\partial\left( \tau_{1}-\tau_{2}\right)
}\delta\left( \tau_{1}-\tau_{2}\right). \label{PBinA2}$$ Therefore, from the classical-quantum correspondence involving the replacement of Poisson brackets by commutators, we find that the quantum operator version $\hat{A}^{\rightleftarrows}$ of the fields satisfy the commutation relation $$\left[ \hat{A}^{\rightleftarrows}\left( x_{1},t_{1}\right) ,\hat
{A}^{\rightleftarrows}\left( x_{2},t_{2}\right) \right] =\frac{i\hbar}{2}\frac{\partial}{\partial\left( \tau_{1}-\tau_{2}\right) }\delta\left(
\tau_{1}-\tau_{2}\right) ,$$ which is analogous to the commutation relation between the electric and magnetic field in 3-D quantum electrodynamics. Note that the fields are Hermitian at this stage. Introducing the Fourier transform, $$\hat{A}^{\rightleftarrows}\left[ \omega\right] =\frac{1}{\sqrt{2\pi}}\int_{-\infty}^{+\infty}\hat{A}^{\rightleftarrows}\left( x=0,\tau\right)
e^{i\omega\tau}\mathrm{d}\tau,$$ where the Fourier components (which are now non-hermitian operators) satisfy $$\hat{A}^{\rightleftarrows}\left[ \omega\right] ^{\dag}=A^{\rightleftarrows
}\left[ -\omega\right] ,$$ we can also write the Hamiltonian as $$\sum_{\sigma=\rightleftarrows}\int_{-\infty}^{+\infty}\hat{A}^{\sigma}\left[
\omega\right] \hat{A}^{\sigma}\left[ -\omega\right] \mathrm{d}\omega.$$ The field operators in the frequency domain satisfy $$\left[ \hat{A}^{\rightleftarrows}\left[ \omega_{1}\right] ,\hat
{A}^{\rightleftarrows}\left[ \omega_{2}\right] \right] =\frac{\hbar}{4}\left( \omega_{1}-\omega_{2}\right) \delta\left( \omega_{1}+\omega
_{2}\right) .$$ We now introduce the usual quantum field annihilation operators $$\begin{aligned}
a^{\rightarrow}\left[ \omega\right] & =\frac{\hat{A}^{\rightarrow}\left[
\omega\right] }{\sqrt{\hbar\left\vert \omega\right\vert /2}}=a^{\rightarrow
}\left[ -\omega\right] ^{\dagger},\\
a^{\leftarrow}\left[ \omega\right] & =\frac{\hat{A}^{\leftarrow}\left[
\omega\right] }{\sqrt{\hbar\left\vert \omega\right\vert /2}}=a^{\leftarrow
}\left[ -\omega\right] ^{\dagger}.\end{aligned}$$ They satisfy the commutation relations $$\left[ a^{\rightleftarrows}\left[ \omega_{1}\right] ,a^{\rightleftarrows
}\left[ \omega_{2}\right] \right] =\mathrm{sgn}\left( \frac{\omega
_{1}-\omega_{2}}{2}\right) \delta\left( \omega_{1}+\omega_{2}\right)
.\label{bosonic_com}$$ It is useful to note that since $$a^{\rightleftarrows}\left[ \omega\right] =a^{\rightleftarrows}\left[
-\omega\right] ^{\dag},$$ Eq. (\[bosonic\_com\]) exhaustively describes all possible commutator cases.
In the thermal state of the line, at arbitrary temperature (including $T=0$), $$\left\langle a^{\rightleftarrows}\left[ \omega_{1}\right]
a^{\rightleftarrows}\left[ \omega_{2}\right] \right\rangle
=S_{a^{\rightleftarrows}a^{\rightleftarrows}}\left[ \frac{\omega_{1}-\omega_{2}}{2}\right] \delta\left( \omega_{1}+\omega_{2}\right) ,$$ where $$S_{a^{\rightleftarrows}a^{\rightleftarrows}}\left[ \omega\right]
=\mathrm{sgn}\left( \omega\right) N_{T}\left( \omega\right) .$$ When $\omega$ is strictly positive $N_{T}\left( \omega\right) $ is the number of available photons per unit bandwidth per unit time travelling on the line in a given direction around frequency $\omega$ $$\begin{aligned}
N_{T}\left( \omega\right) & =\frac{1}{\exp\left( \frac{\hbar\omega}{k_{B}T}\right) -1}\\
& =\frac{1}{2}\left[ \coth\left( \frac{\hbar\omega}{2k_{B}T}\right)
-1\right] .\end{aligned}$$ Negative frequencies $\omega$ correspond to the possibility of emitting photons into the line $$N_{T}\left( -\left\vert \omega\right\vert \right) =-N_{T}\left( \left\vert
\omega\right\vert \right) -1.$$ The Bose-Einstein expression $N_{T}\left( \omega\right) $ is expected from the Hamiltonian of the line, which reads, with the $a$ operators, $$H=\frac{\hbar}{2}\sum_{\sigma=\rightleftarrows}\int_{-\infty}^{+\infty
}\left\vert \omega\right\vert a^{\sigma}\left[ \omega\right] a^{\sigma
}\left[ -\omega\right] \mathrm{d}\omega.$$ We can now give the expression for the anticommutator of the fields $$\begin{aligned}
& \left\langle \left\{ a^{\rightleftarrows}\left[ \omega_{1}\right]
,a^{\rightleftarrows}\left[ \omega_{2}\right] \right\} \right\rangle
_{T}=2\mathcal{N}_{T}\left[ \frac{\omega_{1}-\omega_{2}}{2}\right]
\delta\left( \omega_{1}+\omega_{2}\right) \nonumber\\
& =\mathrm{sgn}\left( \frac{\omega_{1}-\omega_{2}}{2}\right) \coth\left(
\frac{\hbar\left( \omega_{1}-\omega_{2}\right) }{4k_{B}T}\right)
\delta\left( \omega_{1}+\omega_{2}\right) . \label{bos-anticom6}$$ Equation (\[Na\_in\_first\]) with no external drive is identical to Eq. (\[bos-anticom6\]) $$\begin{aligned}
\mathcal{N}_{T}\left[ \omega\right] & =\frac{\mathrm{sgn}\left(
\omega\right) }{2}\coth\left( \frac{\hbar\omega}{2k_{B}T}\right) \\
& =\mathrm{sgn}\left( \omega\right) \left[ N_{T}\left( \left\vert
\omega\right\vert \right) +\frac{1}{2}\right] .\end{aligned}$$ We now introduce the forward-propagating and backward-propagating voltage and current amplitudes obeying $$\begin{aligned}
V^{\rightarrow}\left( x,t\right) & =\sqrt{Z_{c}}A^{\rightarrow}\left(
x,t\right) ,\\
V^{\leftarrow}\left( x,t\right) & =\sqrt{Z_{c}}A^{\leftarrow}\left(
x,t\right) ,\end{aligned}$$$$\begin{aligned}
I^{\rightarrow}\left( x,t\right) & =V^{\rightarrow}\left( x,t\right)
/Z_{c},\\
I^{\leftarrow}\left( x,t\right) & =V^{\leftarrow}\left( x,t\right)
/Z_{c}.\end{aligned}$$ Quantum-mechanically, the voltage and current amplitudes become hermitian operators $$\begin{aligned}
V^{\rightleftarrows}\left( x,t\right) & \rightarrow\hat{V}^{\rightleftarrows}\left( x,t\right) ,\\
I^{\rightleftarrows}\left( x,t\right) & \rightarrow\hat{I}^{\rightleftarrows}\left( x,t\right) .\end{aligned}$$ These operators, in turn, can be expressed in terms of field annihilation operators as $$\begin{aligned}
\hat{V}^{\rightleftarrows}\left( x,t\right) & =\sqrt{\frac{\hbar Z_{c}}{4\pi}}\int_{-\infty}^{+\infty}\mathrm{d}\omega\sqrt{\left\vert
\omega\right\vert }\hat{a}^{\rightleftarrows}\left[ \omega\right]
e^{-i\omega\left( t\,\mp\,x/v_{p}\right) },\\
\hat{I}^{\rightleftarrows}\left( x,t\right) & =\sqrt{\frac{\hbar}{4\pi
Z_{c}}}\int_{-\infty}^{+\infty}\mathrm{d}\omega\sqrt{\left\vert \omega
\right\vert }\hat{a}^{\rightleftarrows}\left[ \omega\right] e^{-i\omega
\left( t\,\mp\,x/v_{p}\right) }.\end{aligned}$$ All physical operators can be deduced from these primary expressions. For instance, the transmission line charge operator, describing the charge in the line brought from one end to the position $x$, is $$\hat{Q}^{\rightleftarrows}\left( x,t\right) =i\sqrt{\frac{\hbar}{4\pi Z_{c}}}\int_{-\infty}^{+\infty}\frac{\mathrm{d}\omega\sqrt{\left\vert
\omega\right\vert }}{\omega}\hat{a}^{\rightleftarrows}\left[ \omega\right]
e^{-i\omega\left( t\,\mp\,x/v_{p}\right) }.$$
Nyquist model of resistance: semi-infinite transmission line {#nyquist-model-of-resistance-semi-infinite-transmission-line .unnumbered}
------------------------------------------------------------
We now are in a position to deal with the semi-infinite line extending from $x=0$ to $x=\infty$, whose terminals at $x=0$ models a resistance $R=Z_{c}$ \[see Fig. \[Nyquist-model\]\]. In that half-line, the left- and right-moving propagating waves are no longer independent. We will now refer to the wave amplitude $A^{\leftarrow}\left( x=0,t\right) $ as $A^{\mathrm{in}}\left(
t\right) $ and $A^{\rightarrow}\left( x=0,t\right) $ as $A^{\mathrm{out}}\left( t\right) $. The quantum-mechanical voltage across the terminal of the resistance and the current flowing into it satisfy the operator relations $$\begin{aligned}
\hat{V}\left( t\right) & =\hat{V}^{\mathrm{out}}\left( t\right) +\hat
{V}^{\mathrm{in}}\left( t\right) ,\\
\hat{I}\left( t\right) & =\hat{I}^{\mathrm{out}}\left( t\right) -\hat
{I}^{\mathrm{in}}\left( t\right) .\end{aligned}$$ These relations can be seen either as continuity equations at the interface between the damped circuit and the resistance/line, or as boundary conditions linking the semi-infinite line quantum fields $\hat{A}^{\mathrm{in}}\left(
t\right) $ and $\hat{A}^{\mathrm{out}}\left( t\right) $. From the transmission line relations, $$\hat{V}^{\mathrm{out},\mathrm{in}}\left( t\right) =R\hat{I}^{\mathrm{out},\mathrm{in}}\left( t\right) ,$$ we obtain $$\begin{aligned}
\hat{I}\left( t\right) & =\frac{1}{R}\hat{V}\left( t\right) -2\hat
{I}^{\mathrm{in}}\left( t\right) ,\\
& =\frac{1}{R}\hat{V}\left( t\right) -\frac{2}{\sqrt{R}}\hat{A}^{\mathrm{in}}\left( t\right) .\end{aligned}$$ For a dissipationless circuit with Hamiltonian $H_{bare}\left( \hat{\Phi
},\hat{Q}\right) $, where $\hat{\Phi}$ is the generalized flux of the node electrically connected to the transmission line, and $\hat{Q}$ its canonically conjugate operator (top panel of Fig. \[Nyquist-model\]), we can write the Langevin equation, $$\begin{aligned}
\frac{\mathrm{d}}{\mathrm{d}t}\hat{Q} & =\frac{i}{\hbar}\left[
H_{bare},\hat{Q}\right] -\hat{I},\nonumber\\
& =\frac{i}{\hbar}\left[ H_{bare},\hat{Q}\right] -\frac{\mathrm{d}}{R\mathrm{d}t}\hat{\Phi}+\frac{2}{\sqrt{R}}\hat{A}^{\mathrm{in}}\left(
t\right) . \label{Langevin-example}$$ The latter equation is just a particular case of the more general quantum Langevin equation giving the time evolution of any operator $\hat{Y}$ of a system with Hamiltonian $H_{bare}$, which is coupled to the semi-infinite transmission line by an Hamiltonian term proportional to another system operator $\hat{X}$, $$\begin{aligned}
\frac{\mathrm{d}}{\mathrm{d}t}\hat{Y} & =\frac{i}{\hbar}\left[
H_{bare},\hat{Y}\right] \nonumber\\
& +\frac{1}{2i\hbar}\left\{ \left[ \hat{X},\hat{Y}\right] ,2R^{\zeta
/2}\hat{A}^{\mathrm{in}}\left( t\right) -R^{\zeta}\frac{\mathrm{d}}{\mathrm{d}t}\hat{X}\right\} .\nonumber\\
& \label{general-QLE1}$$ The value of $\zeta$ in Eq. (\[general-QLE1\]) depends on whether the damping is parallel“ ($\zeta=-1$) or series” type ($\zeta=+1$) \[see Fig. \[Nyquist-model\]\]. In the parallel case, the greater the line impedance the smaller the damping, whereas in the series case the situation is reversed.
Equation (\[general-QLE1\]) should be supplemented by $$\left[ \hat{A}^{\mathrm{in}}\left( t_{1}\right) ,\hat{A}^{\mathrm{in}}\left( t_{2}\right) \right] =\frac{i\hbar}{2}\frac{\partial}{\partial\left( t_{1}-t_{2}\right) }\delta\left( t_{1}-t_{2}\right)$$ and $$\hat{A}^{\mathrm{out}}\left( t\right) =\zeta\left[ \hat{A}^{\mathrm{in}}\left( t\right) -R^{\zeta/2}\frac{\mathrm{d}}{\mathrm{d}t}\hat{X}\right] .$$ It follows from the last three equations that the output fields have the same commutation relation as the input fields $$\left[ \hat{A}^{\mathrm{out}}\left( t_{1}\right) ,\hat{A}^{\mathrm{out}}\left( t_{2}\right) \right] =\frac{i\hbar}{2}\frac{\partial}{\partial\left( t_{1}-t_{2}\right) }\delta\left( t_{1}-t_{2}\right) .$$
Quantum Langevin equation in the RWA approximation {#quantum-langevin-equation-in-the-rwa-approximation .unnumbered}
--------------------------------------------------
We now consider an approximate form of the input-output formalism which is valid when the system degree of freedom consists of an oscillator with very low damping, and for which all the frequencies of interest will lie in a narrow range around the oscillator frequency $\omega_{a}$. We start from Eq. (\[Langevin-example\]) and use $$\begin{aligned}
\hat{\Phi} & =\Phi^{ZPF}\left( a+a^{\dag}\right) ,\\
\hat{Q} & =Q^{ZPF}\frac{\left( a-a^{\dag}\right) }{i},\end{aligned}$$ where $\Phi^{ZPF}=\sqrt{\hbar Z_{a}/2}$ and $Q^{ZPF}=\sqrt{\hbar/2Z_{a}}$.
We then obtain, neglecting the effect of driving terms oscillating at twice the resonance frequency, $$\frac{\mathrm{d}}{\mathrm{d}t}a=\frac{i}{\hbar}\left[ H_{bare},a\right]
-\omega_{a}\frac{Z_{a}}{2R}a+\sqrt{\frac{2Z_{a}}{\hbar R}}\tilde
{A}^{\mathrm{in}}\left( t\right)$$ with $$\tilde{A}^{\mathrm{in}}(t)=\int_{0}^{\infty}\hat{A}^{\mathrm{in}}[\omega]e^{-i\omega t}\mathrm{d}\omega.$$ The field amplitude $\tilde{A}^{\mathrm{in}}(t)$ is non-hermitian and contains only the negative frequency component of $A^{\mathrm{in}}(t)$. For signals in a narrow band of frequencies around the resonance frequency, we can make the substitution $$\sqrt{\frac{2}{\hbar\omega_{a}}}\tilde{A}^{\mathrm{in}}\left( t\right)
\rightarrow\tilde{a}^{\mathrm{in}}\left( t\right) ,$$ where $$\tilde{a}^{\mathrm{in}}(t)=\int_{0}^{\infty}a^{\mathrm{in}}[\omega]e^{-i\omega
t}\mathrm{d}\omega.$$ The input field operator $a^{\mathrm{in}}[\omega]$ is identical to $a^{\leftarrow}[\omega]$ of the infinite line. We finally arrive at the RWA quantum Langevin equation, also referred to in the quantum optics literature as the quantum Langevin equation in the Markov approximation $$\frac{\mathrm{d}}{\mathrm{d}t}a=\frac{i}{\hbar}\left[ H_{bare},a\right]
-\frac{\gamma_{a}}{2}a+\sqrt{\gamma_{a}}\tilde{a}^{\mathrm{in}}\left(
t\right) ,$$ where $$\left[ \tilde{a}^{\mathrm{in}}\left( t\right) ,\tilde{a}^{\mathrm{in}}\left( t^{\prime}\right) ^{\dagger}\right] =\delta\left( t-t^{\prime
}\right) .$$ For any oscillator, the input output relationship is obtained from $$\sqrt{\gamma_{a}}a\left( t\right) =\tilde{a}^{\mathrm{in}}\left( t\right)
-\zeta\tilde{a}^{\mathrm{out}}\left( t\right) . \label{IOT}$$ It is worth noting that although $a^{\mathrm{in}}$ and $a^{\mathrm{out}}$ play the role of $a^{\leftarrow}$ and $a^{\rightarrow}$ in Eq. (\[bosonic\_com\]), only the average values of the moments of $a^{\mathrm{in}}$ can be imposed, $a^{\mathrm{out}}$ being a slave" of the dynamics of $a^{\mathrm{in}}$, as processed by the oscillator.
[99]{}
D. I. Schuster, A. A. Houck, J. A. Schreier, A. Wallraff, J. M. Gambetta, A. Blais, L. Frunzio, J. Majer, B. R. Johnson, M. H. Devoret, S. M. Girvin and R. J. Schoelkopf, Nature **445**, 515 (2007).
M. H. Devoret, S. M. Girvin and R. J. Schoelkopf, Annalen der Physik **16**, 767 (2007).
M. A. Castellanos-Beltran, K. D. Irwin, G. C. Hilton, L. R Vale, and K. W. Lehnert, Nature Phys. **4**, 928 (2008).
N. Bergeal, F. Schackhert, M. Metcalfe, R. Vijay, V. E. Manucharyan, L. Frunzio, D.E. Prober, R. J. Schoelkopf, S. M. Girvin and M. H. Devoret, Nature **465**, 64 (2010).
B. Yurke, P. G. Kaminsky, R. E. Miller, E. A. Whittaker, A. D. Smith, A. H. Silver and R. W. Simon, Phys. Rev. Lett. **60**, 764 (1988).
B. Yurke, L. R. Corruccini, P. G. Kaminsky, L. W. Rupp, A. D. Smith, A. H. Silver, R. W. Simon, and E. A. Whittaker, Phys. Rev. A **39**, 2519 (1989).
R. Movshovich, B. Yurke, P. G. Kaminsky, A. D. Smith, A. H. Silver, R. W. Simon, and M. V. Schneider, Phys. Rev. Lett. **65**, 1419 (1990).
J. D. Teufel, T. Donner, M. A. Castellanos-Beltran, J. M. Harlow and K. W. Lehnert, Nature Nanotechnology **4**, 820 (2009).
R. Vijay, D. H. Slichter, and I. Siddiqi, Phys. Rev. Lett. **106**, 110502 (2011).
M. Hatridge, S. Shankar, M. Mirrahimi, F. Schackert, K. Geerlings, T. Brecht, K. Sliwa, B. Abdo, L. Frunzio, S. Girvin, R. Schoelkopf, M. Devoret, submitted.
N. Bergeal, R. Vijay, V. E. Manucharyan, I. Siddiqi, R. J. Schoelkopf, S. M. Girvin and M. H. Devoret, Nature Physics **6**, 296 (2010).
B. Abdo, F. Schackert, M. Hatridge, C. Rigetti and M. H. Devoret, Appl. Phys. Lett. **99**, 162506 (2011).
J. D. Teufel, T. Donner, Li, Dale, J. W. Harlow, M. S. Allman, K. Cicak, A. J. Sirois, J. D. Whittaker, K. W. Lehnert and R. W. Simmonds, Nature **475**, 359 (2011).
T. Yamamoto, K. Inomata, M. Watanabe, K. Matsuba, T. Miyazaki, W. D. Oliver, Y. Nakamura and J. S. Tsai, Appl. Phys. Lett. **93**, 042510 (2008).
R. Vijay, M. H. Devoret and I. Siddiqi, Rev. Sci. Instrum. **80**, 111101 (2009).
I. Siddiqi, R. Vijay, F. Pierre, C. M. Wilson, M. Metcalfe, C. Rigetti, L. Frunzio, and M. H. Devoret, Phys. Rev. Lett. **93**, 207002 (2004).
M. Hatridge, R. Vijay, D. H. Slichter, J. Clarke and I. Siddiqi, Phys. Rev. B **83**, 134501 (2011).
B. Abdo, E. Segev, O. Shtempluck, and E. Buks, Appl. Phys. Lett. **88**, 022508 (2006).
E. A. Tholen, A. Ergul, E. M. Doherty, F. M. Weber, F. Gregis and D. B. Haviland, Appl. Phys. Lett. **90**, 253509 (2007).
B. H. Eom, P. K. Day, H. G. LeDuc, J. Zmuidzinas, Nature Physics **8**, 623 (2012).
Treatment of a linear dissipative environment by scattering formalism, usually employed in quantum optics. See appendix for details.
D. M. Pozar, *Microwave Engineering*, (Wiley, 2005), Third edition.
E. Flurin, N. Roch, F. Mallet, M. H. Devoret, B. Huard, arXiv:1204.0732v1.
B. Abdo, K. Sliwa, F. Schackert, N. Bergeal, M. Hatridge, L. Frunzio and M. H. Devoret, manuscript in preparation.
N. Bergeal, F. Schackert, L. Frunzio and M. H. Devoret, Phys. Rev. Lett. **108**, 123902 (2012).
C. M. Caves, Phys. Rev. D **26**, 1817 (1982).
B. Huard, N. Bergeal and M. H. Devoret, Proceedings of the International School of Physics Enrico Fermi CLXXI 151 (2009).
R. E. Miller, W. H. Mallison, A. W. Kleinsasser, K. A. Delin and E. M. Macedo, Appl. Phys. Lett. **63**, 1423 (1993).
R. Vijay, E. M. Levenson-Falk, D. H. Slichter, and I. Siddiqi, Appl. Phys. Lett. **96**, 223112 (2010).
J. M. Manley and R. H. Rowe, Proceedings of the IRE 904 (1956).
N. Roch, E. Flurin, F. Nguyen, P. Morfin, P. Campagne-Ibarcq, M. H. Devoret, B. Huard, Phys. Rev. Lett. **108**, 147701 (2012).
J. Koch, T. M. Yu, J. Gambetta, A. Houck, D. Schuster, J. Majer, A. Blais, M. H. Devoret, S. M. Girvin and R. J. Schoelkopf, Phys. Rev. A **76**, 042319 (2007).
V. E. Manucharyan, J. Koch, L. I. Glazman and M. H. Devoret, Science **326**, 113 (2009).
A. Wallraff, D. I. Schuster, A. Blais, L. Frunzio, R.-S Huang, J. Majer, S. Kumar, S. M. Girvin, R. J. Schoelkopf, Nature **431**, 162 (2004).
R. F. Bradley, Nucl. Phys. B (Proc. Suppl.) **72**, 137 (1999).
C. W. Gardiner and P. Zoller, *Quantum Noise* (Springer, 2004), Third edition.
Its value can also be obtained as the sum of the electrostatic and magnetic energy in the line $$\frac{1}{2}\int_{-\infty}^{+\infty}\left\{ C_{\ell}\left[ V\left(
x,t\right) \right] ^{2}+L_{\ell}\left[ I\left( x,t\right) \right]
^{2}\right\} dx.$$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, an analytical framework is provided to analyze the energy coverage performance of unmanned aerial vehicle (UAV) energy harvesting networks with clustered user equipments (UEs). Locations of UEs are modeled as a Poisson Cluster Process (PCP), and UAVs are assumed to be located at a certain height above the center of user clusters. Hence, user-centric UAV deployments are addressed. Two different models are considered for the line-of-sight (LOS) probability function to compare their effects on the network performance. Moreover, antennas with doughnut-shaped radiation patterns are employed at both UAVs and UEs, and the impact of practical 3D antenna radiation patterns on the network performance is also investigated. Initially, the path loss of each tier is statistically described by deriving the complementary cumulative distribution function and probability density function. Following this, association probabilities with each tier are determined, and energy coverage probability of the UAV network is characterized in terms of key system and network parameters for UAV deployments both at a single height level and more generally at multiple heights. Through numerical results, we have shown that cluster size and UAV height play crucial roles on the energy coverage performance. Furthermore, energy coverage probability is significantly affected by the antenna orientation and number of UAVs in the network.'
author:
- '[^1] [^2]'
title: Energy Harvesting in Unmanned Aerial Vehicle Networks with 3D Antenna Radiation Patterns
---
Unmanned aerial vehicles (UAVs), energy harvesting, energy coverage probability, Poisson point processes, Poisson cluster processes, Thomas cluster processes, 3D antenna radiation patterns, stochastic geometry.
Introduction
============
To support the unprecedented growth in demand for mobile data fueled by emerging wireless applications and the increased use of smart mobile devices, new technologies and designs are being incorporated into next-generation cellular networks. One novelty is expected to be the deployment of unmanned aerial vehicle (UAV) base stations (BSs). UAVs have been primarily considered as high-altitude platforms at altitudes of kilometers to provide coverage in rural areas. On the other hand, use of low-altitude UAVs has also become popular recently due to the advantage of having better link quality in shorter-distance line-of-sight (LOS) channels with the ground users. Moreover, owing to the relative flexibility in UAV deployments, UAV BSs can be employed in a variety of scenarios including public safety communications and data collection in Internet of Things (IoT) applications [@Zhang2], [@Zeng2]. Other scenarios include disasters, accidents, and other emergencies and also temporary events requiring substantial network resources in the short-term such as in concerts and sporting events. In such cases, UAVs can be deployed rapidly to provide seamless wireless connectivity [@Motlagh]–[@Bor-Yaliniz].
In addition to growing data traffic, increasing number of devices results in a significant growth in energy demand. RF (radio frequency) energy harvesting where a harvesting device may extract energy from the incident RF signals has emerged as a promising solution to power up low-power consuming devices [@Zeng], [@Bi]. Therefore, the advances in energy harvesting technologies have motivated research in the study of different wireless energy harvesting networks. For example, wireless energy and/or information transfer in large-scale millimeter-wave and microwave networks has been studied in [@Khan]–[@Renzo]. In these works, energy is harvested wirelessly from energy transmitters which are generally deployed at fixed locations. However, low-power consuming devices can potentially be distributed in a large area. In such cases, the performance of energy harvesting will be limited by the low end-to-end power transmission efficiency due to the loss of RF signals over long distances [@Xu].
In order to improve the efficiency, instead of fixed energy transmitters such as ground base stations (BSs), the deployment of mobile energy transmitters is proposed recently. In particular, although the UAVs are typically power-limited, UAV-assisted energy harvesting has become attractive due to the flexibility and relative ease in deploying UAV BSs. In [@Wu], mobility of the UAV with a directional antenna is exploited by jointly optimizing the altitude, trajectory, and transmit beamwidth of the UAV in order to maximize the energy transferred to two energy receivers over a finite charging period. In [@Xu], authors consider a more general scenario with more than two energy receivers where the amount of received energy by all energy receivers is maximized via trajectory control. In [@Hu], a UAV-enabled wireless power transfer network is studied as well. Minimum received energy among all ground nodes is maximized by optimizing the UAV’s one-dimensional trajectory. Both downlink wireless power transfer and uplink information transfer is considered in [@Xie] with one UAV and a set of ground users. The UAV charges the users in downlink and users use the harvested energy to send the information to the UAV in the uplink. Similarly, a wireless-powered communication network with a mobile hybrid access point UAV is considered in [@Cho]. In this work, the UAV performs weighted energy transfer and receives information from the far-apart nodes based on the weighted harvest-then-transmit protocol. The use of UAVs to power up energy constrained sensor nodes has been suggested in [@Caillouet]. UAVs are considered to be equipped with the dedicated chargers facing the ground so that they can recharge the sensors’ batteries using RF energy harvesting. A practical example is AT&T’s flying Cell on Wings (COW) drones [@ATT]. In particular, AT&T’s COW is a UAV equipped with a small base station. Flying COW is tethered to a power source on the ground with a cable. In these types of practical implementations, power constraints are more relaxed and such a UAV can provide both cell service and also supply wireless energy to sensors and small IoT devices.
In a separate line of research in the literature, the performance of UAV-assisted wireless networks is extensively studied recently. Similar to 2D networks, stochastic geometry has been employed in the network level analysis of UAV networks by considering UAVs distributed randomly in 3D space. Effect of different network parameters on the coverage probability is explored in several recent works such as [@Galkin]–[@Zhou]. In [@Liu], authors analyzed the network performance for three different type of LOS probability models. Spectrum sharing in UAV networks is analyzed in [@Zhang]–[@Lyu]. Additionally, optimal deployment of UAVs is investigated in [@Bor-Yaliniz2]–[@Mozaffari1].
It is important to note that the antenna number, type, and orientation are critical factors that affect the performance in UAV-assisted networks. Indeed, several recent studies, e.g., [@Chandhar] and [@Geraci], have addressed scenarios in which antenna arrays are deployed in UAV-assisted cellular networks. Similarly, in [@Lyu2], the authors have considered directional antennas for UAVs. However, a practical antenna pattern which is omnidirectional in the horizontal plane but directional in the vertical plane is employed for ground BSs. Regarding the antenna type, omnidirectional antennas can also be used in UAVs, especially considering their mobility [@Khawaja]. At the same time, since even the UAV’s own body can shadow the antenna and result in a poor link quality, the orientation of the antennas plays an important role on the performance [@Yanmaz]. There has been limited analytical and experimental works studying the effect of three dimensional (3D) antenna radiation patterns on the link quality between the UAV and ground users. In [@Yanmaz], impact of antenna orientation is investigated by placing two antennas on a fixed wing UAV flying on a linear path with 802.11a interface. Similarly, path loss and small-scale fading characteristics of UAV-to-ground user links are analyzed with a simple antenna extension to 802.11 devices in [@Yanmaz2]. In [@Chen], ultra-wideband (UWB) antennas with doughnut-shaped radiation patterns are employed at both UAVs and ground users to analyze the link quality at different link distances, UAV heights, and antenna orientations. The large bandwidth of UWB radio signals is utilized in the measurements to obtain a high temporal resolution of multipath components. Authors develop a simple analytical model to approximate the impact of the 3D antenna radiation pattern on the received signal. However, none of these works study the effect of UAV antenna orientation on the network performance.
Contributions and Organization
------------------------------
In this paper, we consider a UAV network consisting of UAVs operating at a certain altitude above ground. The locations of user equipments (UEs) are modeled as a Poisson cluster process (PCP), and the UAVs are assumed to be located at a certain height above the center of user clusters. Since UAVs are deployed in overloaded scenarios, locations of UAVs and UEs are expected to be correlated and UEs are more likely to form clusters. Hence, modeling the UE locations by PCP is more appropriate and realistic. Moreover, we consider that UWB antennas with doughnut-shaped radiation patterns are employed at both UAVs and UEs, and we study the effect of practical 3D antenna radiation patterns on energy harvesting from UAVs.
More specifically, our main contributions can be summarized as follows:
- An analytical framework is provided to analyze energy coverage performance of a UAV network with clustered UEs. UE locations are assumed to be PCP distributed to capture the correlations between the UAV and UE locations.
- We divide the network into two tiers: $0^{{\text{th}}}$ tier UAV and $1^{{\text{st}}}$ tier UAVs. $0^{{\text{th}}}$ tier UAV is the cluster center UAV around which the typical UE is located, while other UAVs constitute the $1^{{\text{st}}}$ tier.
- Two different LOS probability functions, i.e., a high-altitude model and a low-altitude model, are considered in order to investigate and compare their impact on the network performance.
- Different from the previous studies, more practical antennas with doughnut-shaped radiation patterns are employed at both UAVs and UEs to provide a more realistic performance evaluation for the network.
- We first derive the complementary cumulative distribution functions (CCDFs) and the probability density functions (PDFs) of the path losses for each tier, then obtain the association probabilities by using the averaged received power UAV association rule.
- Average harvested power expression is obtained. Then, total energy coverage probability is determined by deriving the Laplace transforms of the interference terms arising from each tier. Energy coverage is characterized for UAV deployments at a single height level and also at multiple heights.
The rest of the paper is organized as follows. In Section \[sec:system\_model\], system model is introduced, and path loss, blockage, and 3D antenna models are described. Path loss and association probabilities are statistically characterized in Section \[sec:Path Loss and Cell Association\]. In Section \[sec:Energy Coverage\_Probability Analysis\], energy coverage probability of the UAV network is determined. In Section \[sec:Simulation and Numerical Results\], simulations and numerical results are provided, demonstrating the impact of several key parameters on the energy coverage performance of the network. Finally, Section \[sec:Conclusion\] concludes the paper. Proofs are included in the Appendix.
System Model {#sec:system_model}
============
In this section, we describe the system model of the UAV network with clustered UEs. We address a downlink network, in which the spatial distribution of the UAVs is modeled by an independent homogeneous PPP $\Phi_{U}$ with density $\lambda_{U}$ on the Euclidean plane. The height of UAVs is denoted by $H$. Note that UAVs can be used to offload traffic from the ground cellular BSs and reduce congestion around hotspots. In energy harvesting applications, UAVs can be used to transfer energy to e.g., ground sensors and low-power IoT devices, to energize them. They can also be deployed in case of emergencies during which ground infrastructure is strained [@Guvenc]. In our model, the UEs are clustered around the projections of UAVs on the ground, and the locations of the clustered UEs is described by a PCP, denoted by $\Phi_C$. In applications involving UAVs, UEs are expected to be located in high UE density areas, forming clusters. Consequently, modeling of UE distribution as a PCP rather than a homogeneous PPP is more accurate.
In this paper, we model $\Phi_C$ as a Thomas cluster process, where the UEs are symmetrically independently and identically distributed (i.i.d.) around the cluster centers (which are projections of UAVs on the ground), according to a Gaussian distribution with zero mean and variance $\sigma_c^2$. Therefore, the UE’s location is statistically described by the following PDF and CCDF [@Haenggi]: $$\begin{aligned}
f_{D}(d)&=\frac{d}{\sigma_c^2} \exp\left( -\frac{d^2 }{2\sigma_c^2}\right), \quad d \geq 0, \label{PDF_of_d} \\
\bar{F}_{D}(d)&=\exp\left( -\frac{d^2 }{2\sigma_c^2}\right), \quad d \geq 0, \label{Fbar_R}\end{aligned}$$ respectively, where $d$ is the 2D distance of a UE with respect to the cluster center on the ground. Without loss of generality, we perform the analysis for a typical UE which is randomly chosen from a randomly chosen cluster. Since cluster centers; i.e. UAVs, are PPP distributed and the PPP is stationary, location of this typical UE can be transformed to the origin according to Slivnyak’s theorem [@Saha]. The typical UE is assumed to be is associated with the UAV providing the maximum average received power. Although we have only a single-tier network composed of UAVs, we also consider an additional tier, named as $0^{{\text{th}}}$ tier that only includes the cluster center of the typical UE similarly as in [@Esma] and [@Saha]. Therefore, overall UAV density is equal to tier 1 density, and tier 0 has only one UAV, which is the cluster center of the typical UE. Essentially tier 0 is introduced to differentiate the cluster-center UAV from other UAVs, because the distance distribution equations of the typical UE to its own cluster-center UAV given in (\[PDF\_of\_d\]) and (\[Fbar\_R\]) are different from the distribution of the distances to other UAVs. Thus, our network model can be considered as a two-tier network consisting of a $0^{{\text{th}}}$ tier cluster-center UAV and $1^{{\text{st}}}$ tier UAVs. The considered network model is depicted in Fig. \[Fig\_Network\_Model\].
![Network model for a UAV energy harvesting network. BSs are distributed as a PPP, while UEs are Gaussian distributed around the cluster centers (projections of UAVs on the ground). Both BS and UEs are equipped with UWB antennas with different antenna orientations. []{data-label="Fig_Network_Model"}](network.ps){width="\figsize\textwidth"}
Path Loss and Blockage Modeling
-------------------------------
A transmitting UAV can either have a line-of-sight (LOS) or non-line-of-sight (NLOS) link to the typical UE. Consider an arbitrary link of length $r$ between a UE and a UAV, and define the LOS probability function as the probability that the link is LOS. Different LOS probability functions have been used in the literature. In this paper, we consider the two models proposed in [@Al-Hourani] and [@3GPP], which are high-altitude and low-altitude models, respectively.
High-altitude model is widely used especially in satellite communications where the altitude is around hundred of kilometers. It has also been widely employed in UAV-assisted networks recently. LOS probability function for the high-altitude model is given as follows: $$\mathcal{P}_{{\text{L}}}^{{\text{high}}}(r)=\left(\frac{1}{1+b\exp\left(-c\left(\frac{180}{\pi}\sin^{-1}\left(\frac{H}{r}\right)-b\right)\right)}\right), \label{LOS_probability1}$$ where $r$ is the 3D distance between the UE and UAV, $H$ is the UAV height, $b$ and $c$ are constants whose values depend on the environment. It can be easily verified that the LOS probability in (\[LOS\_probability1\]) increases as the UAV height, $H$, increases.
Since practical values for UAV height in certain applications is around $50\sim100$ meters, a more realistic LOS probability function proposed for 3GPP terrestrial communications is employed also for UAV networks in [@Liu]. The height of a macrocell base station is usually around 32 m, which is comparable to the practical UAV height. Therefore, employment of the LOS probability function for 3GPP macrocell-to-UE communciation is also reasonable for the UAV networks in such relatively low-altitude scenarios. For the low-altitude model, LOS probability function is expressed as $$\mathcal{P}_{{\text{L}}}^{{\text{low}}}(r)= \min\left(1,\frac{18}{r}\right)\left(1-\exp\left(-\frac{r}{63}\right)\right)+\exp\left(-\frac{r}{63}\right). \label{LOS_probability2}$$ Note that different from the high-altitude model, LOS probability function in (\[LOS\_probability2\]) decreases with the increase in the 3D distance $r$, independent of the UAV height. In Fig. \[Fig\_LOS\_func\], LOS probability function is plotted using high-altitude and low-altitude models. Solid lines show the LOS probability as a function of the UAV height $H$ when the 2D distance to the UAV is fixed at $d=10$ m, and dashed lines display the LOS probability as a function of the 2D distance to the UAV $d$ when the UAV height is $H=50$ m. As shown in Fig. \[Fig\_LOS\_func\], LOS probability increases with increasing UAV height when the high-altitude model is used, and decreases when the low-altitude model is considered. We observe that the LOS probability decreases for both models as the 2D distance to the UAV increases. We also note that the analysis in the remainder of the paper is general and is applicable to any LOS probability function. Only in Section \[sec:Simulation and Numerical Results\], we employ the LOS probability functions in (\[LOS\_probability1\]) and (\[LOS\_probability2\]) to obtain the numerical results.
[0.45]{} ![LOS probability function for high-altitude and low-altitude models as a function of (a) UAV height $H$ and (b) 2D distance to the UAV $d$. []{data-label="Fig_LOS_func"}](LOS_funcH.eps "fig:"){width="\textwidth"}
[0.45]{} ![LOS probability function for high-altitude and low-altitude models as a function of (a) UAV height $H$ and (b) 2D distance to the UAV $d$. []{data-label="Fig_LOS_func"}](LOS_funcd.eps "fig:"){width="\textwidth"}
In general, NLOS links experience higher path loss than the LOS links due to signals being reflected and scattered in NLOS propagation environments. Therefore, different path loss laws are applied to LOS and NLOS links. Thus, the path loss on each link in tier $k \in \{0,1\}$ can be expressed as $$\label{PL_model}
\begin{split}
L_{k,{\text{L}}}(r)&= r^{\alpha_{{\text{L}}}} \\
L_{k,{\text{N}}}(r)&= r^{\alpha_{{\text{N}}}} ,
\end{split}$$ where $\alpha_{{\text{L}}}$ and $\alpha_{{\text{N}}}$ are the LOS and NLOS path-loss exponents, respectively.
3D Antenna Modeling
-------------------
In this paper, we consider the analytical model developed in [@Chen] for the effect of 3D antenna radiation patterns on the received signal. UWB transmitter and receiver antennas with doughnut-shaped radiation patterns centered at a frequency of 4 GHz are placed at the UAV and UE, respectively, and air-to-ground channel measurements are carried out in order to characterize the impact of the 3D antenna radiation pattern on the received signal for different antenna orientations in [@Chen]. As a result of these measurements, transmitter and receiver antenna gains are modeled analytically for horizontal-horizontal (HH), horizontal-vertical (HV) and vertical-vertical (VV) antenna orientations as follows: $$G_k(\theta)=G_{\text{TX}}(\theta)G_{\text{RX}}(\theta)=\begin{cases}
\sin(\theta)\sin(\theta) & \text{for} \quad \text{HH} \\
\sin(\theta)\cos(\theta) & \text{for} \quad \text{HV} , \\
\cos(\theta)\cos(\theta) & \text{for} \quad \text{VV} \label{Antenna_gain}
\end{cases}$$ where $\theta$ is the elevation angle between the transmitter at the UAV and the receiver at the UE on the ground. In this antenna model, radiation pattern is approximated by a circle in the vertical dimension, while it is assumed to be constant for all horizontal directions. In other words, antenna gains depend only on the elevation angle $\theta$, and are considered as independent of the azimuth angle between the transmitter at the UAV and the receiver at the UE. Approximated antenna radiation patterns of UAV and UE are shown in Fig. \[Fig\_Antenna\_Model\] for HH antenna orientation. They can be plotted for HV and VV orientations as well by rotating the transmitter and/or receiver antennas by $90^{\circ}$. Note that for HH antenna orientation $G_{\text{TX}}(\theta)=G_{\text{RX}} \rightarrow 0$ as $\theta \rightarrow 0$ which happens when the UEs are located far away from the cluster center, i.e. as the $\sigma_c$ increases, and $G_{\text{TX}}(\theta)=G_{\text{RX}} \rightarrow 1$ as $\theta \rightarrow 90^{o}$ which happens when the UEs get closer to the cluster center. Similar observations can be drawn for VH and VV antenna orientations. Effective antenna gain $G_k$ as a function of $r$ can be rewritten in terms of UAV height $H$ and the path loss on each link in tier $k \in \{0,1\}$ as $$\small
G_k(r)=\begin{cases}
H^2 L_{k,s}^{-\frac{2}{\alpha_s}}(r) & \text{for} \quad \text{HH} \\
H \left(\sqrt{L_{k,s}^{\frac{2}{\alpha_s}}(r)-H^2} \right) L_{k,s}^{-\frac{2}{\alpha_s}}(r) & \text{for} \quad \text{HV} \\
\left(L_{k,s}^{\frac{2}{\alpha_s}}(r)-H^2 \right) L_{k,s}^{-\frac{2}{\alpha_s}}(r) & \text{for} \quad \text{VV}.
\end{cases} \normalsize$$
![Approximated antenna radiation pattern for HH antenna orientation. []{data-label="Fig_Antenna_Model"}](antenna_model.eps){width="\figsize\textwidth"}
In the rest of the analysis, we assume that the typical UE and all UAVs in the network have horizontal antenna orientation. Therefore, HH antenna orientation for the main link and interfering links are considered due to its analytical tractability. Moreover, UEs are considered to be clustered around the projections of UAVs on the ground and more UEs are encouraged to be associated with their cluster center UAV. As a result, the angle between the transmitter at the UAV and the receiver at the UE is expected to be large. Therefore, HH antenna orientation is more suitable than the other two orientations. However, in the numerical results section, simulation results for HV and VV orientations are also provided in order to compare their effect on the UAV association and energy coverage probabilities.
Finally, we note that a summary of notations is provided in Table \[Table\_notation\] below.
**Notations** **Description**
------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------
$\Phi_U, \lambda_U, \Phi_C $ PPP of UAVs, the density of $\Phi_U$, PCP of UEs.
$H$ Height of UAVs.
$\sigma_c$ UE distribution’s standard deviation.
$d$ 2D distance of a UE with respect to the cluster center on the ground.
$r$ 3D distance between the UE and UAV.
$\mathcal{P}_{{\text{L}}}^{{\text{high}}}(r), \mathcal{P}_{{\text{L}}}^{{\text{low}}}(r)$ LOS probability functions for high and low altitude models.
$b, c$ Environment dependent constants.
$L_{k,{\text{L}}}(r), L_{k,{\text{N}}}(r)$ Path losses on LOS and NLOS links in the $k^{th}$ tier.
$\alpha_{{\text{L}}}, \alpha_{{\text{N}}}$ LOS and NLOS path-loss exponents.
$G_{\text{TX}}(\theta), G_{\text{RX}}(\theta)$ Transmitter and receiver antenna gains.
$G_k(r)$ Effective antenna gain.
$\theta$ Elevation angle between the transmitter and the receiver.
$\bar{F}_{L_{k}}(x)$ CCDF of the path loss from a typical UE to a $k^{th}$ tier UAV.
$f_{L_{k,s}}(x)$ PDF of the path loss from a typical UE to a $k^{th}$ tier LOS/NLOS UAV.
$P_k$ Transmit power of UAVs in the $k^{th}$ tier.
$h$ Small-scale fading gain.
$\mathcal{A}_{k,s}$ Association probability with a $k^{th}$ tier LOS/NLOS UAV.
$I_{j,k}$ Interference from the $j^{th}$ tier UAV, when the main link is in the $k^{th}$ tier.
$\mathcal{E}_{k,0}$ Exclusion disc of path loss, inside which no interference exists.
$P_{r,k}$ Total received power
${\text{E}^{\text{C}}}_{k,s}(\Gamma_k)$ Conditional energy coverage probability given that the UE is associated with a $k^{{\text{th}}}$ tier LOS/NLOS UAV.
$\Gamma_k$ Energy outage threshold.
$\xi$ Rectifier efficiency.
$\mathcal{L}_{I_{j,k}}(\Gamma_k,\mathcal{E}_{k,0})$ Laplace transform of $I_{j,k}$.
: Table of Notations
\[Table\_notation\]
Path Loss and UAV Association {#sec:Path Loss and Cell Association}
=============================
Statistics of the Path Loss
---------------------------
In this section, we provide statistical characterizations by identifying the CCDF and the PDF of the path loss.
*Lemma 1:* The CCDF of the path loss from a typical UE to a $0^{{\text{th}}}$ tier UAV can be formulated as $$\begin{aligned}
\bar{F}_{L_{0}}(x)&= \hspace{-0.3cm} \sum_{s \in \{{\text{L}},{\text{N}}\}} \bar{F}_{L_{0,s}}(x) \nonumber \\
&= \hspace{-0.3cm} \sum_{s \in \{{\text{L}},{\text{N}}\}} \int_{\sqrt{x^{2/\alpha_{s}}-H^2}}^{\infty} \mathcal{P}_s(\sqrt{d^2+H^2}) f_D(d) \mathrm{d}d, \label{CCDF_0}\end{aligned}$$ where $f_D(d)$ is given in (\[PDF\_of\_d\]), and $\mathcal{P}_{s}(\cdot)$ is the LOS or NLOS probability depending on whether $s = {\text{L}}$ or $s = {\text{N}}$ [^3].
*Proof:* See Appendix \[Proof of Lemma 1\]. $\square$
*Lemma 2:* CCDF of the path loss from a typical UE to a $1^{{\text{st}}}$ tier UAV is $$\bar{F}_{L_1}(x)= \prod_{s \in \{{\text{L}},{\text{N}}\}} \bar{F}_{L_{1,s}}(x) = \prod_{s \in \{{\text{L}},{\text{N}}\}} \exp\big(-\Lambda_{1,s}([0,x))\big), \label{CCDF_1}$$ where $\Lambda_{1,s}([0,x))$ is given by $$\begin{aligned}
\Lambda_{1,s}([0,x))= 2\pi\lambda_U \int_{H}^{x^{\frac{1}{\alpha_{s}}}} \mathcal{P}_{s}(r) r \mathrm{d}r . \label{intensity_function_1}\end{aligned}$$
*Proof:* See Appendix \[Proof of Lemma 2\]. $\square$
*Corollary 1:* The PDF of the path loss from a typical UE to a $0^{{\text{th}}}$ tier LOS/NLOS UAV can be determined by applying the Leibniz integral rule as follows: $$\begin{aligned}
f_{L_{0,s}}(x)&=-\frac{d\bar{F}_{L_{0,s}}(x)}{dx} \nonumber \\
& =\frac{1}{\sigma_c^2} \frac{x^{\frac{2}{\alpha_{s}}-1}}{\alpha_s} \mathcal{P}_{s}\left(x^{\frac{1}{\alpha_{s}}}\right) \exp\left(-\frac{1}{2\sigma_c^2} \left(x^{\frac{2}{\alpha_{s}}}-H^2\right)\right). \label{f_L0s}\end{aligned}$$
*Corollary 2:* The PDF of the path loss from a typical UE to a $1^{{\text{st}}}$ tier LOS/NLOS UAV is $$f_{L_{1,s}}(x)=-\frac{d\bar{F}_{L_{1,s}}(x)}{dx}=\Lambda_{1,s}^{\prime}([0,x)) \exp\big(-\Lambda_{1,s}([0,x))\big), \label{f_L1s}$$ where $\Lambda_{1,s}^{\prime}([0,x))$ is given by $$\Lambda_{1,s}^{\prime}([0,x)) =2\pi\lambda_U \frac{x^{\frac{2}{\alpha_{s}}-1}}{\alpha_{s}}\mathcal{P}_{s}\left(x^{\frac{1}{\alpha_{s}}}\right) \label{Lambda_1s_prime}$$ by again applying the Leibniz integral rule.
In the results above, we have determined the CCDFs and PDFs of the path loss for each tier. They depend on the key network parameters including the variance of the cluster process $\sigma_c^2$, UAV density $\lambda_U$, UAV LOS probability $\mathcal{P}_{s}(\cdot)$, UAV height $H$ and path loss exponents $\alpha_s$. In the following sections, these distributions are utilized in determining the association and energy coverage probabilities.
UAV Association
---------------
In this work, we assume that the UEs are associated with a UAV providing the strongest long-term averaged power. In other words, a typical UE is associated with its cluster center UAV, i.e., the $0^{{\text{th}}}$ tier UAV, if $$P_0 G_0(r) L_0^{-1}(r) \geq P_1 G_1(r) L_{\text{min},1}^{-1}(r),$$ where $P_k$ and $G_k(r)$ denote the transmit power and antenna gain of the link, respectively, in tier $k \in (0,1)$. $L_0(r)$ is the path loss from the $0^{{\text{th}}}$ tier UAV, and $L_{\text{min},1}(r)$ is the path loss from $1^{{\text{st}}}$ tier UAV providing the minimum path loss. In the following lemma, we provide the association probabilities using the result of Lemmas 1, Lemma 2, Corollary 1 and Corollary 2.
*Lemma 3:* The association probabilities with a $0^{{\text{th}}}$ tier LOS/NLOS UAV and $1^{{\text{st}}}$ tier LOS/NLOS UAV are given, respectively, as $$\begin{aligned}
\mathcal{A}_{0,s} = \int_{ H^{\alpha_{s}}}^{\infty} &\prod_{m \in \{{\text{L}},{\text{N}}\}} \bar{F}_{L_{1,m}}\left(\left(\frac{P_1}{P_0}l_{0,s}^{\frac{2}{\alpha_s}+1}\right)^{\frac{\alpha_m}{\alpha_m+2}}\right) \nonumber \\
& \times f_{L_{0,s}}(l_{0,s}) \mathrm{d}l_{0,s}, \label{Association_Prob0}\end{aligned}$$ $$\begin{aligned}
\mathcal{A}_{1,s}= \int_{ H^{\alpha_{s}}}^{\infty} & \sum_{m \in \{{\text{L}},{\text{N}}\}} \bar{F}_{L_{0,m}}\left(\left(\frac{P_0}{P_1}l_{1,s}^{\frac{2}{\alpha_s}+1}\right)^{\frac{\alpha_m}{\alpha_m+2}}\right) \nonumber \\
& \times \bar{F}_{L_{1,s^{\prime}}}(l_{1,s}) f_{L_{1,s}}(l_{1,s})\mathrm{d}l_{1,s}, \label{Association_Prob1}\end{aligned}$$ where $s, s' \in \{{\text{L}},{\text{N}}\}$ and $s \neq s'$.
*Proof*: See Appendix \[Proof of Lemma 3\]. $\square$
In the corollary below, we provide a closed-form expression for the association probability in a special case with which the effects of different parameters on association probability can be easily identified.
*Corollary 3:* Consider the same UAV network with $P_0=P_1$ and the LOS probability $\mathcal{P}_{{\text{L}}}(\cdot)=1$, i.e., all UAVs are LOS to the typical UE. Then, the association probabilities specialize to the following expressions (which also confirm the results in [@Wang]): $$\begin{aligned}
\mathcal{A}_{0,L}&=\frac{1}{1+2\pi\lambda_U\sigma_c^2} \\
\mathcal{A}_{1,L}&=\frac{2\pi\lambda_U\sigma_c^2}{1+2\pi\lambda_U\sigma_c^2}\end{aligned}$$ with $\mathcal{A}_{0,N}=\mathcal{A}_{1,N}=0$. According to the these results, a typical UE obviously prefers to connect to the $0^{{\text{th}}}$ tier UAV when the value of $\sigma_c$ is small, and connect to a $1^{{\text{st}}}$ tier UAV for higher values of $\sigma_c$ and $\lambda_U$.
Energy Coverage Probability Analysis {#sec:Energy Coverage_Probability Analysis}
====================================
In this section, we use stochastic geometry and adopt an analytical framework to characterize the energy coverage probability for a typical UE clustered around the $0^{{\text{th}}}$ tier UAV (i.e., its own cluster-center UAV).
Downlink Power Transfer
-----------------------
The total power received at a typical UE at a random distance $r$ from its associated UAV in the $k^{{\text{th}}}$ tier can be written as $$\begin{aligned}
P_{r,k}&=S_k+ \sum_{j=0}^{1} I_{j,k} \quad \text{for} \quad k=0,1, \label{total_received_power}\end{aligned}$$ where the received power from the serving UAV $S_k$ and the interference power received from the UAVs in the $j^{\text{th}}$ tier $I_{j,k}$ are given as follows: $$\begin{aligned}
S_k&=P_k G_k(r) h_{k,0} L_k^{-1}(r), \label{received_power} \\
I_{0,1}&= P_0 G_0(r)h_{0,0} L_0^{-1}(r), \label{interference0}\\
I_{1,k}&= \sum_{i \in \Phi_{U}\setminus{\mathcal{E}_{k,0}}} P_1 G_{i}(r) h_{1,i} L_{i}^{-1}(r), \label{interference1}\end{aligned}$$ where $h_{k,0}$ and $h_{j,i} $ are the small-scale fading gains from the serving and interfering UAVs, respectively. Note that since only one UAV exists in the $0^{{\text{th}}}$ tier, $I_{0,0}=0$. $h$ denotes the small-scale fading gain and is assumed to be exponentially distributed. From the UAV association policy, when a typical UE is associated with a UAV whose path loss is $L_k(r)$, there exists no UAV within a disc $\mathcal{E}_{k,0}$ centered at the origin, which is also known as the exclusion disc. In this work, we also consider a linear energy harvesting model in which energy can be harvested if the received power is larger than zero. Therefore, the average harvested power at a typical UE is given in the following theorem.
*Lemma 4:* The average harvested power at a typical UE at a random distance $r$ from its associated UAV in the $k^{{\text{th}}}$ tier is given at the top of the next page in (\[lemma4\])
$$\begin{aligned}
P^{{\text{avg}}}&= \sum_{s \in \{{\text{L}},{\text{N}}\}} \Bigg[ \int_{H^{\alpha_{s}}}^{\infty}
\left[P_0H^2 l_{0,s}^{-\left(1+\frac{2}{\alpha_s}\right)}+\Psi_{I_{1,0}}(\mathcal{E}_{0,0})\right] \prod_{m \in \{{\text{L}},{\text{N}}\}} \bar{F}_{L_{1,m}}\left(\left(\frac{P_1}{P_0}l_{0,s}^{\frac{2}{\alpha_s}+1}\right)^{\frac{\alpha_m}{\alpha_m+2}}\right) f_{L_{0,s}}(l_{0,s}) \mathrm{d}l_{0,s} \nonumber \\
&+\int_{H^{\alpha_{s}}}^{\infty}
\left[P_1H^2 l_{1,s}^{-\left(1+\frac{2}{\alpha_s}\right)}+\sum_{j=0}^{1}\Psi_{I_{j,1}}(\mathcal{E}_{1,0})\right] \sum_{m \in \{{\text{L}},{\text{N}}\}} \bar{F}_{L_{0,m}}\left(\left(\frac{P_0}{P_1}l_{1,s}^{\frac{2}{\alpha_s}+1}\right)^{\frac{\alpha_m}{\alpha_m+2}}\right) \bar{F}_{L_{1,s^{\prime}}}(l_{1,s}) f_{L_{1,s}}(l_{1,s}) \mathrm{d}l_{1,s}\Bigg] \label{lemma4}\end{aligned}$$
where $$\begin{aligned}
\Psi_{I_{0,k}}(\mathcal{E}_{k,0}) = \sum_{s^{\prime} \in \{{\text{L}},{\text{N}}\}} \int_{\mathcal{E}_{k,0}}^{\infty} P_0 H^2 x^{-\left(1+\frac{2}{\alpha_{s^{\prime}}}\right)} f_{L_{0,s^{\prime}}}(x) \mathrm{d}x, \label{Psi0}\end{aligned}$$ $$\begin{aligned}
\Psi_{I_{1,k}}(\mathcal{E}_{k,0})= \sum_{s^{\prime} \in \{{\text{L}},{\text{N}}\}} \int_{\mathcal{E}_{k,0}}^{\infty} P_1 H^2 x^{-\left(1+\frac{2}{\alpha_{s^{\prime}}}\right)} \Lambda_{1,s^{\prime}}^{\prime}([0,x)) \mathrm{d}x. \label{Psi1}\end{aligned}$$
*Proof*: See Appendix \[Proof of Lemma 4\]. $\square$
Energy Coverage Probability
---------------------------
The energy harvested at a typical UE in unit time is expressed as $E_k=\xi P_{r,k}$ where $\xi \in (0,1]$ is the rectifier efficiency, and $P_{r,k}$ is the total received power given in (\[total\_received\_power\]). Since the effect of additive noise power is negligibly small relative to the total received power, it is omitted [@Khan]. The conditional energy coverage probability ${\text{E}^{\text{C}}}_k(\Gamma_k)$ is the probability that the harvested energy $E_k$ is larger than the energy outage threshold $\Gamma_k>0$ given that the typical UE is associated with a UAV from the $k^{{\text{th}}}$ tier, i.e., ${\text{E}^{\text{C}}}_k(\Gamma_k)= \mathbb{P}(E_k>\Gamma_k|t=k)$. Therefore, total energy coverage probability ${\text{E}^{\text{C}}}$ for the typical UE can be formulated as $${\text{E}^{\text{C}}}=\sum_{k=0}^1 \sum_{s \in \{{\text{L}},{\text{N}}\}} \left[{\text{E}^{\text{C}}}_{k,s} (\Gamma_k)\mathcal{A}_{k,s}\right], \label{CoverageProbability}$$ where ${\text{E}^{\text{C}}}_{k,s}(\Gamma_k)$ is the conditional energy coverage probability given that the UE is associated with a $k^{{\text{th}}}$ tier LOS/NLOS UAV, and $\mathcal{A}_{k,s}$ denotes the association probability. The following theorem provides our main characterization regarding the total energy coverage probability.
\[theo:energycoverage\] In a UAV network with practical HH antenna radiation patterns and clustered UEs, the total energy coverage probability for the typical UE is approximately given at the top of the next page in (\[total\_energy\_coverage\])
$$\begin{aligned}
{\text{E}^{\text{C}}}&\approx \sum_{s \in \{{\text{L}},{\text{N}}\}} \sum_{n=0}^{\mathcal{N}}(-1)^{n} {\mathcal{N} \choose n} \nonumber \\
& \times \Bigg[ \int_{H^{\alpha_{s}}}^{\infty}\left(1+\hat{a} P_0H^2 l_{0,s}^{-\left(1+\frac{2}{\alpha_s}\right)}\right)^{-1} \mathcal{L}_{I_{1,0}}\left(\Gamma_0,\mathcal{E}_{0,0}\right) \prod_{m \in \{{\text{L}},{\text{N}}\}} \hspace{-0.3cm} \bar{F}_{L_{1,m}}\left(\left(\frac{P_1}{P_0}l_{0,s}^{\frac{2}{\alpha_s}+1}\right)^{\frac{\alpha_m}{\alpha_m+2}}\right) f_{L_{0,s}}(l_{0,s}) \mathrm{d}l_{0,s} \nonumber \\
&+\int_{H^{\alpha_{s}}}^{\infty} \left(1+\hat{a} P_1H^2 l_{1,s}^{-\left(1+\frac{2}{\alpha_s}\right)}\right)^{-1} \left(\prod_{j=0}^{1} \mathcal{L}_{I_{j,1}}\left(\Gamma_1,\mathcal{E}_{1,0}\right)\right) \sum_{m \in \{{\text{L}},{\text{N}}\}} \hspace{-0.3cm} \bar{F}_{L_{0,m}}\left(\left(\frac{P_0}{P_1}l_{1,s}^{\frac{2}{\alpha_s}+1}\right)^{\frac{\alpha_m}{\alpha_m+2}}\right) \bar{F}_{L_{1,s^{\prime}}}(l_{1,s}) f_{L_{1,s}}(l_{1,s})\mathrm{d}l_{1,s} \Bigg] \label{total_energy_coverage}\end{aligned}$$
where $\hat{a}=\frac{n\eta }{\Gamma_k /\xi}$, $\eta=\mathcal{N}(\mathcal{N}!)^{-\frac{1}{\mathcal{N}}}$, $\mathcal{N}$ is the number of terms in the approximation and the Laplace transforms of the interference terms are given by $$\begin{aligned}
&\mathcal{L}_{I_{0,k}}(\Gamma_k,\mathcal{E}_{k,0}) \nonumber \\
& =\sum_{s^{\prime} \in \{{\text{L}},{\text{N}}\}} \int_{\mathcal{E}_{k,0}}^{\infty} \left(1+\hat{a} P_0 H^2 x^{-\left(1+\frac{2}{\alpha_{s^{\prime}}}\right)} \right)^{-1} f_{L_{0,s^{\prime}}}(x) \mathrm{d}x, \label{LT_I0}\end{aligned}$$ $$\begin{aligned}
&\mathcal{L}_{I_{1,k}}(\Gamma_k,\mathcal{E}_{k,0}) \nonumber \\
& = \prod_{s^{\prime} \in \{{\text{L}},{\text{N}}\}} \hspace{-0.3cm} \exp \Bigg (-\int_{\mathcal{E}_{k,0}}^{\infty} \left( 1-\left(1+\hat{a} P_1 H^2 x^{-\left(1+\frac{2}{\alpha_{s^{\prime}}}\right)} \right)^{-1} \right) \nonumber \\
& \times \Lambda_{1,s^{\prime}}^{\prime}([0,x)) \mathrm{d}x\Bigg). \label{LT_I1}\end{aligned}$$
*Proof:* See Appendix \[Proof of Theorem 1\]. $\square$
Note that since $0^{{\text{th}}}$ tier consists of only one UAV, i.e., the cluster center UAV, Laplace transform expression $\mathcal{L}_{I_{0,0}}(\Gamma_0,\mathcal{E}_{0,0})=1$. The total energy coverage probability of the network in Theorem \[theo:energycoverage\] is obtained by first computing the conditional energy coverage probability given that a UE is associated with a $k^{{\text{th}}}$ tier LOS/NLOS UAV, and then summing up the conditional probabilities weighted with their corresponding association probabilities. In order to formulate the conditional energy coverage probabilities, Laplace transforms of the interference terms are determined. We also note that although the energy coverage probability approximation in Theorem \[theo:energycoverage\] involves multiple integrals, we explicitly see the dependence of the energy coverage on, for instance, UAV heights, path loss distributions, path loss exponents, transmission power levels. Moreover, the integrals can be readily computed via numerical integration methods, providing us with additional insight on the impact of key system/network parameters, as demonstrated in the next section.
Extension to a Model with UAVs at Different Heights
===================================================
In the preceding analysis, we analyzed the energy coverage performance of a network in which UAVs are located at a height of $H$ above the ground, and $H$ is assumed to be the same for all UAVs. However, the proposed analytical framework can also be employed to analyze the coverage probability when UAV height is not fixed, i.e., UAVs are assumed to be located at different heights. In the extended model, we consider a more general network in which UAVs are located at different heights. Therefore, we assume that there are $M$ groups of UAVs such that the $\mu^{{\text{th}}}$ UAV group is located at the height level $H_{\mu}$ for $\mu=1,2,\ldots,M$ and UAVs at each height level can be considered as a UAV-tier distributed according to an independent homogeneous PPP with density of $\lambda_{U,\mu}$ and the total density is equal to $\sum_{\mu=1}^{M} \lambda_{U,\mu}=\lambda_U$. Different from the preceding analysis in which we have considered a single typical UE located at the origin and named its cluster center UAV as $0^{{\text{th}}}$ tier UAV, a separate typical UE for each UAV tier needs to be considered in the coverage probability analysis for this model with UAVs at different heights. For example, when we are analyzing the energy coverage probability of the network for a UE clustered around a $\mu^{{\text{th}}}$ tier UAV, we assume that the typical UE is located at the origin and its cluster center UAV is considered as the $0^{{\text{th}}}$ tier UAV similar to the previous model. Therefore, total energy coverage probability of the network given that the typical UE is clustered around a $\mu^{{\text{th}}}$ tier UAV for $\mu=1,2,\ldots,M$ can be computed as follows: $${\text{E}^{\text{C}}}_{\mu} \!\!=\!\!\sum_{k=0}^M \sum_{\substack{s \in \{{\text{L}}, {\text{N}}\}}} \left[{\text{E}^{\text{C}}}_{\mu,k,s} (\Gamma_k)\mathcal{A}_{\mu,k,s}\right], \label{EnergyCoverageProbability_multiheight}$$ where ${\text{E}^{\text{C}}}_{\mu,k,s}(\Gamma_k)$ is the conditional energy coverage probability given that the typical UE is clustered around a $\mu^{{\text{th}}}$ tier UAV and it is associated with a $k^{{\text{th}}}$ tier LOS/NLOS UAV, and $\mathcal{A}_{\mu,k,s}$ is the association probability with a $k^{{\text{th}}}$ tier LOS/NLOS UAV.
In a UAV network with practical HH antenna radiation patterns and clustered UEs, the total energy coverage probability of the network given that the typical UE is clustered around a $\mu^{{\text{th}}}$ tier UAV is approximately given at the top of the next page in (\[total\_energy\_coverage\_multiheight\])
$$\begin{aligned}
{\text{E}^{\text{C}}}_{\mu} & \approx \sum_{s \in \{{\text{L}},{\text{N}}\}} \sum_{n=0}^{\mathcal{N}}(-1)^{n} {\mathcal{N} \choose n} \nonumber \\
& \times \Bigg[ \int_{ H_{\mu}^{\alpha_s}}^{\infty}\left(1+\hat{a} P_0H_{\mu}^2 l_{0,s}^{-\left(1+\frac{2}{\alpha_s}\right)}\right)^{-1} \left(\prod_{j=1}^{M} \mathcal{L}_{I_{j,0}}\left(\Gamma_0,\mathcal{E}_{0,0}\right)\right) \prod_{m \in \{{\text{L}},{\text{N}}\}} \prod_{j=1}^{M} \bar{F}_{L_{j,m}}\left(\left(\frac{P_j}{P_0}l_{0,s}^{\frac{2}{\alpha_s}+1}\right)^{\frac{\alpha_m}{\alpha_m+2}}\right) f_{L_{0,s}}(l_{0,s}) \mathrm{d}l_{0,s} \nonumber \\
&+ \sum_{k=1}^{M} \int_{H_{\mu}^{\alpha_{s}}}^{\infty} \left(1+\hat{a} P_kH_k^2 l_{k,s}^{-\left(1+\frac{2}{\alpha_s}\right)}\right)^{-1} \left(\prod_{j=0}^{M} \mathcal{L}_{I_{j,k}}\left(\Gamma_k,\mathcal{E}_{k,0}\right)\right) \sum_{m \in \{{\text{L}},{\text{N}}\}} \hspace{-0.3cm} \bar{F}_{L_{0,m}}\left(\left(\frac{P_0}{P_k}l_{k,s}^{\frac{2}{\alpha_s}+1}\right)^{\frac{\alpha_m}{\alpha_m+2}}\right) \nonumber \\
& \times \prod_{m \in \{{\text{L}},{\text{N}}\}} \prod_{j=1,j \neq \mu}^{M} \bar{F}_{L_{j,m}}\left(\left(\frac{P_j}{P_k}l_{k,s}^{\frac{2}{\alpha_s}+1}\right)^{\frac{\alpha_m}{\alpha_m+2}}\right) \bar{F}_{L_{k,s^{\prime}}}(l_{k,s}) f_{L_{k,s}}(l_{k,s})\mathrm{d}l_{k,s} \Bigg] \label{total_energy_coverage_multiheight}\end{aligned}$$
*Proof:* Derivation of ${\text{E}^{\text{C}}}_{\mu}$ follows similar steps as that of ${\text{E}^{\text{C}}}$ in (\[total\_energy\_coverage\]). In particular, Laplace transforms $\mathcal{L}_{I_{0,k}}$ and $\mathcal{L}_{I_{j,k}}$ for $j=1,2,\ldots,M$ are computed using the Laplace transform equations given in (\[LT\_I0\]) and (\[LT\_I1\]), respectively, by updating UAV height as $H_j$ and UAV density as $\lambda_j$ for $j=0,1,\ldots,M$. $\bar{F}_{L_{0}}(x)$ and $f_{L_{0,s}}(x)$ are computed using (\[CCDF\_0\]) and (\[f\_L0s\]), respectively, by denoting the UAV height as $H_{\mu}$. Furthermore, $\bar{F}_{L_{k,s}}(x)$ and $f_{L_{k,s}}(x)$ are computed using (\[CCDF\_1\]) and (\[f\_L1s\]), respectively, by updating UAV height as $H_k$ and UAV density as $\lambda_k$ for $k=1,\ldots,M$.
Simulation and Numerical Results {#sec:Simulation and Numerical Results}
================================
In this section, we present the numerical evaluations of theoretical expressions in addition to the simulation results which are provided to confirm the accuracy of the proposed UAV network model and the analytical characterizations. In the numerical computations and simulations, unless stated otherwise, the parameter values listed in Table \[Table\] are used.
**Description** **Parameter** **Value**
-------------------------------------- ---------------------------------------------- ----------------------------
Path-loss exponents $\alpha_{{\text{L}}}$, $\alpha_{{\text{N}}}$ 2, 4
Environment dependent constants $b$, $c$ $11.95$, $0.136$
Height of UAVs $H$ $50$ m
Transmit power $P_k$ $\forall k$ 37 dBm
Energy outage threshold $\Gamma_k$ $\forall k$ 0 dBm
UAV density $\lambda_U$ $10^{-4}$ $(1/\text{m}^2)$
UE distribution’s standard deviation $\sigma_c$ 10
Rectifier efficiency $\xi$ 1
: System Parameters
\[Table\]
Impact of Cluster Size
----------------------
First, we address the effect of UE distribution’s standard deviation $\sigma_c$ on the association probability, average harvested power and the energy coverage probability using the LOS probability functions of high-altitude and low-altitude models of (\[LOS\_probability1\]) and (\[LOS\_probability2\]) in Figs. \[Fig\_AP1\], \[Fig\_AHP1\] and \[Fig\_ECP1\]. As the standard deviation increases, the UEs are spread more widely, resulting in generally larger distances between the cluster-center $0^{{\text{th}}}$ tier UAV and UEs. For example, for $\sigma_c =10$, the average link distance between cluster center UAV and UEs is around 12.5 meters, while it is around 115 meters for $\sigma_c =90$. Consequently, association probability with the $0^{{\text{th}}}$ tier UAV, $\mathcal{A}_{0}$, diminishes, while the association probability with $1^{{\text{st}}}$ tier UAVs, $\mathcal{A}_{1}$, increases for both models. Also, for a fixed height, LOS probability of cluster center UAV decreases for both models with the increasing cluster size, and hence association probabilities exhibit similar trends. Therefore, the average harvested power from the $0^{{\text{th}}}$ tier UAV, $P^{{\text{avg}}}_{0}$, increases while the average harvested power from the $1^{{\text{st}}}$ tier UAVs, $P^{{\text{avg}}}_{1}$, decreases as the cluster size grows in both models. On the other hand, the increase in $P^{{\text{avg}}}_{1}$ cannot compensate the decrease in $P^{{\text{avg}}}_{0}$, and therefore the total average harvested power $P^{{\text{avg}}}$ diminishes. In other words, smaller cluster size, i.e., more compactly distributed UEs results in a higher $P^{{\text{avg}}}$. The energy coverage probability in Fig. \[Fig\_ECP1\] exhibits a very similar behavior as the average harvested power in Fig. \[Fig\_AHP1\]. Also note that, the association probability results of the Corollary 3 closely approximate the association probabilities of the high-altitude model. Finally, we note that simulation results are also plotted in the figure with markers and there is generally an excellent agreement with the analytical results, further validating our analysis.
[[.5]{}]{} {width="100.00000%"}
[0.48]{} {width="100.00000%"}
[0.48]{} {width="100.00000%"}
Impact of UAV Height {#sec:Impact_UAV_height}
--------------------
Next, in Figs. \[Fig\_AP2\] and \[Fig\_ECP2\], we plot the association probability and energy coverage probability as a function of UAV height considering the LOS probability functions of both high-altitude and low-altitude models. For the high-altitude model, since LOS probability increases with the increasing UAV height, association probability with the $0^{{\text{th}}}$ tier UAV increases slightly. On the other hand, LOS probability decreases as a result of the increase in the 3D distance with the increasing UAV height in the low-altitude model. Therefore, more UEs prefer to connect to $1^{{\text{st}}}$ tier UAVs (i.e., UAVs other than the cluster-center one) at higher values of the UAV height. Also note that the result of the Corollary 3 very closely approximates the association probabilities of the high-altitude model especially as the UAV height increases.
Energy coverage probability of the cluster center UAV, ${\text{E}^{\text{C}}}_0$, exhibits similar trends for both types of LOS functions. More specifically, ${\text{E}^{\text{C}}}_0$ increases first then it starts decreasing with the increasing UAV height. Since the effective antenna gain for HH antenna orientation is an increasing function of UAV height for a fixed cluster size, an initial increase in ${\text{E}^{\text{C}}}_0$ is expected. However, further increase in UAV height results in a decrease in ${\text{E}^{\text{C}}}_0$ of both high-altitude and low-altitude models due to the increase in the distance. Therefore, for a fixed cluster size, there exists an optimal UAV height maximizing the network energy coverage, ${\text{E}^{\text{C}}}$, for both models. On the other hand, optimal height maximizing the ${\text{E}^{\text{C}}}$ in the low-altitude model is lower and ${\text{E}^{\text{C}}}$ decreases faster than that in the high-altitude model because the LOS probability function of the low-altitude model is a decreasing function of distance while the LOS probability function of the high-altitude model is an increasing function of the UAV height (e.g., as seen in Fig. \[Fig\_LOS\_func\]). Moreover, since UEs are more compactly distributed around the cluster center UAVs for $\sigma_c=10$, energy coverage probability of the $1^{{\text{st}}}$ tier UAVs, ${\text{E}^{\text{C}}}_1$, is relatively small and changes only very slightly for both models.
[[.5]{}]{} {width="100.00000%"}
[[.5]{}]{} {width="100.00000%"}
Impact of Antenna Orientation and UAV Density
---------------------------------------------
[0.47]{} {width="\textwidth"}
[0.47]{} {width="\textwidth"}
\
[0.5]{} {width="\textwidth"}
![Energy coverage probability as a function of UAV height $H$ for different values of UAV density $\lambda_U$ for different antenna orientations when $\sigma_c=10$. []{data-label="Fig_ECP5"}](ECP5_v3.eps){width="\figsize\textwidth"}
In Figs. \[Fig\_AP4\_1\], \[Fig\_AP4\_2\] and \[Fig\_AP4\_3\], we plot the association probability as a function of UAV height $H$ for different values of UAV density $\lambda_U$ for three different antenna orientations considering the high-altitude LOS probability model. Note that since the analysis for VV and HV antenna orientations seems to be intractable, only simulation results are plotted. Since effective antenna gain depends on the sine function of the angle between the UAVs and UEs for HH antenna orientation, UEs prefer to connect to their cluster center UAV, and hence $\mathcal{A}_0$ is much larger than $\mathcal{A}_1$ even when there is an increase in the number of UAVs (as seen when the UAV density is increased from $\lambda_U = 10^{-5}$ to $\lambda_U = 10^{-4}$) as shown in Fig. \[Fig\_AP4\_1\]. Also note that since both antenna gain and LOS probability is an increasing function with UAV height, increase in them can compensate the increasing path loss and the association probabilities remain almost constant.
For the VV antenna orientation, effective antenna gain depends on the cosine of the angle between the UAVs and UEs. For larger values of UAV density, association probability with the $0^{{\text{th}}}$ tier UAV, $\mathcal{A}_{0}$, slightly increases with increasing UAV height at first as a result of the increase in both the LOS probability and the effective antenna gain. Subsequently, $\mathcal{A}_{0}$ starts decreasing because the increase in the LOS probability cannot compensate the rapid decrease in the effective antenna gain between the UE and the cluster center UAV. For a less dense network, UEs associate with the cluster-center UAV mostly at lower UAV heights. However, with the increasing height, antenna gain with the cluster-center UAV decreases and consequently, the association probability with $1^{{\text{st}}}$ tier UAVs, $\mathcal{A}_1$, increases.
Finally, for the HV antenna orientation, effective antenna gain is a function of both cosine and sine of the angle $\theta$. Association probabilities exhibit similar trends as in the case of VV orientation. However, different from the VV case, since the antenna gain depends on both the cosine and sine functions, association probability with the $0^{{\text{th}}}$ tier UAV, $\mathcal{A}_{0}$, is greater than that of the VV orientation.
We also plot the energy coverage probability for different UAV heights, antenna orientations, and UAV densities in Fig. \[Fig\_ECP5\]. The performance with the HV and VV antenna orientations exhibit similar behaviors as that with the HH antenna orientation which is described in Section \[sec:Impact\_UAV\_height\]. For both higher-density (i.e., $\lambda_U = 10^{-4}$) and lower-density UAV networks (i.e., $\lambda_U = 10^{-5}$), HH orientation leads the best performance compared to the HV and VV cases. We also note that as the UAV density and UAV height are increased, energy coverage probability with the VV orientation starts exceeding that of the HV case. Therefore, energy coverage performance can be improved by varying the antenna orientations depending on the number of UAVs in the network and their height.
Furthermore, we display the average harvested power and the energy coverage probability as a function of the UAV density for three different antenna orientations considering the high-altitude LOS probability model in Figs. \[Fig\_AHP\_lambdaU\] and \[Fig\_IG\_afo\_lambdaU2\], respectively. Both the average harvested power and the energy coverage probability are increasing functions of the UAV density irrespective of the antenna orientation for a fixed UAV height. Expectedly, adding more UAVs to the network results in an increase in the total power received at the typical UE, and hence the average harvested power grows and the energy coverage performance of the network improves. We also note that HH antenna orientation generally leads to larger average harvested power and energy coverage probability. On the other hand, VV antenna orientation results in a higher average harvested power when the UAV density is sufficiently large due to the fact that one can harvest more power from the dense $1^{\text{st}}$-tier UAVs with smaller elevation angles when this antenna orientation is used.
![Average harvested power as a function of UAV density $\lambda_U$. []{data-label="Fig_AHP_lambdaU"}](AHP_lambdaU_dBm2.eps){width="\figsize\textwidth"}
![Energy coverage probability as a function of UAV density $\lambda_U$. []{data-label="Fig_IG_afo_lambdaU2"}](IG_afo_lambdaU2_v2.eps){width="\figsize\textwidth"}
Impact of UAV Transmit Power
----------------------------
In Fig. \[Fig\_ECP\_afo\_tpower\], we plot the energy coverage probability as a function of the UAV transmit power for three different antenna orientations considering the high-altitude LOS probability model. As expected, energy coverage probability is an increasing function of UAV transmit power. Moreover, HH antenna orientation generally results in a higher energy coverage probability than VV and HV antenna orientations.
![Energy coverage probability as a function of UAV transmit power. []{data-label="Fig_ECP_afo_tpower"}](ECP_transmitpower.eps){width="\figsize\textwidth"}
Impact of Energy Outage Threshold
---------------------------------
In Fig. \[Fig\_ECP\], we show the energy coverage probabilities of different tiers (i.e., ${\text{E}^{\text{C}}}_0$ and ${\text{E}^{\text{C}}}_1$) and the total energy coverage probability ${\text{E}^{\text{C}}}$ as a function of the energy outage threshold for both high-altitude and low-altitude models. As seen in Fig. \[Fig\_AP1\] and Fig. \[Fig\_AP2\], UEs are more likely to be associated with the $0^{{\text{th}}}$ tier UAV rather than $1^{{\text{st}}}$ tier UAVs in the high-altitude model when $\sigma_c=10$, and hence ${\text{E}^{\text{C}}}_0$ is much higher than ${\text{E}^{\text{C}}}_1$. On the other hand, for the low-altitude model, since association probabilities with each tier are not very different, more UEs can be covered by $1^{{\text{st}}}$ tier UAVs compared to the high-altitude model. However, ${\text{E}^{\text{C}}}_0$ is still greater than ${\text{E}^{\text{C}}}_1$ due to the relatively smaller distance to the cluster-center UAV. We also observe that as a general trend, energy coverage probabilities expectedly diminish with increasing energy outage threshold.
![Energy coverage probability as a function of energy outage threshold in dBm for LOS probability functions of high-altitude and low-altitude models when $\sigma_c=10$ and $H=50$ m. Simulation results are plotted with markers while dashed/solid curves show theoretical results. []{data-label="Fig_ECP"}](ECP_threshold_dBm.eps){width="\figsize\textwidth"}
Impact of Different UAV Heights
-------------------------------
![Energy coverage probability as a function of energy outage threshold in dBm for high-altitude LOS probability function when $\sigma_c=10$. Solid lines show the energy coverage probabilities when half of the UAVs are located at height $H_1 = 50$ m and the other half are located at height $H_2 = 80$ m, and the typical UE is clustered around a UAV at either height $H_1$ or $H_2$. Simulation results are also plotted with markers. []{data-label="Fig_ECP_multiheight"}](ECP_threshold_multiheight_dBm.eps){width="\figsize\textwidth"}
Finally, in Fig. \[Fig\_ECP\_multiheight\], we plot the total energy coverage probabilities as a function of the energy outage threshold using the high-altitude LOS probability function model when UAVs are assumed to be located at different heights. In this setup, we use the same parameters given in Table \[Table\] with some differences in UAV height and UAV density. More specifically, we consider $M = 2$ groups of UAVs located at altitudes $H_1 = 50$ m and $H_2 = 80$ m with densities $\lambda_{U,1} = \lambda_{U,2}=\lambda_{U}/2$ and transmit powers $P_1 = P_2 = 37$ dBm. Therefore, transmit power of the $0^{{\text{th}}}$ UAV is also equal to $P_0 = 37$ dBm. Similar to [@Liu], we consider practical values for UAV heights which are around $50\sim100$ meters.
In Fig. \[Fig\_ECP\_multiheight\], solid lines plot the total energy coverage probabilities when the height is the same for all UAVs. Dashed lines display the coverage probabilities when half of the UAVs are located at height $H_1$ and the other half are located at height $H_2$, and the typical UE is clustered around a UAV at either height $H_1$ or $H_2$. As seen in the figure, energy coverage performance of the different-height UAV network is very similar to that of the same-height UAV. For example, if a typical UE is clustered around a UAV at height $H_1 = 50$ m, the performance of the different-height UAV network is almost the same with that of the same-height UAV for $H=50$ m. On the other hand, performance of the different-height network when the typical UE is clustered around a UAV at height $H_1 = 80$ m is slightly lower than that of the same-height UAV network with $H=80$ m as a result of growing impact of interferences from the UAVs at height $H_2=50$ m.
Conclusion {#sec:Conclusion}
==========
In this paper, we have analyzed the energy coverage probability of a UAV network with clustered UEs. The UEs are distributed according to a PCP around PPP distributed cluster centers. The UAVs are deployed at a certain height above these cluster centers, which, as a result, represent the projections of UAVs on the ground. UEs are assumed to be associated with the UAV providing the strongest long-term averaged power. In this setting, we have determined the association probabilities and characterized the energy coverage probability. We have analyzed the effect of two different LOS probability functions on the network performance. We have also investigated the impact of practical 3D antenna radiation patterns on the network performance. Furthermore, we have presented energy coverage probability expressions for a more general model in which UAVs are located at different heights.
Via numerical results, we have demonstrated that the standard deviation of UE distribution $\sigma_c$, UAV height $H$, and antenna orientation have considerable influence on UAV association and energy coverage probabilities. For instance, more widely spread UEs result in a decrease in the total energy coverage probability of the network for both LOS probability models. For a certain cluster size, there exists an optimal UAV height that maximizes the network energy coverage. However, this optimal height depends on the type of the LOS probability model. In particular, low-altitude model features a lower optimal height than the high-altitude model. Although the maximum value of the energy coverage reached at the optimal height is almost the same for both models, low-altitude model has better energy coverage performance at lower heights. On the other hand, high-altitude model performs better at higher heights. These observations indicate that energy coverage performance is greatly affected by the environment. We have also shown that antenna orientation has a significant impact on the energy coverage probability depending on the UAV density. Specifically, better performance can be achieved by changing the antenna orientations according to the number of UAVs in the network and their height. We have also evaluated the average harvested power levels for different antenna orientations at various UAV density levels and computed the energy coverage probability as UAV density, UAV transmission power or energy outage threshold varies and we identified the impact of these different settings on the performance. Finally, we have addressed the performance of the different-height network and noted that UAVs at lower heights can lead to increased interference. Analyzing the performance of a UAV network with simultaneous information and energy transfer remains as future work.
Proof of Lemma 1 {#Proof of Lemma 1}
----------------
The CCDF of the path loss $L_{0,s}$ from a typical UE to a $0^{th}$ tier LOS/NLOS UAV is $$\begin{aligned}
&\bar{F}_{L_{0,s}}(x) \nonumber \\
&= \mathbb{E}_r \left[\mathbb{P}\left(L_{0,s}(r)\geq x \right) \mathcal{P}_s(r)\right] \nonumber \\
&= \mathbb{E}_d \left[\mathbb{P}\left( (d^2+H^2)^{\alpha_{s}/2}\geq x \right) \mathcal{P}_s(\sqrt{d^2+H^2}) \right]\label{Aeq1} \\
&= \int_{0}^{\infty} \mathbb{P}\left(d \geq \sqrt{x^{2/\alpha_{s}}-H^2} \right) \mathcal{P}_s(\sqrt{d^2+H^2}) f_D(d) \mathrm{d}d \nonumber \\
&= \int_{\sqrt{x^{2/\alpha_{s}}-H^2}}^{\infty} \mathcal{P}_s(\sqrt{d^2+H^2}) f_D(d) \mathrm{d}d\end{aligned}$$ $\text{for } s \in \{{\text{L}}, {\text{N}}\}$ where $f_{D}(d)$ is given in (\[PDF\_of\_d\]), and $\mathcal{P}_s(\cdot)$ is the LOS or NLOS probability depending on whether $s = {\text{L}}$ or $s = {\text{N}}$. With this, the CCDF of the path loss $L_{0}$ from a typical UE to a $0^{{\text{th}}}$ tier UAV given in Lemma 1 can be obtained by summing up over $s$.
Proof of Lemma 2 {#Proof of Lemma 2}
----------------
Intensity function for the path loss model from a typical UE to a $1^{{\text{st}}}$ tier UAV for $s \in \{{\text{L}},{\text{N}}\}$ can be formulated as follows: $$\begin{aligned}
\Lambda_{1,s}([0,x))&=\int_{\mathbb{R}^2} \mathbb{P}\left(L_1(r)<x\right)\mathrm{d}r \label{Aeq3} \\
&=2\pi\lambda_U \int_H^{\infty}\mathbb{P}\left(r^{\alpha_{s}}<x\right) \mathcal{P}_{s}(r) r \mathrm{d}r \label{Aeq31} \\
&= 2\pi\lambda_U \int_H^{\infty} \mathbb{P}\left(r< x^{1/\alpha_{s}}\right)\mathcal{P}_{s}(r) r \mathrm{d}r \\
&= 2\pi\lambda_U \int_{H}^{x^{1/\alpha_{s}}} \mathcal{P}_{s}(r) r \mathrm{d}r \label{app:Lemma1}\end{aligned}$$ where (\[Aeq3\]) is due to the definition of intensity function for the point process of the path loss. CCDF of the path loss $L_{1}$ from a typical UE to a $1^{{\text{st}}}$ tier UAV given in Lemma 2 can be obtained by summing up $\Lambda_{1,s}([0,x))$ over $s$.
Proof of Lemma 3 {#Proof of Lemma 3}
----------------
Association probability with a $0^{{\text{th}}}$ tier LOS/NLOS UAV can be expressed as $$\begin{aligned}
\mathcal{A}_{0,s} &=\prod_{m \in \{{\text{L}},{\text{N}}\}} \mathbb{P}(P_0 G_0(r) L_{0,s}^{-1} \geq P_1 G_1(r) L_{1,m}^{-1}) \label{Aeq7} \\
&= \prod_{m \in \{{\text{L}},{\text{N}}\}} \mathbb{P}\left(P_0 \frac{H^2}{L_{0,s}^{\frac{2}{\alpha_s}}} L_{0,s}^{-1} \geq P_1 \frac{H^2}{L_{1,m}^{\frac{2}{\alpha_L}}} L_{1,m}^{-1}\right) \label{Aeq71} \\
&= \prod_{m \in \{{\text{L}},{\text{N}}\}} \mathbb{P}\left( L_{1,m} \geq \left(\frac{P_1}{P_0}L_{0,s}^{\frac{2}{\alpha_s}+1}\right)^{\frac{\alpha_m}{\alpha_m+2}} \right) \nonumber \\
&=\int_{ H^{\alpha_{s}}}^{\infty} \prod_{m \in \{{\text{L}},{\text{N}}\}} \bar{F}_{L_m}\left(\left(\frac{P_1}{P_0}l_{0,s}^{\frac{2}{\alpha_s}+1}\right)^{\frac{\alpha_m}{\alpha_m+2}}\right) \nonumber \\
& \times f_{L_{0,s}}(l_{0,s}) \mathrm{d}l_{0,s}\label{app:Lemma3_0}\end{aligned}$$ where (\[Aeq7\]) uses the fact that LOS and NLOS links in the $1^{{\text{st}}}$ tier are independent, and (\[app:Lemma3\_0\]) incorporates the definition of the CCDF of the path loss. Since the distance between UEs and UAVs is at least $H$, the lower limit of the integration is $l_{0,s}= H^{\alpha_{s}}$.
Association probability with a $1^{{\text{st}}}$ tier LOS/NLOS UAV is given by $$\begin{aligned}
&\mathcal{A}_{1,s} \nonumber \\
&= \mathbb{P} (L_{1,s^{\prime}}>L_{1,s}) \hspace{-0.3cm} \prod_{m \in \{{\text{L}},{\text{N}}\}} \hspace{-0.3cm} \mathbb{P}(P_1 G_1(r) L_{1,s}^{-1} \geq P_0 G_0(r) L_{0,m}^{-1}) \label{Aeq4} \\
&= \mathbb{P} (L_{1,s^{\prime}}>L_{1,s}) \hspace{-0.3cm} \prod_{m \in \{{\text{L}},{\text{N}}\}} \hspace{-0.3cm} \mathbb{P}\left(P_1 \frac{H^2}{L_{1,s}^{\frac{2}{\alpha_s}}} L_{1,s}^{-1} \geq P_0 \frac{H^2}{L_{0,m}^{\frac{2}{\alpha_m}}} L_{0,m}^{-1}\right) \nonumber \\
&=\mathbb{P} (L_{1,s^{\prime}}>L_{1,s}) \hspace{-0.3cm} \prod_{m \in \{{\text{L}},{\text{N}}\}} \hspace{-0.3cm} \mathbb{P}\left( L_{0,m} \geq \left(\frac{P_0}{P_1}L_{1,s}^{\frac{2}{\alpha_s}+1}\right)^{\frac{\alpha_m}{\alpha_m+2}} \right) \nonumber \\
&=\int_{ H^{\alpha_{s}}}^{\infty} \bar{F}_{L_{1,s^{\prime}}}(l_{1,s}) \hspace{-0.3cm} \prod_{m \in \{{\text{L}},{\text{N}}\}} \hspace{-0.3cm} \bar{F}_{L_{0,m}}\left(\left(\frac{P_0}{P_1}l_{1,s}^{\frac{2}{\alpha_s}+1}\right)^{\frac{\alpha_m}{\alpha_m+2}}\right) \nonumber \\
& \times f_{L_{1,s}}(l_{1,s})\mathrm{d}l_{1,s}, \label{app:Lemma3_1}\end{aligned}$$ where $s, s' \in \{{\text{L}},{\text{N}}\}$, and $s \neq s'$. (\[Aeq4\]) makes use of the definition of the association probability and the fact that LOS and NLOS links in the $0^{{\text{th}}}$ tier are independent, and $\mathbb{P} (L_{1,s^{\prime}}>L_{1,s})=\bar{F}_{L_{1,s^{\prime}}}(l_{1,s})$.
Proof of Lemma 4 {#Proof of Lemma 4}
----------------
The total average harvested power at a typical UE can be formulated as $$\begin{aligned}
P^{{\text{avg}}}=\sum_{k=0}^{1} \sum_{s \in \{{\text{L}},{\text{N}}\}} \left [ P^{{\text{avg}}}_{k,s} \mathcal{A}_{k,s}\right], \label{app:Lemma4_3}\end{aligned}$$ where $P^{{\text{avg}}}_{k,s}$ is the conditional average harvested power given that the UE is associated with a LOS/NLOS UAV in tier $k\in\{0,1\}$, and $\mathcal{A}_{k,s}$ denotes the association probability. Conditional average harvested power can be obtained as follows: $$\begin{aligned}
P^{{\text{avg}}}_{k,s} & = \mathbb{E}_{S_{k,s},I_{tot}} \left[ S_{k,s} + \sum_{j=0}^{1} I_{j,k}\right] \nonumber \\
&= \mathbb{E}_{L_{k,s}}\Bigg[ P_k H^2 L_{k,s}^{-\left(1+\frac{2}{\alpha_s}\right)}\nonumber \\
&+ \sum_{s^{\prime} \in \{{\text{L}},{\text{N}}\}} \int_{\mathcal{E}_{k,0}}^{\infty} P_0 H^2 x^{-\left(1+\frac{2}{\alpha_{s^{\prime}}}\right)}f_{L_{0,s^{\prime}}}(x) \mathrm{d}x \nonumber \\
& + \sum_{s^{\prime} \in \{{\text{L}},{\text{N}}\}} \int_{\mathcal{E}_{k,0}}^{\infty} P_1 H^2 x^{-\left(1+\frac{2}{\alpha_{s^{\prime}}}\right)}\Lambda_{1,s^{\prime}}^{\prime}([0,x))\mathrm{d}x \Bigg] \label{app:Lemma4_0} \\
&=\mathbb{E}_{L_{k,s}}\left[ P_k H^2 L_{k,s}^{-\left(1+\frac{2}{\alpha_s}\right)}+ \sum_{j=0}^{1}\Psi_{I_{j,k}}(\mathcal{E}_{k,0}) \right] \label{app:Lemma4_1}
$$ where (\[app:Lemma4\_0\]) follows from the averaging over the fading distribution, inserting the antenna gain $G_k=H^2 L_{k,s}^{\frac{2}{\alpha_s}}$ for HH antenna orientation and employing Campbell’s theorem, (\[app:Lemma4\_1\]) follows from the definitions of $\Psi_{I_{0,k}}$ and $\Psi_{I_{1,k}}$ provided in (\[Psi0\]) and (\[Psi1\]), respectively. Note that the interfering $0^{{\text{th}}}$ tier UAV and $1^{{\text{st}}}$ tier UAVs lie outside the exclusion disc $\mathcal{E}_{k,0}$ with radius $\left( \frac{P_0}{P_k}l_{k,s}^{1+\frac{2}{\alpha_s}}\right)^{\frac{\alpha_{s^{\prime}}}{\alpha_{s^{\prime}}+2}}$ and $\mathcal{E}_{k,0}$ with radius $\left( \frac{P_1}{P_k}l_{k,s}^{1+\frac{2}{\alpha_s}}\right)^{\frac{\alpha_{s^{\prime}}}{\alpha_{s^{\prime}}+2}}$, respectively. Finally, by inserting (\[Association\_Prob0\]), (\[Association\_Prob1\]), (\[app:Lemma4\_1\]) into (\[app:Lemma4\_3\]), the average harvested power expression in (\[lemma4\]) can be obtained.
Proof of Theorem 1 {#Proof of Theorem 1}
------------------
Given that the UE is associated with a LOS/NLOS UAV in tier $k\in\{0,1\}$, the conditional energy coverage probability ${\text{E}^{\text{C}}}_{k,s}(\Gamma_k)$ is $$\begin{aligned}
&{\text{E}^{\text{C}}}_{k,s}(\Gamma_k) \nonumber \\
&=\mathbb{P}(\xi \left( S_{k,s} +I_{tot}\right) >\Gamma_k) \label{ATh11} \\
&\approx \sum_{n=0}^{\mathcal{N}}(-1)^n {\mathcal{N} \choose n} \mathbb{E}_{S_{k,s},I_{tot}}\left[ e^{-\hat{a}(S_{k,s}+I_{tot})}\right] \label{ATh12} \\
&=\sum_{n=0}^{\mathcal{N}}(-1)^n {\mathcal{N} \choose n} \mathbb{E}_{S_{k,s}}\left[ e^{-\hat{a}S_{k,s}} \mathbb{E}_{I_{tot}|S_{k,s}}\left[ e^{-\hat{a}I_{tot}} \right] \right] \nonumber \\
&=\sum_{n=0}^{\mathcal{N}}(-1)^n {\mathcal{N} \choose n} \mathbb{E}_{L_{k,s}} \Bigg[ \left(1+\hat{a}P_kG_kL_{k,s}^{-1}\right)^{-1} \nonumber \\
& \times \prod_{j=0}^{1}\mathbb{E}_{I_{j,k}|L_{k,s}}\left[ e^{-\hat{a}I_{j,k}}\right]\Bigg] \label{Ath13} \\
&=\sum_{n=0}^{\mathcal{N}}(-1)^n {\mathcal{N} \choose n} \mathbb{E}_{L_{k,s}} \Bigg[ \left(1+\hat{a}P_k H^2 L_{k,s}^{-\left(1+\frac{2}{\alpha_s} \right)}\right)^{-1} \nonumber \\
& \times \prod_{j=0}^{1}\mathcal{L}_{I_{j,k}}(\Gamma_k,\mathcal{E}_{k,0}) \Bigg] \label{app:Theorem1}\end{aligned}$$ where $\hat{a}=\frac{n\eta }{\Gamma_k /\xi}$, $\eta=\mathcal{N}(\mathcal{N}!)^{-\frac{1}{\mathcal{N}}}$, $\mathcal{N}$ is the number of terms in the approximation, $\mathcal{L}_{I_{j,k}}(\Gamma_k,\mathcal{E}_{k,0})$ is the Laplace transform of $I_{j,k}$, and (\[ATh12\]) is approximated by following the similar steps in [@Khan]. In (\[Ath13\]), we inserted the antenna gain $G_k=H^2 L_{k,s}^{\frac{2}{\alpha_s}}$, and the last step in (\[app:Theorem1\]) follows from $h_{k,0}$ $\sim$ $\exp(1)$ and by noting that Laplace transforms of interference at the UE from different tier UAVs are independent.
Laplace transforms can be determined by employing key characterizations from stochastic geometry. Recall that $0^{{\text{th}}}$ tier interference arises from the UAV at the cluster center of the typical UE. When the typical UE is associated with a UAV in the $1^{{\text{st}}}$ tier, Laplace transform of the interference from $0^{{\text{th}}}$ tier UAV can be formulated as $$\begin{aligned}
&\mathcal{L}_{I_{0,k}}(u) \nonumber \\
&\hspace{-0.25cm} = \mathbb{E}_{I_{0,k}}\left[e^{-\hat{a}I_{0,k}}\right]\nonumber\\
& \hspace{-0.25cm} = \hspace{-0.4cm} \sum_{s^{\prime} \in \{{\text{L}},{\text{N}}\}} \hspace{-0.3cm} \mathbb{E}_x\left[\mathbb{E}_{h_{0,0}}\left[\exp\left(-\hat{a} P_0G_0h_{0,0}x^{-1}\right)|P_0G_0x^{-1}<P_kG_kl_k^{-1} \right] \right] \label{Aeq9} \\
&\hspace{-0.25cm}=\hspace{-0.4cm} \sum_{s^{\prime} \in \{{\text{L}},{\text{N}}\}}\hspace{-0.3cm} \mathbb{E}_x\left[\left(1+\hat{a} P_0H^2x^{-\left(1+\frac{2}{\alpha_{s^{\prime}}}\right)}\right)^{-1}\middle |x>\left( \frac{P_0}{P_k}l_{k,s}^{1+\frac{2}{\alpha_s}}\right)^{\frac{\alpha_{s^{\prime}}}{\alpha_{s^{\prime}}+2}} \right] \label{Aeq10} \\
&\hspace{-0.25cm}=\hspace{-0.4cm}\sum_{s^{\prime} \in \{{\text{L}},{\text{N}}\}} \int_{\mathcal{E}_{k,0}}^{\infty} \left(1+\hat{a} P_0 H^2 x^{-\left(1+\frac{2}{\alpha_{s^{\prime}}}\right)} \right)^{-1} f_{L_{0,s^{\prime}}}(x) \mathrm{d}x\end{aligned}$$ where conditioning in (\[Aeq9\]) is a result of the UAV association policy, i.e., the received power from the interfering $0^{{\text{th}}}$ tier UAV is less than the received power from the associated UAV, (\[Aeq10\]) follows from $h_{0,0}$ $\sim$ $\exp(1)$ and inserting the antenna gains, in the last step the exclusion disc $\mathcal{E}_{k,0}=\left( \frac{P_0}{P_k}l_{k,s}^{1+\frac{2}{\alpha_s}}\right)^{\frac{\alpha_{s^{\prime}}}{\alpha_{s^{\prime}}+2}}$. Also note that $\mathcal{L}_{I_{0,k}}(u) = 1$, if the typical UE is associated with $0^{{\text{th}}}$ tier UAV.
Laplace transform of the interference from $1^{{\text{st}}}$ tier UAVs is $$\begin{aligned}
&\mathcal{L}_{I_{1,k}}(u)= \mathbb{E}_{I_{1,k}}\left[e^{-\hat{a}I_{1,k}}\right] \nonumber \\
&=\prod_{s^{\prime} \in \{{\text{L}},{\text{N}}\}} \exp\Bigg(-\int_{\mathcal{E}_{k,0}}^{\infty}\!\!\!\left(1-\mathbb{E}_{h_{1,i}} \left[ e^{-\hat{a} P_1 H^2 h_{1,i} x^{-\left(1+\frac{2}{\alpha_{s^{\prime}}}\right)} }\right]\right) \nonumber \\
& \times \Lambda_{1,s^{\prime}}^{\prime}([0,x))\mathrm{d}x \Bigg) \label{Aeq11} \\
&=\prod_{s^{\prime} \in \{{\text{L}},{\text{N}}\}} \hspace{-0.3cm} \exp \Bigg (-\int_{\mathcal{E}_{k,0}}^{\infty} \left( 1-\left(1+\hat{a} P_1 H^2 x^{-\left(1+\frac{2}{\alpha_{s^{\prime}}}\right)} \right)^{-1} \right) \nonumber \\
& \times \Lambda_{1,s^{\prime}}^{\prime}([0,x)) \mathrm{d}x\Bigg) \label{Aeq12}\end{aligned}$$ where (\[Aeq11\]) is obtained by computing the probability generating functional of the PPP, and (\[Aeq12\]) follows by computing the moment generating function of the exponentially distributed random variable $h$. Note that the interfering $1^{{\text{st}}}$ tier UAVs lie outside the exclusion disc $\mathcal{E}_{k,0}$ with radius $\left( \frac{P_1}{P_k}l_{k,s}^{1+\frac{2}{\alpha_s}}\right)^{\frac{\alpha_{s^{\prime}}}{\alpha_{s^{\prime}}+2}}$. Finally, by inserting (\[Association\_Prob0\]), (\[Association\_Prob1\]), (\[LT\_I0\]), (\[LT\_I1\]) into (\[CoverageProbability\]), energy coverage probability expression in (\[total\_energy\_coverage\]) can be obtained.
[99]{}
L. Zhang, H. Zhao, S. Hou, Z. Zhao, H. Xu, X. Wu, Q. Wu, and R. Zhang, “A survey on 5G millimeter wave communications for UAV-assisted wireless networks," *IEEE Access*, Early Access, Jul. 2019.
Y. Zeng, Q. Wu, and R. Zhang, “Accessing from the sky: A tutorial on UAV communications for 5G and beyond," *Proceedings of the IEEE*, vol. 107, no. 12, pp. 2327–2375, Dec. 2019.
N. H. Motlagh, T. Taleb, and O. Arouk, “Low-altitude unmanned aerial vehicles-based internet of things services: comprehensive survey and future perspectives," *IEEE Internet of Things J.*, vol. 3, no. 6, pp. 899–922, Dec. 2016.
A. Merwaday, A. Tuncer, A. Kumbhar, and I. Guvenc, “Improved throughput coverage in natural disasters: Unmanned aerial base stations for public-safety communications," *IEEE Veh. Techno. Mag.*, vol. 11, no. 4, pp. 53-60, Dec. 2016.
R. I. Bor-Yaliniz, and H. Yanikomeroglu, “The new frontier in RAN heterogeneity: Multi-tier drone-cells," *IEEE Commun. Mag.*, vol. 54, no. 11, pp. 48–55, Nov. 2016.
Y. Zeng, B. Clerckx, and R. Zhang, “Communications and signals design for wireless power transmission," *IEEE Trans. Commun.*, vol. 65, no. 5, pp. 2264–2290, May 2017.
S. Bi, C. K. Ho, and R. Zhang, “Wireless powered communication: Opportunities and challenges," *IEEE Commun. Mag.*, vol. 53, no. 4, pp. 117–125, Apr. 2015.
T. A. Khan, A. Alkhateeb, and R. W. Heath, “Millimeter wave energy harvesting," *IEEE Trans. Wireless Commun.*, vol. 15, no. 9, pp. 6048-6062, Sep. 2016.
A. Sakr and E. Hossain, “Cognitive and energy harvesting-based D2D communication in cellular networks: Stochastic geometry modeling and analysis," *IEEE Trans. Commun.*, vol. 63, pp. 1867-1880, May 2015.
K. Huang and V. Lau, “Enabling wireless power transfer in cellular networks: Architecture, modeling and deployment," *IEEE Trans. Wireless Commun.*, vol. 13, pp. 902-912, Feb. 2014.
M. Di Renzo, and W. Lu, “System-level analysis and optimization of cellular networks with simultaneous wireless information and power transfer: Stochastic geometry modeling," *IEEE Trans. Veh. Techn.*, vol. 66, no. 3, pp. 2251-2275, Mar. 2017.
J. Xu, Y. Zeng, and R. Zhang, “UAV-enabled wireless power transfer: Trajectory design and energy optimization," *IEEE Trans. Wireless Commun.*, vol. 17, no. 8, pp. 5092–5106, Aug. 2018.
Y. Wu, J. Xu, and L. Qiu, “UAV-enabled wireless power transfer with directional antenna: A two-user case (invited paper)," *Intern. Symp. on Wireless Commun. Syst. (ISWCS)*, pp. 1–6, Aug. 2018.
Y. Hu, X. Yuan, J. Xu, and A. Schmeink, “Optimal 1D trajectory design for UAV-enabled multiuser wireless power transfer," *IEEE Trans. Commun.*, vol. 67, no. 8, pp. 5674-5688, Aug. 2019.
L. Xie, J. Xu, and R. Zhang, “Throughput maximization for UAV-enabled wireless powered communication networks," *IEEE Internet of Things J.*, Early Access, Oct. 2018.
S. Cho, K. Lee, B. Kang, K. Koo, and I. Joe, “Weighted harvest-then-transmit: UAV-enabled wireless powered communication networks," *IEEE Access*, vol. 6, pp. 72212-72224, Nov. 2018.
C. Caillouet, T. Razafindralambo, and D. Zorbas, “Optimal placement of drones for fast sensor energy replenishment using wireless power transfer," *Wireless Days 2019*, Apr. 2019.
“ When COWs Fly: AT&T Sending LTE Signals from Drones," https://about.att.com/innovationblog/cows\_fly, Feb. 2017.
B. Galkin, J. Kibiłda, and L. A. DaSilva, “Coverage Analysis for Low-Altitude UAV Networks in Urban Environments," *Proc. of IEEE Globecom*, Dec. 2017.
V. V. Chetlur, and H. S. Dhillon, “Downlink coverage analysis for a finite 3-D wireless network of unmanned aerial vehicles," *IEEE Trans. Commun.*, vol. 65, no. 10, pp. 4543-4558, Oct. 2017.
E. Turgut and M. Cenk Gursoy, “Downlink analysis in unmanned aerial vehicle (UAV) assisted cellular networks with clustered users," *IEEE Access*, vol. 6, pp. 36313-36324, May 2018.
P. K. Sharma and D. I. Kim, “Coverage probability of 3D mobile UAV networks," *IEEE Wireless Commun. Letters*, vol. 8, no. 1, pp. 97-100, Feb. 2019.
L. Zhou, Z. Yang, S. Zhou, and W. Zhang, “Coverage probability analysis of UAV cellular networks in urban environments," *IEEE Inter. Conf. on Commun. Workshops (ICC Workshops)*, pp. 1–6, May 2018.
C. Liu, M. Ding, C. Ma, Q. Li, Z. Lin, and Y.-C. Liang, “Performance analysis for practical unmanned aerial vehicle networks with LoS/NLoS transmissions," *IEEE Inter. Conf. on Commun. Workshops (ICC Workshops)*, pp. 1–6, May 2018.
C. Zhang, and W. Zhang, “Spectrum sharing for drone networks," *IEEE J. Sel. Areas Commun.*, vol. 35, no. 1, pp. 136-144, Jan. 2017.
M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Unmanned aerial vehicle with underlaid device-to-device communications: Performance and tradeoffs," *IEEE Trans. Wireless Commun.*, vol. 15, no. 6, Jun. 2016.
J. Lyu, Y. Zeng, and R. Zhang, “UAV-aided cellular offloading for cellular hotspot," *IEEE Trans. Wireless Commun.*, vol. 17, no. 6, pp. 3988-4001, Jun. 2018.
R. I. Bor-Yaliniz, A. El-Keyi, and H. Yanikomeroglu, “Efficient 3-D placement of an aerial base station in next generation cellular networks," *Proc. of IEEE Int. Conf. Commun. (ICC)*, pp. 1-5, May 2016.
M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Efficient deployment of multiple unmanned aerial vehicles for optimal wireless coverage," *IEEE Wireless Commun. Letters*, vol. 20, no. 8, Aug. 2016.
M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Optimal transport theory for power-efficient deployment of unmanned aerial vehicles," *Proc. IEEE Int. Conf. Commun. (ICC)*, May 2016.
P. Chandhar, D. Danev, and E. G. Larsson, “Massive MIMO for communications with drone swarms," *IEEE Trans. Wireless Commun.*, vol. 17, no. 3, pp. 1604-1629, Mar. 2018.
G. Geraci, A. Garcia-Rodriguez, L. G. Giordano, D. Lopez-Perez, and E. Bjornson “Understanding UAV Cellular Communications: From Existing Networks to Massive MIMO," *IEEE Access*, vol. 6, pp. 67853-67865, Nov. 2018.
J. Lyu and R. Zhang, “Network-connected UAV: 3D system modeling and coverage performance analysis," *IEEE Internet of Things J.*, vol. 6, no. 4, pp. 7048-7060, Aug. 2019.
W. Khawaja, I. Guvenc, D. Matolak, U.-C. Fiebig, and N. Schneckenberger, “A survey of air-to -ground propagation channel modeling for unmanned aerial vehicles," *IEEE Commun. Surveys & Tutorials*, vol. 21, no. 3, pp. 2361-2391, May 2019.
E. Yanmaz, R. Kuschnig, and C. Bettstetter, “Channel measurements over 802.11a-based UAV-to-ground links," *Proc. IEEE GLOBECOM Workshops (GC Wkshps)*, pp. 1280–1284, Dec. 2011.
E. Yanmaz, R. Kuschnig, and C. Bettstetter, “Achieving air-ground communications in 802.11 networks with three-dimensional aerial mobility," *Proc. IEEE INFOCOM*, pp. 120–124, Apr. 2013.
J. Chen, D. Raye, W. Khawaja, P. Sinha, and I. Guvenc, “Impact of 3D UWB antenna radiation pattern on air-to-ground drone connectivity," *Proc. of the IEEE Vehicular Techno. Conf. (VTC)-Fall*, Aug. 2018.
A. Kumbhar, I. Guvenc, S. Singh, and A. Tuncer, “Exploiting LTE-advanced HetNets and FeICIC for UAV-assisted public safety communications," *IEEE Access*, vol. 6, pp. 783-796, Dec. 2018.
M. Haenggi, *Stochastic Geometry for Wireless Networks*. Cambridge, U.K.: Cambridge Univ. Press, 2012.
C. Saha, M. Afshang, and H. S. Dhillon, “Enriched K-Tier HetNet model to enable the analysis of user-centric small cell deployments," *IEEE Trans. Wireless Commun.*, vol. 16, no. 3, Mar. 2017.
A. Al-Hourani, S. Kandeepan, and S. Lardner, “Optimal LAP altitude for maximum coverage," *IEEE Wireless Commun. Letters*, vol. 3, no. 6, Dec. 2014.
3GPP, “Tr 36.828 (v11.0.0): Further enhancements to lte time division duplex (tdd) for downlink-uplink (dl-ul) interference management and traffic adaptation," 2016.
X. Wang, E. Turgut, and M. Cenk Gursoy, “Coverage in downlink heterogeneous mmWave cellular networks with user-centric small cell deployment," *IEEE Transactions on Vehicular Technology*, vol. 68, no. 4, pp. 3513-3533, Apr. 2019.
E. Turgut, M. Cenk Gursoy, and I. Guvenc, “Energy harvesting in unmanned aerial vehicle networks with 3D antenna radiation patterns," *Proc. of the IEEE Vehicular Technology Conference (VTC)-Fall*, pp. 1-5, Sep. 2019.
[^1]: E. Turgut and M. C. Gursoy are with the Department of Electrical Engineering and Computer Science, Syracuse University, Syracuse, NY, 13244. I. Guvenc is with the Department of Electrical and Computer Engineering, North Carolina State University, Raleigh, NC 27695. (e-mail: [email protected], [email protected], [email protected] ).
[^2]: This work has been supported in part by the National Science Foundation (CCF-1618615). The material in this paper was presented in part at the IEEE Vehicular Technology Conference (VTC)-Fall, Honolulu, HI, Sep. 2019. [@Turgut_VTC].
[^3]: For instance, LOS probability is given by (\[LOS\_probability1\]) and (\[LOS\_probability2\]) for the high-altitude and low-altitude models, respectively, and NLOS probability is $\mathcal{P}_{{\text{N}}}(\cdot) = 1-\mathcal{P}_{{\text{L}}}(\cdot)$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Dynamics of a particle is formulated from classical principles that are amended by the uncertainty principle. Two best known quantum effects: interference and tunneling are discussed from these principles. It is shown that identical to quantum results are obtained by solving only classical equations of motion. Within the context of interference Aharonov-Bohm effect is solved as a local action of magnetic force on the particle. On the example of tunneling it is demonstrated how uncertainty principle amends traditional classical mechanics: it allows the momentum of the particle to change without the force being the cause of it.'
author:
- |
Nenad Klipa\
Department of Physics, University of Zagreb, Bijenička\
32, Zagreb, Republic of Croatia
- |
S. Danko Bosanac\
Institut R. Boškovi' c, Zagreb, Republic of Croatia
title: Understanding quantum effects from classical principles
---
Introduction
============
The phrase *quantum effect* was introduced into the language of physics to signify everything that cannot be understood in terms of traditional classical mechanics. It is used to emphasize deep division between the two views: one of the classical and the other of the quantum world. From the failures to explain them from classical principles a very important conclusion was reached: quantum dynamics is fundamental and classical is derived in the limit $\hbar \rightarrow 0$. This view can be relaxed somewhat by accepting that the approximation $\hbar \sim 0$ is sufficient for classical dynamics to be a suitable alternative to quantum, and it is called semi-classical. In this approximation one combines classical and quantum principles, but certain caution should be exercised because deciding when it is applicable is not sometime clear. Quantum effects in this approximation are partially described, and this is the closest to what one can achieve with the classical principles.
Which are the quantum effects? In the first place this is interference and tunnelling, but there are many more: zero point energy, spin, discrete bound states, photo-electricity, ... to mention only few. As it was mentioned classical mechanics could not explain their physical origin from its principles, but the question is does the quantum mechanics? It does, but in saying so one thing should not be overlooked: in order to explain them the wave-particle dualism must be accepted without being able to rationally derive it from other principles. In fact the problem is not to comprehend either of the components: the concept of particle is well understood and so is the concept of the wave, but the union of the two. However, by accepting this logical system, the quantum mechanics, in which the wave-particle dualism is the postulate (and few others), one derives equation the solutions of which describe all the phenomena (not entirely, e.g. to describe the spontaneous emission one needs more elaborate theory). While there is no problem with the mathematical aspect of obtaining the correct result, there are problems with the physical understanding of them. However, the epithet “physical” must be defined. Within the logical system of quantum mechanics “physical” means that the essential features of the phenomena can be described by the properties of either the waves or the particles. In this respect the physics of interference is easily described by the wave-like nature of particles, but, for example, when it comes to the effect of tunneling there are problems. It cannot be explained from the particle-like nature of particles, but neither it can be from the wave-like nature. In other words, there is no explanation how the waves get through the potential barrier, except that the solution of the wave equation predicts it.
Are there any alternative formulations of quantum mechanics that is based on the wave-particle dualism? This question inevitable rises the following one: why there is necessity for alternative formulations? The answer to the second is that alternative formulations offer a different viewpoint of physics, and as such they are very important, and the answer to the first is that there are. One alternative formulation is based on strictly abstract approach, in which the essential postulate is that there is observable-operator connection [@neumann]. Apart from that one there are other alternative formulations of quantum mechanics. The path integral method is the best known [@feynman; @hibbs], in which instead of Schroedinger equation one postulates its integral equation form. The other is the Bohm‘s method of quantum potentials [@bohm], which in essence is not new formulation because classical trajectory equations are solved in the effective potential that is obtained from solution of Schroedinger equation. The third formulation is random classical mechanics [@nelson], which departs from the usual approach to quantum mechanics by introducing the concept of probability into classical mechanics, however Schroedinger equation is retained indirectly by postulating imaginary diffusion coefficient for the probability. Characteristic of all the alternative formulations is that Schroedinger equation is postulated, in one form or the other, while classical mechanics plays no important role.
Alternative formulations are also known in classical mechanics, for example Lagrange or Hamilton formulations [@goldstein]. They do not offer, in essence, anything different than the Newton equations of motion, but in many circumstances are more useful and emphasize different concepts in physics (e.g. the energy conservation law). However, one obvious formulation of classical mechanics had been entirely neglected. For centuries the basic ingredient of the scientific method was the concept of error, because no result of experiment is considered reliable if the error margin is not cited. Yet to the best of knowledge no reference work undertook to discuss the following question: given the error margins for initial conditions how this error propagates in time? This question is of utmost importance for theoretical predictions, because it can be shown (but not elaborated in details here) that the assumption of the precise initial conditions is academic in most circumstances. The meaning of this is that even the tiniest error, often in a relatively short time, increases to such an extent that no reliable predictions could be made. Therefore instead of asking whereabouts of the particle if its initial conditions are known, more appropriate question is to ask for probability of whereabouts of the particle if the probability of its initial conditions is known. This shift in emphases means that instead of treating dynamics of a point in the phase space one should treat dynamics of a phase space density. In other words, the emphases is shifted from the Newton equations of motion to the Liouville equation. Again, the two formulations are equivalent, although analysis of the phase space density provides additional insight into the dynamics of the system.
Importance of formulating classical dynamics in terms of the Liouville equation is that the uncertainty principle can be imposed on its solutions, and this condition can be treated as the additional postulate in classical mechanics [@bos3]. It says that if the standard deviations for the coordinate and the momentum are $\Delta x$ and $\Delta p$, respectively, then $$\Delta x\;\Delta p\succeq c \label{un}$$ where $c$ is a constant (determination of the constant, which has the value $%
c=\hbar /2,$ is not discussed, but it can be done in the same way as from the black-body radiation law, by fitting theoretical predictions to the experimental data). By implementing the postulate it remains to find solution of the Liouville equation with that property. It is anticipated that quantum mechanical results will be obtained, and if this is the case then this approach could be treated as alternative formulation of quantum mechanics. The most important difference with the previous formulations is that the starting point is classical dynamics, which is formulated with the Liouville equation but amended with the uncertainty principle. The other formulations start, in one form or the other, by postulating Schroedinger equation, however, if the suggested formulation is correct Schroedinger equation should be derived. Therefore, this approach does not replace quantum mechanics but derives it from different principles that do not incorporate the wave-particle dualism. It should be pointed out that the uncertainty principle in quantum mechanics is the law, i.e. it is derived from more basic principles, e.g. the wave-particle dualism, while in this formulation it acquires the status of postulate.
The problem of implementing the uncertainty postulate into classical mechanics is purely a mathematical task. The main problem is to find a suitable parametrization of the phase space density that ensures that the amendment is fulfilled at all times. The solution was demonstrated in several instances [@bos4; @bos5; @bos6; @bos7], but for the sake of completeness it will be described in the following section. Once this is done then the time evolution of the probability densities is obtained by solving the Liouville equation, but this essentially means solving Newton‘s equations of motion. This is the essence of what it will be called the *classical solution* for dynamics of particles. In short, suitable parametrization of the phase space density that ensures the uncertainty principle, plus the Newton’s equations of motion, is *classical dynamics*. In contrast the *traditional classical dynamics* is based on the concept of trajectory and without the uncertainty principle. *Quantum dynamics* implies starting from the same initial conditions as in *classical dynamics*, but solving Schroedinger equation instead of Newton’s. As it will be shown classical dynamics describes two very important quantum effects: interference and tunneling (in this context the problem of the zero point energy will also be discussed).
Formulation of classical dynamics {#sec:theory}
=================================
The suggestion in Introduction of incorporating the uncertainty principle into classical mechanics seems contradiction with the concept of trajectory, the concept that is an integral part of traditional classical mechanics. This is indeed correct but it is no longer that if instead of deterministic view one assumes the probabilistic one. The arguments for this change were mentioned in Introduction. The shift of emphases in classical mechanics from the concept of trajectory to the concept of probability means that formally one replaces the Newton equations of motion $$\frac{d\vec{p}}{dt}=\vec{F}\quad ;\quad \frac{d\vec{r}}{dt}=\frac{\vec{p}}{m}
\label{New}$$ with the Liouville equation for the probability density in the phase space $$\frac{\partial \rho }{\partial t}\,+\,\frac{\vec{p}}{m}\cdot \nabla _{r}\rho
\,+\,\vec{F}\cdot \nabla _{p}\rho \;=\;0 \label{liouv}$$ where $m$ is mass of particle and $\vec{F}$ is force acting on it. One step towards implementing the uncertainty principle (\[un\]) is to change the meaning of the function $\rho $. Instead of being treated as the probability density one should accept that it is a general, but real, function that satisfies the Liouville equation. The reason for this change is the observation that $\rho $ cannot be measured accurately because that implies accurate measurement of both the position and momentum of a particle, and this would violate the uncertainty principle. On the other hand, for the averages $$P(\vec{r},t)\;=\;\int d^{3}p\,\,\rho (\vec{r},\vec{p},t)\;\;\;\;;\;\;\;\;Q(%
\vec{p},t)\;=\;\int d^{3}r\,\;\rho (\vec{r},\vec{p},t) \label{pq}$$ this restriction is not applicable because, for example, for the probability $P(\vec{r},t)$ to be measured it is not necessary to know the momentum. Therefore the phase space density (not the probability density) is treated as an auxiliary function that satisfies the Liouville equation, and whose initial value is obtained from the quantities such as (\[pq\]), or from the probability current $$\vec{J}(\vec{r},t)\;=\;\frac{1}{m}\int d^{3}p\,\,\vec{p}\;\rho (\vec{r},\vec{%
p},t)\; \label{curr}$$ Therefore the quantities that have physical significance are the probabilities (\[pq\]) and the probability current (\[curr\]), and not the phase space density $\rho $, although the time evolution of the former is derived from the latter.
The uncertainty principle requires that $P(\vec{r},t)$ and $Q(\vec{p},t)$ are related by the inequality (\[un\]), which puts a constraint on the possible phase space densities, the solutions of Liouville equation. The problem is, therefore, to select the family of functions with that requirement, which can be readily solved if certain rules from the Fourier analysis (for a reference see [@korner]) are recalled. According to these rules the probability densities are written as $$P(\vec{r},t)\;=\;|f(\vec{r},t)|^{2}\quad ;\quad Q(\vec{p},t)\;=\;|g(\vec{p}%
,t)|^{2} \label{pqq}$$ and if the two functions are interrelated by $$f(\vec{r},t)\;=\;\frac{1}{\sqrt{\left( 2\pi \hbar \right) ^{3}}}\int dp\,e^{i%
\vec{p}\cdot \vec{r}/\hbar }\,g(\vec{p},t) \label{ft}$$ then the inequality (\[un\]) is ensured. The relationships (\[pqq\]) and (\[ft\]) are known from quantum mechanics, in fact they are integral part of it. It would appear therefore that in this way quantum mechanics is introduced through the ”back door”, however this is not correct. The mentioned relationships are a mathematical way of selecting those probabilities $P(\vec{r},t)$ and $Q(\vec{p},t)$ that obey the uncertainty principle, and before being used in quantum mechanics they were known in the Fourier analysis. The same relationships are also used in the signal theory, although not for the probabilites but for the intensities of pulses in the time and frequency domains.
The next step is to find how the phase space density is related to the amplitude $f$, because in this way the required constraint on the solutions of Liouville equation would be ensured. The relationship should be obtained by using the definitions (\[pq\]), which are familiar rules for convolutions in Fourier analysis, and from that observation one obtains $$\rho (\vec{r},\vec{p},t)\;=\;\frac{1}{\left( \pi \hbar \right) ^{3}}\,\int
dq\,e^{2i\vec{p}\cdot \vec{q}/\hbar }\,f^{*}(\vec{r}+\vec{q},t)\,f(\vec{r}-%
\vec{q},t) \label{w}$$ In order to prevent possible misunderstandings few comments about the function (\[w\])[@wig1; @wig2] are in order. It is known as the Wigner function, but it should not be considered here as the Wigner quasi-probability distribution i.e., the Weyl transform (up to constant) of a pure state density operator, because we did not introduced any quantum operator or state. Remember that $f$ is only an auxiliary function used for the parametrization of the classical probability for the coordinate of the particle. The Wigner function [@wig1] is one of the many quasi-probability distribution functions invented to express quantum mechanical averages in the classical (phase space) manner, which is not the subject under discussion in the present work.
The phase space density should be solution of the Liouville equation (\[liouv\]), and if the constraint is the parametrization (\[w\]) then one derives the equation for $f$. It is straightforward to show [@dos1] that for the polynomial potentials up to the second degree the equation that one derives for $f$ is $$i\hbar \frac{\partial f}{\partial t}\;=\;-\frac{\hbar ^{2}}{2m}\Delta
f\,+\,V\,f \label{schr}$$ in which Schroedinger equation is recognized. Therefore for the potentials of this kind the equation (\[schr\]) is classical solution for the problem of implementing the uncertainty principle. This means that the problems such as free particle, charged particle in the homogeneous (in general a time varying) electric or magnetic field, harmonic oscillator, etc. are exactly described by the amended classical theory. However, it can be shown that the same conclusion is valid in general [@bos8], and the proof is based on observation that any potential can be divided-up into the segments of constant value. In each segment the classical solution that is based on solving (\[schr\]) is in order, and by conveniently adjusting the boundary conditions between segments one derives again (\[schr\]), where now the potential $V$ is a general function. Consequently, free particle trajectories are used for the time evolution of the phase space density, a rather complicated procedure but in principle exact. However, if one is not interested in the solution in the phase space, only in the coordinate subspace, then it is sufficient to solve the equation (\[schr\]), which is a much simpler task. The price, which is paid for this benefit, is the loss of information that the phase space provides, which will be demonstrated on the following examples.
It would appear that the steps that were taken here are in the reverse order as it is done when deriving the Liouville equation from Schroedinger equation, but in what way they are different is discussed in the Conclusion of this paper.
Interference
============
The best-known quantum effect is interference that results from a particle having a choice to get through two slits. The setup is the following. In the y-z plane there is a screen with two circular slits that are centered on the y-axes at $\pm \;y_{0}$, having the width $\Delta $. The screen ends at $x=0$. The particle is sufficiently de-localized before the screen so that there is equal probability to enter either of the slits. In the slits, and just before exiting them, the probability of finding particle in the plane $x=0$ is a sum of the form $P(x,y,z)=P_{1}(y,z)g_{1}(x)+P_{2}(y,z)g_{2}(x)$, where the index designates the probability centered at a particular hole. The probability $P_{i}(y,z)$ is not zero in a circle of the radius $\Delta $, and for simplicity it will be assumed to have the form of a Gaussian with that width. The probability $g_{i}(x)$ is determined essentially by the length of de-localization of the incoming particle along the x-axes, and for simplicity it will also be assumed to be Gaussian of the width $\Delta $. The average velocity of the particle in the x direction is $v_{0}$. The fact that the particle enters each slit with the same probability implies that the functional forms for $P_{1}(y,z)g_{1}(x)$ and $P_{2}(y,z)g_{2}(x)$ are the same. The moment when the maximum of the probability $P(x,y,z)$ is at $%
x=0$ will be $t=0$. Propagation of this probability density from this moment on will be as for a free particle, and the impact that the screen has on the motion in $x>0$ will be neglected. Without considering the uncertainty principle the motion of the probability density can only be deduced if in addition to the probability $P(x,y,z)$ one also knows the velocity distribution of the particle. In the traditional classical mechanics this distribution is arbitrary, but with the uncertainty principle included it is no longer that. In fact the velocity distribution is not the main problem, it is the initial phase space density from which the initial conditions for classical trajectories are selected. The phase space density, if the uncertainty principle is included, is calculated from (\[w\]) where the function $f$ is defined by (\[pqq\]). It is obvious that the definition (\[pqq\]) does not determine this function uniquely, because $f$ is in general complex. The phase of $f$ is determined from the probability current (\[curr\]), and if (\[w\]) is replaced for the phase space density it can be easily verified that $$P\;\nabla Arg(f)=\vec{J}$$ By assuming that the probability current is known the phase is calculated as the line integral of the function $\vec{J}/P$. In the example with the two slits the current in the y-z plane is zero, while in the x direction it is given by $J=v_{0}g(x)$.
In this way the initial conditions are determined and the function $f$ is $$f(x,y,z)=\left[ \sqrt{P_{1}(y,z)}+\sqrt{P_{2}(y,z)}\right] \sqrt{g(x)}%
\;e^{iv_{0}x}$$ where from now on it will be assumed that $m=\hbar =1$ (the square root of a sum of two functions is equal to the sum of the square roots of these functions only if they do not overlap). For a particular example of the Gaussian probabilities the function $f$ is (non-essential factors are omitted for convenience) $$f(x,y,z)=\left[ e^{-\frac{1}{2\Delta ^{2}}(y-y_{0})^{2}}+e^{-\frac{1}{%
2\Delta ^{2}}(y+y_{0})^{2}}\right] e^{-\frac{1}{2\Delta ^{2}}%
(x^{2}+y^{2})+iv_{0}x}$$ from which the initial phase space density is $$\begin{aligned}
\rho _{0}(x,y,z,v_{x},v_{y},v_{z}) &=&\left[ e^{-\frac{1}{\Delta ^{2}}%
(y-y_{0})^{2}}+e^{-\frac{1}{\Delta ^{2}}(y+y_{0})^{2}}+2\cos
(2v_{y}y_{0})e^{-\frac{1}{\Delta ^{2}}y^{2}}\right] \label{ro2in} \\
&&e^{-\Delta ^{2}[(v_{x}-v_{0})^{2}+v_{y}^{2}+v_{z}^{2}]-\frac{1}{2\Delta
^{2}}(x^{2}+z^{2})} \nonumber\end{aligned}$$
At any time later, and if the particle is free, the phase space density is $$\rho (\vec{r},\vec{v},t)=\rho _{0}(\vec{r}-\vec{v}\;t,\vec{v}) \label{ro2t}$$ from which the probability $P(\vec{r},t)$ is $$\begin{aligned}
P(\vec{r},t) &=&\int d^{3}v\;\rho (\vec{r},\vec{v},t)= \\
&&\frac{1}{\left( \Delta ^{4}+t^{2}\right) ^{3/2}}\left[ \cosh \left( \frac{%
2\Delta ^{2}yy_{0}}{\Delta ^{4}+t^{2}}\right) +\cos \left( \frac{2tyy_{0}}{%
\Delta ^{4}+t^{2}}\right) \right] e^{-\frac{\Delta ^{2}\left[ \left(
x-tv_{0}\right) ^{2}+y^{2}+y_{0}^{2}+z^{2}\right] }{\Delta ^{4}+t^{2}}}\end{aligned}$$ The screen where the probability density is measured is at $x=X$ and as a function of the z-y coordinates its typical form is shown in Figure 1 (left pattern).
The parameters where chosen arbitrarily, for the demonstration purpose only, and their values are: $y_{0}=1000$, $\Delta =100$, and $v_{0}=1$. The screen is located at the distance $X=10^{5}$, and the time at which the probability is observed is $t=10^{5}$. Typical interference pattern is obtained, which is the same as if the particle is treated as a wave. However, in the treatment here the interference pattern is obtained by propagating the phase space density by classical trajectories, the equation (\[ro2t\]), and by the classical rules of probability addition. This is made possible by having additional term in the initial phase space density (the third term in (\[ro2in\])), besides those that correspond to the typical classical probability densities that are centered around the slits. The interference term in the phase space density, as the additional term can be called, has two distinctive features: one is that it has both positive and negative values, and the other that it is centered at a totally ”non-physical” place, in between the two slits. The first feature is essential if by the classical rules of addition of probabilities one can describe the oscillatory structure of the probability density on the screen at $x=X$. The negative values of the phase space density rule out the possibility to attach to it physical significance of the probability density. However, as it was mentioned, this ”non-physical” character of the phase space density is explained by impossibility to measure it in experiment.
While one could accept the possibility to work with the non-positive phase space densities, the location of the interference term in between the two silts rises at least two important questions. One is if it has physical significance, because its location would imply that it is only a ”mathematical trick” by which the correct result is obtained. The other question is why is it placed in a region where it does not overlap with the space where the particle is certainly located, around the slits?
The physicality of the interference term of the phase space density can be tested, by applying the force on the particle that is only localized in the region between the two slits. For example this can be homogenous magnetic field that is localized in a tube of the radius smaller than $y_{0}$, centered at $y=0$ and oriented in the z direction. In the traditional view the phase space density has zero value in this region and hence the force would not have any effect on the pattern on the screen at $x=X$. However, the phase space density that is in accordance with the uncertainty principle has a non-zero value around $y=0$ (the interference term), therefore all trajectories that originate there are affected by the magnetic field (it is tacitly assumed that the particle is charged). They will be affected for a time $t=T$ until they exit the magnetic tube, and for simplicity it will be assumed that this time is independent of the initial conditions of trajectory. The equation for these trajectories is (for simplicity the mass of particle is assumed to be unity) $$d_{t}^{2}\vec{r}=\vec{v}\times \vec{H}$$ and for the assumed magnetic field, i.e. $\vec{H}=h_{0}\hat{z}$, the solution is know exactly. After time $T$ the trajectory exits the magnetic field, and after that it goes as if no force in applied on the particle. It can be assumed that the magnetic field acts for a short time, meaning that $%
Th_{0}\ll 1$, in which case the trajectory is $$\vec{r}=\vec{r}_{0}+\vec{v}_{0}t+\vec{v}_{0}\times \vec{H}\;\left( t-\frac{1%
}{2}T\right)$$
These trajectories are used to propagate the phase space density, but only the interference part, because the trajectories that originate around the two slits are not affected by the magnetic force. The time evolution of the interference part is $$\rho ^{(int)}(\vec{r},\vec{v},t)=\rho _{0}^{(int)}\left[ \vec{r}-\vec{v}t-%
\vec{v}\times \vec{H}\;\left( t+\frac{1}{2}T\right) ,\vec{v}+\vec{v}\times
\vec{H}\;\right]$$ where $\rho _{0}^{(int)}$ is the third term in (\[ro2in\]). The probability $P(\vec{r},t)$ is now obtained by integrating the phase space density over the momentum (velocity) variables, which is finally $$\begin{aligned}
P(\vec{r},t) &=&\frac{1}{\left( \Delta ^{4}+t^{2}\right) ^{3/2}}\left[ \cosh
\left( \frac{2\Delta ^{2}yy_{0}}{\Delta ^{4}+t^{2}}\right) +\cos \left(
\frac{2ty+h_{0}T^{2}x}{\Delta ^{4}+t^{2}}y_{0}\right) e^{-\frac{\Delta
^{2}T^{2}h_{0}v_{0}y}{\Delta ^{4}+t^{2}}}\right] \\
&&e^{-\frac{\Delta ^{2}\left[ \left( x-tv_{0}\right)
^{2}+y^{2}+y_{0}^{2}+z^{2}\right] }{\Delta ^{4}+t^{2}}}\end{aligned}$$ The effect of the localized magnetic field is real, despite the fact that it only affects what appears to be ”non-physical” part of the phase space density. It is manifested essentially as the shift of the interference pattern, as it is shown in Figure 1 (right pattern), where the parameters are the same as before, but in addition $T=1000$ and $h_{0}=0.0002$. Therefore, if the effect is confirmed then one can indeed argue that the interference part of the phase space density is physically real.
The effect is known as the Aharonov-Bohm [@aharonov; @silverman], and its main point was to show that the concept of potential is more physical than the concept of force. This force-potential dilemma is historic, but it was always thought that the latter is just a convenient mathematical simplification of treating electromagnetic field. In quantum mechanics it is the vector potential that enters the Schroedinger equation explicitly, and not the magnetic force, and based on the Aharonov-Bohm effect it was argued that with the concept of field one needs a non-local theory to explain it. Namely, the magnetic field in this effect is confined in a localized region in space while the vector potential is spread all over it, including both slits. Therefore the latter has local effect on the charged particle, while the magnetic force does not have. However, by starting from the classical principles the force was reinstated as a legitimate concept, because it was shown that the effect can be explained as a local event on the interference term. It could be argued that the problem of non-locality with the concept of field is replaced by the problem of non-locality in the phase space density. After all, there is always a question (the second mentioned earlier) why in the phase space there is a contribution that does not have any direct relationship to where the particle indeed is, i.e. around one of the slits? In fact the question is not so much why is it there (it is there because of the uncertainty principle), but how it comes to be there?
In order to answer this question one needs to understand the meaning of the initial conditions for the phase space density, and how they are formed. For the two slit problem the initial probability was single centered, i.e. a wide distribution that overlaps both slits. If the walls of the screen are very thin, and for the particle they are infinitely high potential barrier, then almost instantaneously the initial single centered distribution splits into two, well separated ones. This means that before overlapping with the screen the phase space density is single centered around $y=0$ and $p_{y}=0$ (the other degrees of freedom are not essential for discussion), but after exiting the slits the phase space density is non-zero around three centers. For the two, centered around $y=\pm y_{0}$, one can easily give arguments why they are there. However, there exists the third, around $y=0$ (the interference term), which is a surprise because it is behind the infinite wall with respect to the original phase space density and it appears instantaneously. The choice is now either to accept this fact, but then one should accept the view that this classical approach is non-local theory, or to argue that formation of the interference term takes time. Intuitively the latter is the more acceptable view, but in order to prove it one would really need to work with the relativistic theory, in which correlation effects (and this is what one talks about in the formation of the interference term) cannot travel faster than the speed of light. In other words, sudden appearance of the interference term is nothing but an artifact of non-relativistic theory, where the signals can travel at arbitrary speed.
Solving the relativistic two slit experiment is quite demanding, but without the loss of generality one can treat simpler but analogous one dimensional problem. At $t=0$ a probability distribution is formed around $y=0$ in such a way that its momentum distribution contains two disjoined components, one with the average momentum $-m_{0}$ and the other with $m_{0}$. It is expected that the probability in the coordinate $y$ would split into two components traveling in the opposite directions (provided that $m_{0}$ is larger than the width of the momentum probability). After certain time one would have two disjoined probability densities, of the sort as in the two slits experiment, the only difference being that the two probabilities travel in the opposite directions. By applying a force on each probability distribution one can reduce their average momentum to zero, in which case this would be precisely the initial conditions for two slits (one neglects finer details that make this statement not entirely correct). This problem can be treated by relativistic mechanics, in which case one observes how the phase space density is formed.
The relativistic phase space density, from which Dirac equation is derived, in single dimension is [@bos9] $$\begin{aligned}
\rho (y,t,p,p_{0}) &=&\int dq\;dq_{0}\;\left[
f^{*}(y+q,t+q_{0})f(y-q,t-q_{0})+g^{*}(y+q,t+q_{0})g(y-q,t-q_{0})\right] \\
&&e^{2ipq-2ip_{0}q_{0}}\end{aligned}$$ where it was assumed that only positive energy components are present. At $%
t=0$ the phase space density is $$\begin{aligned}
\rho _{0}(y,p,p_{0}) &=&\int dq\;dq_{0\;}\left[
f_{0}^{*}(y+q,q_{0})f_{0}(y-q,-q_{0})+g_{0}^{*}(y+q,q_{0})g_{0}(y-q,-q_{0})%
\right] \\
&&e^{2ipq-2ip_{0}q_{0}}\end{aligned}$$ where the functions $f_{0}$ and $g_{0}$ are defined in the momentum space as $$f_{0}(y,q_{0})=\int
dk\;A(k)e^{iky-iq_{0}e(k)}\;\;\;;\;\;\;g_{0}(y,q_{0})=\int
dk\;w(k)A(k)e^{iky-iq_{0}e(k)}$$ $A(k)$ is the momentum space amplitude for the initial conditions (the units are $m=c=\hbar =1$), $w(k)=k/[1+e(k)]$ and $e(k)=\sqrt{1+k^{2}}$ . By evaluating the integrals in the variables $q$ and $q_{0}$ the initial phase space density is (non-essential factors are omitted) $$\begin{aligned}
\rho _{0}(y,p,p_{0}) &=&\int dk\;\delta \left[ 2p_{0}-e(p-k)-e(p+k)\right] %
\left[ 1+w(p-k)w(p+k)\right] \\
&&A^{*}(p-k)A(p+k)\;e^{2iky}\end{aligned}$$ and at any later time the phase space density is $$\begin{aligned}
\rho (y,t,p,p_{0}) &=&\rho _{0}(y-\frac{p}{p_{0}}t,p,p_{0}) \\
&=&\int dk\;\delta \left[ 2p_{0}-e(p-k)-e(p+k)\right] \left[ 1+w(p-k)w(p+k)%
\right] \\
&&A^{*}(p-k)A(p+k)\;e^{2ik(y-\frac{p}{p_{0}}t)}\end{aligned}$$ The phase space density in only the coordinate-momentum variables is obtained by integrating in the variable $p_{0}$ (the fourth component of the four-momentum), in which case $$\rho (y,p,t)=\int dk\;\left[ 1+w(p-k)w(p+k)\right] A^{*}(p-k)A(p+k)%
\;e^{2iky-ie(p+k)t+ie(p-k)t} \label{relro}$$
For the momentum amplitude it is now assumed to have the form $$A(p)=e^{-\frac{(p-m_{0})^{2}}{2\Delta ^{2}}}+e^{-\frac{(p+m_{0})^{2}}{%
2\Delta ^{2}}}$$ where $m_{0}>>\Delta $ (the two distribution do not overlap). When in the phase space (\[relro\]) the product of the amplitudes is evaluated one gets three terms that are centered around $y=0$: two terms that are centered around $p=\pm m_{0}$ and one around $p=0$. Between all three contributions there is no overlap (or it is negligible). The first two contribute to the phase space density that is centered around $y=\pm \frac{m_{0}}{e(m_{0})}t$ and $p=\pm m_{0}$, which is analogous to the phase space density centered around $y=\pm y_{0}$ in the two slit setup. These two contributions recede, each traveling at the speed $m_{0}$ (in the non-relativistic limit when $%
m_{0}\ll 1$) or at nearly the speed of light (in the relativistic limit when $m_{0}\gg 1$). They are of no interest for what is the intention to show. The third term, which is analogous to the interference term, should be analyzed in details. In the phase space its contribution is $$\begin{aligned}
\rho _{int}(y,p,t) &=&e^{-\frac{p^{2}}{\Delta ^{2}}}\int dk\;\left[
1+w(p-k)w(p+k)\right] \left[ e^{-\frac{(k-m_{0})^{2}}{\Delta ^{2}}}+e^{-%
\frac{(k+m_{0})^{2}}{\Delta ^{2}}}\right] \\
&&\;e^{2iky-ie(p+k)t+ie(p-k)t}\end{aligned}$$
If $m_{0}$ is small (and so is $\Delta $ by assumption) then the function $%
w(k)$ is small and can be neglected. Also one can write $e(k)\approx 1+\frac{%
1}{2}k^{2}$, in which case the interference term in the phase space density is $$\rho _{int}(y,p,t)=e^{-\frac{p^{2}}{\Delta ^{2}}-\Delta ^{2}(x-pt)^{2}}\cos %
\left[ 2m_{0}(x-pt)\right]$$ Its typical form is shown in Figure 2 for the parameters $m_{0}=0.05$ and $%
\Delta =0.01$, and for two time instants: $t=0$ (left figure) and $t=5000$ (right figure). The main feature of the interference term is that its modulus is independent of time and $m_{0}$. Furthermore it changes its shape in unison with the rate at which the two main peaks in the phase space density separate from each other.
On the other hand if $m_{0}$ is large (but the width $\Delta $ is again small), and because most of contribution to the interference term comes from $k=\pm m_{0}$, one can write $e(k)\approx |k|+\frac{1}{2|k|}$ and $%
1+w(p-k)w(p+k)\sim 2m_{0}^{-1}$. This means that the interference term diminishes in the limit $m_{0}\rightarrow \infty $, and it is approximately $$\rho _{int}(y,p,t)\sim \frac{1}{m_{0}}e^{-\frac{p^{2}}{\Delta ^{2}}-\Delta
^{2}x^{2}}\cos \left[ 2m_{0}\left( x-\frac{p}{m_{0}}t\right) \right]$$ which is valid in the time interval $t\ll m_{0}^{4}\Delta ^{-3}$. For longer times it diminishes as $t^{-1}$. Besides being small contribution in the overall phase space density the interference term has additional feature that indicates its dependence on the time it takes the correlation to have effect on it (the two receding peaks travel at nearly the speed of light). The exponential term is ”frozen” meaning that it is time independent, and the oscillatory term is changing but with a great time lag, in fact it is nearly constant also. If the two receding peaks are stopped by, say, a potential step, then the situation would be similar to the two slit problem. However, at that instant the interference part of the phase space density would be very small in amplitude and not having the adequate shape. By stopping the two peaks the phase space density would redistribute itself in order to match this situation but it is obvious that process would take some time. This indicates that the interference term in the two slits setup does not form itself instantaneously but in reality it takes some time, and therefore its source is in a physical process.
Zero point energy
=================
Among the quantum effects is the so-called *zero point energy*, which is the lowest possible of all stationary states of a particle in a potential (ground state energy). There are several reasons why it is in this group of effects, but the one with the greatest weight comes from the interpretation of the energy of the stationary states. According to the standard approach to the quantum-classical relationship the correspondence between the quantum and classical stationary states is only possible in the limit $\hbar
\rightarrow 0$, or equivalently for large quantum numbers (Bohr‘s principle of correspondence). As a consequence, it can be shown relatively easily from the WKB approximation that the energy of stationary states have the following interpretation. If position of a particle is random then the modulus of its momentum is not, it is determined from the energy conservation law. In other words, if $E_{0}$ is the energy of a stationary state then the phase space density should parametrize, in the classical limit, as $$\rho (\vec{r},\vec{p})\;=\;\delta \left( \frac{p^{2}}{2m}+V-E_{0}\right) =%
\frac{1}{\sqrt{2m(E_{0}-V)}}\delta (p-p^{\prime }\;) \label{ro1}$$ Indeed this limit is approached, on average, for the stationary states with large quantum numbers. This means that for the ground state there is not even approximate agreement between this classical limit and the quantum ground state, as shown in Figure 3. Classical probability curve (dotted line) has singularity at the points where momentum of the particle is zero. Large portion of the quantum probability, however, is outside these classical bounds, and the standard interpretation is that this is due to tunneling, and therefore classical interpretation is not possible.
If the analysis starts from the Liouville equation (\[liouv\]) then the stationary solutions are obtained by requiring that $\partial _{t}\rho =0$, which has a general solution in the form $\rho =F\left( \frac{p^{2}}{2m}%
+V\right) $, where $F$ is any function that has finite norm. Therefore there is infinite number of stationary states, but one particular is $$\rho (x,p)=e^{-a\left( \frac{p^{2}}{2m}+\frac{1}{2}m\omega ^{2}x^{2}\right) }
\label{ro2}$$ for, say, a one dimensional harmonic oscillator. The constant $a$ is arbitrary and can be fixed by requiring that the average energy of the oscillator is equal to the energy of the ground state of the quantum oscillator, i.e. $$<E>=\int dp\;\frac{p^{2}}{2m}\;Q(p)\;+\;\int dx\;\frac{1}{2}m\omega
^{2}x^{2}\;P(x)=E_{0}$$ The result for the probability density $P(x)$ is identical curve as in Figure 3 for the quantum solution.
The major departure from the standard classical analysis is in the interpretation of what the energy of the stationary states represents. The difference is summarized in the two expressions for the phase space densities, (\[ro1\]) and (\[ro2\]). Thus according to the solution that is based on the Liouville equation, both the coordinate and momentum of particle are chosen randomly, according to the prescribed phase space distribution. For each pair of these points in the phase space the appropriate energy of the particle does not coincide with the energy of the ground state, but on average it is equal to it. Therefore, the points in Figure 3 that appear to be in the classical forbidden region, and hence being interpreted as tunneling, are in fact manifestation of entirely classical effect. The points that are considering tunneling express probability of finding particle with the energy that is larger than $E_{0}$, which according to (\[ro2\]) is not zero. In the phase space this is manifestly clear but from the solution of the Schroedinger equation (\[schr\]) it is not, because one works with the function that is the average over all momentum part of the phase space. This in essence is the meaning of the comment at the end of the previous section.
Previous analysis was entirely classical, and the identity with the quantum result is accidental because the choice of the function $F$ was arbitrary, however, the analysis explains the physical content of the zero point energy. The only contribution of the uncertainty principle is to select from the functions $F$ only those that satisfy it, and as it turns out there is only one.
Tunneling {#sec:tun}
=========
The effect of tunneling is one of the most intriguing quantum effects [@tunn], but its proper understanding requires careful analysis. As it has already been indicated what is considered to be the tunneling effect in the case of the zero point energy it is in fact a classical effect. It is manifestation of the phase space density when it is averaged over the momentum subspace. However, that is not a typical tunneling effect, more appropriately it is described in a scattering of particle on a potential barrier. Discussion of this example starts by considering a much simpler one, which is scattering of a particle from an infinitely high potential barrier.
In this example the particle moves in the space on the negative x axes, and the barrier is positioned at $x=0.$ If initial probability density $P_{0}(x)$ is given, and the initial probability current $j_{0}(x)$, then the problem to find time evolution of the phase space density $\rho (x,p,t)$, and from that the probability density $P(x,t)$, is well defined. There are several reasons why the problem cannot be solved in the same way as for the interference on two slits, which was described in the previous section. One is that the potential is not of the parabolic type, which is a necessary condition that the equation (\[schr\]) is derived from the parametrization (\[w\]). Derivation of the equation (\[schr\]) does not appear a necessary condition for solving time evolution of the phase space density because this is done by propagating trajectories in a potential. The weakness of this argument is that by simple solution of this kind the uncertainty principle may not be satisfied at arbitrary times, although at the initial instant it is by the parametrization (\[w\]). Therefore to ensure that this is the case the function $f$ in the initial (\[w\]) should incorporate all the information about the potential, including the possible boundary conditions, and this is achieved by solving the equation (\[schr\]). This argument appears non-physical because there is no reason why at the initial instant the particle, if it is well localized away from the barrier, should ”know” of its existence. The same argument was applied for the existence of the interference term in the two slits experiment, and the answer was that this is an artifact of the non-relativistic theory, where infinite velocities are possible. At the initial instant one can ensure that the particle does not have ”contact” with the barrier, but due to the infinite dispersion of momenta at any short time after the particle will be everywhere in the space, and therefore will know of the barrier. It can be shown, but not elaborated here, that if the relativistic theory is used, with the restriction that the probability density is zero outside certain boundary, then indeed the particle does not know of the existence of the barrier until this boundary reaches it. Because of that the boundary propagates as in the traditional classical mechanics, i.e. no quantum effects are observed. Therefore the initial (\[w\]) should be determined by using time independent solutions of the equation (\[schr\]) as the basis in which $P_{0}(x)$ is represented. However, there is a problem, as mentioned earlier, the equation (\[schr\]) cannot be derived for the potentials other than harmonic. There is solution to this difficulty for potentials of the step-like character, and the infinite barrier is of this kind, which is to replace potential with the boundary condition. In other words, one can neglect the potential barrier and treat the particle as being free on the whole x axes. The barrier is mimicked by imposing the boundary condition on the probability and the probability current at $x=0$ by demanding that at each instant they should be equal to zero. The net effect is the same, although the physics of the problem is not. The difference is that physics demands that for $x>0$ the probability $P(x,t)$ is zero, while from the imposed boundary condition this is not necessarily the case. In other words, the problem is solved mathematically formally, but the physical content should be extracted.
The initial function $f$, from which the initial phase space density is calculated, is therefore obtained from a linear combination (again the units are used in which $m=\hbar =1)$$$f(x)=\frac{1}{2i}\int dk\;A(k)\;\left( e^{ikx}-e^{-ikx}\right)$$ where the amplitude $A(k)$ should be determined from $P_{0}(x)$ and $%
j_{0}(x) $. The plane waves $e^{\pm ikx}$ are solutions of the stationary equation (\[schr\]) and their combination ensures that the proper boundary condition at $x=0$ is satisfied. For simplicity the initial conditions are $$P_{0}(x)=e^{-\frac{(x-x_{0})^{2}}{\Delta ^{2}}}\;\;\;;\;\;\;j_{0}(x)=p_{0}%
\;P_{0}(x) \label{initp0}$$ where $x_{0}$ is negative and chosen so that $|x_{0}|\gg \Delta $. It can be shown that (constant pre-factors are omitted for convenience) $$A(k)=e^{-(k-p_{0})^{2}\Delta ^{2}/2-ikx_{0}} \label{ampla}$$ in which case the initial phase space density is $$\begin{aligned}
\rho _{0}(x,p) &=&\int_{-\infty }^{\infty }dq\;e^{2ipq}\;f^{*}(x+q)f(x-q)
\label{robarr} \\
&=&e^{-\frac{(x-x_{0})^{2}}{\Delta ^{2}}-(p-p_{0})^{2}\Delta ^{2}}+e^{-\frac{%
(x+x_{0})^{2}}{\Delta ^{2}}-(p+p_{0})^{2}\Delta ^{2}}-2\cos \left[
2(p_{0}x-px_{0})\right] e^{-\frac{x^{2}}{\Delta ^{2}}-p^{2}\Delta ^{2}}
\nonumber\end{aligned}$$ which consists of three terms: one centered around $x_{0}$ and $p_{0}$, the other around $-x_{0}$ and $-p_{0}$, and the third around $x=p=0$ which represents the interference term. The structure of the phase space density is very similar to the one in the problem with two slits. However, the interest here is not to discuss the effect of interference but to note one important property of the momentum distribution, which has important repercussions for understanding the tunneling effect.
At any later time the phase space density is given by (\[ro2t\]), and the probability for the coordinates is obtained from $$P(x,t)=\int_{-\infty }^{\infty }dp\;\rho _{0}(x-pt,p)$$ which is not zero for $x>0$, but for $x<0$ it coincides with the true solution of the problem. The physical solution is therefore obtained by disregarding the probability for $x>0$. However, the momentum distribution is obtained by integrating the physical solution for $f$, which is zero for $%
x>0$, but the phase space density (\[robarr\]) was derived under the assumption that this function extends over the whole x axes. Therefore, by formally calculating the integral $$Q(p,t)=\int_{-\infty }^{0}dx\;\rho _{0}(x-pt,p)$$ to obtain the momentum distribution is not a legitimate procedure. One needs to extract $f(x,t)$ from the probability $P(x,t)$ and the probability current $j(x,t)$, and then obtain $Q(p,t)$ from (\[ft\]) and (\[pqq\]). This is the price one pays by replacing the potential barrier with the boundary condition. For the case under discussion the function $f(x,t)$ is relatively easily extracted, and the result is $$f(x,t)=\frac{1}{\sqrt{\Delta ^{2}+it}}\left[ e^{\frac{-\Delta
^{2}(x-x_{0}-p_{0}t)^{2}+2i(\Delta ^{4}p_{0}-x_{0}t)x}{2(\Delta ^{4}+t^{2})}%
}-e^{\frac{-\Delta ^{2}(x+x_{0}+p_{0}t)^{2}-2i(\Delta ^{4}p_{0}-x_{0}t)x}{%
2(\Delta ^{4}+t^{2})}}\right]$$ The momentum function is $$g(p,t)=\int_{-\infty }^{0}dx\;f(x,t)e^{-ipx}$$ which is not given explicitly because it is a rather lengthy expression, instead a typical probability $Q(p,t)$ (solid line) is shown in Figure 4 for four typical times: initial instant (a), just before the maximum of the Gaussian probability reaches the barrier ($t_{0}=|x_{0}|/p_{0}$) (b), just after this instant (c) and long after that (d). Its shape is not what one would expect from the intuitive reasoning, and which is based on the fundamental law of classical mechanics that any change in momentum is caused by force.
In this particular example the force is of a special kind, it only changes the sign of the momentum, and therefore the width of $Q(p,t)$ should not change in time because the modulus of the momentum is constant. The expected probability is the following: in the beginning particle moves towards the barrier and its momentum distribution is centered around $p_{0}$, and after long time the particle moves away from the barrier and its momentum distribution is centered around $-p_{0}$. These two probabilities should be the mirror images of each other. At any other time, in particular around $%
t=t_{0}$, the function $Q(p,t)$ is a combination of the two extreme cases. This is indeed the case if $Q(p,t)$ is calculated from the phase space distribution in which the interference term in (\[robarr\]) is neglected (from the traditional classical mechanics). The resulting probability is shown in Figure 4 by the dotted line, which deviates considerably from the probability when the uncertainty principle is implemented. The essence of the difference is in the change of the modulus of the momentum, which in traditional classical explanation it is attributed to the action of a force. However, there is no such force only the infinite barrier. In fact the momentum distribution changes dramatically between the initial instant and $%
t\sim t_{0}$, which is shown by calculating $g(p,t)$ for large $p$. In this limit (the non-essential factors are omitted) $$g(p,t)\sim \frac{1}{p^{2}}e^{\frac{-\Delta ^{2}(x_{0}+p_{0}t)^{2}}{2(\Delta
^{4}+t^{2})}}$$ which means that the modulus of the function changes from $g(p,t)=\exp
(-p^{2}\Delta ^{2}/2)$ to $g(p,t)\sim p^{-2}$ between these two instants. The widening effect of the momentum distribution has no source in the dynamics, because it is entirely the consequence of the change in the width of the probability in the coordinate. In other words, as the width of the probability in the coordinate of the particle changes the distribution of the momenta also changes, meaning that no classical dynamics can explain this effect because it is not caused by the action of a force.
The widening effect of the momentum distribution can be tested by assuming that the barrier is not infinitely high, say it has the value $V_{0}$. One consequence of the finite height is that all the phase space density for which $p>\sqrt{2V_{0}}$ would ”leak” into the half space $x>0$ and will manifest itself as the non-zero probability $P(x,t)$. However, for a very high barrier the estimate of this probability, under the assumption that there is no effect due to the widening of the momentum distribution, gives $$P(x,t)\sim e^{-2V_{0}\Delta ^{2}}$$ which is negligible small. Therefore, the prediction is, which is based on considering only the initial distribution of momenta, that the probability for particle to get over the barrier is negligible. On the other hand, if widening is taken into account then the estimate of the probability $P(x,t)$ is obtained by first calculating the function $f(x,t)$ from $$f(x,t)=\int dp\;e^{ipx}\;g(p,t)\sim e^{\frac{-\Delta ^{2}(x_{0}+p_{0}t)^{2}}{%
2(\Delta ^{4}+t^{2})}}\int dk\;\;\frac{e^{ikx}}{k^{2}+2V_{0}}$$ where $p$ was replaced by $\sqrt{k^{2}+2V_{0}}$ so that it is explicitly taken into account that $p>2V_{0}$. By evaluating the integral one gets the estimate $$P(x,t)\sim e^{\frac{-\Delta ^{2}(x_{0}+p_{0}t)^{2}}{(\Delta ^{4}+t^{2})}%
}\;e^{-2x\sqrt{2V_{0}}} \label{cltunn}$$ which has two important features. One, the probability for over the barrier transmission is incomparable larger than the estimate based only on the initial distribution of momenta, however, this happens when $t\sim
|x_{0}/p_{0}|$. Second, the probability decays exponentially for increasing $%
x$, but there is no time dependence of it, except in the factor that indicates arrival of the incident probability. In other words, the probability for over the barrier transmission spreads instantaneously in the whole $x>0$ half space. This, again, as an artifact of the non-relativistic dynamics, which can be shown by the relativistic treatment of this problem (because of its rather lengthy treatment the details are not given here).
The effect for over the barrier transmission, and the exponential dependence of its probability with the coordinate, can be tested by making the potential zero at some distance $x=\delta >0$. If there is relatively large probability to observe the particle in the space $x>\delta $ then that would be direct test of the uncertainty principle, because this is the only way to explain its appearance. It would be called a paradox, and it is called the *tunneling effect,* because the only conclusion from the initial conditions is that such events are not possible (or with the negligible probability), and classical dynamics cannot account for such a large dispersion of momenta. However, it will be shown that once uncertainty principle is implemented the tunneling effect has classical explanation as the over the barrier transmission.
If the effect of tunneling is manifestation of the uncertainty principle, and not dynamics of the particle, then the question is whether there is any meaning in saying that the solutions of (time dependent) Schroedinger equation can be obtained by solving classical equations of motion. The answer is affirmative because inability to describe the change in the momentum distribution, which is due to the uncertainty principle, is replaced by the unusual initial phase space distribution. The meaning of this will be demonstrated on the more exact calculation for the over the barrier transmission, i.e. the tunneling probability on a step potential. Scattering on the step potential has been analyzed in details, and it was showed that classical and quantum calculations for the probability $P(x,t)$ produce identical results. Therefore the analysis of this example will not be analyzed in details, only the essentials points in the part that is relevant for this discussion.
The idea is, as discussed previously, to replace the potential by the boundary condition, but for the step potential its implementation is somewhat more elaborate than in the case of the infinite barrier. Complication is caused by the fact that the phase space density is not confined to the space $x<0$, but it is also transmitted into the space $x>0$. Because of that it is necessary to analyze two separate sets of trajectories: one in the zero potential and the other in the potential $%
V_{0} $. This means that if the same idea as for the infinite barrier is used then two separate phase space densities should be defined.: one when the potential is zero ($\rho _{1}$) and the other ($\rho _{2}$) when it has the value $V_{0}$. Both are defined on the whole x axes, but the phase space density $\rho _{1}(x,p,t)$ is only meaningful in the space $x<0$ while $\rho
_{2}(x,p,t)$ in the space $x>0$. The initial conditions are set on the negative x axes, which means that the phase space density $\rho _{1}(x,p,t)$ is defined from them. On the other hand, the phase space density $\rho
_{2}(x,p,t)$ does not have direct relationship to them, only indirectly through the boundary condition at $x=0$ for the two quantities: the probability densities $P_{1}(x,t)$ and $P_{2}(x,t)$, and the probability currents $j_{1}(x,t)$ and $j_{2}(x,t)$ that are defined for the phase space densities $\rho _{1}(x,p,t)$ and $\rho _{2}(x,p,t)$, respectively. At the boundary it is required that $P_{1}(x,t)$ $=P_{2}(x,t)$ and $j_{1}(x,t)=$ $%
j_{2}(x,t)$.
The details of deriving the phase space densities are omitted, because that was shown elsewhere [@bos8]. The analysis is relatively complicated for the review, and so only the final result will be cited, for the particular case of interest: the tunneling. The initial phase space density $\rho
_{2}^{0}(x,p)$ is given by $$\rho _{2}^{0}(x,p)=\;16\func{Re}\left\{ \int dk\,A(k)A^{*}(k_{p}^{-})\;\frac{%
k\left[ 2p-K(k)\right] }{\left[ k+K(k)\right] \left[ k+K(k_{p}^{-})\right] }%
\;e^{2ix[K(k)-p]}\right\} \,$$ where $K(k)=\sqrt{k^{2}-2V_{0}}$. The variable $k_{p}^{-}$ is defined as $$k_{p}^{-}\;=\pm \;\left[ K(k)-2p+i\sqrt{2V_{0}}\right] ^{1/2}\left[ K(k)-2p-i%
\sqrt{2V_{0}}\right] ^{1/2}$$ where the sign is selected from the requirement that $\func{Im}[K(k_{p}^{-})%
]<0$. The integration path is in the upper half of the complex k-plane, which has important feature to avoid two Riemann cuts that are defined there, and preferably it should go through the stationary point of the phase of the integrand (this is particularly important for the study of tunneling). For the specific case when $\sqrt{2V_{0}}\gg p_{0}$ the initial phase space density is approximately (the non-essential factors are again omitted) $$\rho _{2}^{0}(x,p)\sim e^{-2x\sqrt{2V_{0}}}\;\func{Re}\left[ e^{-2ixp}\int
dk\,A(k)A^{*}(k_{p}^{-})\;k\;\right]$$ which appears not to make physical sense: it is unbounded on the negative x axes. This is very unpleasant because at later time the phase space density is $$\rho _{2}(x,p,t)=\rho _{2}^{0}(x-pt,p)\sim e^{-2(x-pt)\sqrt{2V_{0}}}\;\func{%
Re}\left[ e^{-2i(x-pt)p}\int dk\,A(k)A^{*}(k_{p}^{-})\;k\right]$$ and in the space $x>0$, where by definition it is meaningful, its amplitude increases without bounds in the limit $t\rightarrow \infty $. Therefore it could be rejected as non-physical. However, it should be recalled that it is the probability $P_{2}(x,t)$ that has physical meaning, and this should be finite for all time. In other words, from the physics of the problem it should follow that $$\stackunder{t\rightarrow \infty }{\lim }P_{2}(x,t)=\stackunder{t\rightarrow
\infty }{\lim }\int dp\;\rho _{2}^{0}(x-pt,p)=0 \label{limp2}$$ For the amplitude (\[ampla\]) this can be explicitly proved by evaluating two integrals, in the variables $k$ and $p$, by the stationary phase method, where the phase of the integrand is $$\vartheta (p,k)=-\frac{1}{2}(k_{p}^{-}-p_{0})^{2}\Delta ^{2}+ik_{p}^{-}x_{0}-%
\frac{1}{2}(k-p_{0})^{2}\Delta ^{2}-ikx_{0}-2(x-pt)\left( \sqrt{2V_{0}}%
+ip\right)$$ The set $$\partial _{p}\vartheta (p,k)=0\;\;\;\;\;;\;\;\;\;\;\partial _{k}\vartheta
(p,k)=0$$ defines the stationary points, which are obtained by first making the replacement $p=r/(\Delta ^{2}\sqrt{2V_{0}})$ and then in the equations retain the leading terms in the powers of $V_{0}$. The solution of this approximate set of equations is (the details are not shown, because obtaining it is straightforward but relatively lengthy) $$k_{st}=\frac{p_{0}\Delta ^{2}-ix_{0}}{\Delta ^{2}+it}\;\;\;\;;\;\;\;p_{st}=%
\Delta ^{2}\left( -p_{0}\Delta ^{4}+tx_{0}\right) \frac{x_{0}+tp_{0}}{\sqrt{%
2V_{0}}\left( \Delta ^{4}+t^{2}\right) ^{2}}$$ The probability $P_{2}(x,t)$ is therefore $$P_{2}(x,t)\sim \;\func{Re}\left[ e^{\vartheta (p_{st},k_{st})}\right]$$ which, it can be shown, is equal to (\[cltunn\]), and also to the solution from the quantum treatment. The limit $t\rightarrow \infty $ is finite, but not zero, however it is very small. The fact that this limit is not zero is an artifact of the choice for the initial probability, which will be discussed shortly. Therefore, despite the fact that the phase space density increases without bounds in the space $x<0$ the probability $P_{2}(x,t)$ has all prerogatives to be physically acceptable.
The stationary value $\;p_{st}$ for the momentum plays the role of the average momentum $p_{0}$ for the free particle, but in this case it is measured with respect to the potential $V_{0}$, and its initial value is $%
p_{st}=-\frac{p_{0}x_{0}}{\sqrt{V_{0}}}$, i.e it is very small. One confirms this by calculating the probability current $j_{2}(x,t)$ in the space $x>0$, and from the definition (\[curr\]) the velocity of the particle is $%
v_{tunn}=j_{2}(x,t)/P_{2}(x,t)$. This calculation produces the identity $%
v_{tunn}=p_{st}$. Therefore, if one can talk about the *tunneling velocity* of the particle then this would be $v_{tunn}$, and it follows that it is very small.
As it was shown classical mechanics describes tunneling but not because it is result of dynamics, but because of the specific initial phase space density that results from implementing the uncertainty principle. From this phase space density, and by using solutions of classical equations of motion, one describes tunneling effect, and the result is the same as by solving time dependent Schroedinger equation. One aspect of this solution is quite intriguing, and needs to be understood properly. Time dependence of the tunneling probability (\[cltunn\]) has a very specific form: time variation of the probability at $x=0$ is instantaneously transmitted to all points $x>0$. This means that at some point $x=\delta $ the time variation of the probability is identical with that at the point $x=0$. If the potential is cut at $x=\delta $ then propagation of the probability in the space $x>\delta $ is the same as for a free particle. In fact it has the same time dependence as if one chooses for it the initial $P_{0}(x)$, and as if there is no gap. The only difference is that the amplitude of this probability is scaled by the factor $e^{-2\delta \sqrt{2V_{0}}}$. The net effect is that the probability travels from the space $x<0$ to $x>\delta $ as if there is no gap, i.e. as if the tunneling velocity is infinite. This finding directly contradicts what had been shown before, that the tunneling velocity is very small. The controversy can be resolved by noting one important physical aspect of tunneling: the interval within which the tunneling probability is significant is of the order $x_{tunn}\sim
(2V_{0})^{-1/2}$. This means that the time it takes for the particle to travel this distance, at the velocity $v_{tunn}$, is $$t_{tunn}=\frac{x_{tunn}}{v_{tunn}}\sim \frac{\left( \Delta ^{4}+t^{2}\right)
^{2}}{\Delta ^{2}\left( -p_{0}\Delta ^{4}+tx_{0}\right) \left(
x_{0}+tp_{0}\right) }\sim -\frac{\Delta ^{2}}{p_{0}x_{0}}$$ which is independent of $V_{0}$. In the last step time dependence was neglected. Physical circumstances require that $|x_{0}|\gg \Delta $, and also that during the time of scattering the shape of the probability is nearly constant, which means that $t\ll \Delta ^{2}$. From the characteristic collision time $t\sim |x_{0}|/p_{0}$ it follows that $%
p_{0}=Mx_{0}/\Delta ^{2}$, where $M$ is a large number. The tunneling time is then $$t_{tunn}\sim \left( \frac{\Delta }{x_{0}}\right) ^{2}\frac{\Delta ^{2}}{M}$$ which is very short compared with the characteristic time variation of the probability for the free particle. Indeed, if the tunneling time is multiplied by $p_{0}$ then the distance that the free probability travels during this interval is $$x=p_{0}t_{tunn}\sim x_{0}\left( \frac{\Delta }{x_{0}}\right) ^{2}=\Delta
\frac{\Delta }{x_{0}}$$ which is small. Therefore, the reason why the probability has time dependence (\[cltunn\]) is that the tunneling velocity, although in absolute magnitude is very small, is sufficiently large so that any change at one end is transmitted ”instantaneously” to the other.
There are other questions in connection with the tunneling effect, e.g. how the specific form of the spatial dependence (\[cltunn\]) is formed, but they cannot be answered without considering more precise theoretical model. In particular to answer these questions the initial probability (\[initp0\]) is not adequate, because it would be more accurate to work with the one that is strictly zero outside certain boundary. In connection with this one should also use relativistic theory, because any cutoff in the probability on the x axes makes distribution of momenta very wide, which also includes relativistic values, i.e. for a given momentum $p$ the velocity of particle $%
p/\sqrt{1+p^{2}}$ is nearly the speed of light. However, considering these issues would require more extensive discussion, but results would not contribute in an essential way to the understanding of the tunneling effect.
Conclusion
==========
The aim of the previous discussion was to show that the two basic quantum effects: interference and tunneling, can be explained and quantitatively described by formulating dynamics of a particle from the classical principles, with addition of the uncertainty principle. The results are identical to those obtained from quantum mechanics, and by that it is meant solutions of Schroedinger equation. The two approaches are different ways of seeing the same effects, in many respects analogous to analyzing the motion of classical particles from either the Newton‘s equations of motion or from the Lagrange principle of least action. Or, solving the harmonic oscillator problem starting from matrix mechanics or from the differential equation. The advantage of analyzing quantum effects from the classical principles is that one works in the phase space, and therefore sees the problem with additional degrees of freedom. In this respect more information is available about the system, which is lost if one only works in, say, the coordinate space. The last is characteristic of quantum mechanics, and although one can switch between the coordinate and momentum spaces, one never works in both at the same time. Crudely speaking, quantum mechanics works with the averaged quantities in one of the phase space coordinates and because of that one easily makes erroneous conclusions. This is best observed in the analysis of the zero point energy in one of the previous sections. From quantum mechanics one makes conclusion that relatively large portion of the probability for the zero point energy is due to tunneling, but in fact its shape is explained entirely classically.
There was attempt to overcome this drawback of quantum mechanics by formulating *quantum phase space density*, which would enable to study dynamics of the particle in all the phase space variables. The transform that extends quantum dynamics into the phase space is the Wigner function $w(\vec{r},\vec{p},t)$, which was mentioned in Section 2. For the wave function $\psi (\vec{r},t)$ it is defined through the property $$|\psi (\vec{r},t)|^{2}\;=\;\int d^{3}p\,\,w(\vec{r},\vec{p}%
,t)\;\;\;\;;\;\;\;\;|\phi (\vec{p},t)|^{2}\;=\;\int d^{3}r\,\;w(\vec{r},\vec{%
p},t)$$ where $\phi (\vec{p},t)$ is the wave function in the momentum space. This, however, is the only connection with the phase space density that was used throughout the paper. Namely, the Wigner function, or the quantum phase space density, does not satisfy any simple equation, and definitely not the Liouville equation, except in the special case of the harmonic-type forces. The Wigner function is used as the mean to study the limit $\hbar
\rightarrow 0$ of quantum mechanics, and as the proof that this is classical mechanics it is shown that Liouville equation is obtained. Based on such arguments the quantum phase space density was used in the study of the quantum-classical relationship [@muga1; @muga2; @balazs; @heller; @lee; @bonci; @berry; @bonasera; @royer; @ghosh; @smerzi; @kasperkovitz]. The quantum phase space density is mentioned in the context of this work because it is the base of a dilemma: is classical mechanics the limit of quantum when $\hbar \rightarrow 0$ or is quantum mechanics derived from classical by taking into account the uncertainty principle? This point was discussed elsewhere [@bos8] and therefore will not be discussed here in details. The resime of this discussion is that by strictly taking the limit $\hbar \rightarrow 0$ in quantum mechanics one obtains classical but with a special property, in which the phase space density is parametrized as (\[ro1\]). On the other hand, by starting from classical mechanics one obtains correct result for, say, the ground state of harmonic oscillator.
The disadvantage of working in the phase space (or from classical principles) is that solving problems is more difficult, and often not straightforward. As discussed on the example of tunneling the essential new feature that the uncertainty principle introduces into dynamics is that the change in the momentum is not necessarily caused by the action of a force (it can be called *non-dynamic effect*). Therefore, action of a force on the particle is not sufficient to reproduce quantum results, one needs to take into account that momentum changes from the other cause, the implementation of the uncertainty principle. The exception is harmonic force, for which it can be shown not to affect the phase space density in a way that would change the momentum distribution that cannot be explained by the force itself. This can be demonstrated by solving scattering problem on inverted parabolic potential $V(x)=-\frac{1}{2}wx^{2}$. Given initial conditions $x_{i}$ and p$_{i}$ for classical trajectory in this potential its time dependence is $$x=x_{i}\,cosh(\omega t)+\frac{p_{i}}{\omega }\,sinh(\omega
t)\;\;\;;\;\;p=p_{i}\,cosh(\omega t)+\omega x_{i}\,sinh(\omega t)\;$$ and the phase space density $\rho (x,p,t)$, if it is initial $\rho _{0}(x,p)$, is $$\rho (x,p,t)=\rho _{0}\left[ x\;cosh(\omega t)-\frac{p}{\omega }%
\,sinh(\omega t),-p\;\,cosh(\omega t)+\omega x\,sinh(\omega t)\right] \;$$ From this phase space density one calculates the probability $P(x,t)$ from the definition (\[pq\]). This is classical solution for arbitrary $\rho
_{0}(x,p)$, and specifically if the initial conditions (\[initp0\]) are chosen then the result is $$P(x,t)=\frac{1}{\sqrt{\pi \Delta _{t}^{2}}}\,e^{-(x-x_{t})^{2}/\Delta
_{t}^{2}}$$ where $$\begin{aligned}
\Delta _{t} &=&\Delta \left[ cosh^{2}(\omega t)\,+\,\frac{sinh^{2}(\omega t)%
}{\omega ^{2}\Delta ^{4}}\right] ^{1/2} \\
x_{t} &=&x_{0}\,cosh(\omega t)\,+\,\frac{p_{0}}{\omega }\,sinh(\omega t)\end{aligned}$$ The same result is obtained if the probability $P(x,t)$ is calculated from quantum mechanics, i.e. by solving Schroedinger equation for the time evolution of the wave function. This result is independent of the particular choice of the initial conditions, but for the Gaussian type (\[initp0\]) it has convenient analytic form.
For other than harmonic potentials the contribution of the non-dynamic effects may be significant, but sometime negligible. However, the problem can be avoided by working with the step like potentials, as mentioned in Section 2, in which case the change in potential is replaced by the boundary conditions on the phase space density. By doing that one also includes the non-dynamic effects into account. The procedure is exact, but the result for the time evolution of the phase space density is relatively complicated. The solution simplifies considerably if one is only interested in the time evolution of the probability $P(x,t)$, when it is sufficient to solve the equation (\[schr\]), the procedure that is valid for any potential. That it is a correct one can be demonstrated on one example. It will be assumed that the potential is a delta function at the origin, i.e. $V(x)=W_{0}\delta
(x)$, and the initial conditions are (\[initp0\]). The potential divides the space into $x<0$ and $x>0$, and in each one the particle moves in zero potential. Time evolution of the phase space density $\rho ^{>}(x,p,t)$ in $%
x>0$ can be thought to originate from some initial $\rho _{0}^{>}(x,p)$. That initial phase space density is obtained by requiring that at $x=0$ the phase space density $\rho ^{<}(x,p,t)$ for the space $x<0$ changes smoothly into $\rho ^{>}(x,p,t)$. The initial phase space density for $\rho
^{<}(x,p,t)$ is obtained from (\[initp0\]) by the same procedure as described in the section on tunneling, with a slight modification due to the fact that the phase space density penetrates into the space $x>0$. The proper connection between the two phase space densities is ensured by the proper choice of the function $f$ that enters the phase space density (\[w\]). It can be shown, without giving the details, that this function for the two spaces is $$f^{<}(x)=\int dk\;A(k)\left[ e^{ikx}+R\left( k\right) e^{-ikx}\right]
\;\;\;;\;\;\;f^{>}(x)=\int dk\;A(k)T(k)e^{ikx}$$ where the coefficients are $$R(k)=-\frac{iW_{0}}{k+iW_{0}}\;\;\;;\;\;\;T(k)=\frac{k}{k+iW_{0}}$$ The phase space density $\rho _{0}^{>}(x,p)$ has analytic form if $%
W_{0}>>p_{0}$, in which case $$\rho _{0}^{>}(x,p)\sim \frac{1}{W_{0}^{2}}e^{-\Delta ^{2}(p-p_{0})^{2}-\frac{%
1}{\Delta 2}(x-x_{0})^{2}}\left[ 2(x-x_{0})^{2}+2\Delta ^{4}p^{2}-\Delta ^{2}%
\right]$$ and its time evolution is $\rho ^{>}(x,p,t)=\rho _{0}^{>}(x-pt,p)$. The probability $P(x,t)$ of finding the particle in the space $x>0$ (the tunneling probability) is then (normalization is omitted) $$P(x,t)=\int dp\;\rho ^{>}(x,p,t)=\frac{(x-x_{0})^{2}+\Delta ^{4}p_{0}^{2}}{%
W_{0}^{2}\left( \Delta ^{4}+t^{2}\right) ^{3/2}}\;e^{\frac{\Delta ^{2}}{%
\Delta ^{4}+t^{2}}(x-x_{0}-p_{0}t)^{2}}$$ which is exactly the same result as if the problem was solved by quantum mechanics.
In conclusion one can say that the alternative formulation of dynamics of particle, which is based on the classical principles with the amendment of the uncertainty principle, gives identical results as the original formulation in terms of the wave-particle dualism. In other words, Schroedinger equation is derived from these classical principles, which was confirm in the analysis of two quantum effects: interference and tunneling.
[99]{} J. von Neumann, *Mathematical Foundations of Quantum Mechanics,* Princeton University Press (1955)
R. P. Feynman, Rev. Mod. Phys. **20**, 367 (1948)
R. P. Feynman and A. R. Hibbs, *Quantum Mechanics and Path Integrals*, McGraw-Hill (1965)
D. Bohm, Phys. Rev. **85**, 166, 180 (1952)
E. Nelson, Phys. Rev. **150**, 1079 (1966)
H. Goldstein, *Classical Mechanics*, Addison-Wesley (1981)
S. D. Bosanac, *Classical Dynamics with the Uncertainty Principle*, in *From Simplicity to Complexity: Information, Interaction, Emergence*, Edited by: A. Mueller, K. Mainzer and W. Saltzer, Vieweg-Verlag, Wiesbaden (1997)
H. Skenderović and S. D. Bosanac, Zeit. f. Phys. **D35**, 107 (1995)
N. Došlić and S. D. Bosanac, Zeit. f. Phys. **D32**, 261 (1995)
S. D. Bosanac, Phys. Rev. **A50**, 2899 (1994)
N. Došlić and S. D. Bosanac, Phys. Rev. **A51**, 3485 (1995)
T. W. Koerner, *Fourier Analysis*, Cambridge University Press (1988)
M. Hillery, R. F. O‘Connell, M. O. Scully and E. P. Wigner, Phys. Rep. **106**, 122 (1984)
P. Carruthers and F. Zachariasen, Rev. Mod. Phys. **55**, 245 (1983)
N. Došlić and S. D. Bosanac, Mol. Phys. **A90**, 599 (1997)
S. D. Bosanac, Physica Scripta **57**, 171 (1998)
Y. Aharonov and D. Bohm, Phys. Rev. **115**, 485 (1959)
M. P. Silverman, *More than One Mystery,* Springer (1995)
S. D. Bosanac, J. Math. Phys. **38**, 3895 (1997)
D. Mugnai, A. Ranfongi and L. S. Schulman, *Tunneling and its Implications,* World Scientific (1997)
J. G. Muga, R. Sala and R. F. Snider, Phys. Scr. **47**, 732, (1993)
R. Sala, S. Brouard and J. G. Muga, J. Chem. Phys. **99**, 2708, (1993)
N. L. Balazs and A. Voros, Ann. Phys. **199**, 123, (1990)
E. J. Heller, J. Chem. Phys. **65**, 1289, (1976)
H. W. Lee and M. O. Scully, J. Chem. Phys. **77**, 4604, (1982)
L. Bonci, R. Roncaglia, B. J. West and P. Grigolini, Phys. Rev. **A45**, 8490, (1992)
M. V. Berry, Phil. Trans. Roy. Soc. **287**, 237, (1977)
A. Bonasera, V. N. Kondratyev, A. Smerzi and E. A. Remler, Phys. Rev. Lett. **71**, 505, (1993)
A. Royer, Phys. Rev. **A43**, 44, (1991)
S. K. Ghosh and A. K. Dhara, Phys. Rev. **A44**, 65, (1991)
A. Smerzi, Phys. Rev. **A52**, 4365, (1995)
P. Kasperkovitz and M. Peev, Phys. Rev. Lett. **75**, 990, (1995)
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Francesco Di Palma$^{1}$, Francesco Colizzi$^{1}$ and Giovanni Bussi$^{1*}$'
date: 'SISSA - Scuola Internazionale Superiore di Studi Avanzati, via Bonomea 265, 34136 Trieste, Italy'
title: 'Ligand-induced stabilization of the aptamer terminal helix in the *add* adenine riboswitch'
---
Stabilization of the P1 stem in *add* riboswitch
keywords: P1 stem, RNA aptamer, adenine riboswitch, molecular dynamics simulation, free energy calculation.
**ABSTRACT**
Riboswitches are structured mRNA elements that modulate gene expression. They undergo conformational changes triggered by highly-specific interactions with sensed metabolites. Among the structural rearrangements engaged by riboswitches, the forming and melting of the aptamer terminal helix, the so-called P1 stem, is essential for genetic control. The structural mechanisms by which this conformational change is modulated upon ligand binding mostly remain to be elucidated. Here we used pulling molecular dynamics simulations to study the thermodynamics of the P1 stem in the *add* adenine riboswitch. The P1 ligand-dependent stabilization was quantified in terms of free energy and compared with thermodynamic data. This comparison suggests a model for the aptamer folding in which direct P1-ligand interactions play a minor role on the conformational switch when compared with those related to the ligand-induced aptamer preorganization.
INTRODUCTION
============
Riboswitches are ligand-responsive regulatory elements located in untranslated regions of messenger RNAs [@serganov2013alexander]. They change their conformation in response to specific metabolite binding [@roth2009structural; @edwards2010riboswitches; @serganov2012metabolite] and they have been proposed as modern descendants of an ancient sensory and regulatory system in the RNA world [@breaker2012riboswitches]. Many pathogenic bacteria use riboswitches to control essential metabolic pathways and they are currently regarded as promising antibacterial drug targets [@blount2006riboswitches; @mulhbacher2010novel; @deigan2011riboswitches]. Riboswitches consist of an aptamer domain that binds the effector ligand and an expression platform that transduces the ligand-induced conformational “switch” into a modulation of gene expression [@barrick2007distributions; @garst2009switch]. Among more than 20 natural aptamer classes [@breaker2012riboswitches], purine-sensing riboswitches have the peculiarity to recognize the targeted purine by utilizing a conserved pyrimidine [@kim2008purine; @batey2012structure]. One of the most characterized members of this class is the adenine sensing riboswitch (A-riboswitch) *cis*-regulating the *add* gene in *Vibrio vulnificus* [@mandal2003adenine]. The ligand-bound structure of its aptamer is a three-way junction composed of three stems (P1, P2, P3) with the ligand completely encapsulated into the structure (Fig. 1) [@serganov2004structural; @mandal2004gene]. The specificity for adenine is ensured by canonical Watson-Crick (WC) base pairing established between a uracil in conserved position and the ligand [@noeske2005intermolecular; @gilbert2006thermodynamic].
The A-riboswitch acts as a translational regulator [@serganov2004structural; @lemay2011comparative]. In the absence of adenine the ribosome binding site and the initiation codon, which are portions of the expression platform, are sequestered by pairing with a portion of the aptamer (OFF-state, Fig. 2B). The presence of adenine stabilizes an aptamer conformation in which the terminal P1 helix is well structured and both the regulatory sequences are available for ribosomal binding thus enabling mRNA translation (ON-state, Fig. 2A) [@rieder2007ligand; @lee2010real; @leipply2011effects]. The structural mechanism regulating the switch between the ON- and the OFF-state upon ligand binding mostly remains to be elucidated. The P1 stem is formed in the ON-state and it is disrupted in the OFF-state [@mandal2003adenine; @serganov2004structural].
It has been proposed that P1 is stabilized by the ligand [@batey2004structure] and that this could be a common feature in many riboswitch classes [@montange2006structure]. The role of ligand binding in the structural organization of the aptamer has been investigated with single-molecule spectroscopy providing an insightful overview on the folding dynamics [@neupane2011single], yet lacking the critical atomistic details needed for an accurate structural characterization of the process [as extensively discussed by [@lin2012rna].]{} Although *in silico* techniques have been used to investigate the ligand role [@lin2008relative; @sharma2009md; @priyakumar2010role; @gong2011role; @allnr2013loop], a quantitative estimation of the energetic contributions associated to ligand binding, in particular regarding the role of direct P1-ligand interactions, has not yet been provided. In this context, state-of-the-art free-energy methods combined with atomistic simulations can bridge the gap providing an unparalleled perspective on the mechanism and dynamics of the biomolecular process of interest [@dellago2009transition]. In this work we used steered molecular dynamics (SMD) [@grubmuller1996ligand; @sotomayor2007single] simulations to study the thermodynamics of the P1 stem formation in the presence and in the absence of the cognate ligand. We enforced the breaking of the P1 stem base pairs (bp) and then using a recently developed reweighting scheme [@colizzi2012rna] we quantitatively estimated the ligand-induced stabilization of the helix. The A9-U63 bp which directly stacks with adenine was used as a proxy for the P1 stability. Our non-equilibrium simulations provide measurements of the stability of the A9-U63 bp and quantify the direct ligand-dependent stabilization of the pairing. In the following our results are presented and compared with melting and single-molecule experiments. A structural model for the conformational switch emerging from the combination of our results and previous experimental data is also discussed.
RESULTS
=======
We carried out the simulations of the aptamer domain of the *add* A-riboswitch in different forms, namely the entire aptamer (PDB id 1Y26) [@serganov2004structural] has been simulated in the presence (Holo) and in the absence (Apo) of the cognate ligand, the adenine; additionally, to better estimate the ligand-induced stabilization, we also simulated a truncated aptamer ($\triangle$1-8/64-71), both in the Apo and Holo forms. Long unbiased molecular dynamics (MD) for all the four systems were performed to test the stability of the aptamer in different conditions. In the truncated systems the terminal bp was restrained in its initial configuration to mimic the presence of the rest of the stem. Furthermore, the full-length systems were pulled from the terminal bases to disrupt the entire P1 stem (Fig. 3) thus allowing its different stability between in the Holo and in the Apo forms to be qualitatively inferred. At last, to quantify this difference, SMD simulations of both the $\triangle$1-8/64-71 systems were done enforcing the breaking of the A9-U63 bp that directly stacks with the ligand (Fig. 4).
The stability of both the Apo and Holo systems was evaluated monitoring the root mean square deviation (RMSD) from the native structure along 48 ns MD runs (Fig. 5A-B). Ligand removal (see Materials and Methods for details) did not affect the overall stability of the Apo aptamer in this time-scale, and secondary and tertiary structures were substantially unchanged.
The analysis of the trajectories obtained by pulling the P1 stem showed that the secondary and tertiary structure elements of the rest of the aptamer were not affected by the opening of the helix (data not shown). Focusing our attention on the P1 stem we observed that in the Apo form the A9-U63 bp (Fig. 3) was broken when the distance between the centers of mass of the terminal bases reached a value of $\approx$9.8 nm. Differently, in presence of the ligand a longer pulling was needed and the rupture only happened at a distance of $\approx$11.5 nm (Fig. 6). This behavior is compatible with the picture in which the ligand stabilizes the P1 stem [@montange2006structure]. It was however difficult to extract quantitative information on the ligand-P1 interaction from these simulations because the rupture is a stochastic event and extensive sampling would be required. [Moreover, as pointed out in a recent paper [@lin2012rna], the end-to-end distance could be a non-optimal CV for pulling experiments or simulations since local bp formation plays an important role in global stem folding.]{}
The quantitative analysis of the P1-ligand interaction was better obtained from the simulation of both the $\triangle$1-8/64-71 systems. We verified that, when the P1 stem is replaced with the A9-U63 bp restrained to be in canonical WC pairing, the aptamer remains stable (Fig. 5C-D). Remarkably, fluorescence experiments have shown that the aptamer can also fold and bind adenine when large fractions of the P1 stem are removed [@lemay2007core]. This validates the possibility of using the two structures, $\triangle$1-8/64-71 Holo and $\triangle$1-8/64-71 Apo, to investigate the direct P1 stabilization given by the adenine. In the following we focus on the SMD simulations performed on these truncated forms. Typical initial and final conformations from the SMD are shown in Fig. 4.
Analysis of work profiles
-------------------------
The unbinding event of the A9-U63 bp is described as a function of the value of the steered RMSD in Fig. 7. The initial value corresponds to the configuration with the WC pairing formed, whereas at the final value (0.35 nm) the pairing is completely broken. Even if the ensembles of work profiles for the two forms are broadly spread and superposed, the free-energy profiles computed using the Jarzynski equality [@jarzynski1997nonequilibrium] as the exponential average of the two sets of data are clearly distinguishable (Fig. 7A). Qualitatively it is worth highlighting that during the breaking of the A9-U63 bp the Apo form (red line) is always lower in free energy than the Holo form (blue line). It follows that the breaking of the monitored bp in the Apo form was unambiguously more probable than in the Holo one (Fig. 7A). However, such an approach was still a long way off from quantitatively accounting for the energetic stabilization of the A9-U63 bp related to the presence of adenine in the binding site. Within this framework there was no way to automatically detect when the nucleobases reached the unbound configuration. It was thus difficult to avoid systematic errors in the comparison of the two systems. Furthermore, few low-work realizations occurred during the unpairing in the presence of adenine. In these low-work realizations the number of hydrogen bonds was non-zero at large RMSD values and the structural analysis of the trajectories revealed the transient formation of a cis-sugar edge pair (data not shown) [@leontis2002non]. Due to the exponential nature of the Jarzynski average, these low-work realizations dominated the free energy profile for the Holo form further compromising the possibility of a quantitative comparison with the Apo form.
Energetics of hydrogen bond breaking
------------------------------------
We thus analyzed the trajectories in terms of number of hydrogen bonds formed between A9 and U63, a discrete variable that more strictly reported on the breaking of the pairing. In this metrics, the bound (1 or 2 hydrogen bonds) and unbound (0 hydrogen bond) ensembles could be clearly and unambiguously identified thus allowing a quantitative comparison between the Apo and the Holo system. Additionally, the configurations from the outlier trajectories could be assigned properly to one or the other ensemble in spite of their atypical RMSD value.
The differences in free energy ($\Delta F$) between the ensembles, with and without hydrogen bonds, was computed using a reweighting scheme [@colizzi2012rna]. The values and the associated standard errors were estimated for both systems. For the Apo form $\Delta F=-2.5\pm1.4\ kJ/mol$ suggesting that the bp could spontaneously break in the absence of adenine. For the Holo form $\Delta F=1.9\pm1.7\ kJ/mol$ implying that the presence of the ligand and its pairing with U63 stabilized the stacked bp. The $\Delta\Delta F$ between the two forms is equal to $-4.4\pm2\ kJ/mol$. This value quantifies the thermodynamic stabilization to the formation of the base pair which directly interacts with adenine in the P1 stem.
DISCUSSION AND CONCLUSIONS
==========================
Our simulations at atomistic detail provide for the first time the free-energy contribution of ligand stacking to the formation of the P1 stem in a riboswitch. In particular, the presented *in silico* approach allows the energetics involved in the aptamer stabilization upon ligand binding to be dissected in detail. Below we compare our results with single-molecule manipulation, both *in vitro* and *in silico*, and thermodynamic data from dsRNA melting experiments. We also provide a model for ligand-modulated co-transcriptional folding of the *add* riboswitch.
Comparison with related works
-----------------------------
Our results are in nice agreement with thermodynamic data based on dsRNA melting experiments [@mathews2004incorporating; @turner2010nndb]. The comparison between our simulations and those experiments can be straightforwardly achieved by considering the pairing between U62 and the sensed adenine as an additional terminal bp of the P1 stem. The direct stabilization of the P1 stem due to the cognate-ligand binding should be thus equivalent to that given by adding one further AU bp to the P1 helix. Using the most recent nearest neighbor energy parameters for the comparison of RNA secondary structures [@mathews2004incorporating; @turner2010nndb], the free-energy difference between the sequence of the P1 stem with and without the additional AU base pair,
---------------------------------------
$_{\textrm{5'-CGCUUCAUA\emph{A}-3'}}$
$^{\textrm{3'-GUGAAGUAU\emph{U}-5'}}$
---------------------------------------
and
-------------------------------
$_{\textrm{5'-CGCUUCAUA-3'}}$
$^{\textrm{3'-GUGAAGUAU-5'}}$
-------------------------------
, can be computed [@hofacker1994fast; @lorenz2011viennarna] as $\Delta\Delta F=-3.7\ kJ/mol$, consistently with our results.
Our free-energy estimates complement previously reported investigations in which the role of the ligand in the folding process of the A-riboswitch has always been referred to the whole aptamer [@lin2008relative; @neupane2011single] and never specifically to the P1 stem. Using a one-bead-per-nucleotide coarse-grained model, the $\Delta\Delta F$ has been computed as approximately $-15\ kJ/mol$ [@lin2008relative]. [Notably also this calculation has been done using a shortened P1 stem, possibly affecting the $\Delta F$ estimation.]{} Single-molecule force spectroscopy experiments have been also performed to characterize the folding pathway of the aptamer with an estimated $\Delta\Delta F\simeq-33\ kJ/mol$ [@neupane2011single]. However in both these works the separated contributions of the P1-ligand stacking, of the interaction between the ligand and the junctions J1-2, J2-3 and J3-1, and of the interaction between loops L2-L3 could not be discerned (secondary structure elements labeled as in Fig. 8).
From the comparison of our data with the above mentioned experimental and computational works, a twofold modular role for the ligand emerges. On the one hand, the binding of adenine can contribute to the aptamer preorganization and it could allow the long-range induction of the tight hydrogen-bonding and base-stacking networks observed in the native state [@rieder2007ligand; @lee2010real]. This preorganization would reduce the distance between A9 and U63, thus increasing the probability of their pairing. A similar mechanism has been proposed also for the SAM-I riboswitch [@whitford2009nonlocal]. On the other hand, adenine binding enhances the P1 formation by direct stacking interaction, mimicking the extension of the helix by an additional bp. Notably, the energetic contribution of the direct stacking is smaller than that involved in the aptamer preorganization. The latter can be estimated as the difference between the global ligand-induced aptamer stabilization [@lin2008relative; @neupane2011single] and the stacking contribution dissected in our work.
Folding model
-------------
Our work provides atomistic details and energetic estimates to the currently accepted model for the folding of the *add* riboswitch upon ligand binding [@rieder2007ligand; @lee2010real; @leipply2011effects]. Altogether, our data and the related experimental works suggest a folding model as depicted in Fig. 8. Initially, only the P2 and P3 stems and the corresponding loops (L2, L3, still not interacting each other) are formed and not fully stable (Fig. 8A). Then, adenine binding allows for a preorganization of the aptamer where the three junctions arrange around the ligand (Fig. 8$\textrm{B}_{\textrm{1}}$), stabilizing also the previously formed helices [@rieder2007ligand]. It has not been established clear if the loop-loop interaction is formed before or after ligand binding [@leipply2011effects]. Thus, an alternative pathway, the junctions and the P1 could acquire a partially folded conformation also in the absence of adenine (Fig. 8$\textrm{B}_{\textrm{2}}$) [@lee2010real]. Finally, the P1 helix becomes fully structured and stabilized by the ligand (Fig. 8C), to the detriment of the expression platform (see Fig. 2) [@lemay2011comparative]. This step is mandatory for translation to be initiated. We quantified the ligand contribution to the P1 stem formation due to direct interactions to be approximately -4 kJ/mol.
Our result is compatible with both the folding pathways (Figs. 8$\textrm{B}_{\textrm{1}}$ and 8$\textrm{B}_{\textrm{2}}$) irrespectively of their relative population and cannot discriminate among them. The relative probability between the two paths can be modulated by the ligand concentration and its binding affinity. On the one hand, the intermediate shown in Fig. 8$\textrm{B}_{\textrm{1}}$ could be relevant for ligand-RNA binding in an early transcriptional context in which the last 9 nucleobases (i.e those allowing P1 formation) of the aptamer have not yet been synthesized. Indeed, it has been shown that also an aptamer missing a large portion of the P1 stem is able to bind adenine [@lemay2007core]. On the other hand, the intermediate shown in Fig. 8$\textrm{B}_{\textrm{2}}$ could be populated at low-ligand concentration once the nucleobases allowing P1 formation are synthesized. Later on, after the synthesis of the expression platform, ligand binding could shift the thermodynamic equilibrium towards one of two competing riboswitch conformations (P1 formed and non formed).
Conclusion
----------
Ligand-induced stabilization of the P1 stem is crucial for A-riboswitch regulation and function. Here we quantified the direct interaction between adenine and P1 stem and analyzed it at atomistic detail. Our results suggest a model for the aptamer folding in which the direct P1-ligand interactions play a minor role on the conformational switch when compared with those related to the ligand-induced aptamer preorganization. Because the structural/functional role of the aptamer terminal helix is a common feature in the “straight junctional” riboswitches [@serganov2013alexander], we foresee a wider validity of the model presented herein.
MATERIALS AND METHODS
=====================
System description and set-up
-----------------------------
We simulated the Holo and the Apo form of the A-riboswitch aptamer domain both composed of 71 nucleotides. The Apo form was generated by adenine removal from the ligand-bound (Holo) crystal structure (PDB id: 1Y26) [@serganov2004structural]. This deletion is justified by the fact that the Apo and Holo form have been shown experimentally to share an overall similar secondary structure [@lemay2011comparative]. This is at variance with e.g. the *pbuE* adenine riboswitch in which the two structures are different. The generation of the Apo form by simply removing the ligand has been adopted also in a recent work [@allnr2013loop]. Molecular dynamics (MD) simulations were performed using the Amber99 force field [@wang2000well] refined with the *parmbsc0* corrections [@perez2007refinement]. From the analysis of the SMD trajectories we do not expect the refinement on the $\chi$ torsional parameters [@zgarbova2011refinement] to affect the results. Adenine was parametrized using the general Amber force field (gaff) [@wang2004development]. Partial atomic charges were assigned using the restricted electrostatic potential fit method [@bayly1993well] based on an electronic structure calculation at the HF/6-31G[\*]{} level of theory performed with Gaussian03 [@g03]. Bond-lengths were constrained with LINCS [@hess1997lincs] and the electrostatic interactions were calculated using the particle-mesh Ewald method [@darden1993particle]. For both forms, the following protocol was used to prepare the systems (Table \[tabS1\] for details) for MD simulations: steepest descent minimization (200 steps) starting from the X-ray structure. Solvation with $\approx$13000 TIP3P water molecules [@jorgensen1983comparison] and NaCl at 0.15 M concentration (plus extra Na$^{+}$ counter-ions to neutralize the charges of the systems) in a hexagonal prism (lattice vectors in nm [\[]{}(10,0,0), (0,7,0), (0,$\frac{7}{2}$,$\frac{7\cdot\sqrt{3}}{2}$)[\]]{}) that was created orienting the major length of the aptamer along the X axis. Steepest descent minimization (200 steps) for ions and solvent; the systems were thermalized at 300 K, initially for 200 ps with frozen solute positions and then for 5 ns in NPT ensemble (1 atm) with stochastic velocity rescaling [@bussi2007canonical] and Berendsen barostat [@berendsen1984molecular]; to maintain the systems oriented along the largest lattice vector (X) a restraint was imposed with a force constant of 4$\cdot$10^3^$\unitfrac{(\unitfrac{kJ}{mol})}{nm^{2}}$ on the Y and Z components of the distance between phosphate atoms of A52 and G71. Each system was simulated for 48 ns in NVT ensemble to assess the stability of the aptamer.
Steered molecular dynamics
--------------------------
To perform SMD simulations inducing the opening of the whole P1 stem the systems were solvated again with $\approx$39500 water molecules in a larger rhombic dodecahedral box with distance between periodic images equal to 12 nm, adding ions to maintain the same ionic strengh (P1-SMD systems in Table \[tabS1\]). The same protocol described above was applied for the minimization, thermalization and equilibration of this larger Holo and Apo systems for the pulling simulations. An incremental separation between the centers of mass of the terminal nucleotides (C1 and G71) was imposed from an initial value of 1.05 nm to a value sufficient to completely unfold the 9 bp of the P1 helix (Apo 10.05 nm, Holo 11.75 nm) at a speed of 0.56 nm/ns (see Fig. 3). The spring constant was set to 3.9$\cdot$10^4^$\unitfrac{(\unitfrac{kJ}{mol})}{nm^{2}}$.
The first eight bp of the P1 stem (i.e. whole stem from C1 to U8 and from A64 to G71, except for A9-U63 bp) were then cut in both systems creating the $\triangle$1-8/64-71 Holo and $\triangle$1-8/64-71 Apo structures (Table \[tabS1\] for details). Water molecules were allowed to relax filling the space left by the 16 removed bases through an additional 1 ns equilibration in which the positions of aptamer atoms were frozen followed by 5 ns of unrestrained NPT simulation. Then the systems were simulated for 48 ns in the NVT ensemble restraining the terminal bases in the initial state to avoid any spontaneous flipping. The pairs deletion is not biologically meaningless, because it has been shown experimentally that a series of aptamer variants with shorter P1 helix are still able to bind the ligand [@lemay2007core]. The deletion reduced the noise during the pulling allowing to focus the calculation on the influence of the ligand on the A9-U63 pairing. This bp rupture was here enforced by pulling on the RMSD between the heavy atoms of A9 and U63 with reference to the crystal structure. This collective variable (CV) was chosen as it identifies the native conformation (RMSD $\approx$ 0) of the A9-U63 bp, which is necessary for the initiation of the P1 stem formation. The steered CV was pulled at constant velocity of 0.175 from 0 to 0.35 nm in 2 ns. This pulling induced the complete opening of the A9-U63 bp in presence and absence of the ligand (Fig. 4). The spring constant was set to 3.9$\cdot$10^4^$\unitfrac{(\unitfrac{kJ}{mol})}{nm^{2}}$. The starting points were extracted equidistantly (one every 16 ps) from a 8.192 ns run NVT ensemble restraining the RMSD value of those atoms at 0. For the two system 512 independent SMD simulations were performed, corresponding to an aggregate time of approximately 1$\mu$s each. Simulations were carried out with the Gromacs 4.0.7 program package [@hess2008gromacs] combined with the PLUMED 1.3 plug-in [@bonomi2009plumed].
Analysis
--------
The Jarzynski equality [@jarzynski1997nonequilibrium] was used to estimate the equilibrium free-energy landscape of the process from the collected work profiles. The simulations were then analyzed using a recently proposed scheme [@colizzi2012rna] which combines an identity by Jarzynski [@PhysRevE.56.5018] with the weighted-histogram analysis method [@kumar1992weighted]. The algorithm allows the free-energy profiles to be projected onto any *a posteriori* chosen CV. [It is well known that free-energy calculations using Jarzynski-based relationships are difficult to converge. Statistical errors were thus estimated by the bootstrapping procedure described in [@do2013rna] indicating that our results were converged within $\approx k_BT$.]{} The VIENNA RNA package [@hofacker1994fast; @lorenz2011viennarna] was used to compare our results with the thermodynamic data based on dsRNA melting experiments [@mathews2004incorporating; @turner2010nndb].
ACKNOWLEDGMENTS
===============
We thank Daniel Lafontaine and Gabriele Varani for reading the manuscript and providing critical comments. Sandro Bottaro is also acknowledged for carefully reading the manuscript and suggesting several improvements. We acknowledge the CINECA award no. HP10B2G6OF, 2012 under the ISCRA initiative for the availability of high performance computing resources.
The research leading to these results has received funding from the European Research Council under the European Union’s Seventh Framework Programme (FP/2007-2013) / ERC Grant Agreement n. 306662, S-RNA-S.
[58]{} \[1\][\#1]{} urlstyle \[1\][doi:\#1]{}
Alln[é]{}r O, Nilsson L, Villa A. 2013. Loop-loop interaction in adenine-sensing riboswitch: A molecular dynamics study. *RNA* **19**:916–926.
Barrick JE, Breaker RR. 2007. The distributions, mechanisms, and structures of metabolite-binding riboswitches. *Genome Biol* **8**:R239.
Batey RT. 2012. Structure and mechanism of purine-binding riboswitches. *Q Rev Biophys* **1**:1–37.
Batey RT, Gilbert SD, Montange RK. 2004. Structure of a natural guanine-responsive riboswitch complexed with the metabolite hypoxanthine. *Nature* **432**:411–415.
Bayly CI, Cieplak P, Cornell W, Kollman PA. 1993. A well-behaved electrostatic potential based method using charge restraints for deriving atomic charges: the resp model. *J Phys Chem* **97**:10269–10280.
Berendsen HJC, Postma JPM, van Gunsteren WF, DiNola A, Haak JR. 1984. Molecular dynamics with coupling to an external bath. *J Chem Phys* **81**:3684.
Blount KF, Breaker RR. 2006. Riboswitches as antibacterial drug targets. *Nat Biotechnol* **24**:1558–1564.
Bonomi M, Branduardi D, Bussi G, Camilloni C, Provasi D, Raiteri P, Donadio D, Marinelli F, Pietrucci F, Broglia RA, Parrinello M. 2009. Plumed: A portable plugin for free-energy calculations with molecular dynamics. *Comput Phys Commun* **180**:1961–1972.
Breaker RR. 2012. Riboswitches and the [RNA]{} world. *Cold Spring Harb Perspect Biol* **4**.
Bussi G, Donadio D, Parrinello M. 2007. Canonical sampling through velocity rescaling. *J Chem Phys* **126**:014101–014107.
Colizzi F, Bussi G. 2012. R[NA]{} unwinding from reweighted pulling simulations. *J Am Chem Soc* **134**:5173–5179.
Darden T, York D, Pedersen L. 1993. Particle mesh ewald: An n$\cdot$log(n) method for ewald sums in large systems. *J Chem Phys* **98**:10089.
Deigan KE, Ferr[é]{}-D’Amar[é]{} AR. 2011. Riboswitches: Discovery of drugs that target bacterial gene-regulatory [RNA]{}s. *Accounts Chem Res* **44**:1329–1338.
Dellago C, Bolhuis PG. 2009. Transition path sampling and other advanced simulation techniques for rare events. *Adv Polym Sci* **221**:167–233.
Do TN, Carloni P, Varani G, Bussi G. 2013. /peptide binding driven by electrostatics - [I]{}nsight from bidirectional pulling simulations. *J Chem Theory Comput* **9**:1720–1730.
Edwards AL, Batey RT. 2010. Riboswitches: A common [RNA]{} regulatory element. *Nat Edu* **3**:9.
Frisch MJ, Trucks GW, Schlegel HB, Scuseria GE, Robb MA, Cheeseman JR, Montgomery JA Jr, Vreven T, Kudin KN, Burant JC, Millam JM, et al. 2004. Gaussian 03, evision .02. aussian, Inc., Wallingford, CT.
Garst AD, Batey RT. 2009. A switch in time: detailing the life of a riboswitch. *Biochim Biophys Acta* **1789**:584.
Gilbert SD, Stoddard CD, Wise SJ, Batey RT. 2006. Thermodynamic and kinetic characterization of ligand binding to the purine riboswitch aptamer domain. *J Mol Biol* **359**:754.
Gong Z, Zhao Y, Chen C, Xiao Y. 2011. Role of ligand binding in structural organization of add a-riboswitch aptamer: a molecular dynamics simulation. *J Biomol Struct Dyn* **29**:403.
Grubmuller H, Heymann B, Tavan P. 1996. Ligand-binding-molecular mechanics calculation of the streptavidin biotin rupture force. *Science* **271**:997–999.
Hess B, Bekker H, Berendsen HJC, Fraaije JGEM. 1997. Lincs: a linear constraint solver for molecular simulations. *J Comput Chem* **18**:1463–1472.
Hess B, Kutzner C, van der Spoel D, Lindahl E. 2008. Gromacs 4: Algorithms for highly efficient, load-balanced, and scalable molecular simulation. *J Chem Theory Comput* **4**:435–447.
Hofacker IL, Fontana W, Stadler PF, Bonhoeffer LS, Tacker M, Schuster P. 1994. Fast folding and comparison of [RNA]{} secondary structures. *Monatsh Chem* **125**:167–188.
Jarzynski C. 1997. [N]{}onequilibrium equality for free energy differences. *Phys Rev Lett* **78**:2690.
Jarzynski C. 1997. Equilibrium free-energy differences from nonequilibrium measurements: A master-equation approach. *Phys Rev E* **56**:5018–5035.
Jorgensen WL, Chandrasekhar J, Madura JD, Impey RW, Klein ML. 1983. Comparison of simple potential functions for simulating liquid water. *J Chem Phys* **79**:926.
Kim J, Breaker R. 2008. Purine sensing by riboswitches. *Biol Cell* **100**:1–11.
Kumar S, Rosenberg JM, Bouzida D, Swendsen RH, Kollman PA. 1992. The weighted histogram analysis method for free-energy calculations on biomolecules. i. the method. *J Comput Chem* **13**:1011–1021.
Lee MK, Gal M, Frydman L, Varani G. 2010. Real-time multidimensional [NMR]{} follows [RNA]{} folding with second resolution. *P Natl Acad Sci U S A* **107**:9192–9197.
Leipply D, Draper DE. 2011. Effects of [M]{}g$^{2}+$ on the free energy landscape for folding a purine riboswitch [RNA]{}. *Biochemistry* **50**:2790–2799.
Lemay JC, Lafontaine DA. 2007. Core requirements of the adenine riboswitch aptamer for ligand binding. *RNA* **13**:339–350.
Lemay JF, Desnoyers G, Blouin S, Heppell B, Bastet L, St-Pierre P, Mass[é]{} E, Lafontaine DA. 2011. Comparative study between transcriptionally-and translationally-acting adenine riboswitches reveals key differences in riboswitch regulatory mechanisms. *PLoS Genet* **7**:e1001278.
Leontis NB, Stombaugh J, Westhof E. 2002. The non-watson–crick base pairs and their associated isostericity matrices. *Nucleic Acids Res* **30**:3497–3531.
Lin JC, Hyeon C, Thirumalai D. 2012. under tension: Folding landscapes, kinetic partitioning mechanism, and molecular tensegrity. *J Phys Chem Lett* **3**:3616–3625.
Lin JC, Thirumalai D. 2008. Relative stability of helices determines the folding landscape of adenine riboswitch aptamers. *J Am Chem Soc* **130**:14080–14081.
Lorenz R, Bernhart SH, Zu Siederdissen CH, Tafer H, Flamm C, Stadler PF, Hofacker I. 2011. Vienna[RNA]{} [P]{}ackage 2.0. *Algorithm Mol Biol* **6**:26.
Mandal M, Breaker RR. 2003. Adenine riboswitches and gene activation by disruption of a transcription terminator. *Nat Struct Mol Biol* **11**:29–35.
Mandal M, Breaker RR. 2004. Gene regulation by riboswitches. *Nat Rev Mol Cell Biol* **5**:451–463.
Mathews DH, Disney MD, Childs JL, Schroeder SJ, Zuker M, Turner DH. 2004. Incorporating chemical modification constraints into a dynamic programming algorithm for prediction of [RNA]{} secondary structure. *P Natl Acad Sci U S A* **101**:7287–7292.
Montange RK, Batey RT. 2006. Structure of the s-adenosylmethionine riboswitch regulatory m[RNA]{} element. *Nature* **441**:1172–1175.
Mulhbacher J, Brouillette E, Allard M, Fortier LC, Malouin F, Lafontaine DA. 2010. Novel riboswitch ligand analogs as selective inhibitors of guanine-related metabolic pathways. *PLoS Pathog* **6**:e1000865.
Neupane K, Yu H, Foster DAN, Wang F, Woodside MT. 2011. Single-molecule force spectroscopy of the add adenine riboswitch relates folding to regulatory mechanism. *Nucleic Acids Res* **39**:7677–7687.
Noeske J, Richter C, Grundl MA, Nasiri HR, Schwalbe H, W[ö]{}hnert J. 2005. An intermolecular base triple as the basis of ligand specificity and affinity in the guanine-and adenine-sensing riboswitch [RNA]{}s. *P Natl Acad Sci U S A* **102**:1372–1377.
P[é]{}rez A, March[á]{}n I, Svozil D, Sponer J, Cheatham III TE, Laughton CA, Orozco M. 2007. Refinement of the amber force field for nucleic acids: Improving the description of $\alpha$/$\gamma$ conformers. *Biophys J* **92**:3817–3829.
Priyakumar U, MacKerell AD. 2010. Role of the adenine ligand on the stabilization of the secondary and tertiary interactions in the adenine riboswitch. *J Mol Biol* **396**:1422–1438.
Rieder R, Lang K, Graber D, Micura R. 2007. Ligand-induced folding of the adenosine deaminase a-riboswitch and implications on riboswitch translational control. *Chembiochem* **8**:896–902.
Roth A, Breaker RR. 2009. The structural and functional diversity of metabolite-binding riboswitches. *Annu Rev Biochem* **78**:305–334.
Serganov A, Nudler E. 2013. A decade of riboswitches. *Cell* **152**:17–24.
Serganov A, Patel DJ. 2012. Metabolite recognition principles and molecular mechanisms underlying riboswitch function. *Ann Rev Biophys* **41**:343.
Serganov A, Yuan YR, Pikovskaya O, Polonskaia A, Malinina L, Phan AT, Hobartner C, Micura R, Breaker RR, Patel DJ. 2004. Structural basis for discriminative regulation of gene expression by adenine-and guanine-sensing mrnas. *Chem Biol* **11**:1729–1741.
Sharma M, Bulusu G, Mitra A. 2009. Md simulations of ligand-bound and ligand-free aptamer: Molecular level insights into the binding and switching mechanism of the add a-riboswitch. *RNA* **15**:1673–1692.
Sotomayor M, Schulten K. 2007. Single-molecule experiments in vitro and in silico. *Science* **316**:1144–1148.
Turner DH, Mathews DH. 2010. N[NDB]{}: the nearest neighbor parameter database for predicting stability of nucleic acid secondary structure. *Nucleic Acids Res* **38**:D280–D282.
Wang J, Cieplak P, Kollman PA. 2000. How well does a restrained electrostatic potential (resp) model perform in calculating conformational energies of organic and biological molecules? *J Comput Chem* **21**:1049–1074.
Wang J, Wolf RM, Caldwell JW, Kollman PA, Case DA. 2004. Development and testing of a general amber force field. *J Comput Chem* **25**:1157–1174.
Whitford PC, Schug A, Saunders J, Hennelly SP, Onuchic JN, Sanbonmatsu KY. 2009. Nonlocal helix formation is key to understanding s-adenosylmethionine-1 riboswitch function. *Biophy J* **96**:L7–L9.
Zgarbov[á]{} M, Otyepka M, Šponer J, Ml[á]{}dek A, Ban[á]{}š P, Cheatham III TE, Jurečka P. 2011. Refinement of the [C]{}ornell et al. nucleic acids force field based on reference quantum chemical calculations of glycosidic torsion profiles. *J Chem Theory Comput* **7**:2886–2902.
System Total atoms Water molecules aptamer atoms Ions (Na+Cl)
------------------------ ------------- ----------------- --------------- ---------------
Holo form 41628 13078 2257 122 (96+26)
Apo form 41676 13099 2257 122 (96+26)
Holo (P1-SMD) 120654 39364 2257 290 (180+110)
Apo (P1-SMD) 120588 39347 2257 290 (180+110)
Holo $\Delta$1-8/64-71 41101 13078 1746 106 (80+26)
Apo $\Delta$1-8/64-71 41149 13099 1746 106 (80+26)
: Specifications for the simulated systems
\[tabS1\]
![ Adenine riboswitch aptamer and binding site. A) Secondary structure elements and B) 3-dimensional structure with bound adenine. The P1 stem is grey, the other stems and loops are black. C) Cartoon representation of the binding site; the two dotted lines represent the hydrogen bonds of the WC pairing between the U62 and the ligand. []{data-label="fig1"}](Figure1){width="8cm"}
![ Secondary structure representation of the *add* riboswitch in the ON (A) and OFF (B) states. The ligand, the initiation codon and the Shine-Dalgarno sequence are labeled.[]{data-label="fig3new"}](Figure2){width="8cm"}
{width="8cm"}
{width="8cm"}
{width="8cm"}
{width="8cm"}
{width="8cm"}
{width="8cm"}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, we define a certain *proportional volume property* for an unit vector field on a spherical domain in $\mathbb{S}^{3}$. We prove that the volume of these vector fields has an absolute minimum and this value is equal to the volume of the Hopf vector field. Some examples of such vector fields are given. We also study the minimum energy of solenoidal vector fields which coincides with a Hopf flow along the boundary of a spherical domain of $\mathbb{S}^{2k+1}$.'
author:
- |
Fabiano G. Brito\
Centro de Matemática, Computação e Cognição\
Universidade Federal do ABC\
09.210-170 Santo André, Brazil\
E-mail: [email protected]
- |
André O. Gomes\
Departamento de Matemática\
Instituto de Matemática e Estatística da USP\
Universidade de São Paulo\
05508-090 São Paulo, Brazil\
E-mail: [email protected]
- |
Robson M. Mesquita\
Campus Universitário de Arraias\
Universidade Federal do Tocantins\
77330-000 Tocantins, Brazil\
E-mail: [email protected]
title: 'A theorem about vector fields with the “proportional volume property"'
---
[^1]
[^2]
Introduction
============
The *volume* of an unit vector field $X$ on a compact and oriented Riemannian $n$-manifold $K$ can be defined as the volume of the submanifold $X(K)$ of the unit tangent bundle equipped with the restriction of the Sasaki metric. It is given by $$\begin{aligned}
\mathrm{vol}(X)=\int_{K}\sqrt{\mathrm{det}(Id+(\nabla X)^{T}\nabla X)}d\nu\end{aligned}$$ where $d\nu$ is the volume element determined by the metric and $\nabla$ is the Levi-Civita connection. On the other hand, the *energy* of an unit vector field $X$ defined on $K$ is given by $$\begin{aligned}
\mathcal{E}(X)=\frac{n}{2}\mathrm{vol}(K)+\frac{1}{2}\int_{K}\left\|\nabla X\right\|^{2}d\nu\end{aligned}$$ The following theorem is well known
$[\mathrm{GZ}]$ The unit vector fields of minimum volume on $\mathbb{S}^{3}$ are precisely the Hopf vector fields, and no others.
There is an analogue of Theorem 1.1 for the energy functional
$[\mathrm{B}]$ Hopf vector fields on $\mathbb{S}^{3}$ minimize the functional $\mathcal{E}$ and Hopf vector fields are the unique unit vector fields on $\mathbb{S}^{3}$ to minimize $\mathcal{E}$.
The case where K is a submanifold with boundary of $\mathbb{S}^{2k+1}$ was treated first in \[BGN\] and the following general “boundary version" was obtained
$[\mathrm{BGN}]$ Let $U$ be an open set of the (2k+1)-dimensional unit sphere $\mathbb{S}^{2k+1}$, let $K\subset U$ be a connected (2k+1)-submanifold with boundary of the sphere $\mathbb{S}^{2k+1}$ and let $\vec{v}$ be an unit vector field on $U$ which coincides with a Hopf flow H along the boundary of $K$. Then $$\begin{aligned}
\mathrm{vol}(\vec{v})\!\!\!&\geq&\!\!\!\frac{4^{k}}{{2k \choose k}}\mathrm{vol}(K)
\\
\\
\mathcal{E}(\vec{v})\!\!\!&\geq&\!\!\!(\frac{2k+1}{2}+\frac{k}{2k-1})\mathrm{vol}(K)\end{aligned}$$
If $k=1$ we obtain that $\mathrm{vol}(\vec{v})\geq 2\mathrm{vol}(K)$, the volume of the Hopf vector field. In the proof of Theorem 1.2 was used the following important map $\varphi_{t}^{\vec{v}}:\mathbb{S}^{2k+1}\longrightarrow \mathbb{S}^{2k+1}(\sqrt{1+t^{2}})$ given by $\varphi_{t}^{\vec{v}}(x)=x+t\vec{v}(x)$ which was first used in the Milnor’s paper \[M\].\
\
In this work we continue to study the volume and energy of vector fields defined on submanifolds with boundary $K\subset \mathbb{S}^{3}$. In order to find similar results, we say that an unit vector field $\vec{v}$ satisfies the “proportional volume property" if the following inequality hold for some $t>0$ $$\begin{aligned}
\frac{\mathrm{vol}(\varphi_{t}^{\vec{v}}(K))}{\mathrm{vol}(\mathbb{S}^{3}(\sqrt{1+t^{2}}))}\geq\frac{\mathrm{vol}(K)}{\mathrm{vol}(\mathbb{S}^{3})}\end{aligned}$$ This condition is obviously equivalent to $$\begin{aligned}
\mathrm{vol}(\varphi_{t}^{\vec{v}}(K))\geq\mathrm{vol}(K)(1+t^{2})^{3/2}\end{aligned}$$ We will show examples of vector fields satisfying the proportional volume property. Our goal is also to prove the following theorem
Let $K$ be a connected $3$-submanifold with boundary of the unit sphere $\mathbb{S}^{3}$ and let $\vec{v}$ be an unit vector field defined on $\mathbb{S}^{3}$. Consider the map $\varphi_{t}^{\vec{v}}:\mathbb{S}^{3}\longrightarrow \mathbb{S}^{3}(\sqrt{1+t^{2}})$ given by $\varphi_{t}^{\vec{v}}(x)=x+t\vec{v}(x)$ and a real number $t>0$ small enough so that the map $\varphi_{t}^{\vec{v}}$ is a diffeomorphism. Suppose also that the following conditions hold:
1. $\displaystyle{\int_{\partial K}\left\langle \vec{v},\eta\right\rangle=0}$, where $\eta$ is the conormal vector to the boundary $\partial K$.
2. $\displaystyle{\frac{\mathrm{vol}(\varphi_{t}^{\vec{v}}(K))}{\mathrm{vol}(\mathbb{S}^{3}(\sqrt{1+t^{2}}))}\geq\frac{\mathrm{vol}(K)}{\mathrm{vol}(\mathbb{S}^{3})}}$.
Then $\mathrm{vol}(\vec{v}) \geq \mathrm{vol}(H)$ and $\mathcal{E}(\vec{v})\geq \mathcal{E}(H)$, where $H$ is the Hopf flow.
Clearly, if the vector field $\vec{v}$ is solenoidal, the condition *1* of Theorem 1.4 is satisfied. As a corollary of Proposition 1 in \[BS\], there is a lower limit for the energy of solenoidal fields defined on odd-dimensional Euclidean spheres
$[\mathrm{BS}]$ The Hopf vector fields has minimum energy among all solenoidal unit vector fields on the sphere $\mathbb{S}^{2k+1}$.
Using the same techniques of \[BGN\] we also prove the following boundary version of Theorem 1.5
Let $U$ be an open set of the $(2k+1)$-dimensional unit sphere $\mathbb{S}^{2k+1}$ and let $K\subset U$ be a connected $(2k+1)$-submanifold with boundary of the sphere $\mathbb{S}^{2k+1}$. Let $\vec{v}$ be a solenoidal unit vector field on $U$ which coincides with a Hopf flow $H$ along the boundary of K. Then $$\begin{aligned}
\mathcal{E}(\vec{v})\geq \left(\frac{2k+1}{2}+k\right)\mathrm{vol}(K)= \mathcal{E}(H)\end{aligned}$$
Preliminaries
=============
Using an orthonormal local frame $\left\{e_{1},\ldots, e_{n-1}, e_{n}=\vec{v}\right\}$ of $K$, the volume of the unit vector field $\vec{v}$ is given by
$$\begin{aligned}
\mathrm{vol}(\vec{v})=\int_{K} (1+\sum\limits_{a=1}^{n}\left\|\nabla_{e_{a}}\vec{v}\right\|^{2}+\sum\limits_{a<b}\left\|\nabla_{e_{a}}\vec{v}\wedge\nabla_{e_{b}}\vec{v}\right\|^{2}+\ldots
\\
\ldots+\sum\limits_{a_{1}<\cdots<a_{n-1}}\left\|\nabla_{e_{a_{1}}}\vec{v}\wedge\cdots\wedge\nabla_{e_{a_{n-1}}}\vec{v}\right\|^{2})^{1/2}\: d\nu\end{aligned}$$
Now, let $U\subset\mathbb{S}^{2k+1}$ be an open set of the unit sphere and let $K\subset U$ be a connected $(2k+1)$-submanifold with boundary of $\mathbb{S}^{2k+1}$. Suppose that $\vec{v}$ is an unit vector field defined on $U$. We also consider the map $\varphi_{t}^{\vec{v}}:U\longrightarrow \mathbb{S}^{2k+1}(\sqrt{1+t^{2}})$ given by $\varphi_{t}^{\vec{v}}(x)=x+t\vec{v}(x)$. We assume that $t>0$ is small enough so that the map $\varphi_{t}^{\vec{v}}$ is a diffeomorphism.\
\
In \[BGN\], the Authors showed that the determinant of the Jacobian matrix of $\varphi_{t}^{\vec{v}}$ can be express in the form $$\begin{aligned}
\det(d\varphi_{t}^{\vec{v}})=\sqrt{1+t^{2}}(1+\sum\limits_{i=1}^{2k}\sigma_{i}(\vec{v})t^{i})\end{aligned}$$ where, by definition, $h_{ij}(\vec{v}):=\left\langle \nabla_{e_{i}}\vec{v},e_{j}\right\rangle$ (with $i,j\in\left\{1,\ldots,2k\right\}$) and the functions $\sigma_{i}$ are the $i$-symmetric functions of the $h_{ij}$. For instance, if $k=1$ $$\begin{aligned}
\sigma_{1}(\vec{v})\!\!\!&=&\!\!\!h_{11}(\vec{v})+h_{22}(\vec{v})\\ \sigma_{2}(\vec{v})\!\!\!&=&\!\!\!h_{11}(\vec{v})h_{22}(\vec{v})-h_{12}(\vec{v})h_{21}(\vec{v})\end{aligned}$$ and the volume $\mathrm{vol}(\vec{v})$ has the following form $$\begin{aligned}
\mathrm{vol}(\vec{v})=\int_{K}\left(\sqrt{1+\sum\limits_{i,j=1}^{2}h_{ij}^{2}+(\mathrm{det}(h_{ij}))^{2}+\cdots}\right)\: d\nu\end{aligned}$$ For an arbitrary integer $k \geq 1$ the energy $\mathcal{E}(\vec{v})$ has the following form $$\begin{aligned}
\mathcal{E}(\vec{v})=\frac{2k+1}{2}\mathrm{vol}(K)+\frac{1}{2}\int_{K}[\sum\limits_{i,j=1}^{2k}(h_{ij}(\vec{v}))^{2}+\sum\limits_{i=1}^{2k}(\left\langle\nabla_{\vec{v}}\vec{v},e_{i}\right\rangle)^{2}]\end{aligned}$$ and, by change of variables theorem, we have the following expression for the volume of the spherical domain $\varphi_{t}^{\vec{v}}(K)\subset \mathbb{S}^{3}(\sqrt{1+t^{2}})$ $$\begin{aligned}
\mathrm{vol}[\varphi_{t}^{\vec{v}}(K)]=\int_{K}\sqrt{1+t^{2}}(1+\sum\limits_{i=1}^{2k}\sigma_{i}(\vec{v})t^{i})\end{aligned}$$
Proof of theorem 1.4
====================
By hypothesis *2*) and by equality (2.3), we have $$\begin{aligned}
\mathrm{vol}[\varphi_{t}^{\vec{v}}(K)]=\int_{K}\sqrt{1+t^{2}}(1+\sigma_{1}(\vec{v})t+\sigma_{2}(\vec{v})t^{2})\geq \int_{K}(\sqrt{1+t^{2}})^{3}\end{aligned}$$ On the other hand, by hypothesis *1*) and by divergence theorem we have $\displaystyle{\int_{K}\sigma_{1}(\vec{v})=0}$. Then (3.1) has the form $$\begin{aligned}
\int_{K}(1+\sigma_{2}(\vec{v})t^{2})\geq \int_{K}(1+t^{2})\end{aligned}$$ and then $$\begin{aligned}
\int_{K}\sigma_{2}(\vec{v})\geq\mathrm{vol}(K)\end{aligned}$$ Now, observing that $$\begin{aligned}
\mathrm{vol}(H)=2\mathrm{vol}(K) \ \ \ \ \& \ \ \ \ \sum\limits_{i,j=1}^{2}h_{ij}^{2}\geq 2\sigma_{2}(\vec{v})\end{aligned}$$ we have by equation (2.1) that $$\begin{aligned}
\mathrm{vol}(\vec{v})&=\int_{K}\sqrt{1+\sum h_{ij}^{2}+(\mathrm{det}(h_{ij}))^{2}+\ldots}\\
&\geq \int_{K}\sqrt{1+2\sigma_{2}(\vec{v})+\sigma_{2}(\vec{v})^{2}}\\
&\geq \int_{K}(1+\sigma_{2}(\vec{v}))\geq 2\mathrm{vol}(K)=\mathrm{vol}(H)\end{aligned}$$ In a similar way, using the equation (2.2) for $k=1$ we can write $$\begin{aligned}
\mathcal{E}(\vec{v})\geq \frac{3}{2}\mathrm{vol}(K)+\frac{1}{2}\int_{K}\sum\limits_{i,j=1}^{2}(h_{ij}(\vec{v}))^{2}\end{aligned}$$ and then $$\begin{aligned}
\mathcal{E}(\vec{v})\geq\frac{3}{2}\mathrm{vol}(K)+\int_{K}\!\!\sigma_{2}(\vec{v})\geq \frac{3}{2}\mathrm{vol}(K)+\mathrm{vol}(K)=\mathcal{E}(H)\end{aligned}$$
Proof of theorem 1.6
====================
Once again, as a consequence of equation (2.2) we have $$\begin{aligned}
\mathcal{E}(\vec{v})\geq \frac{2k+1}{2}\mathrm{vol}(K)+\frac{1}{2}\int_{K}\sum\limits_{i,j=1}^{2k}(h_{ij}(\vec{v}))^{2}\end{aligned}$$ Now observe that $$\begin{aligned}
\sum\limits_{i<j}(h_{ii}-h_{jj})^{2}=(2k-1)\sum\limits_{i}h_{ii}^{2}-2\sum\limits_{i<j}h_{ii}h_{jj}\end{aligned}$$ and as $\vec{v}$ is a solenoidal vector field $$\begin{aligned}
0=[\sigma_{1}(\vec{v})]^{2}=(\sum\limits_{i}h_{ii})^{2}=\sum\limits_{i}h_{ii}^{2}+2\sum\limits_{i<j}h_{ii}h_{jj}\end{aligned}$$ in other words $$\begin{aligned}
-2\sum\limits_{i<j}h_{ii}h_{jj}=\sum\limits_{i}h_{ii}^{2}\end{aligned}$$ Substituting equation (4.4) in (4.2) we obtain $$\begin{aligned}
\sum\limits_{i<j}(h_{ii}-h_{jj})^{2}=-4k\sum\limits_{i<j}h_{ii}h_{jj}\end{aligned}$$ Further, we also have the following equation $$\begin{aligned}
\sum\limits_{i<j}(h_{ij}+h_{ji})^{2}=\sum\limits_{i\neq j}h_{ij}^{2}+2\sum\limits_{i<j}h_{ij}h_{ji}\end{aligned}$$ and then $$\begin{aligned}
2k\sum\limits_{i<j}(h_{ij}+h_{ji})^{2}=2k\sum\limits_{i\neq j}h_{ij}^{2}+4k\sum\limits_{i<j}h_{ij}h_{ji}\end{aligned}$$ Adding equations (4.5) and (4.7), we have $$\begin{aligned}
\sum\limits_{i\neq j}h_{ij}^{2}\geq 2\sigma_{2}\end{aligned}$$ and $$\begin{aligned}
\sum\limits_{i,j=1}^{2k}h_{ij}^{2}=\sum\limits_{i}h_{ii}^{2}+\sum\limits_{i\neq j}h_{ij}^{2}\geq 2\sigma_{2}\end{aligned}$$ Using the inequalities (4.1) and (4.9), we find $$\begin{aligned}
\mathcal{E}(\vec{v})&\geq&\!\!\! \frac{2k+1}{2}\mathrm{vol}(K)+\int_{K}\sigma_{2}(\vec{v})\end{aligned}$$ On the other hand, by a similar argument used in \[BGN\] we can say that $$\begin{aligned}
\int_{K}\sigma_{2}(\vec{v})=k\mathrm{vol}(K)\end{aligned}$$ and then we have $$\begin{aligned}
\mathcal{E}(\vec{v})\geq \frac{2k+1}{2}\mathrm{vol}(K)+k\mathrm{vol}(K)=\left(\frac{2k+1}{2}+k\right)\mathrm{vol}(K)\end{aligned}$$
Examples
========
Let $K\subset \mathbb{S}^{3}$ be a solid torus whose boundary is the Clifford torus $T$. We can take the following parametrization for $K$ $$\begin{aligned}
x(\theta, \alpha, \delta)=\left(\delta\cos(\theta),\delta\sin(\theta),\sqrt{1-\delta^{2}}\cos(\alpha),\sqrt{1-\delta^{2}}\sin(\alpha)\right)
\end{aligned}$$ with $0\leq \theta,\alpha \leq 2\pi$ and $0\leq \delta \leq \frac{1}{\sqrt{2}}$.
Consider an unit vector field $\vec{v}$ defined on $\mathbb{S}^{3}$ and tangent to the $\partial K$. For example, an element of the family of solenoidal unit vector fields $$\begin{aligned}
\vec{v}_{\lambda}(x,y,z,w)=\frac{1}{\sqrt{1+(\lambda^{2} -1)(x^{2}+y^{2})}}\left(-\lambda y,\lambda x,-w,z\right)\end{aligned}$$ where $\lambda >1$. Then it satisfies the first condition of Theorem 1.4. For the second condition, we have two possibilities: $$\begin{aligned}
\frac{\mathrm{vol}(\varphi_{t}^{\vec{v}}(K))}{\mathrm{vol}(\mathbb{S}^{3}(\sqrt{1+t^{2}}))}\geq\frac{\mathrm{vol}(K)}{\mathrm{vol}(\mathbb{S}^{3})}\end{aligned}$$ or $$\begin{aligned}
\frac{\mathrm{vol}(\varphi_{t}^{\vec{v}}(K))}{\mathrm{vol}(\mathbb{S}^{3}(\sqrt{1+t^{2}}))}<\frac{\mathrm{vol}(K)}{\mathrm{vol}(\mathbb{S}^{3})}\end{aligned}$$ In the first alternative, we find the example. In the second alternative, we consider the complementar of $K$, that is, the solid torus $K^{c}\subset\mathbb{S}^{3}$ with boundary $T$ such that $K \cup K^{c}=\mathbb{S}^{3}$ and $K\cap K^{c}=T$.\
\
We claim that: $\displaystyle{\frac{\mathrm{vol}(\varphi_{t}^{\vec{v}}(K^{c}))}{\mathrm{vol}(\mathbb{S}^{3}(\sqrt{1+t^{2}}))}\geq\frac{\mathrm{vol}(K^{c})}{\mathrm{vol}(\mathbb{S}^{3})}}$.\
\
In fact, if $\displaystyle{\frac{\mathrm{vol}(\varphi_{t}^{\vec{v}}(K))}{\mathrm{vol}(\mathbb{S}^{3}(\sqrt{1+t^{2}}))}<\frac{\mathrm{vol}(K)}{\mathrm{vol}(\mathbb{S}^{3})}}$ and $\displaystyle{\frac{\mathrm{vol}(\varphi_{t}^{\vec{v}}(K^{c}))}{\mathrm{vol}(\mathbb{S}^{3}(\sqrt{1+t^{2}}))}<\frac{\mathrm{vol}(K^{c})}{\mathrm{vol}(\mathbb{S}^{3})}}$\
\
then, adding the two inequalities we arrive at a contradiction $$\begin{aligned}
1=\frac{\mathrm{vol}(\varphi_{t}^{\vec{v}}(K))+\mathrm{vol}(\varphi_{t}^{\vec{v}}(K^{c}))}{\mathrm{vol}(\mathbb{S}^{3}(\sqrt{1+t^{2}}))}<\frac{\mathrm{vol}(K)+\mathrm{vol}(K^{c})}{\mathrm{vol}(\mathbb{S}^{3})}=1\end{aligned}$$ Therefore we find an example after substituting $K$ for $K^{c}$.
Dedication {#dedication .unnumbered}
----------
Dedicated to Professor Antonio Gervasio Colares on his 80th birthday.
[HD]{}
F. G. B. Brito, *Total bending of flows with mean curvature correction*, Diff. Geom. Appl. [12]{} (2000), 157–163.
F. Brito, A. Gomes and G. Nunes, *Energy and volume of vector fields on spherical domains*, Pacific Journal of Math. [257]{} (2012), no. 1, 1–7.
F. Brito, M. Salvai, *Solenoidal unit vector fields with minimum energy*, Osaka J. Math. [41]{} (2004), 533–544.
H. Gluck and W. Ziller, *On the volume of an unit vector field on the three-sphere*, Comment. Math. Helv. [61]{} (1986), 177–192.
J. Milnor, *Analytic proofs of the “hairy ball theorem" and the Brouwer fixed-point theorem*, Amer. Math. Monthly 85, (1978), no. 7, 521–524.
[^1]: 2010 *Mathematics Subject Classification*: Primary .
[^2]: *Key words and phrases*: volume of vector fields
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present the calculation of the cross section for Higgs boson production in association with a top quark pair plus one jet, at next-to-leading-order (NLO) accuracy in QCD. All mass dependence is retained without recurring to any approximation. After including the complete NLO QCD corrections, we observe a strong reduction in the scale dependence of the result. We also show distributions for the invariant mass of the top quark pair, with and without the additional jet, and for the transverse momentum and the pseudorapidity of the Higgs boson. Results for the virtual contributions are obtained with a novel reduction approach based on integrand decomposition via Laurent expansion, as implemented in the library . Cross sections and differential distributions are obtained with an automated setup which combines the and frameworks.'
author:
- 'H. van Deurzen'
- 'G. Luisoni'
- 'P. Mastrolia'
- 'E. Mirabella'
- 'G. Ossola'
- 'T. Peraro'
bibliography:
- 'references.bib'
title: |
NLO QCD corrections to Higgs boson production in association\
with a top quark pair and a jet
---
Introduction {#Sec:intro}
============
The evidence of the existence of a new particle of mass between 125 and 126 GeV, initially reported about one year ago by the ATLAS and CMS collaborations [@Aad:2012tfa; @Chatrchyan:2012ufa], has been confirmed with very high confidence level by more recent analyses, thus providing more stringent arguments in favor of the validity of the electroweak symmetry breaking mechanism. It is interesting to observe that all the analyses performed so far are in good agreement with the hypothesis that the new particle is the Higgs boson predicted by the Standard Model (SM). Indeed, rates and distributions are compatible with the assumption that the new particle is a scalar that couples to other SM particles with a strength proportional to their mass [@CMS:yva; @Aad:2013wqa; @Aaltonen:2013kxa]. Accurate predictions are necessary and will play a crucial role for the complete determination of the nature of the Higgs boson [@Heinemeyer:2013tqa], in particular to shed light on the structure of its couplings to the other particles.
The production rate for a Higgs boson associated with a top-antitop pair ($t \bar t H$) is particularly interesting in this context, since it is directly proportional to the SM Yukawa coupling of the Higgs boson to the top quark. The study of differential observables and distributions will bring information on the coupling structure and on the parity of the Higgs particle [@Frederix:2011zi; @Degrande:2012gr].
The difficulties related to the analysis of the $t \bar t H$ channel are well known. The combined production of three heavy particles requires a large center-of-mass energy for the initial partons, which is strongly suppressed by parton distribution functions. Furthermore, additional difficulties are represented by the presence of various challenging backgrounds and by the complexity of the final state, which make its kinematic reconstruction far from straightforward [@Artoisenet:2013vfa].
![Sample of one-loop diagrams contributing to the NLO corrections to $g g \to t {\bar t} H g $ and $q {\bar q} \to t {\bar t} H g$.[]{data-label="hexagons"}](diagrams5.eps){width="6cm"}
At the parton level, the $t \bar t H$ production at next-to-leading order (NLO) in QCD has been known for some time [@Beenakker:2001rj; @Beenakker:2002nc; @Dawson:2002tg; @Dawson:2003zu; @Dittmaier:2003ej]. More recently, this process has been employed in a number of studies, motivated by the new analyses performed at the LHC [@Plehn:2009rk; @Frederix:2011zi; @Degrande:2012gr; @Artoisenet:2013vfa].
In this letter, we present the complete NLO QCD corrections to the process $ pp \to t \bar t H + 1$ jet ($t \bar t H j$) at the LHC. Examples of contributing one-loop diagrams are depicted in Fig. \[hexagons\]. We illustrate the outcome of our calculation by showing the total cross section, and a selection of differential distributions.
The goal of the considered calculation is twofold. On the one hand, it is important for the phenomenological analyses at the LHC, in particular for the high-$p_T$ region, where the presence of the additional jet can be sensibly relevant. On the other hand, $t \bar t H j$ constitutes the first application of a novel reduction algorithm for the evaluation of one-loop amplitudes, which strengthens the performances of the integrand decomposition [@Mastrolia:2012bu], in particular in the presence of massive particles.
Computational set-up {#Sec:calc}
====================
In perturbation theory, computations at the NLO accuracy require, aside from the evaluation of leading-order (LO) contributions, the calculation of both virtual and real-emission corrections. The Born and the real emission matrix elements are computed using [@Gleisberg:2008ta] and the library [@Krauss:2001iv], which implements the Catani-Seymour dipole formalism [@Catani:2002hc; @Gleisberg:2007md]. also performs the integration over the phase space and the analysis. The virtual corrections are generated with the package [@Cullen:2011ac], which combines automated diagram generation and algebraic manipulation [@Nogueira:1991ex; @Vermaseren:2000nd; @Reiter:2009ts; @Cullen:2010jv; @Kuipers:2012rf] with $d$-dimensional integrand-level reduction techniques [@Ossola:2006us; @Ossola:2007bb; @Ellis:2007br; @Ossola:2008xq; @Mastrolia:2008jb; @Mastrolia:2010nb; @Heinrich:2010ax]. The master integrals (MIs) are computed using [@vanHameren:2010cp]. The code generated by is linked to by means of the Binoth Les Houches Accord (BLHA) [@Binoth:2010xt] interface, which uses a system of [*order*]{} and [*contract*]{} files and allows for a direct communication between the two codes at running time. The same setup has been recently employed for the computation of NLO QCD corrections to $p p \to H j j$ [@vanDeurzen:2013rv] and $p p \to H j j j $ [@Cullen:2013saa] (for the latter, in combination with [MadDipole/MadEvent]{} [@Maltoni:2002qb; @Frederix:2008hu; @Frederix:2010cj]) and also for the analysis of the $t \bar t$ forward-backward asymmetry [@Hoeche:2013mua].
For $t \bar t H j$ production, the basic partonic processes identified by the - contract file are: q |q t |t H g , g g t |t H g , \[Eq:Pproc\] while the remaining subprocesses can be obtained by proper crossings. The ultraviolet, the infrared, and the collinear singularities are regularized using dimensional reduction. The renormalization conditions are fixed along the lines of [@Beenakker:2002nc; @Dawson:2003zu], where the top mass is renormalized on-shell, while the strong coupling is renormalized in the $\overline{\mbox{MS}}$ scheme, decoupling the top quark from the running. In the case of LO \[NLO\] contributions, we describe the running of the strong coupling constant with one-loop \[two-loop\] accuracy. The wave functions of the gluon and of the quarks are renormalized on-shell, [*i.e.*]{} the corresponding renormalization constants cancel the external self-energy corrections exactly.
The virtual amplitudes of $t {\bar t}Hj$ have been decomposed in terms of MIs using for the first time the [*integrand reduction via Laurent expansion*]{} [@Mastrolia:2012bu], implemented in the library . This new algorithm exploits the complete knowledge of the analytic expression of the integrand and of the residues at the multiple cut to ameliorate the determination of the coefficients of the MIs with respect to the canonical integrand reduction [@Ossola:2006us]. Elaborating on the techniques introduced in [@Forde:2007mi; @Kilgore:2007qr; @Badger:2008cm], the series expansion combined with the integrand decomposition lowers the computational load and improves the accuracy of the results. Within this new algorithm, the sampling of the numerators and the subtractions of the higher-point residues, characterizing the [*triangular*]{} system-solving approach of the original integrand-reduction procedure, are avoided. Instead, the series expansion allows for a [*diagonal*]{} system-solving strategy, where the polynomial subtractions of the residues, when needed, are replaced by universal correction terms which have to be added to the coefficients of the Laurent series. These universal corrections, required only for the determination of the coefficients of 2-point and 1-point MIs, are obtained, once and for all, from the expansions of the generic polynomial forms of the residues at the triple and double cuts.
![Scale dependence of the total cross section at LO and NLO.[]{data-label="Fig:scalevars"}](scalevars.eps){width="8cm"}
The library, which has been interfaced to , implements the integrand reduction via Laurent expansion using a semi-analytic algorithm. The coefficients of the Laurent expansion of a generic integrand are efficiently computed by performing a [*polynomial division*]{} between the numerator and the set of uncut denominators [@Mastrolia:2012bu].
The calculation of the NLO virtual corrections performed with has been checked using the independent reduction algorithm implemented in the library [@Mastrolia:2010nb]. We verified the agreement of the virtual corrections obtained with the two reduction procedures in ten thousand phase-space points. The values of double and the single poles, for each individual subprocess, conform to the universal singular behavior of dimensionally regulated one-loop amplitudes [@Catani:2000ef]. Our results fulfill gauge invariance, verified through the vanishing of the amplitudes when substituting the polarization vector of one or more gluons with the corresponding momentum.
The reduction algorithm proved to be numerically more efficient and stable. In fact, for the highly non-trivial process under consideration, only a small set of phase-space points, of the order of few per mill, were detected as unstable. All these points have been recovered using the tensorial reduction provided by [Golem95]{} [@Binoth:2008uq; @Cullen:2011kv], thus avoiding the necessity of higher precision routines, which are extremely time consuming.
![Invariant mass distributions of the $t\bar{t}$-pairs for $t \bar t H$ and $t \bar t H j$ at NLO relative to the $t \bar t H j$ at LO for $\mu=2\times\GAT$.[]{data-label="Fig:inv1"}](histo_Massttb.eps){width="7.5cm"}
The time required for the computation of the full color- and helicity-summed amplitudes in one phase-space point is about $2.5$ seconds. The numerical values of the one-loop amplitudes for the two partonic processes listed in Eq. (\[Eq:Pproc\]) in a non-exceptional phase-space point are collected in the Appendix.
In view of the later comparison between the processes $pp \to t {\bar t} H$ and $pp \to t {\bar t} Hj$ at NLO QCD accuracy, we also used the [/]{}+ framework to compute the cross section for $t {\bar t} H$ production. We found excellent agreement with the results presented in Refs. [@Hirschi:2011pa; @Frederix:2011zi].
Numerical results {#Sec:resu}
=================
In the following, we present results for the integrated cross section for a center-of-mass energy of $8$ TeV. The mass of the Higgs boson is set to $m_H=126$ GeV and the top quark mass is set to $m_t=172.5$ GeV. The parameters of the electroweak sector are fixed by setting $M_W=80.419$ GeV, $M_Z=91.1876$ GeV and $\alpha_{EW}^{-1}=132.50698$.
To cluster the jets we use the [antikt]{}-algorithm implemented in [FastJet]{} [@Cacciari:2005hq; @Cacciari:2008gp; @Cacciari:2011ma] with radius $R=0.5$, a minimum transverse momentum of $p_{T,jet}>15$ GeV and pseudorapidity $|\eta|<4.0$. The LO cross sections are computed with the LO parton-distribution functions cteq6L1 [@Pumplin:2002vw], whereas at NLO we use CT10 [@Lai:2010vv].
In order to study the scale dependence of the total cross section, we employ two different choices of the renormalization and factorization scales $\mu_{R} = \mu_{F} = \mu_{0}$, namely $\mu_0 = H_T$ and $\mu_0 = 2\times\GAT$ with $$\begin{aligned}
H_T & = \sum_{
\tiny{
\begin{array}{cc} \mbox{\tiny final } \label{eq:ht} \\
\mbox{ \tiny states } f \end{array} }
}
\left | p_{T,f}\right | \; , \\
\GAT & =\sqrt[3]{m_{T,H}\, m_{T,t}\, m_{T,\bar t}} + \sum_{ \tiny{ \mbox{\tiny jets } j }} |p_{T,j}| \label{eq:ga} \, .\end{aligned}$$
![Transverse momentum distribution of the Higgs boson at LO and NLO for $\mu=H_T$.[]{data-label="Fig:pth"}](histo_PTh0.eps){width="7.5cm"}
![Pseudorapidity $\eta$ of the Higgs boson at LO and NLO accuracy for $\mu=H_T$.[]{data-label="Fig:etah"}](histo_Etah0.eps){width="7.5cm"}
Within this setup, for the two scale choices, we obtain the total LO and NLO cross sections reported in Table \[Tab:xs\].
[MMM]{} & $\sigma_{LO}$ \[fb\] & $\sigma_{NLO}$ \[fb\]\
\
\[-1.7ex\] 2$\times\GAT$ & $80.03^{+35.64}_{-23.02}$ & $100.6^{+0.00}_{-9.43}$\
\
$H_T$ & $88.93^{+41.41}_{-26.13}$ & $102.3^{+0.00}_{-15.82}$\
\
\[-1.7ex\]
The scale dependence of the total cross section, depicted in Fig. \[Fig:scalevars\], is strongly reduced by the inclusion of the NLO contributions. It is worthwhile to notice that both choices for the central value of the scale provide an adequate description, being close to the physical scale of the process.
In Fig. \[Fig:inv1\], we compare the distributions for the invariant mass of the top quark pair in $pp \to t {\bar t} H j$ at LO and NLO with the NLO curve for $pp \to t {\bar t} H$. For $t {\bar t} H j$, going from LO to NLO accuracy, we observe an increase in the distribution by 20–35% over the full kinematical range. On the other hand, when comparing the NLO $t {\bar t} H$ prediction with the NLO $t {\bar t} H j$ curve, the cross section decreases due to the presence of the additional jet which takes away energy from the $t\bar{t}$ system. This is particularly evident near the $t\bar{t}$ production threshold, while for high values of the $t {\bar t}$ invariant mass the two NLO curves get closer. The scale for this comparison is set to $\mu=2\times\GAT$.
In Fig. \[Fig:pth\] and Fig. \[Fig:etah\], we display the distributions of the transverse momentum $p_T$ and the pseudorapidity $\eta$ of the Higgs boson, respectively. Each plot contains the distributions at LO and NLO accuracy, for a value of the scale set to $\mu=H_T$. The NLO corrections are particularly important for high values of the $p_T$, which are the kinematical regions involved in the boosted analyses [@Butterworth:2008iy; @Plehn:2009rk].
These distributions show the potential of the framework obtained combining [/]{} with , which can be successfully used to compute NLO predictions for multi-leg processes involving massive particles. Moreover they shed some light on the impact of further jet activity in $p p \to t \bar t H$, one of the most important processes for the direct determination of the coupling of the Higgs boson to fermions. The NLO QCD corrections reduce the scale uncertainty and their numerical impact can be sizable. Therefore they could be helpful for an accurate simulation of the signal in the experimental searches looking for Higgs production in association with a top-antitop pair at the LHC.
#### Acknowledgments – {#acknowledgments .unnumbered}
We thank all the other members of the project for collaboration on the common development of the code. We also thank Jan Winter for interesting discussions. The work of H.v.D., G.L., P.M., and T.P. was supported by the Alexander von Humboldt Foundation, in the framework of the Sofja Kovaleskaja Award Project “Advanced Mathematical Methods for Particle Physics”, endowed by the German Federal Ministry of Education and Research. G.O. was supported in part by the National Science Foundation under Grant PHY-1068550. Computing resources were provided by the CTP cluster of the New York City College of Technology.
Benchmark phase-space point {#App:Bench}
===========================
In this appendix we collect numerical results for the renormalized virtual contributions to the processes (\[Eq:Pproc\]), in correspondence to the phase-space point in Table \[pspoint\]. The results are collected in Table \[benchmark\] and are computed using dimensional reduction. The coefficients $a_i$ are defined as follows: + + a\_0 . \[Eq:AI\] The reconstruction of the renormalized pole can be checked against the value of $a_{-1}$ and $a_{-2}$ obtained by the universal singular behavior of the dimensionally regularized one-loop amplitudes [@Catani:2000ef], while the precision of the finite parts is estimated by re-evaluating the amplitudes for a set of momenta rotated by an arbitrary angle about the axis of collision.
particle $E$ $p_x$ $p_y$ $p_z$
---------- -------------------- --------------------- --------------------- ---------------------
$p_1$ 250.00000000000000 0.0000000000000000 0.0000000000000000 250.00000000000000
$p_2$ 250.00000000000000 0.0000000000000000 0.0000000000000000 -250.00000000000000
$p_3$ 177.22342332868467 -31.917865771774753 -19.543909461587205 -15.848571666570733
$p_4$ 174.89951284907735 13.440699620020803 24.174898117950033 -8.2771667589629576
$p_5$ 126.37478917634435 6.8355633672742222 -3.2652801590882752 6.0992096455298030
$p_6$ 21.502274645893632 11.641602784479652 -1.3657084972745175 18.026528780003872
[ l r r ]{}
------------------------------------------------------------------------
& $q \bar q \to t \bar t H g$ & $g g \to t \bar t H g$\
------------------------------------------------------------------------
$a_0$ & 8274311 & 3354295\
$a_{-1}$ & 180535 & 376358\
$a_{-2}$ & 3801 & 56112\
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We propose a neural reranking system for named entity recognition (NER). The basic idea is to leverage recurrent neural network models to learn sentence-level patterns that involve named entity mentions. In particular, given an output sentence produced by a baseline NER model, we replace all entity mentions, such as *Barack Obama*, into their entity types, such as *PER*. The resulting sentence patterns contain direct output information, yet is less sparse without specific named entities. For example, “PER was born in LOC” can be such a pattern. LSTM and CNN structures are utilised for learning deep representations of such sentences for reranking. Results show that our system can significantly improve the NER accuracies over two different baselines, giving the best reported results on a standard benchmark.'
author:
- Jie Yang
- Yue Zhang
- |
Fei Dong\
Singapore University of Technology and Design\
[{jie\_yang, fei\_dong}@mymail.sutd.edu.sg]{}\
[yue\[email protected]]{}\
bibliography:
- 'ranlp2017.bib'
title: Neural Reranking for Named Entity Recognition
---
Introduction
============
Shown in Figure \[fig:nerdemo\], named entity recognition aims to detect the entity mentions in a sentence and classify each entity mention into one out of a given set of categories. NER is typically solved as a sequence labeling problem. Traditional NER systems use Hidden Markov Models (HMM) [@zhou2002named] and Conditional Random Fields (CRF) [@lafferty2001conditional] with manually defined discrete features. External resources such as gazetteers and human defined complex global features are also incorporated to improve system performance [@ratinov2009design; @che2013named]. Recently, deep neural network models have shown the ability of learning more abstract features compared with traditional statistical models with indicator features for NER [@zhang2015neural].
Recurrent Neural Network (RNN), in particular Long Short-Term Memory (LSTM) [@hochreiter1997long], shows the ability to automatically capture history information over input sequences, which makes LSTM a proper automatic feature extractor for sequence labeling tasks. Different methods have been proposed by stacking CRF over LSTM in NER task [@chiu2015named; @huang2015bidirectional; @lample2016neural; @ma2016end]. In addition, it is possible to combine discrete and neural features for enriched information, which helps improve sequence labeling preformance [@zhang2016libn3l].
{width="6.0in"}
Reranking is a framework to improve system performance by utilizing more abstract features. A reranking system can take full advantage of global features, which are intractable in baseline sequence labelling systems that use exact decoding. The reranking method has been used in many NLP tasks, such as parsing [@collins2005discriminative], QAs [@chen2006reranking] and machine translation [@wang2007reranking; @shen2004discriminative].
Some work has adopted the reranking strategy for NER. tried both a boosting algorithm and a voted perceptron algorithm as reranking models on named-entity boundaries (without classification of entities). applied Support Vector Machine (SVM) with kernels to reranking model, obtaining significant improvements in F-measure on CoNLL 2003 datasets. used a simple log-linear rerank model on a biomedical NER task, also obtaining slight improvemts. All the above methods use sparse manual features. To the best of our knowledge, there has been no neural reranking model for NER task.
In this paper, we propose a simple neural reranking model for NER. The model learns sentence patterns that involve output named entities automatically, using neural network. Take the sentence ** as an example, Figure \[fig:collapse\] illustrates several candidate sentence patterns such as ** ($C_3$) and ** ($C_2$), where *PER* represents entity type *persons* and *LOC* means *locations*. It is obvious that $C_3$ is a much more reasonable sentence pattern compared to $C_2$. To generate the sentence patterns above, we replace predicted entities in candidate sequences with their entity type names. This can effectively reduce the sparsity of candidate sequences, as each entity type contains open vocabulary names (e.g. *PER* can be *Donald Trump*, *Hillary Clinton* etc.), which can bring noise when learning the sentence patterns. In addition, since the learned sentence patterns are global over output structures, it is difficult for baseline sequence labeling systems to capture such patterns.
We develop a neural reranking model which captures candidate pattern features using LSTM and auxilliary neural structures, including Convolution Neural Network (CNN) [@kim2014convolutional; @kalchbrenner2014convolutional] and character based neural features. The learned global sentence pattern representations are then used as features for scoring by the reranker. Results over a state-of-the-art discrete baseline using CRF and a state-of-the-art neural baseline using LSTM-CRF show significant improvements. On CoNLL 2003 test data, our model achieves the best reported result.
Our main contributions include (a) leveraging global sentence patterns that involve entity type information for NER raranking, (b) exploiting auxilliary neural features to enrich basic LSTM sequence representation and (c) achieving the best F1 result on CoNLL 2003 data. The source codes of this paper are released under GPL at <https://github.com/jiesutd/RerankNER>.
Baselines
=========
Formally, given a sentence $S$ with $t$ words: $S=\{w_1,w_2,...,w_t\}$, the task of NER is to find out all the named entity mentions from $S$. The dominate approach takes the task as a sequence labelling problem, where the goal is to generate a label sequence $L = \{l_1, l_2, ..., l_t\}$, where $l_i = p_i e_i$. Here $p_i$ is an entity label, $p_i \in \{B, I, O\}$, where $B$ indicates the beginning of an entity mention, $I$ denotes a non-beginning word of a named entity mention and $O$ denotes a non-named-entity word [^1]. $e_i$ indicates the entity type. In the CoNLL dataset that we use for our experiments, $e_i \in\{ \it PER, ORG, LOC, MISC\} $, where “$PER$” indicates a *person* name; “$LOC$”, “$ORG$”, “$MISC$” represent *location*, *organization* and *miscellaneous*, respectively.
We choose two baseline systems, one using dicrete CRF with handcrafted features and one using neural CRF model with bidirectional LSTM structure, both baselines giving the state-of-the-art accuracies among their respective category of models.
Discrete CRF
------------
We choose a basic discrete CRF model as our baseline tagger. As shown in Figure \[fig:discretecrf\], discrete word features are first extracted as binary vectors (black and white circles) and then fed into a CRF layer. Taking those discrete features as input, the CRF layer can give *n-best* predicted sequences as well as their probabilities. Table \[tab:features\] shows the discrete features that we used, which follow the definition of [@Jie2016combining]. Here *shape* means whether characters in word are belonging to number, English character or not. *capital* is the indication if word starts with *upper-case* English character, *connect words* include five types: “of”, “and”, “for”, “-” and other. Prefix and suffix include the 4-level prefixes and suffixes of each words.
Neural CRF
----------
A neural CRF with bidirectional LSTM structure is used as our second baseline, which is shown in Figure \[fig:neuralcrf\]. Word representations are represented with continious vectors (gray circles), which are fed into a bidirectional LSTM layer to extract neural features. A CRF layer with *n-best* output is stacked on top of the LSTM layer to decode the label sequences based on the neural features. We use the neural structure of @ma2016end, where the word representation is the concatenation of word embedding and a CNN output on the character sequence of the word.
Reranking Algorithms
====================
Collapsed Sentence Representation {#rule}
---------------------------------
Given the *n-best* output label sequences of a baseline system $\{L_1,L_2,...,L_i,...,L_n\}$, where $L_i = \{l_{i1},l_{i2},...,l_{it}\}$, we learn a reranking score $s(L_i)$ for $L_i$ by firsting converting $L_i$ into a sequence pattern $C_i$, and then learning a representation $h(C_i)$ as its dense representation. To convert candidate sequence $L_i$ to collapsed sequence $C_i$. We use the following rules to convert each label sequence $L_i$ into a collapsed sentence pattern $C_i$.
If the $L_i$ include entity labels (e.g. $l_{i1}$ =*B-PER*, $l_{i2}$=*I-PER*), then the entity labels are replaced with the corresponding entity type name (e.g. $\{l_{i1},l_{i2}\}$ $\rightarrow$ *PER*, $C_{i1}$ = *PER*), else labels are replaced by its corresponding words ($C_{ix} = w_x$). In the example shown in Figure \[fig:collapse\], $S=$ *{Barack Obama was born in hawaii .}* and $L_3=$ *{B-PER I-PER O O O B-LOC O}*. The corresponding collapsed sequence is $C_3=$ *{PER was born in LOC .}*, *Barack Obama* and *hawaii* are regarded as entities and hence are replaced by the entity names, i.e. *PER* and *LOC*, respectively.
Neural Features
---------------
Given a collapsed sentence representation $C_i$, we use neural network to learn its overall representation vector $h(C_i)$, which is used for the scoring of $C_i$.
***Word Representation:*** We use *SENNA* [@collobert2011natural] embedding to initialize the word embedding of our reranking system. For out of vocabulary words , embeddings are randomly initialized within $(-\sqrt{\frac{3.0}{wordDim}},\sqrt{\frac{3.0}{wordDim}})$, where $wordDim$ is the word dimension size [@ma2016end].
Character features are proved useful in capturing morphological features, such as word similarity and dealing with the out-of-vocabulary problem [@ling2015finding]. As shown in Figure \[fig:cnnchar\], we follow @ma2016end by utilizing CNN to extract character-level representation [^2]. Input character sequences are firstly passed through the embedding layer to lookup the character embeddings. To extract local features, a *convolution layer* with a fixed window-size is applied on top of the embedding layer. Then we use a *max-pooling* layer to map varying length vectors into a fixed size output vector. Finally, word representation is the concatenation of character CNN output vectors and word embeddings.
***LSTM features:*** We choose a word-based LSTM as the main network, using it for capturing global sentence pattern information. For input sequence vectors $\{x_1,x_2,...,x_t\}$, our LSTM model is implemented as follows: $$\begin{aligned}
h_t& \:=\:tanh(M_t)\odot o_t\\
i_t&\:=\:\sigma(W_1h_{t-1}+W_2x_t+\mu_1 \odot M_{t-1}+b_1)\\
f_t&\:=\:\sigma(W_3h_{t-1}+W_4x_t+\mu_2 \odot M_{t-1}+b_2)\\
\widetilde{M_i}&\:=\:tanh(W_5y_{t-1}+W_6x_i+b_3)\\
M_t&\:=\:i_i\odot\widetilde{M_i}+f_i\odot M_{t-1}\\
o_t&\:=\:\sigma(W_7h_{t-1}+W_8x_t +b_4),\\
\end{aligned}$$ where $\odot$ is the element-wise multiply operator, $\sigma$ is the sigmoid function, and $\{W,b,\mu\} \in \Theta$ are parameters. $i_t, f_t, M_t$ and $o_t$ are the *input gate, forget gate, memory cell* and *output gate*, respectively. $h_t$ is the hidden vector at step $t$ in the input sentence. As shown in Figure \[fig:lstmword\], word representations are the concatenation of word embeddings and character CNN output (red block). We choose the hidden vector in last word $h_{LSTM}$ as the representation of the input sequence.
***CNN features:*** We introduce CNN to capture local features of the candidate sequences. It consists of a *filter* $W \in R^{h\times k}$ which operates on a context of $k$ words to produce local order features. Max pooling layer is employed over the convolutional layer to extract the most salient features. Assume $u_j$ is the concatenation of word representations in Eq. (1) centralized in the embedding $z_j$ in a given sequence $u_1, u_2, ..., u_L$, CNN applies a matrix-vector operation to each window of size $k$ successive window along the sequence in Eq. (2). $$\begin{aligned}
u_j &= (z_{j-(k-1)/2}, ..., z_{j+(k-1)/2}) \\
r_i &= max_{1<j<L} {(W u_j + b)_i}, i=1, ..., d ,\end{aligned}$$ where $z_j$ is the $j$-th word embedding in the given sequence, $d$ is the output dimension of the CNN. $h=[r_1, ..., r_i,..., r_d]$ is the fixed-size feature representation for the sequence after pooling.
The *CNN* representation structure is similar to Figure \[fig:cnnchar\] but its input is word representations rather than character embeddings. We define the CNN features of word sequence as $h_{CNN}$.
Score Calculation
-----------------
After the LSTM and CNN features of collapsed sequence $C_i$ are extracted, we concatenate them together and feed the result into a *softmax* layer.
$$\begin{aligned}
h(C_i) & \:=\:h_{LSTM}\oplus h_{CNN}\\
s(C_i) & \:= \: \sigma(W h(C_i) + b) ,
\end{aligned}$$
where $\oplus$ represents the concatenating operation, $h(C_i)$ is the final representation of collapsed sequence $C_i$ and $s(C_i)$ is the output score of $C_i$.
Decoding
--------
We use a mixture reranking strategy during decoding. Denote the candidate label sequence set on sentence $S$ as $C(S) = \{C_1,C_2,...,C_n\}$ . We take advantage of both the reranker prediction score and the baseline tagger’s output probability, using the score $$\label{equ:mixture}
\hat{y_i} = \arg\max_{C_i\in C(S)} (\alpha s(C_i) + (1-\alpha)p(L_i)),$$ where $\alpha \in [0,1]$ is an interpolation weight, which is a hyperparameter tuned on the development set. $p(L_i)$ is the probability of label sequence $L_i$ in the baseline tagger.
Training
--------
For each training triplet $\{S,L_i, C_i\}$, given the golden sequence $L_{golden}$, we calculate the tag accuracy $y_i \in [0,1]$ of each candidate sequence based on $L_i$ and $L_{golden}$. The same decoding process is applied to each collapsed sequence $(C_i, y_i)$. We use a logistic regression model with mean square error (MSE) as the loss function, with a $l_2$-regulation term [^3] : $$J(\Theta) \:=\: \frac{1}{|\mathcal{D}|}\sum_{(C_i, y_i)\in \mathcal{D}}(y_i-s(C_i))^2 + \frac{\lambda}{2}||\Theta||_2^2$$ where $\Theta$ are all the parameters to be trained, $\mathcal{D}$ is the training set and $\lambda$ is the regulation factor.
*Adam* [@kingma2014adam] is used to update model parameters.
Experiments
===========
Settings
--------
We use *CRF++* [^4] as our discrete baseline CRF implementation and default parameters are used. For neural baseline, we follow the same structure and settings of the state-of-the-art system [@ma2016end]. When building the neural reranking system, *SENNA* embedding with 50 dimensions is used to initialize word embeddings. Hyperparameters of reranking system are listed in Table \[tab:hyperparameter\].
As we use the mixture strategy in Eq. (\[equ:mixture\]) during decoding, we search the ideal interpolation weight $\alpha$ within $[0,1]$ in a step of *0.005* based on the preformance under the development set.
Reranking Data
--------------
All of our experiments are evaluated on the standard CoNLL 2003 English dataset [@tjong2003introduction], which is a collection of Reuters newswire articles. The CoNLL 2003 English dataset includes 14,987 training sentences, 3,466 development sentences and 3,684 test sentences, annotated into 4 entity types, i.e. *persons(PER)*, *locations(LOC)*, *organizations(ORG)* and *miscellaneous(MISC)*.
To construct the reranking training data, we conduct five-fold , spliting the training set into 5 equal parts. In each case, the baseline tagger trains the model with 4/5 of the data and decode the remaining 1/5 to generate *n-best* candidate label sequences. For the reranking development and test data, the full training set is used to encode baseline tagger and decode development/test sentences with *n-best* output. All the *n-best* candidate sequences are converted into collapsed sequences following Section \[rule\].
![Oracle scores in baseline outputs.[]{data-label="fig:baseline"}](baselineanalysis.pdf){width="2.5in"}
Baseline Oracle Results
-----------------------
The discrete baseline achieves 92.13% of F1-measure in development set and 88.15% in test set. Our neural baseline gives 94.58% and 91.25% on development and test data, respectively. The discrete baseline for example. Figure \[fig:baseline\] shows different oracle scores varying with *n-best* in discrete baseline. The oracle best is obtained by always chooseing the best sequence in the *n-best* candidates, and *vice versa* for the oracle worst. The orcale best sentence accuracy (OBA) [^5] represents the accuracy of the sequence choice within the *n-best* candidates under oracle best assumption, and the orcale best F1-value (OBF) is the corresponding F1-value similarly. Orcale worst F1-value (OWF) is the F1-value under the worst choice situation.
As the figure shows, the larger $n$ is, the better is the OBA, which means that a potentially better reranking result is possible. On the other hand, the OWF also drops, which means that the reranking task is more difficult. In our experiments, *n-best* is set as 10, the orcale best F1-value of test set achieves 97.13% (+8.98%) while its orcale worst F1-value drops 49.07% to 39.08%.
![SSA with sentence length.[]{data-label="fig:sentAcc"}](sentAcc.pdf){width="3in"}
Influence of Sentence Length
----------------------------
We perform development experiments to evaluate model performance on various sentence lengths. Figure \[fig:sentAcc\] shows results by reranking the discrete baseline. Here sentence select accuracy (SSA) is calculated using the corrected number of sentences divided by the total number of sentences. The *x-axis* is the sentence length range (e.g. 10 means sentence length range from 5 to 10), while the *y-axis* corresponds to SSA within 10-best candidates before the mixture strategy (without mixing baseline output probability).
As shown in the Figure \[fig:sentAcc\], the accuracies of all model settings drop as the size of the sentence increases, which demonstrates that longer sentences are more challenging to our neural rerankers as they are to the baseline models. Both character information and CNN local features are useful for enhancing the SSA over a LSTM-only baseline. With the integration of character information and CNN features, our full model reranker can improve its performance on all sentence length ranges, especially for long sentences.
![F1-value comparision by entity types.[]{data-label="fig:entityAnalysis"}](entityAnalysis.pdf){width="3in"}
Influence of Entity Type
------------------------
Figure \[fig:entityAnalysis\] shows the comparision of models on different entity types. Compared with the baseline, entities with type of *PER* and *ORG* receive the most improvements, showing that sentence patterns are useful for those types. Table \[tab:finalresult\] shows the F1-value and SSA (after mixing baseline output probability) of our reranker on test data with different neural features on the discrete baseline. The word based LSTM reranker achieves the F1-value of 88.75%, with 0.6% absolute improvement over the baseline tagger. Cooperating with CNN features on word only does not make much improvement, while character CNN features are more effective (+0.78%). However, the full combination of character representation and word CNN features improves the F-value to 89.25% (+1.10%) with the significance level of $p<0.05$ with *t-test*. The trend of SSA is the same as the F1-value, the accuracy is improved from the baseline 83.31% to 85.12% using the full model reranker, with an absolute improvement of 1.82%.
Effectiveness of Reranking
--------------------------
Table \[tab:bestresult\] shows our rerank results on two baselines and the comparison with state-of-the-art systems. Our reranker on discrete baseline compares favourably to the best discrete models, including the use of external corpus [@kazama2007exploiting; @suzuki2008semi]. It also outperforms @nguyen2010kernel which builds a discrete reranking model by utilizing SVM with kernels. @ratinov2009design\* achieves 90.57% in discrete model by combining global features and abundant external lexicons, while its performance drops to 88.55% when removing the global features [@ratinov2009design]. @luo2015joint gives the best discrete result (91.20%) by jointing NER with disambiguation task together.
--------------------------------- --
**Discrete Model (%)& **F1\
@kazama2007exploiting&88.02\
@suzuki2008semi&89.92\
@nguyen2010kernel&88.16\
@ratinov2009design&88.55\
@ratinov2009design\*&90.57\
@luo2015joint & 91.20\
Discrete baseline& 88.13\
Our reranker& 89.25\
**Neural Model (%)& **F1\
@collobert2011natural&89.59\
@passos2014lexicon & 90.90\
@huang2015bidirectional& 90.10\
@chiu2015named & 90.77\
@lample2016neural & 90.94\
@ma2016end & 91.21\
Neural baseline& 91.25\
Our reranker& **91.62\
**********
--------------------------------- --
: Comparison of state-of-the-art systems.[]{data-label="tab:bestresult"}
@collobert2011natural builds a first neural NER model with comparable performance to discrete models on CoNLL 2003 corpus. Most state-of-the-art neural NER models utilize bidirectional LSTM with a CRF layer [@huang2015bidirectional]. @lample2016neural and @ma2016end concatenate character representation with word embedding and @chiu2015named even merge lexicon features into word representation. @passos2014lexicon obtain a 90.90% by combining discrete features and neural word embeddings in a CRF model. Our neural baseline, which takes the same features as @ma2016end, achieves 91.25% in F-value. Our reranker on this baseline outperforms all the previous models with the F-value of 91.62%, which is the best reported F-score on CoNLL 2003.
Examples
--------
Figure \[fig:resultexample\] gives some example outputs on the development dataset for which discrete baseline gives incorrect outputs yet the reranker corrects the mistake. Our reranker learns better sentence patterns by correcting both named entity boundary errors and named entity type errors.
In the first case, example 1 shows that *“U.N. Ambassador Albright”* in sentence ** is incorrectly tagged as a *organization* by the baseline and the entity boundary is incorrect either. By building the collapsed sentences as the input of our reranker, entities such as *“U.N. Ambassador Albright”* are replaced as a single entity name *“ORG”*. Our reranking model learns that ** is more possible compared to **, thereby the candidate with the reasonable entity boundary and type is picked by our reranker. For the second case, the entity type of *“EL SALVADOR”* in example 3 ** is incorrectly recognized as *organization* by baseline. Our reranker corrects this entity type error by giving higher score to sentence pattern ** rather than pattern **.
Conclusion
==========
We proposed a neural reranking architecture for NER by exploiting neural structure to learn sentence patterns. Given the candidate label sequences generated from a baseline tagger, we replace the predicted entity words with the corresponding entity type names to build collapsed sentences, which are used as inputs of a neural reranking model. A mixture reranking strategy is used to combine both the knowledge of the probability from the baseline tagger and the reranker score. Experiments on both discrete and neural baselines show our reranking system improves NER performance significantly, obtaining the best results on CoNLL 2003 English task .
Acknowledgments {#acknowledgments .unnumbered}
===============
[^1]: When $p_i=O$, $e_i$ equals to NULL.
[^2]: Characters are padded into a fixed length by using a special token *Pad*.
[^3]: We also tried *max-margin* criterion like [@zhu2015re], while the results are similar with regression model.
[^4]: https://taku910.github.io/crfpp/
[^5]: Notice this is different with accuracy which represents the correct rate of tags, OBA represents the correct rate in sentence level.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present the first comparison of virial masses of galaxy clusters with their Sunyaev-Zel’dovich Effect (SZE) signals. We study 15 clusters from the Hectospec Cluster Survey (HeCS) with MMT/Hectospec spectroscopy and published SZE signals. We measure virial masses of these clusters from an average of 90 member redshifts inside the radius $r_{100}$. The virial masses of the clusters are strongly correlated with their SZE signals (at the 99% confidence level using a Spearman rank-sum test). This correlation suggests that $Y_{SZ}$ can be used as a measure of virial mass. Simulations predict a powerlaw scaling of $Y_{SZ}\propto M_{200}^\alpha$ with $\alpha\approx$1.6. Observationally, we find $\alpha$=1.11$\pm$0.16, significantly shallower (given the formal uncertainty) than the theoretical prediction. However, the selection function of our sample is unknown and a bias against less massive clusters cannot be excluded (such a selection bias could artificially flatten the slope). Moreover, our sample indicates that the relation between velocity dispersion (or virial mass estimate) and SZE signal has significant intrinsic scatter, comparable to the range of our current sample. More detailed studies of scaling relations are therefore needed to derive a robust determination of the relation between cluster mass and SZE.'
author:
- 'Kenneth Rines, Margaret J. Geller, and Antonaldo Diaferio'
title: Comparison of Hectospec Virial Masses with SZE Measurements
---
Introduction
============
Clusters of galaxies are the most massive virialized systems in the universe. The normalization and evolution of the cluster mass function is therefore a sensitive probe of the growth of structure and thus cosmology [e.g., @cirsmf; @rines08; @vikhlinin09b; @henry09; @mantz08; @rozo08 and references therein]. Many methods exist to estimate cluster masses, including dynamical masses from either galaxies [@zwicky1937] or intracluster gas [e.g., @flg80], gravitational lensing [e.g., @smith05; @richard10], and the Sunyaev-Zel’dovich effect [SZE @sz72]. In practice, these estimates are often made using simple observables, such as velocity dispersion for galaxy dynamics or X-ray temperature for the intracluster gas. If one of these observable properties of clusters has a well-defined relation to the cluster mass, a large survey can yield tight constraints on cosmological parameters [e.g., @majumdar04]. There is thus much interest in identifying cluster observables that exhibit tight scaling relations with mass [@kravtsov06; @rozo08]. Numerical simulations indicate that X-ray gas observables [@nagai07] and SZE signals [@motl05] are both candidates for tight scaling relations. Both methods are beginning to gain observational support [e.g., @henry09; @lopes09b; @mantz09b; @locutushuang09]. Dynamical masses from galaxy velocities are unbiased in numerical simulations [@diaferio1999; @evrard07], and recent results from hydrodynamical simulations indicate that virial masses may have scatter as small as $\sim$5% [@lau10].
Previous studies have compared SZE signals to hydrostatic X-ray masses [@bonamente08; @plagge10] and gravitational lensing masses [@marrone09 hereafter M09]. Here, we make the first comparison between virial masses of galaxy clusters and their SZE signals. We use SZE measurements from the literature and newly-measured virial masses of 15 clusters from extensive MMT/Hectospec spectroscopy. This comparison tests the robustness of the SZE as a proxy for cluster mass and the physical relationship between the SZE signal and cluster mass. Large SZ cluster surveys are underway and are beginning to yield cosmological constraints [@carlstrom10; @hincks10; @staniszewski09].
We assume a cosmology of $\Omega_m$=0.3, $\Omega_\Lambda$=0.7, and $H_0$=70 km s$^{-1}$ Mpc$^{-1}$ for all calculations.
Observations
============
Optical Photometry and Spectroscopy
-----------------------------------
We are completing the Hectospec Cluster Survey (HeCS), a study of an X-ray flux-limited sample of 53 galaxy clusters at moderate redshift with extensive spectroscopy from MMT/Hectospec. HeCS includes all clusters with ROSAT X-ray fluxes of $f_X>5\times10^{-12}$erg s$^{-1}$ at \[0.5-2.0\]keV from the Bright Cluster Survey [BCS @bcs] or REFLEX survey [@reflex] with optical imaging in the Sixth Data Release (DR6) of SDSS [@dr6]. We use DR6 photometry to select Hectospec targets. The HeCS targets are all brighter than $r$=20.8 (SDSS catalogs are 95% complete for point sources to $r$$\approx$22.2). Out of the HeCS sample, 15 clusters have published SZ measurements.
### Spectroscopy: MMT/Hectospec and SDSS
HeCS is a spectroscopic survey of clusters in the redshift range 0.10$\leq$$z$$\leq$0.30. We measure spectra with the Hectospec instrument [@hectospec] on the MMT 6.5m telescope. Hectospec provides simultaneous spectroscopy of up to 300 objects across a diameter of 1$^\circ$. This telescope and instrument combination is ideal for studying the virial regions and outskirts of clusters at these redshifts. We use the red sequence to preselect likely cluster members as primary targets, and we fill fibers with bluer targets (Rines et al. in prep. describes the details of target selection). We eliminate all targets with existing SDSS spectroscopy from our target lists but include these in our final redshift catalogs.
Of the 15 clusters studied here, one was observed with a single Hectospec pointing and the remaining 14 were observed with two pointings. Using multiple pointings and incorporating SDSS redshifts of brighter objects mitigate fiber collision issues. Because the galaxy targets are relatively bright ($r$$\leq$20.8), the spectra were obtained with relatively short exposure times of 3x600s to 4x900s under a variety of observing conditions.
Figure \[hecsyszcaus\] shows the redshifts of galaxies versus their projected clustrocentric radii for the 15 clusters studied here. The infall patterns are clearly present in all clusters. We use the caustic technique [@diaferio1999] to determine cluster membership. Briefly, the caustic technique uses a redshift-radius diagram to isolate cluster members in phase space by using an adaptive kernel estimator to smooth out the galaxies in phase space, and then determining the edges of this distribution [see @diaferio09 for a recent review]. This technique has been successfully applied to optical studies of X-ray clusters, and yields cluster mass estimates in agreement with estimates from X-ray observations and gravitational lensing [e.g., @cairnsi; @bg03; @diaferio05; @cirsi; @cirsmf and references therein].
We apply the prescription of @danese to determine the mean redshift $cz_\odot$ and projected velocity dispersion $\sigma_p$ of each cluster from all galaxies within the caustics. We calculate $\sigma_p$ using only the cluster members projected within $r_{100}$ estimated from the caustic mass profile.
SZE Measurements
----------------
The SZE detections are primarily from @bonamente08 [hereafter B08], supplemented by three measurements from @marrone09 [hereafter M09]. Most of the SZ data were obtained with the OVRO/BIMA arrays; the additional clusters from M09 were observed with the Sunyaev-Zel’dovich Array [SZA; e.g., @muchovej07].
Numerical simulations indicate that the integrated Compton y-parameter $Y_{SZ}$ has smaller scatter than the peak y-decrement $y_{peak}$ [@motl05], so B08 and M09 report only $Y_{SZ}$. Although $y_{peak}$ should be nearly independent of redshift, $Y_{SZ}$ depends on the angular size of the cluster. The quantity $Y_{SZ}D_A^2$ removes this dependence. Thus, we compare our dynamical mass estimates to this quantity rather than $y_{peak}$ or $Y_{SZ}$. Table \[hecsysztab\] summarizes the SZ data and optical spectroscopy.
It is also critical to determine the radius within which $Y_{SZ}$ is determined. B08 use $r_{2500}$, the radius that encloses an average density of 2500 times the critical density at the cluster’s redshift; $r_{2500}$ has physical values of 300-700 kpc for the massive clusters studied by B08 (470-670 kpc for the subsample studied here). M09 use a physical radius of 350 kpc because this radius best matches their lensing data.
To use both sets of data, we must estimate the conversion between $Y_{SZ}(r_{2500})$ measured within $r_{2500}$ and $Y_{SZ}(r=350~\mbox{kpc})$ measured within the smaller radius $r=$350 kpc. There are 8 clusters analyzed in both B08 and M09 (5 of which are in HeCS). We perform a least-squares fit to $Y_{SZ}(r_{2500})-Y_{SZ}(r=350\mbox{kpc})$ to determine an approximate aperture correction for the M09 clusters. We list both quantities in Table 1.
[lcccccccc]{} Cluster & $z$ & $\sigma_p$ & $M_{100,v}$ & $M_{100,c}$ & $Y_{SZ}D_A^2$ & $Y_{SZ}D_A^2$ & SZE\
& & & & & (350 kpc) & $(r_{2500})$ &\
& & ${{\rm km~s}^{-1}}$ & $10^{14} M_\odot$ & $10^{14} M_\odot$ & $10^{-5}$Mpc$^{-2}$ & $10^{-4}$Mpc$^2$ & Ref.\
A267 & 0.2288 & $ 743 ^{+81} _{-61}$ & 6.86$\pm$0.82 & 4.26$\pm$0.14 & 3.08$\pm$0.34 & 0.42$\pm$0.06 & 1\
A697 & 0.2812 & $ 784 ^{+77} _{-59}$ & 6.11$\pm$0.69 & 5.96$\pm$3.51 & – & 1.29$\pm$0.15 & 1\
A773 & 0.2174 & $ 1066 ^{+77} _{-63}$ & 18.4$\pm$1.7 & 16.3$\pm$0.7 & 5.40$\pm$0.57 & 0.90$\pm$0.10 & 1\
Zw2701 & 0.2160 & $ 564 ^{+63} _{-47}$ & 3.47$\pm$0.42 & 2.69$\pm$0.30 & 1.46$\pm$0.016 & [0.17$\pm$0.02]{}$^\tablenotemark{a}$ & 2\
Zw3146 & 0.2895 & $ 752 ^{+92} _{-67}$ & 6.87$\pm$0.89 & 4.96$\pm$0.91 & – & 0.71$\pm$0.09 & 1\
A1413 & 0.1419 & $ 674 ^{+81} _{-60}$ & 6.60$\pm$0.85 & 3.49$\pm$0.15 & 3.47$\pm$0.24 & 0.81$\pm$0.12 & 1\
A1689 & 0.1844 & $ 886 ^{+63} _{-52}$ & 15.3$\pm$1.4 & 9.44$\pm$5.66 & 7.51$\pm$0.60 & 1.50$\pm$0.14 & 1\
A1763 & 0.2315 & $ 1042 ^{+79} _{-64}$ & 16.9$\pm$1.6 & 12.6$\pm$1.5 & 3.10$\pm$0.32 & [0.46$\pm$0.05]{}$^\tablenotemark{a}$ & 2\
A1835 & 0.2507 & $ 1046 ^{+66} _{-55}$ & 19.6$\pm$1.6 & 20.6$\pm$0.3 & 6.82$\pm$0.48 & 1.37$\pm$0.11 & 1\
A1914 & 0.1659 & $ 698 ^{+46} _{-38}$ & 6.70$\pm$0.57 & 6.21$\pm$0.21 & – & 1.08$\pm$0.09 & 1\
A2111 & 0.2290 & $ 661 ^{+57} _{-45}$ & 4.01$\pm$0.41 & 4.77$\pm$1.23 & – & 0.55$\pm$0.12 & 1\
A2219 & 0.2256 & $ 915 ^{+53} _{-45}$ & 12.8$\pm$1.0 & 12.0$\pm$4.7 & 6.27$\pm$0.26 & [1.19$\pm$0.05]{}$^\tablenotemark{a}$ & 2\
A2259 & 0.1606 & $ 735 ^{+67} _{-53}$ & 5.59$\pm$0.60 & 4.90$\pm$1.69 & – & 0.27$\pm$0.10 & 1\
A2261 & 0.2249 & $ 725 ^{+75} _{-57}$ & 7.13$\pm$0.83 & 5.10$\pm$2.07 & – & 0.71$\pm$0.09 & 1\
RXJ2129 & 0.2338 & $ 684 ^{+88} _{-64}$ & 4.31$\pm$0.57 & 2.94$\pm$0.13 & – & 0.40$\pm$0.07 & 1\
Results
=======
We examine two issues: (1) the strength of the correlation between SZE signal and the dynamical mass and (2) the slope of the relationship between them. Figure \[hecsysz\] shows the $Y_{SZ}-\sigma_p$ relation. Here, we compute $\sigma_p$ for all galaxies inside both the caustics and the radius $r_{100,c}$ defined by the caustic mass profile \[$r_\delta$ is the radius within which the enclosed density is $\delta$ times the critical density $\rho_c(z)$\].
Because we make the first comparison of dynamical properties and SZE signals, we first confirm that these two variables are well correlated. A nonparametric Spearman rank-sum test (one-tailed) rejects the hypothesis of uncorrelated data at the 98.4% confidence level. The strong correlation in the data suggests that both $\sigma_p$ and $Y_{SZ}D_A^2$ increase with increasing cluster mass.
Hydrodynamic numerical simulations indicate that $Y_{SZ}$ (integrated to $r_{500}$) scales with cluster mass as $Y_{SZ}\propto
M_{500}^\alpha$, where $\alpha$=1.60 with radiative cooling and star formation, and 1.61 for simulations with radiative cooling, star formation, and AGN feedback [$\alpha$=1.70 for non-radiative simulations, @motl05]. Combining this result with the virial scaling relation of dark matter particles, $\sigma_p\propto
M_{200}^{0.336\pm0.003}$ [@evrard07], the expected scaling is $Y_{SZ}\propto \sigma^{4.76}$ (we assume that $M_{100}\propto
M_{500}$). The right panels of Figure \[hecsysz\] shows this predicted slope (dashed lines).
The bisector of the least-squares fits to the data has a slope of $2.94\pm0.74$, significantly shallower than the predicted slope of 4.8.
We recompute the velocity dispersions $\sigma_{p,A}$ for all galaxies within one Abell radius (2.14 Mpc) and inside the caustics. Surprisingly, the correlation is slightly stronger (99.4% confidence level). This result supports the idea that velocity dispersions computed within a fixed physical radius retain strong correlations with other cluster observables, even though we measure the velocity dispersion inside different fractions of the virial radius for clusters of different masses. Because cluster velocity dispersions decline with radius [e.g. @cairnsi; @cirsi], $\sigma_{p,A}$ may be smaller than $\sigma_{p,100}$ (measured within $r_{100,c}$) for low-mass clusters, perhaps exaggerating the difference in measured velocity dispersions relative to the differences in virial mass (i.e., $\sigma_{p,A}$ of a low-mass cluster may be measured within 2$r_{100}$ while $\sigma_{p,A}$ of a high-mass cluster may be measured within $r_{100}$; the ratio $\sigma_{p,A}$ of these clusters would be exaggerated relative to the ratio $\sigma_{p,100}$). Future cluster surveys with enough redshifts to estimate velocity dispersions but too few to perform a caustic analysis should still be sufficient for analyzing scaling relations.
Because of random errors in the mass estimation, the virial mass and the caustic mass within a given radius do not necessarily coincide. Therefore, the radius $r_{100}$ depends on the mass estimator used. Figure \[hecsysz\] shows the scaling relations for two estimated masses $M_{100,c}$ and $M_{100,v}$; $M_{100,c}$ is the mass estimated within $r_{100,c}$ (where both quantities are defined from the caustic mass profile), and $M_{100,v}$ is the mass estimated within $r_{100,v}$ [both quantities are estimated with the virial theorem, e.g., @cirsi]. including galaxies projected inside $r_{100,v}$. Similar to $\sigma_p$, there is a clear correlation between $M_{100,v}$ and $Y_{SZ}D_A^2$ (99.0% confidence with a Spearman test). The strong correlation of dynamical mass with SZE also holds for $M_{100,c}$ estimated directly from the caustic technique (99.8% confidence).
The bisector of the least-squares fits has a slope of $1.11\pm0.16$, again significantly shallower than the predicted slope of 1.6. This discrepancy has two distinct origins. By looking at the distribution of the SZE signals in Figure \[hecsysz\], we see that, at a given velocity dispersion or mass, the SZE signals have a scatter which is a factor of $\sim$2. Alternatively, at fixed SZE signal, there is a scatter of a factor of $\sim$2 in estimated virial mass. Unless the observational uncertainties are significantly underestimated, the data show substantial intrinsic scatter. Moreover, this scatter is comparable to the range of our sample and, therefore, the error on the slope derived from our least-squares fit to the data is likely to be underestimated [see @andreon10 for a detailed discussion of a Bayesian approach to fitting relations with measurement uncertainties and intrinsic scatter in both quantities].
Our shallow slopes may also arise in part from the fact that our sample, which has been assembled from the literature and whose selection function is difficult to determine, is likely to be biased against clusters with small mass and low SZE signal. Larger samples should determine whether unknown observational biases or issues in the physical understanding of the relation account for this discrepancy.
Discussion
==========
The strong correlation between masses from galaxy dynamics and SZE signals indicates that the SZE is a reasonable proxy for cluster mass. B08 compare SZE signals to X-ray observables, in particular the temperature $T_X$ of the intracluster medium and $Y_X=M_{gas}T_X$, where $M_{gas}$ is the mass of the ICM [see also @plagge10]. Both of these quantities are measured within $r_{500}$, a significantly smaller radius than $r_{100}$ where we measure virial mass. M09 compare SZE signals to masses estimated from gravitational lensing measurements. The lensing masses are measured within a radius of 350 kpc. For the clusters studied here, this radius is smaller than $r_{2500}$ and much smaller than $r_{100}$. Numerical simulations indicate that the scatter in masses measured within an overdensity $\delta$ decreases as $\delta$ decreases [@white02], largely because variations in cluster cores are averaged out at larger radii. Thus, the dynamical measurement reaching to larger radius may provide a more robust indication of the relationship between the SZE measurements and cluster mass.
The $Y_{SZ}D_A^2-M_{lens}$ data presented in M09 show a weaker correlation than our optical dynamical properties. A Spearman test rejects the hypothesis of uncorrelated data for the M09 data at only the 94.8% confidence level, compared to the 98.4-99.8% confidence levels for our optical dynamical properties. One possibility is that $M_{lens}$ is more strongly affected by substructure in cluster cores and by line-of-sight structures than are the virial masses and velocity dispersions we derive.
Few measurements of SZE at large radii ($>r_{500}$) are currently available. Hopefully, future SZ data will allow a comparison between virial mass and $Y_{SZ}$ within similar apertures.
Conclusions
===========
Our first direct comparison of virial masses, velocity dispersions, and SZ measurements for a sizable cluster sample demonstrates a strong correlation between these observables (98.4-99.8% confidence). The SZE signal increases with cluster mass. However, the slopes of both the $Y_{SZ}-\sigma$ relation ($Y_{SZ}\propto
\sigma_p^{2.94\pm0.74}$) and the $Y_{SZ}-M_{100}$ relation ($Y_{SZ}\propto M_{100}^{1.11\pm0.16}$) are significantly shallower (given the formal uncertainties) than the slopes predicted by numerical simulations (4.76 and 1.60 respectively).
This result may be partly explained by a bias against less massive clusters that could artificially flatten our measured slopes. Unfortunately, the selection function of our sample is unknown and we are unable to quantify the size of this effect. More importantly, our sample indicates that the relation between SZE and virial mass estimates (or velocity dispersion) has a non-negligible intrinsic scatter. A complete, representative cluster sample is required to robustly determine the size of this scatter, its origin, and its possible effect on the SZE as a mass proxy.
Curiously, $Y_{SZ}$ is more strongly correlated with both $\sigma_p$ and $M_{100}$ than with $M_{lens}$ (M09). Comparison of lensing masses and cluster velocity dispersions (and virial masses) for larger, complete, objectively selected samples of clusters may resolve these differences.
The full HeCS sample of 53 clusters will provide a large sample of clusters with robustly measured velocity dispersions and virial masses as a partial foundation for these comparisons.
We thank Stefano Andreon for fruitful discussions about fitting scaling relations with measurement errors and intrinsic scatter in both quantities. AD gratefully acknowledges partial support from INFN grant PD51. We thank Susan Tokarz for reducing the spectroscopic data and Perry Berlind and Mike Calkins for assisting with the observations.
[*Facilities:*]{}
[40]{} , J. K. [et al.]{} 2008, , 175, 297
, S. & [Hurn]{}, M. A. 2010, in press, arXiv:1001.4639
, H. [et al.]{} 2004, , 425, 367
, A. & [Girardi]{}, M. 2003, , 585, 205
, M., [Joy]{}, M., [LaRoque]{}, S. J., [Carlstrom]{}, J. E., [Nagai]{}, D., & [Marrone]{}, D. P. 2008, , 675, 106
, J. [et al.]{} 2010, ArXiv e-prints
, L., [de Zotti]{}, G., & [di Tullio]{}, G. 1980, , 82, 322
, A. 1999, , 309, 610
—. 2009, ArXiv e-prints
, A., [Geller]{}, M. J., & [Rines]{}, K. J. 2005, , 628, L97
, H., [Edge]{}, A. C., [Allen]{}, S. W., [Crawford]{}, C. S., [Fabian]{}, A. C., & [Huchra]{}, J. P. 1998, , 301, 881
, A. E. [et al.]{} 2008, , 672, 122
, D., [Lecar]{}, M., & [Gorenstein]{}, P. 1980, , 241, 552
, D. [et al.]{} 2005, , 117, 1411
, J. P., [Evrard]{}, A. E., [Hoekstra]{}, H., [Babul]{}, A., & [Mahdavi]{}, A. 2009, , 691, 1307
, A. D. [et al.]{} 2009, ArXiv e-prints
, A. V., [Vikhlinin]{}, A., & [Nagai]{}, D. 2006, , 650, 128
, E. T., [Nagai]{}, D. & [Kravtsov]{}, A. V. 2010, , 708, 1419
, C. [et al.]{} 2009, ArXiv e-prints
, P. A. A., [de Carvalho]{}, R. R., [Kohl-Moreira]{}, J. L., & [Jones]{}, C. 2009, ArXiv e-prints
, S. & [Mohr]{}, J. J. 2004, , 613, 41
, A., [Allen]{}, S. W., [Ebeling]{}, H., & [Rapetti]{}, D. 2008, , 387, 1179
, A., [Allen]{}, S. W., [Ebeling]{}, H., [Rapetti]{}, D., & [Drlica-Wagner]{}, A. 2009, ArXiv e-prints
, D. P. [et al.]{} 2009, , 701, L114
, P. M., [Hallman]{}, E. J., [Burns]{}, J. O., & [Norman]{}, M. L. 2005, , 623, L63
, S. [et al.]{} 2007, , 663, 708
, D., [Vikhlinin]{}, A., & [Kravtsov]{}, A. V. 2007, , 655, 98
, T. [et al.]{} 2010, ArXiv e-prints
, J. [et al.]{} 2010, ArXiv e-prints
, K. & [Diaferio]{}, A. 2006, , 132, 1275
, K., [Diaferio]{}, A., & [Natarajan]{}, P. 2007, , 657, 183
—. 2008, , 679, L1
, K., [Geller]{}, M. J., [Kurtz]{}, M. J., & [Diaferio]{}, A. 2003, , 126, 2152
, E. [et al.]{} 2008, astro-ph/0703571
, G. P., [Kneib]{}, J., [Smail]{}, I., [Mazzotta]{}, P., [Ebeling]{}, H., & [Czoske]{}, O. 2005, , 359, 417
, Z. [et al.]{} 2009, , 701, 32
, R. A. & [Zeldovich]{}, Y. B. 1972, Comments on Astrophysics and Space Physics, 4, 173
, A. [et al.]{} 2009, , 692, 1060
, M. 2002, , 143, 241
, F. 1937, , 86, 217
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We analytically and numerically study the effects of pulsed control on the decoherence of a qubit coupled to a quantum spin bath. When the environment is critical, decoherence is faster and we show that the control is relatively more effective. Two coupling models are investigated, namely a qubit coupled to a bath via a single link and a spin star model, yielding results that are similar and consistent.'
author:
- 'D. Rossini'
- 'P. Facchi'
- Rosario Fazio
- 'G. Florio'
- 'D.A. Lidar'
- 'S. Pascazio'
- 'F. Plastina'
- 'P. Zanardi'
title: 'Bang-Bang control of a qubit coupled to a quantum critical spin bath'
---
Introduction {#sec:intro}
============
Decoherence results from the unavoidable coupling between any quantum system and its environment, and is responsible for the dynamical destruction of quantum superpositions. It is detrimental for quantum information processing [@nielsen] since it leads to a loss of the quantum parallelism that is implicit in the superposition principle. The possibility of preventing or avoiding decoherence is hence of significant importance for any technological use of quantum systems, aimed at processing, communicating or storing information. To this end, one must understand and model all of the relevant features characterizing the environment of the physical system to be protected. Understanding decoherence is also of fundamental interest in its own right, since it is at the basis of the description of the quantum-classical transition [@decoherencereview].
The study of open quantum systems has a long history, and many ingenious models have been proposed in order to describe the action of the environment in a quantum dynamical framework (see, e.g., Ref. [@FKM]). Paradigmatic models represent the environment as a set of harmonic oscillators [@weiss] or spins [@prokofiev]. Recently there has been a renewed interest in the analysis of decoherence induced by such spin baths [@paganelli; @tessieri; @khveshchenko; @Breuer:04; @cucchietti; @dawson; @lages05; @quan; @cucchietti2; @rossini07; @borz07; @zurek07; @hamdouni07; @camalet07; @alvarez07; @Krovi:07; @cormick07; @relano07; @yuan07]; these are clearly relevant in a number of physically important situations, such as NMR [@decNMR] or spin qubits [@loss], where loss of coherence is induced by the coupling to nuclear spins [@decnuclear]. Several questions have been addressed so far and the picture that emerges is rather rich. A possible, monogamy-like, relation between the entanglement in the bath and decoherence has been put forward in Ref. [@dawson] and subsequently analyzed in different papers. The signatures of criticality of the environment in decoherence have been discussed through the study of solvable one-dimensional model systems [@quan; @rossini07; @camalet07]. A universal regime exists, in the strong coupling limit, in which the decay of the Loschmidt echo [@lereview] does not depend on the system-bath coupling [@cucchietti2; @cormick07].
Several different protocols have been designed to protect quantum information. These include passive correction techniques, in which quantum information is encoded in such a way as to suppress the coupling with the environment [@topological; @dfs], and active approaches such as quantum error correction [@nielsen] and dynamical decoupling techniques [@viola98; @DD; @FLP; @theorem] (for an overview see, e.g., [@BBreview], for historical references see [@history]). Dynamical decoupling strategies aim, by means of a dynamical control field, at averaging to zero the unwanted interaction with the environment. In its simplest version, which we consider here, the control field comprises a train of instantaneous pulses (bang-bang control). While previous work on dynamical decoupling has made clear distinctions between different environments, in particular bosonic baths [@viola98] versus spin baths [@KL:07; @Zhang:07], and fast versus slow $1/f$ noise [@bangoneoverf], no attention has been paid so far to the impact a quantum critical environment might have on the efficacy of decoupling protocols. This is our goal in the present work: here we study bang-bang decoupling in the case where the quantum environment can become critical.
Many-body environments displaying critical behavior have been recently investigated in great detail, in order to study the sensitivity of decoherence to environmental dynamics (see, e.g., [@quan; @rossini07]). Close to a quantum critical point the environment becomes increasingly slower (a phenomenon known as critical slowing down). We analyze the decoherence process of a two level system (qubit) coupled to an environment modeled as a one-dimensional lattice of spins interacting through an Ising-like coupling. We focus on the suppression of qubit decoherence through a bang-bang control procedure, and study how the occurrence of a quantum phase transition (QPT) in the bath modifies the effectiveness of the control procedure. Our analysis is focused on the behavior of the Loschmidt Echo (LE) [@lereview], whose study has given new insights into the decoherence process of quantum spin chains. We discuss the application of a pulse train to the qubit and show its effectiveness in quenching qubit dephasing, especially at the critical point.
The paper is organized as follows. In Sec. \[sec:model\] we introduce the model and pertinent notation. The control procedure, based on a sequence of pulses that repeatedly flip the state of the system, is described in Sec. \[sec:control\], where we also derive an explicit expression for the LE in the presence of such control. We then provide a detailed analysis of its effects in the limiting cases of a single qubit-bath link (Subsec. \[sec:controlsingle\]) and a spin-star model (Subsec. \[sec:controlbath\]). Finally, in Sec. \[sec:conclusion\] we discuss our results. In the Appendices we provide an analytical formula for evaluating the LE in the presence of control (App. \[sec:Appendixechofree\]), we perform a perturbative analysis in the pulse frequency of the LE (App. \[sec: AppendixLE\]), and discuss in detail a closed-form formula for the LE in the spin-star model (App. \[sec:Appendixeffectivedecay\]).
Model and notation {#sec:model}
==================
We consider a two level quantum system $S$ (qubit) coupled to an interacting spin bath $E$ (environment), comprising a linear chain of $N$ spin-$1/2$ particles, modeled by a transverse field Ising model. The Hamiltonian reads $$\mathcal{H}_{0}=\mathcal{H}_{S}+\mathcal{H}_{E}+\mathcal{H}_{\mathrm{int}}\,,
\label{eq:HamFree}$$ where $\mathcal{H}_{S}$ and $\mathcal{H}_{E}$ are the free Hamiltonians of $S$ and $E$: $$\begin{aligned}
\mathcal{H}_{S} &=&-\frac{\omega _{0}}{2}\,(\mathbbm{1}-\tau ^{z})=-\omega
_{0}\left\vert \downarrow \right\rangle \left\langle \downarrow \right\vert
\,, \\
\mathcal{H}_{E} &=&-J\sum_{j=1}^{N}\Big(\sigma _{j}^{x}\sigma
_{j+1}^{x}+\lambda \sigma _{j}^{z}\Big)\,; \label{eq:systemhamilt}\end{aligned}$$ here $\sigma _{i}^{\alpha }$ and $\tau ^{\alpha }$ (with $\alpha =x,y,z$) indicate, respectively, the Pauli matrices of the $i$th spin of the chain $E$ and of the qubit $S$, whose ground and excited states are denoted by $\left\vert \uparrow \right\rangle $ and $\left\vert \downarrow \right\rangle$. In this work we will use periodic boundary conditions, therefore we assume $\sigma _{N+1}^{\alpha}\equiv \sigma _{1}^{\alpha }$. The constants $J$ and $\lambda $ are the interaction strength between neighboring spins of the bath and an external transverse magnetic field, respectively (in the following, the energy and the time scale are taken in units of $J$, therefore, when not specified, we will implicitly assume $J=1$). We suppose that the system is coupled to a given number of bath spins [@rossini07]: $$\mathcal{H}_{\mathrm{int}}=-\epsilon \,\left\vert \downarrow \right\rangle
\left\langle \downarrow \right\vert \otimes \sum_{j=j_{1}}^{j_{m}}\sigma
_{j}^{z}\,, \label{eq:interactionhamilt}$$ where $\epsilon $ is the coupling constant and $m$ the number of environmental spins to which $S$ is coupled. The LE can be calculated for a generic sequence $\{j_{1},\ldots j_{m}\}$ of system-bath links. In the following, however, we consider the cases $m=1$ and $m=N$. We expect that the generic case will be a quantitative interpolation between these two extremes but no new qualitative features should emerge.
With the above choice of $\mathcal{H}_S$ and $\mathcal{H}_{\mathrm{int}}$, the populations of the ground and excited state of the qubit do not evolve, since $[\tau^z, \mathcal{H}_0 ] =0$, and we can study a model of pure dephasing.
As usual, we assume that the initial global state of the system is factorized: $${\left\vert\Psi (0)\right\rangle} = \left( c_{\uparrow} {\left\vert\uparrow\right\rangle} +
c_{\downarrow} {\left\vert\downarrow\right\rangle} \right) \otimes {\left\vertG\right\rangle} \, ,
\label{eq:InitialState}$$ so that the qubit $S$ is in a generic superposition of the ground and excited state, while the bath $E$ is in its ground state (i.e., ${\left\vertG\right\rangle}$ is the ground state of the Hamiltonian $\mathcal{H}_{E}$). The evolution of such a state under the Hamiltonian is dictated by the unitary operator $U_{0} =
e^{-i\mathcal{H}_{0}t}$ and yields, at time $t$, the state $${\left\vert\Psi (t)\right\rangle} = c_\uparrow {\left\vert\uparrow\right\rangle} {\left\vert\varphi_{0}(t)\right\rangle}
+ c_\downarrow \, e^{i\omega_{0}t} {\left\vert\downarrow\right\rangle} {\left\vert\varphi_{1}(t)\right\rangle} \, ,$$ where ${\left\vert\varphi_{0}(t)\right\rangle} \equiv e^{-i {\mathcal H}_\uparrow t} {\left\vertG\right\rangle}$ and ${\left\vert\varphi_{1}(t)\right\rangle} \equiv e^{-i {\mathcal H}_\downarrow t} {\left\vertG\right\rangle}$ are the environment states evolved under an “unperturbed” and a “perturbed” Hamiltonian given, respectively, by $$\mathcal{H}_{\uparrow }\equiv \mathcal{H}_{E} \, , \qquad
\mathcal{H}_{\downarrow} \equiv \mathcal{H}_{E} +
\left\langle \downarrow \right\vert \mathcal{H}_{\mathrm{int}}
\left\vert \downarrow \right\rangle \, .
\label{eq:Hud}$$
The density matrix of the qubit is $\rho =\mathrm{Tr}_{E}\left\vert \Psi \right\rangle \left\langle \Psi
\right\vert $. Its diagonal elements are constant, while off-diagonal elements decay in time as $$\rho _{\downarrow \uparrow }(t)=\rho _{\downarrow \uparrow }(0)\,e^{i\omega
_{0}t}D(t), \label{eq:offred}$$ with $$D(t)=\langle {\varphi _{0}(t)}|\varphi _{1}(t)\rangle =\left\langle G\right\vert
e^{i\mathcal{H}_{\uparrow }t}e^{-i\mathcal{H}_{\downarrow }t}\left\vert
G\right\rangle . \label{eq:D}$$ The decoherence of the qubit is then fully characterized by the so called *Loschmidt echo* $\mathcal{L}_{0}(t)\in \lbrack 0,1]$ of the environment: $$\mathcal{L}_{0}(t)\equiv |D(t)|^{2}= |\left\langle G\right\vert
e^{-i(\mathcal{H}_{E}+\left\langle \downarrow \right\vert \mathcal{H}_{\mathrm{int}}
\left\vert \downarrow \right\rangle )t}\left\vert G\right\rangle |^{2} \, .
\label{eq:LEchoFree}$$
The decay of the LE in the model - with $m=N$ (spin-star model) was first studied in detail in Ref. [@quan]; an extension to the more general case $m\neq N $, and for other spin baths – including the $XY$ and Heisenberg models – can be found in Ref. [@rossini07]. It was pointed out that the echo decay is enhanced at criticality, due to the hypersensitivity to perturbations of the (time-evolved) unperturbed ground state $\left\vert \varphi _{0}(t)\right\rangle$. Indeed, at criticality the perturbation $\mathcal{H}_{\mathrm{int}}$ is very effective at making the unperturbed state ${\left\vert\varphi_{0}(t)\right\rangle}$ orthogonal to ${\left\vert\varphi_{1}(t)\right\rangle}$, thus leading to a strong decay of the echo. Away from criticality, the perturbation is not so effective at orthogonalizing ${\left\vert\varphi _{0}(t)\right\rangle}$ and ${\left\vert\varphi_{1}(t)\right\rangle}$, whence the echo decays more slowly. In the following we investigate these effects when a control is also present. Details on how to evaluate the LE in both the absence and presence of such a control are given in Appendix \[sec:Appendixechofree\].
Controlled dynamics {#sec:control}
===================
Quantum dynamical decoupling procedures aimed at actively fighting decoherence hinge either on the action of frequent interruptions of the evolution or on the effect of a strong continuous coupling to an external field. These procedures are known to be physically and, to a large extent, mathematically equivalent [@FLP]. Here we focus on one possible procedure, based on multipulse control [@viola98]. Let us formally introduce the control scheme as $$\mathcal{H}(\omega _{0},t)=\mathcal{H}_{0}+\mathcal{H}_{P}(\omega _{0},t)\,,
\label{eq:HamGlobal}$$ where $\mathcal{H}_{P}$ is an additional time-dependent Hamiltonian that causes spin flips of the qubit at regular time intervals through a monochromatic alternating magnetic field at resonance: $$\begin{aligned}
\mathcal{H}_{P}(\omega _{0},t) & = & \sum_{n} V^{(n)}(t)
\Big[\cos \big( \omega_{0}(t-n\Delta t)\big) \,\tau _{x} \notag \\
& & + \sin \big(\omega _{0}(t-n\Delta t)\big) \,\tau _{y} \Big] \,.\end{aligned}$$ Here $V^{(n)}(t)$ is constant and equal to $V$ for the entire duration $\tau_{P}$ of the $n$th pulse (i.e., for $n\Delta t\leq t\leq n\Delta t+\tau _{P}$), $\Delta t$ being the time interval between two consecutive pulses. In this work we only deal with $\pi$ pulses, satisfying the condition $2 V \tau_P = \pm \pi$, and suppose that $V$ is large enough to yield almost instantaneous spin flips, i.e., we take $\tau_P \ll \Delta t$. Therefore, in the ideal limit of instantaneous kicks of infinite strength ($\tau_P \to 0, \, V \to \infty$ such that $V \tau_P = \pm \pi / 2$), the effect of each pulse on the qubit is simply a flip, that is described by the operator $$U_P = \pm i \, \tau^x.$$
The evolution of the initial state under the Hamiltonian in one spin-flip cycle \[i.e., two flips, from time $t=0$ to time $t_{1}=2(\Delta t+\tau _{P})\simeq 2\,\Delta t$\] is dictated by the unitary operator $$U_{\mathcal{C}}\equiv e^{-2i\mathcal{H}\Delta t}=U_{P}\;U_{0}(\Delta t)
\; U_{P} \; U_{0}(\Delta t)$$ and it is such that $$\begin{aligned}
{\left\vert\Psi (2 \Delta t)\right\rangle} & = & - c_\uparrow e^{- i \omega_0 \Delta t} {\left\vert\uparrow\right\rangle}
e^{- i {\mathcal H}_\downarrow \Delta t} e^{- i {\mathcal H}_\uparrow \Delta t} {\left\vertG\right\rangle} \nonumber \\
& & - c_\downarrow e^{- i \omega_0 \Delta t} {\left\vert\downarrow\right\rangle}
e^{- i {\mathcal H}_\uparrow \Delta t} e^{- i {\mathcal H}_\downarrow \Delta t} {\left\vertG\right\rangle} .\nonumber \\\end{aligned}$$ This is again a pure dephasing phenomenon, so that all relevant information is contained in the off-diagonal element of the system reduced density matrix. The behavior of decoherence is then fully captured by the LE: $$\mathcal{L}(2\Delta t) = \Big\vert \left\langle G\right\vert
\left( e^{i \mathcal{H}_{\downarrow} \Delta t}
e^{i\mathcal{H}_{\uparrow }\Delta t}\right) \cdot
\left( e^{-i\mathcal{H}_{\downarrow }\Delta t} e^{-i\mathcal{H}_{\uparrow }\Delta t}\right)
\left\vert G\right\rangle \Big\vert^{2}\,.$$ In general, at a certain time $t=2M \Delta t+\tilde{t}$, the evolution operator of the global system is given by: $$U=\left\{
\begin{array}{ll}
U_{0}(\tilde{t}) \: [U_{\mathcal{C}}]^{M} & \mathrm{if} \quad
\tilde{t} < \Delta t \vspace{1mm} \\
U_{0}(\tilde{t}-\Delta t) \, U_{P} \, U_{0}(\Delta t) \: [U_{\mathcal{C}}]^{M}
& \mathrm{if} \quad \tilde{t} \geq \Delta t
\end{array}
\right.$$ where $M=[\frac{t}{2\Delta t}]$, $[\cdot ]$ denotes the integer part and $\tilde{t}\equiv t- 2M \Delta t$ is the residual time after $M$ cycles. It is now easy to write down the LE at a generic time t:
$$\mathcal{L} (t) = \left\{ \begin{array}{ll}
\Big\vert \, \langle G \vert \left( e^{i {\mathcal H}_\downarrow \Delta t}
e^{i {\mathcal H}_\uparrow \Delta t} \right)^M e^{i {\mathcal H}_\downarrow \tilde{t}}
\cdot e^{- i {\mathcal H}_\uparrow \tilde{t}} \left( e^{-i {\mathcal H}_\downarrow \Delta t}
e^{-i {\mathcal H}_\uparrow \Delta t} \right)^M {\left\vertG\right\rangle} \, \Big\vert^2 & \quad
\mathrm{if} \;\; \tilde{t} < \Delta t \\
\Big\vert \, \langle G \vert \left( e^{i {\mathcal H}_\downarrow \Delta t}
e^{i {\mathcal H}_\uparrow \Delta t} \right)^M e^{i {\mathcal H}_\downarrow \Delta t}
e^{i {\mathcal H}_\uparrow \tilde{t}} \cdot e^{- i {\mathcal H}_\downarrow \tilde{t}}
e^{- i {\mathcal H}_\uparrow \Delta t} \left( e^{-i {\mathcal H}_\downarrow \Delta t}
e^{-i {\mathcal H}_\uparrow \Delta t} \right)^M {\left\vertG\right\rangle} \, \Big\vert^2 & \quad
\mathrm{if} \;\; \tilde{t} \geq \Delta t \end{array} \right.
\label{eq:LEchoPulse}$$
An explicit formula for evaluating the LE, also in the presence of pulses, is given in Appendix \[sec:Appendixechofree\].
In the limit of short pulse intervals, and when $t$ is an integer multiple of the duration of a single spin-flip cycle, $t=2 M\Delta
t$, one can show (see Appendix \[sec: AppendixLE\]) that Eq. can be rewritten as $$\mathcal{L}(t = 2 M\Delta t)=\Big\vert \left\langle G \right\vert
e^{i t \mathcal{H}_{\mathrm{eff}}} \left\vert G\right\rangle
\Big\vert^2 + M \, O(\Delta t^{2}) \, ,
\label{eq:LE-eff}$$ where $$\mathcal{H}_{\mathrm{eff}}\equiv i\frac{\Delta t}{2}[\mathcal{H}_{\downarrow },
\mathcal{H}_{\uparrow }]=i\frac{\Delta t}{2}\,[\left\langle \downarrow
\right\vert \mathcal{H}_{\mathrm{int}}\left\vert \downarrow \right\rangle
,\mathcal{H}_{E}]
\label{eq:H-eff}$$ is an effective Hamiltonian. By noting that $\left\langle \downarrow
\right\vert \mathcal{H}_{\mathrm{int}}\left\vert \downarrow \right\rangle
=-\epsilon \sum_{j=j_{1}}^{j_{m}}\sigma _{j}^{z}$, we have, for arbitrary $\lambda$ $$\mathcal{H}_{\mathrm{eff}} = i\epsilon _{\mathrm{eff}} \bigg[
\sum_{j=j_{1}}^{j_{m}}\sigma _{j}^{z},\sum_{j=1}^{N}\sigma _{j}^{x}\sigma
_{j+1}^{x}\bigg] \, ,
\label{eq:Heff}$$ where $$\epsilon _{\mathrm{eff}}\equiv \epsilon J\frac{\Delta t}{2}. \label{eq:renormcc}$$ This is the renormalized system-bath coupling constant in the presence of multipulse control. We notice that $\mathcal{H}_{\mathrm{eff}}$ does not depend on $\lambda$ (which would appear at $O(\Delta t^{3})$ through the double commutator $[[\mathcal{H}_{\downarrow },\mathcal{H}_{\uparrow }], \mathcal{H}_{\uparrow }]$). Therefore, in the small $\Delta t$ limit, the criticality of the model can manifest itself only through ${\left\vertG\right\rangle}$ in the LE expression .
In the next two subsections we turn to a numerical study of the LE for the cases of a qubit coupled to one spin of the chain \[$m=1$ in Eq. \], and the spin-star model \[$m=N$ in Eq. \].
Qubit coupled to a single bath spin {#sec:controlsingle}
-----------------------------------
When $m=1$, the system-bath Hamiltonian of Eqs. - can be rewritten as: $$\mathcal{H}_{0}=-\left\vert \downarrow \right\rangle \left\langle \downarrow
\right\vert \big(\omega _{0}+\epsilon \,\sigma _{1}^{z}\big)-J\sum_{j=1}^{N}
\Big( \sigma _{j}^{x}\sigma _{j+1}^{x}+\lambda \sigma _{j}^{z}\Big)
\label{eq:onespin}$$ and corresponds to a situation in which the qubit is directly coupled to only one spin of an Ising chain with periodic boundary conditions (the coupled bath-spin qubit is assumed for simplicity and with no loss of generality to be the first one).
In Fig. \[fig:lecho\_time\] we show the behavior of the LE in Eq. as a function of time, for different values of the pulse frequency $\Delta t$. The three panels refer to different values of the transverse magnetic field $\lambda $; the thick dashed lines represent the case $\mathcal{L}_{0}(t)$ with no external control \[$\Delta t\rightarrow \infty$ in Eq. , or simply Eq. \]. Here the environment consists of $N=100$ Ising spins, and the system-bath coupling has been set at $\epsilon =0.25$.
We notice a very different behavior as $\lambda $ is varied. Away from criticality (i.e., for $\lambda =0.5$ Fig. \[fig:lecho\_time\](a), and $\lambda = 1.5$ Fig. \[fig:lecho\_time\](c)) the LE in absence of control quickly reaches its asymptotic (saturation) value $\mathcal{L}_{\infty}$, as indicated by the dashed black lines. Very fast control pulses do improve the situation, but only in the sense that this asymptotic value becomes slightly closer to unity. In contrast, slow pulses make the situation much worse: when $J \Delta t$ is larger than a certain value, the pulses act as an additional source of noise and, as a consequence, the coherence decays (exponentially). On the other hand, when the chain is critical ($\lambda =1$ Fig. \[fig:lecho\_time\](b)) and there is no control, the LE decays (albeit only logarithmically [@rossini07]), as can be seen from the dashed curve. In this case the pulses can be very effective, as a control procedure: when $J \Delta t\lesssim 0.375$ decay is suppressed. Again, when $\Delta t$ exceeds this threshold, decay is enhanced. This situation is reminiscent of the transition between a quantum Zeno and an inverse Zeno effect [@invZeno].
In Fig. \[fig:plat\_dt\] we show the values of the LE at a fixed time $t^{\ast}$ (we performed an average of $\mathcal{L}(t)$ for $Jt\in \lbrack Jt^{\ast }-5,Jt^{\ast }+5]$ in order to eliminate fast oscillations), as a function of $\Delta
t$. The different curves are obtained for different values of the transverse field $\lambda $. We set $Jt^{\ast }=25$ so that: i) in the absence of pulse control and for noncritical $\lambda $, $\mathcal{L}_{0}$ has already reached its saturation value $\mathcal{L}_{\infty }$; ii) at criticality, the minimum of $\mathcal{L}_{0}(t)$ is found exactly at $Jt^{\ast }\simeq N/4$ (in this case $N=100$) [@rossini07]. In the panel (b), bars denote the corresponding value of $\mathcal{L}_{0}(t^{\ast })$ without external control.
The behavior at large pulse intervals $\Delta t$ is non-trivial and rather interesting: we note that the echo has a minimum and has an almost complete recovery, and that the LE for $\lambda=0.5$ rises higher than for $\lambda=0.9,1,1.1$. The large $\Delta t$ regime is non-perturbative (in the sense of the perturbation theory of Section \[sec:control\] and Appendix \[sec: AppendixLE\]). Nevertheless, the rise of the LE for large $\Delta t$ can be understood as being due to the fact that, after a short transient time $\bar{t}$, the LE *without control* saturates around a constant value (see the black dashed curves in the insets of Fig. \[fig:lecho\_time\], or Ref. [@rossini07]). Therefore, if the pulse frequency is such that $\Delta t > \bar{t}$, the effect of the bang-bang control procedure will be progressively reduced as $\Delta t$ grows, until, in the limit $\Delta t \to + \infty$, it will completely disappear. In other words, the detrimental effect of the control for large $\Delta$ is offset by the gradual diminishing of its effect as $\Delta t$ grows, which allows the LE to recover to its saturation value. Moreover, as the insets of Fig. \[fig:lecho\_time\] show, for $\lambda=1.5$ the saturation is truly at a constant value; for $\lambda=0.5$ the saturation is an oscillation around a constant value; at criticality ($\lambda=1$) there is a logarithmic decay of the LE, but for a finite system size this decay will eventually stop and revivals of quantum coherence will appear. The oscillation at $\lambda = 0.5$ explains why this curve rises higher than the other curves in Fig. \[fig:plat\_dt\](a); at a time $t^* \approx 1.5$, the uncontrolled LE in Fig. \[fig:lecho\_time\] at $\lambda = 0.5$ is larger than for other values of $\lambda$.
The panels (a)-(b) of Fig. \[fig:frac\] display $\mathcal{L}(t^{\ast})$ as a function of $\lambda $, for different values of $\Delta t$. In panels (c)-(d) we plot the rescaled quantities, obtained by dividing $\mathcal{L}(t^{\ast })$ by the corresponding value in absence of pulse control, $\mathcal{L}_{0}(t^{\ast})$. The LE has a maximum not at $\lambda =1$ but at $\lambda >1$, while at $\lambda
=1$ there is an inflexion point. At criticality, the rescaled LE displays a cusp. The cusp disappears at $J \Delta t\gtrsim 0.375$, in agreement with Fig. \[fig:lecho\_time\](b), where we observed, at the same value of $\Delta t$, an increase of the LE when the control is present. A qualitative explanation of this phenomenon is straightforward: for short time pulses, the renormalized coupling constant $\epsilon_{\mathrm{eff}}$ in Eq. , and therefore the LE, are only weakly dependent on $\lambda$ at leading order in the perturbative expansion. In contrast, the free echo $\mathcal{L}_{0}$ has a downward cusp [@rossini07] (present also in Fig. \[fig:SpinStar\](a) for the spin-star case). The ratio must therefore display an upward cusp, as seen in Fig. \[fig:frac\]. Another way to state this explanation is the following. For sufficiently small values of $\Delta t$ the bang-bang protocol succeeds at effectively eliminating the environment action. The only remnant of criticality is then the weak signature of an inflexion point seen in Fig. \[fig:frac\](a). The echo of the uncontrolled system, however, is hypersensitive to criticality, as indicated by the cusp. On the other hand, when $\Delta t$ is too large (Fig. \[fig:frac\](b)-(d)), the bang-bang protocol fails at removing the coupling of the qubit to the environment, and the controlled and uncontrolled echos behave similarly.
There are other interesting features in Fig. \[fig:frac\]. Panels (a)-(b) show that the LE rises for sufficiently large $\lambda$, and (c)-(d) show that the ratio between the decoupled and free echoes approaches unity for large $\lambda$. This can be understood as being due to the dominance of the uniform magnetic field term $\lambda \sum_{j=1}^{N}\sigma _{j}^{z}$ over the transverse Ising term $\sum_{j=1}^{N}\sigma _{j}^{x}\sigma
_{j+1}^{x}$ in Eq. . Indeed, in the limit of large $\lambda$, this means that $\mathcal{H}_{\downarrow }\simeq
\mathcal{H}_{E}$ \[recall Eq. \], so that $[\mathcal{H}_{\uparrow }, \mathcal{H}_{\downarrow}]\simeq 0$ and the LE $\simeq 1$ by Eqs. and . Thus, at large $\lambda$, decoupling is not needed to obtain a large LE.
More interesting is the monotonic rise of the LE visible in panel (a) as a function of $\lambda $ for $J \Delta t=0.1,0.2$, in contrast to the maximum around $\lambda \sim 1.25$ for $J \Delta t=0.3$. Indeed, panel (c) shows that decoupling makes the situation worse for $J \Delta t=0.3$ and $\lambda \gtrsim 1.25$, and a similar trend continues in panels (b)-(d). Thus, in our model decoupling is fully effective (i.e., for all values of $\lambda $) for $J \Delta t\lesssim 0.2$.
Spin-star model {#sec:controlbath}
---------------
The “spin-star” model corresponds to the case when the qubit is equally coupled to all the spins of the chain \[$m=N$ in Eq. \]. This situation is opposite to the one considered in the previous subsection. Interestingly, in this limit the model is almost solvable. The system-bath Hamiltonian of Eq. reads: $$\mathcal{H}_{0}= -\omega_0 {\left\vert\downarrow\right\rangle} {\left\langle\downarrow\right\vert}
- J \sum_{j=1}^{N} \left[ \sigma_{j}^{x} \sigma_{j+1}^{x} +
\left(\lambda +\frac{\epsilon}{J} \right) \sigma_{j}^{z} \right] \, .$$ We first notice that ${\mathcal H}_\downarrow (\lambda) = {\mathcal H}_\uparrow (\tilde{\lambda})
\equiv {\mathcal H}_E (\tilde{\lambda})$, where $\tilde{\lambda} = \lambda +\epsilon /J$. Therefore, both the perturbed and the unperturbed Hamiltonians describe an Ising model with a uniform transverse field, and can be diagonalized analytically by means of a standard Jordan-Wigner-Fourier transformation, followed by a Bogoliubov rotation. Details on how to evaluate the LE of Eq. for a spin-star model can be found in Appendix \[sec:Appendixeffectivedecay\], where we show that $$\left\langle G\right\vert e^{it\mathcal{H}_{\mathrm{eff}}} {\left\vertG\right\rangle} =
\prod_{k>0} \cos \left( 8 t \, \epsilon _{\mathrm{eff}}\Delta _{k} \right)\,,
\label{eq:LE-final}$$ with $\Delta_k = \sin (2 \pi k /N)$ and $\epsilon_{\mathrm{eff}}$ defined in Eq. . In the limit of small $\epsilon _{\mathrm{eff}}$, while keeping $t$ finite, we can approximate this as $$\left\langle G\right\vert e^{it\mathcal{H}_{\mathrm{eff}}} {\left\vertG\right\rangle} \simeq
\prod_{k>0} e^{-\frac{1}{2}(8t \epsilon _{\mathrm{eff}}\Delta _{k})^{2}}
= e^{-\frac{\Gamma }{2}(t\epsilon _{\mathrm{eff}})^{2}},
\label{eq:G-ss}$$ where we have defined $$\Gamma \equiv 64 \sum_{k>0}\Delta _{k}^{2} \, .$$
We notice that the dependence on $\lambda$ in Eq. has entirely disappeared. This means that, to leading order in the pulse interval $\Delta t$, dynamical decoupling is not sensitive to criticality \[Eq. for the LE and Eq. lead to $\mathcal{L} \approx e^{- \Gamma (t
\epsilon_{\rm eff})^2}$; we explicitly checked that, for small $\Delta t$ and at short times, this formula exactly reproduces the data obtained from numerical simulations, which are completely insensitive to $\lambda$ in that regime\]. This is consistent with the data shown in Fig. \[fig:SpinStar\]. In panel (a) we see the behavior of the LE in absence of control; we notice a slight dip at $\lambda =1$. In panels (b)-(c) we study the LE for various $\Delta t$; we observe strong similarities with Fig. \[fig:frac\], in particular the weak dependence of the LE on $\lambda$ for very small $\Delta t$. In panel (d) the rescaled LE again displays a cusp.
It is remarkable how similar the results are for $m=1$ (qubit coupled to a single spin of the chain) and $m=N$ (spin-star model). The consistency of these results and the analogies between these two opposite situations lead us to conclude that general features of the decoherence of the qubit under bang-bang control are largely independent of the number of chain spins coupled to it, at least when the chain is close to criticality.
Discussion and conclusions {#sec:conclusion}
==========================
We have studied the efficacy of pulsed control of a qubit when it is coupled to a spin bath. It is well known that, without control pulses, the qubit decoheres particularly fast in the vicinity of the critical point. The reason for this is that the evolution takes the initial state ${\left\vert\Psi(0)\right\rangle}$, in the form of Eq. , into a superposition of the type ${\left\vert\uparrow\right\rangle}
{\left\vert\varphi_{0}(t)\right\rangle} + {\left\vert\downarrow\right\rangle} {\left\vert\varphi_{1}(t)\right\rangle}$ and the two bath states become rapidly orthogonal near the critical point. The application of decoupling pulses to the qubit removes the dependence of decoherence on the criticality of the environment. On the other hand, we also found a regime (larger interval $\Delta t$ between pulses) such that the control can increase the effects of decoherence. Away from criticality the perturbation is not as effective at orthogonalizing ${\left\vert\varphi_{0}(t)\right\rangle}$ and ${\left\vert\varphi_{1}(t)\right\rangle}$, leading to a slow decay of the echo and to relatively less effective control. Therefore, we can conclude that in general decoupling is *relatively* more effective near the critical point, since there it results in the largest enhancement of coherence.
From the quantum information processing perspective, there is another positive message in these results: suppose we are trying to preserve the coherence of a qubit in the presence of a spin bath. Without decoupling we know that the spin decoheres particularly fast in the vicinity of the critical point. Therefore not knowing whether we are close to criticality when trying to operate a quantum computer coupled to a spin bath, is a problem. But in light of the results presented here, it follows that application of dynamical decoupling pulses removes this concern: for sufficiently frequent pulses, decoupling works independently of the value of the system-bath coupling $\lambda$, so closeness to criticality does not matter.
Our analytical and numerical calculations suggest that these results seem to be largely independent of the details of the model of qubit-environment coupling. Indeed, we have considered two extreme situations (qubit coupled to a single spin of the chain and qubit coupled to all spins in the chain), and obtained the same qualitative behavior.
Finally, a comparison of different control strategies (Zeno effect, decoupling pulses and strong continuous coupling) [@FTPNTL2005] has shown that, although these procedures are physically equivalent, there are important practical differences among them. Future attention will be directed towards the exploration of these similarities and differences in the context of coupling of a qubit to a critical system.
This work is partly supported by the European Community through the Integrated Project EuroSQIP. DAL was sponsored by the United States Department of Defense. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressly or implied, of the U.S. Government.
\[sec:Appendixechofree\]
We explain here how to evaluate the LE for the Hamiltonian in Eq. , and then extend some of these results to the case of pulsed control, Eq. . This technique can be easily generalized to the case of an $XY$ spin bath, as has been done in Ref. [@rossini07].
By means of the Jordan-Wigner transformation $$\sigma _{j}^{+} = c_{j}^{\dagger }\exp \bigg( i \pi
\sum_{k=1}^{j-1} c_{k}^{\dagger}c_{k} \bigg) \,, \quad
\sigma _{j}^{z} = 2 c_{j}^{\dagger }c_{j}-1,
\label{eq:JWT}$$ we first map the Hamiltonians $\mathcal{H}_{\downarrow }$ and $\mathcal{H}_{\uparrow }$ of the spin bath onto a free fermion model that can be expressed in the form $$\mathcal{H}_{\uparrow /\downarrow }=\frac{1}{2}\mathbf{\Psi ^{\dagger}C\Psi},$$ where $\mathbf{\Psi ^{\dagger }}=(c_{1}^{\dagger }\ldots c_{N}^{\dagger}
\, c_{1}\ldots c_{N})$ ($c_{i}$ being the corresponding spinless fermion operators) and $$\mathbf{C}=\sigma ^{z}\otimes \mathbf{A}+i\sigma ^{y}\otimes \mathbf{B}$$ is a tridiagonal block matrix with $$\begin{aligned}
A_{j,k} &=&-J(\delta _{k,j+1}+\delta _{j,k+1})-2(\lambda +\epsilon
_{j})\delta _{j,k} \\
B_{j,k} &=&-J(\delta _{k,j+1}-\delta _{j,k+1})\end{aligned}$$ such that $\epsilon _{j}=0$ for $\mathcal{H}_{\uparrow }$, while $\epsilon_{j}=\epsilon \,\delta _{j,j_{m}}$ for $\mathcal{H}_{\downarrow }$. The LE can then be evaluated exactly, by rewriting it in terms of the determinant of a $2N\times 2N$ matrix [@rossini07]: $$\mathcal{L}_{0}(t)=\left| \mathrm{det}\left( \mathbbm{1} -\mathbf{r}+\mathbf{r}\,
e^{i\mathbf{C_{\downarrow }}t}\right) \right| \;, \label{eq:LEcho}$$ where $\mathbf{r}$ is a matrix whose elements $r_{i,j}=\langle \Psi_{i}^{\dagger}
\Psi _{j}\rangle $ are the two-point correlation functions of the spin chain, evaluated in the ground state of the Hamiltonian $\mathcal{H}_{\uparrow }$. Eq. can be obtained from the following trace formula [@klich03]: $$\mathrm{Tr} \big( e^{\Gamma (A)}e^{\Gamma (B)} \big) =\mathrm{det}\left(
\mathbbm{1} +e^{\mathbf{A}}e^{\mathbf{B}}\right) \,, \label{eq:trace}$$ where $\Gamma (A)=\sum_{i,j}\mathbf{A}_{ij}\,a_{i}^{\dagger }a_{j}$ and $a_{i}^{\dagger },a_{i}$ are the creation and annihilation operators for a fermion particle state $i$.
In the presence of pulsed control, in analogy with the free evolution case, Eq. , we can rewrite the formula for the LE in Eq. in terms of the determinant of a $2N\times 2N$ matrix. Indeed the trace formula is straightforwardly generalized to products of more than two operators [@klich03] by using the following identity: $$\begin{aligned}
&&\left\langle \psi _{0}\right\vert e^{-i\mathcal{H}_{1}t}
e^{-i\mathcal{H}_{2}t}\ldots e^{-i\mathcal{H}_{n}t}\left\vert
\psi _{0}\right\rangle \notag \\
&=&\mathrm{det}\left( \mathbbm{1}-\mathbf{r_{0}}+\mathbf{r_{0}}
e^{-i\mathbf{C_{1}}t} e^{-i\mathbf{C_{2}}t}\ldots e^{-i\mathbf{C_{n}}t}\right) ,\end{aligned}$$ where we supposed that $\mathcal{H}_{k}=\sum_{i,j}[\mathbf{C_{k}}]_{ij} \:
a_{i}^{\dagger }a_{j}$ and $\mathbf{r_{0}}=\Gamma (\mathcal{N})$ with $\mathcal{N}$ occupation number operator \[i.e. $(\mathbf{r_{0}})_{ij}=\left\langle \psi_0 \right\vert
a_{i}^{\dagger }a_{j}\left\vert \psi_0 \right\rangle $\].
\[sec: AppendixLE\]
We evaluate here the leading order expansion of the LE in Eq. in terms of the pulse interval $\Delta t$, in the limit of short pulses. To simplify the notations, let us define $A\equiv i\mathcal{H}_{\downarrow}$, $B\equiv i\mathcal{H}_{\uparrow }$, and $\varepsilon \equiv \Delta t$. We consider Eq. at integer multiples of a spin-flip cycle, i.e., $t=2 M \Delta t$: $$\mathcal{L}(t)=\Big\vert \mathrm{Tr} \left[\left\vert G\right\rangle \left\langle
G\right\vert (e^{\varepsilon A} e^{\varepsilon B})^{M} (e^{-\varepsilon
A}e^{-\varepsilon B})^{M}\right] \Big\vert^2$$ Now recall the (approximate) Lie sum and product formulas $$\begin{aligned}
e^{\varepsilon A}e^{\varepsilon B} & = & e^{\varepsilon
(A+B)}+O(\varepsilon^{2}), \\
e^{\varepsilon A}e^{\varepsilon B}e^{-\varepsilon A}e^{-\varepsilon B} & = &
e^{\varepsilon^{2} [A,B]}+O(\varepsilon^{3}).\end{aligned}$$ Using this we have $$\begin{aligned}
& & (e^{\varepsilon A} e^{\varepsilon B})^{M} (e^{-\varepsilon A}
e^{-\varepsilon B})^{M} = \notag \\
& = & (e^{\varepsilon A} e^{\varepsilon B})^{M-1} [e^{\varepsilon^{2} [A,B]}
+ O(\varepsilon^{3})] (e^{-\varepsilon A} e^{-\varepsilon B})^{M-1} \notag
\\
& = & (e^{\varepsilon A} e^{\varepsilon B})^{M-2} [e^{\varepsilon^{2} [A,B]}
e^{\varepsilon A} e^{\varepsilon B} + O(\varepsilon^{2})] (e^{-\varepsilon
A} e^{-\varepsilon B})^{M-1} \notag \\
& & + O(\varepsilon^{3}) (e^{\varepsilon A} e^{\varepsilon B})^{M-1}
(e^{-\varepsilon A} e^{-\varepsilon B})^{M-1}.\end{aligned}$$ Keeping terms only to leading order $O(\varepsilon ^{2})$ we can neglect the last line, yielding:
$$\begin{aligned}
(e^{\varepsilon A} e^{\varepsilon B})^{M}(e^{-\varepsilon A} e^{-\varepsilon B})^{M}
& = & (e^{\varepsilon A}e^{\varepsilon B})^{M-2} [e^{\varepsilon ^{2} [A,B]} e^{\varepsilon A} e^{\varepsilon B}
e^{-\varepsilon A} e^{-\varepsilon B} + O(\varepsilon^{2})e^{-\varepsilon A}e^{-\varepsilon B}]
(e^{-\varepsilon A} e^{-\varepsilon B})^{M-2} \nonumber\\
& = & (e^{\varepsilon A} e^{\varepsilon B})^{M-2} [e^{2\varepsilon^{2} [A,B]} + O(\varepsilon ^{3})
+ O(\varepsilon ^{2}) (\mathbbm{1}-\varepsilon (A+B))] (e^{-\varepsilon A} e^{-\varepsilon B})^{M-2} \nonumber\\
& = & (e^{\varepsilon A} e^{\varepsilon B})^{M-2}[e^{2 \varepsilon ^{2} [A,B]} + O(\varepsilon ^{2}) \mathbbm{1}]
(e^{-\varepsilon A} e^{-\varepsilon B})^{M-2},
\end{aligned}$$
where in the last line we again neglected $O(\varepsilon^{3})$ terms. Continuing in this manner we have $$(e^{\varepsilon A} e^{\varepsilon B})^{M}(e^{-\varepsilon A} e^{-\varepsilon
B})^{M} = e^{M\varepsilon^{2} [A,B]} + M O(\varepsilon^{2}) \mathbbm{1},$$ which yields Eqs. -.
\[sec:Appendixeffectivedecay\]
Here we derive Eq. . We first notice that, in the spin-star case, both ${\mathcal H}_\uparrow (\lambda) \equiv {\mathcal H}_E (\lambda)$ and ${\mathcal H}_\downarrow (\lambda) \equiv {\mathcal H}_E (\tilde{\lambda})$ can be written in momentum space, by using the Jordan-Wigner transformation followed by a Fourier transform, in this way: $$\begin{array}{cl}
\displaystyle {\mathcal H}_E (\lambda) = 2 J \sum_{k>0}
& \hspace{-1.5mm} \displaystyle \big[
\varepsilon_{k}(\lambda)(c_{k}^{\dag} c_{k} +c_{-k}^{\dag} c_{-k}) \\
& \displaystyle - i\Delta_{k} (c_{k}^{\dag }c_{-k}^{\dag}-c_{-k}c_{k})
\big] \end{array}
\label{eq:Hferm}$$ where $\varepsilon_{k}(\lambda) = \lambda - \cos ( 2 \pi k / N)$ and $\Delta_{k} = \sin (2 \pi k /N)$, and the sum over $k$ runs from $1$ to $N/2$.
The ground state of the Hamiltonian in Eq. is $${\left\vertG(\lambda)\right\rangle} =\bigotimes_{k>0} \left[
\cos \Big( \frac{\theta_{k}}{2} \Big) {\left\vert00\right\rangle}_{k,-k}
+ i \sin \Big( \frac{\theta_{k}}{2} \Big) {\left\vert11\right\rangle}_{k,-k} \right] ,$$ where $\theta_{k}=\arctan ( \Delta _{k} / \varepsilon_{k}(\lambda) )$, and the kets refer to fermion occupation numbers in the two modes $k$ and $-k$. Consider now the space $$\nonumber
\mathbb{H}_{k} \otimes \mathbb{H}_{-k} = \mathrm{Sp}
\{ {\left\vert00\right\rangle}_{k,-k},{\left\vert01\right\rangle}_{k,-k}, {\left\vert10\right\rangle}_{k,-k}, {\left\vert11\right\rangle}_{k,-k} \} \, .$$ Since the subspaces $\mathrm{Sp}\{ {\left\vert00\right\rangle}_{k,-k}, {\left\vert11\right\rangle}_{k,-k} \}$ and $\mathrm{Sp}\{ {\left\vert01\right\rangle}_{k,-k}, {\left\vert10\right\rangle}_{k,-k} \}$ are not coupled by $\mathcal{H}_E$, and since ${\left\vertG(\lambda )\right\rangle}$ lives in the former two-dimensional subspace, we can rewrite the Hamiltonian over the $\mathrm{Sp}\{{\left\vert00\right\rangle}_{k,-k},{\left\vert11\right\rangle}_{k,-k}\}$ subspace, up to a constant, as $$\mathcal{H}_E (\lambda) = 2J \sum_{k>0} \left[ \varepsilon_{k}(\lambda) \Sigma_{k}^{z}
+ \Delta_{k} \Sigma_{k}^{y} \right]\equiv \sum_{k>0} \mathcal{H}_{E,k}(\lambda) \, ,$$ where $\Sigma_{k}^{z}$ and $\Sigma_{k}^{x}$ generate an SU$(2)$ algebra and are defined as $$\begin{aligned}
\Sigma_{k}^{x} & = & c_{k}^{\dag} c_{-k}^{\dag} + c_{-k} c_{k} , \\
\Sigma_{k}^{y} & = & -i(c_{k}^{\dag} c_{-k}^{\dag} - c_{-k} c_{k}) , \\
\Sigma_{k}^{z} & = & c_{k}^{\dag} c_{k} + c_{-k}^{\dag} c_{-k}-1 \, .\end{aligned}$$ The problem of evaluating $\left\langle G\right\vert
e^{it\mathcal{H}_{\mathrm{eff}}}\left\vert G\right\rangle $ is now reduced to computing the action of the $2\times 2$ matrix $[\mathcal{H}_{\downarrow }, \mathcal{H}_{\uparrow }]$ over the subspace $\mathrm{Sp}\{{\left\vert00\right\rangle}_{k,-k}, {\left\vert11\right\rangle}_{k,-k}\}$. We can rewrite the ground state as $$\begin{aligned}
{\left\vertG(\lambda)\right\rangle} & = & \bigotimes_{k>0} \left[
\cos \Big( \frac{\theta_{k}}{2} \Big) {\left\vert0\right\rangle}_{k}
+ i \sin \Big( \frac{\theta_{k}}{2} \Big) {\left\vert1\right\rangle}_{k} \right] \notag \\
& \equiv & \otimes_{k>0} {\left\vertG_{k}(\lambda)\right\rangle} \, ,\end{aligned}$$ where now ${\left\vert0\right\rangle}_{k}$ and ${\left\vert1\right\rangle}_{k}$ are the standard $\pm 1$ eigenstates of $\Sigma_{k}^{z}$. Over this subspace, using the fact that ${\mathcal H}_\downarrow (\lambda) = {\mathcal H}_\uparrow (\lambda) + \epsilon \Sigma^z$, with $\Sigma^z = \sum_{j=1}^N \sigma^z_j$, we have that $$\begin{aligned}
\mathcal{H}_{\mathrm{eff}} & = & i \epsilon \frac{\Delta t}{2}
[\Sigma^{z}, \mathcal{H}_{\uparrow}] =
4 i \, \epsilon_{\mathrm{eff}} \sum_{k>0} \Delta_{k}[\Sigma_{k}^{z},\Sigma_{k}^{y}] \notag\\
& = & 8 \epsilon_{\mathrm{eff}} \sum_{k>0} \Delta_{k} \Sigma_{k}^{x}
\equiv \sum_{k>0} \mathcal{H}_{\mathrm{eff}, k} \, .\end{aligned}$$ Now, $$\Sigma_{k}^{x} {\left\vertG_{k}(\lambda)\right\rangle} =\bigg[ \cos \Big( \frac{\theta_{k}}{2} \Big)
{\left\vert1\right\rangle}_{k} + i\sin \Big( \frac{\theta_{k}}{2} \Big) {\left\vert0\right\rangle}_{k} \bigg],$$ so that $$\left\langle G_{k}\right\vert \Sigma_{k}^{x} \left\vert G_{k}\right\rangle = 0,$$ and $$\begin{aligned}
\left\langle G_{k}\right\vert e^{i t \mathcal{H}_{\mathrm{eff}, k}}
\left\vert G_{k}\right\rangle & = & \left\langle G_{k}\right\vert e^{8i t \,
\epsilon_{\mathrm{eff}} \Delta _{k} \Sigma_{k}^{x}} \left\vert G_{k}\right\rangle
\notag \\
& = & \left\langle G_{k}\right\vert \cos (8 t \, \epsilon _{\mathrm{eff}} \Delta
_{k}) \mathbbm{1} \notag \\
& & - i\sin (8 t \, \epsilon _{\mathrm{eff}} \Delta _{k}) \, \Sigma_{k}^{x} \left\vert
G_{k}\right\rangle \notag \\
& = & \cos (8 t \, \epsilon_{\mathrm{eff}} \Delta_{k})\end{aligned}$$ Therefore $$\begin{aligned}
\left\langle G\right\vert e^{it\mathcal{H}_{\mathrm{eff}}} \left\vert
G\right\rangle & = & \Pi_{k>0} \left\langle G_{k}\right\vert e^{i t
\mathcal{H}_{\mathrm{eff}, k}} \left\vert G_{k}\right\rangle \notag \\
& = & \Pi_{k>0} \cos (8 t \, \epsilon_{\mathrm{eff}} \Delta _{k}),
\label{eq:LE-finalapp}\end{aligned}$$ which is Eq. .
[99]{}
M.A. Nielsen and I.L. Chuang, *Quantum Computation and Quantum Information*, (Cambridge University Press, Cambridge, 2000).
W.H. Zurek, Phys. Today **44**, 36 (1991); D. Giulini, E. Joos, C. Kiefer, J. Kupsch, I.-O. Stamatescu and H.-D. Zeh, *Decoherence and the Appearance of a Classical World in Quantum Theory*, (Springer, Berlin, 1996); M. Namiki, S. Pascazio and H. Nakazato, *Decoherence and Quantum Measurements*, (World Scientific, Singapore, 1997); W.H. Zurek, Rev. Mod. Phys. **75**, 715 (2003).
V.B. Magalinskij, Zh. Eksp. Teor. Fiz. **36**, 1942 (1959) \[Sov. Phys. JETP **9**, 1381 (1959)\]; R.P. Feynman and F.L. Vernon, Ann. Phys. (N.Y.) **24**, 118 (1963); G.W. Ford, M. Kac and P. Mazur, J. Math. Phys. **6**, 504 (1965); P. Ullersma, Physica **32**, 27 (1966); R. Zwanzig, J. Stat. Phys. **9**, 215 (1973); A.O. Caldeira and A.J. Leggett, Phys. Rev. Lett. **46**, 211 (1981); Physica A **121**, 587 (1983); Phys. Rev. A **31**, 1059 (1985); W.H. Zurek, Phys. Rev. D **26**, 1862 (1982); V. Hakim and V. Ambegaokar, Phys. Rev. A, **32**, 423 (1985); G.W. Ford and M. Kac, J. Stat. Phys. **46**, 803 (1987); H. Grabert, P. Schramm and G.-L. Ingold, Phys. Rev. Lett. **58**, 1285 (1987); G.W. Ford, J.T. Lewis and R.F. O’Connell, Phys. Rev. A **37**, 4419 (1988).
U. Weiss, *Quantum dissipative systems*, 2nd ed., World Scientific, Singapore, 1999.
N.V. Prokof’ev and P.C.E. Stamp, Rep. Prog. Phys. **63**, 669 (2000).
S. Paganelli, F. de Pasquale, and S.M. Giampaolo, Phys. Rev. A **66**, 052317 (2002).
L. Tessieri and J. Wilkie, J. Phys. A: Math. Gen. **36**, 12305 (2003).
D.V. Khveshchenko, Phys. Rev. B **68**, 193307 (2003).
H.-P. Breuer, D. Burgarth, and F. Petruccione, Phys. Rev. B **70**, 045323 (2004).
F.M. Cucchietti, J.P. Paz, and W.H. Zurek, Phys. Rev. A **72**, 052113 (2005).
C.M. Dawson, A.P. Hines, R.H. McKenzie, and G.J. Milburn, Phys. Rev. A **71**, 052321 (2005).
J. Lages, V.V. Dobrovitski, M.I. Katsnelson, H.A. De Raedt, and B.N. Harmon, Phys. Rev. E **72**, 026225 (2005).
H.T. Quan, Z. Song, X.F. Liu, P. Zanardi, and C.P. Sun, Phys. Rev. Lett. **96**, 140604 (2006).
F.M. Cucchietti, S. Fernandez-Vidal, and J.P. Paz, Phys. Rev. A **75**, 032337 (2007).
D. Rossini, T. Calarco, V. Giovannetti, S. Montangero, and R. Fazio, Phys. Rev. A **75**, 032333 (2007); J. Phys. A: Math. Theor. **40**, 8033 (2007).
M. Bortz and J. Stolze, Phys. Rev. B **76**, 014304 (2007).
W.H. Zurek, F.M. Cucchietti, and J.P. Paz, Acta Physica Polonica B, **38**, 1685 (2007).
Y. Hamdouni and F. Petruccione, Phys. Rev. B [**76**]{}, 174306 (2007).
S. Camalet and R. Chitra, Phys. Rev. B **75**, 094434 (2007).
G.A. Álvarez, E.P. Danieli, P.R. Levstein, and H.M. Pastawski, Phys. Rev. A **75**, 062116 (2007).
H. Krovi, O. Oreshkov, M. Ryazanov, and D.A. Lidar, Phys. Rev. A **76**, 052117 (2007).
C. Cormick and J.P. Paz, Phys. Rev. A [**77**]{}, 022317 (2008); eprint arXiv:0709.2643.
A. Relaño, J. Dukelsky, and R.A. Molina, Phys. Rev. E [**76**]{}, 046223 (2007).
S. Yuan, M.I. Katsnelson, and H. De Raedt, eprint arXiv:0711.2483.
H.G. Krojanski and D. Suter, Phys. Rev. A **74**, 062319 (2006).
D. Loss and D.P. DiVincenzo, Phys. Rev A **57**, 120 (1998).
R. de Sousa and S. Das Sarma, Phys. Rev. B **67**, 033301 (2003).
T. Gorin, T. Prosen, T.H. Seligman, and M. Žnidarič, Phys. Rep. **435**, 33 (2006).
A. Kitaev, Ann. Phys. **303**, 2 (2003).
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). For a review, see D.A. Lidar and K.B Whaley, in *Irreversible Quantum Dynamics*, Springer Lecture Notes in Physics **622**, 83, F. Benatti and R. Floreanini (eds.), (Springer, Berlin, 2003); eprint arXiv:quant-ph/0301032.
L. Viola and S. Lloyd, Phys. Rev. A **58**, 2733 (1998).
P. Zanardi, Phys. Lett. A **258**, 77 (1999); L. Viola, E. Knill, and S. Lloyd, Phys. Rev. Lett. **82**, 2417 (1999); L. Viola and E. Knill, Phys. Rev. Lett. **90**, 037901 (2003).
P. Facchi, D.A. Lidar, and S. Pascazio, Phys. Rev. A **69**, 032314 (2004).
P. Facchi and S. Pascazio, Phys. Rev. Lett. **89**, 080401 (2002).
M.S. Byrd, L.-A. Wu, and D.A. Lidar, J. Mod. Opt. **51**, 2449 (2004).
W.A. Anderson and F.A. Nelson, J. Chem. Phys. **39**, 183 (1963); R.R. Ernst, J. Chem. Phys. **45**, 3845 (1966); R. Freeman, S.P. Kempsell, and M.H. Levitt, J. Magn. Reson. **35**, 447 (1979); M.H. Levitt, R. Freeman, and T.A. Frenkiel, J. Magn. Reson. **47**, 328 (1982); J.S. Waugh, J. Magn. Reson. **50**, 30 (1982); A.J. Shaka, J. Keeler, and R. Freeman, J. Magn. Reson. **53**, 313 (1983); M.H. Levitt, R. Freeman, and T.A. Frenkiel, Adv. Magn. Reson. **11**, 47 (1983).
K. Khodjasteh and D.A. Lidar, Phys. Rev A **75**, 062310 (2007).
W. Zhang, V.V. Dobrovitski, L.F. Santos, L. Viola, and B.N. Harmon, Phys. Rev. B **75**, 201302(R) (2007).
K. Shiokawa and D.A. Lidar, Phys. Rev. A **69**, 030302(R) (2004); L. Faoro and L. Viola, Phys. Rev. Lett. **92**, 117905 (2004); G. Falci, A. D’Arrigo, A. Mastellone, and E. Paladino, Phys. Rev. A **70**, 040101(R) (2004).
A.M. Lane, Phys. Lett. A **99**, 359 (1983); W.C. Schieve, L.P. Horwitz, and J. Levitan, Phys. Lett. A **136**, 264 (1989); S.A. Gurvitz, Phys. Rev. B **56**, 15215 (1997); A.G. Kofman and G. Kurizki, Nature **405**, 546 (2000); P. Facchi, H. Nakazato, and S. Pascazio, Phys. Rev. Lett. **86**, 2699 (2001); M.C. Fischer, B. Gutiérrez-Medina, and M.G. Raizen, Phys. Rev. Lett. **87**, 040402 (2001).
P. Facchi, S. Tasaki, S. Pascazio, H. Nakazato, A. Tokuse, and D.A. Lidar, Phys. Rev. A **71**, 022302 (2005).
I. Klich, in *Quantum Noise in Mesoscopic Physics*, Nazarov Yu.V. Ed., NATO Science Series, Vol. 97 (Kluwer Academic Press, Dordrecht, 2003), eprint arXiv:cond-mat/0209642.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Within the framework of a five-dimensional brane world with a stabilized radion, we compute the cosmological perturbations generated during inflation and show that the perturbations are a powerful tool to probe the physics of extra dimensions. While we find that the power spectrum of scalar perturbations is unchanged, we show that the existence of the fifth dimension is imprinted on the spectrum of gravitational waves generated during inflation. In particular, we find that the tensor perturbations receive a correction proportional to $(HR)^2$, where $H$ is the Hubble expansion rate during inflation and $R$ is the size of the extra dimension. We also generalize our findings to the case of several extra dimensions as well as to warped geometries.'
author:
- 'Gian F. Giudice'
- 'Edward W. Kolb'
- Julien Lesgourgues
- Antonio Riotto
date: July 2002
title: Transdimensional physics and inflation
---
Introduction
============
The expanding Universe, especially if it underwent a primordial inflationary phase [@review], represents the most powerful probe of small distance scales at our disposal. Present-day astronomical length scales were extremely tiny at early epochs and were sensitive to short-distance physics. This simple observation has recently generated a lot of excitement about the possibility of opening a window on transplanckian or stringy physics in Cosmic Microwave Background (CMB) anisotropies [@tp]. Unfortunately, in the absence of a quantum theory of gravity, uncontrollable nonlinear effects may dominate at transplanckian distances, and the behavior of the cosmological perturbations and crucial related issues such as the definition of the vacuum remain unknown. This makes it difficult to predict on firm grounds the signatures of transplanckian physics on present-day cosmological scales [@tpf].
In this paper we will demonstrate that cosmological perturbations generated during inflation may nevertheless provide a powerful probe of another important aspect of many modern theories of particle physics: the existence of extra dimensions. The presence of extra dimensions is a crucial ingredient in theories explaining the unification of gravity and gauge forces. A typical example is string theory, where more than three spatial dimensions are necessary for the consistency of the theory. It has recently become clear that extra dimensions may be very large and could even be testable in accelerator experiments.
In theories with $n$ compactified extra dimensions with typical radii $R$, the four-dimensional Planck mass, $M_{P}$, is just a derived quantity, while the fundamental scale is the gravitational mass, $M_*$, of the $(n+4)$-dimensional theory. The mass scale $M_*$ is a free parameter and can range from a TeV to $M_{P}$, with $M_{P}^2\sim M_*^{n+2}\, R^n$. The size of extra dimensions can range from macroscopic scales down to Planckian distances.
In general there is a large hierarchy between the size of extra dimensions, $R$, and $M_*^{-1}$, with $R\gg M_*^{-1}$. This means that perturbations that are currently observable on cosmological scales might have been generated at early times on scales much smaller than the size of extra dimensions, but still on scales larger than the fundamental Planck mass so that the $(4+n)$-dimensional Einstein equations should describe gravity and the behavior of the quantum vacuum is more certain. This provides a unique probe of the physics of extra dimensions without the necessity of dealing with unknown effects at energies larger than $M_*$. This is particularly relevant in brane-world scenarios where gravity propagates in a higher-dimensional space while our visible Universe is a three-dimensional brane in the bulk of extra dimensions [@Arkani-Hamed:1998rs].
In this paper we initially assume a five-dimensional world where our visible Universe is a three-dimensional brane located at a given point in the fifth dimension. We consider the simplest possibility that inflation is a brane effect, [*i.e.,*]{} it is driven by a scalar field living on our three-dimensional brane, and study the effects of the transdimensional physics on the spectrum of the primordial density perturbations produced during the epoch of inflation.
Our findings indicate that despite the fact that the power spectrum of scalar perturbations remains unchanged, the existence of the fifth dimension is imprinted on the spectrum of gravitational waves generated during inflation. The tensor spectrum receives a correction proportional to $(HR)^2$, where $H$ is the Hubble rate during inflation and $R$ is the size of the extra dimension. Generalizing our results to the case of more than one extra dimension and to warped geometries, we show that the numerical coefficient of the correction term depends upon the details of the spacetime geometry of the extra dimensions. In four-dimensional single-field models of inflation there exists a consistency relation relating the amplitude of the scalar perturbations, the amplitude of the tensor perturbations, and the tensor spectral index. We compute the correction to such a consistency relation from transdimensional physics. Surprisingly enough, we find that at lowest order in the slow roll expansion, the four-dimensional relation is quite robust and does not suffer corrections from extra-dimensional physics, at least in not the cases addressed in this paper.
Some similar conclusions have been reached in Refs. [@lmw; @mwbh] for a particular five-dimensional setup in which the expansion law on the brane has a non-standard expression. Instead, we will focus on the case where the radius of the extra dimension is stabilized, leading to an ordinary Friedmann law. So any effect should be attributed to the non-trivial geometry along the extra dimension, rather than any modified cosmology on the brane. However, the formalism used in Ref. [@lmw] has many similarities with ours.
Our paper is organized as follows. In Section II we study the five-dimensional background with a stabilized radius. In Section III we compute the power spectrum of the tensor modes generated during inflation, while in Section IV we calculate the power spectrum of scalar perturbations. Section V is devoted to the consistency relation and Section VI to a generalization of our findings to more than one extra dimension and to warped geometries. Finally, in Section VII we draw our conclusions. The paper also contains an Appendix where we collect the background and perturbed Einstein equations.
A Five-dimensional background with a stabilized radion
======================================================
We consider a framework consisting of a $(3+1)$-dimensional brane embedded in a five-dimensional bulk with a stabilized radius. The coordinate along the extra dimension is taken to be $0 \leq y < 2 \pi R$ (eventually, one may consider to orbifold the circle by a $Z_2$-symmetry that identifies $y$ with $-y+2\pi R$ obtaining the segment $S^1/ Z_2$), and the brane is located at $y=0$ at zeroth order in the perturbations. Latin indices ($i=1,2,3$) label the ordinary three space dimensions; Greek indices ($\mu=0,1,2,3,5$) run over time, the three ordinary spatial dimensions, and the extra dimension $\mu=5$ (there is no $\mu=4$). The background metric may be taken to be of the form $$ds^2 = n^2(t,y) \, dt^2 - a^2(t,y) \delta_{ij} d x^i \, d x^j - dy^2 .
\label{backgroundmetric}$$ Through a redefinition of time we can always impose the condition $n(t,0)=1$ in order to obtain the familiar equations on the brane, where the induced metric is simply $ds^2 = dt^2 - a_0^2(t)\, \delta_{ij}\,dx^i\,dx^j$ [@footnotea0]. In the background metric there is no time-dependent $b^2$ term multiplying $dy^2$ because we assume that the radion is stabilized by some unknown high-energy mechanism, and we are free to set $b^2=1$. Then for consistency we must assume that the bulk energy-momentum tensor has a non-vanishing $(55)$ component [@Kanti:2000rd] that accounts for the radion stabilizing mechanism, while for simplicity we take the other components of the bulk energy-momentum tensor to be zero. This situation can be achieved by introducing a potential for the radion in the bulk that vanishes at the minimum and whose mass parameter is much larger than the other relevant mass scales. We will see in the following that our results can be generalized to cases with a non-vanishing bulk cosmological constant and a brane tension (like in the Randall–Sundrum framework [@rs]).
We suppose that the vacuum energy driving inflation is localized on our three-brane at $y=0$ so that the brane energy-momentum tensor provided by the inflaton brane-field $\varphi$ is of the form $T^\mu_{\ \ \nu}=\delta(y)\,{\rm
diag} (\rho,-p,-p,-p,0)$. This might be considered the simplest higher-dimensional scenario to investigate the effects of extra dimensions on cosmological scales. Of course, one may envisage extensions of our set up, such as assuming that the inflaton field $\varphi$ lives in the bulk made of one or more than one extra dimension, or that the spacetime geometry is warped. We will comment of these generalizations at the end of the paper.
We look for solutions of the background (unperturbed) Einstein equations [@ex]: $$\begin{aligned}
G^0_{\ 0} & = &
\frac{3}{n^2} \left( \frac{\dot{a}}{a} \right)^2 - 3 \left[ \frac{a''}{a}
+ \left( \frac{a'}{a} \right)^2 \right] = M_*^{-3} \ \delta(y) \ \rho(t) ,
\label{g00} \\
G^i_{\ j} & = & \left\{
\frac{1}{n^2}\left[2 \frac{\ddot{a}}{a}+\frac{\dot{a}}{a}
\left(\frac{\dot{a}}{a}-
2\frac{\dot{n}}{n}\right)\right] -2\frac{a''}{a}
-\frac{a'}{a}\left(\frac{a'}{a}+2\frac{n'}{n}\right)-\frac{n''}{n}
\right\}\delta_{ij} \nonumber \\
& = & - M_*^{-3} \ \delta(y) \ p(t) \ \delta_{ij} ,
\label{gij} \\
G^0_{\ 5} & = &
\frac{3}{n^2}
\left( \frac{\dot{a}}{a} \frac{n'}{n} - \frac{\dot{a}'}{a} \right) = 0 ,
\label{g05}\end{aligned}$$ where $M_*$ is the fundamental gravitational mass, $\rho$ is the energy density on the brane, and $p$ is the pressure on the brane. An overdot denotes derivation with respect to $t$, while a prime superscript denotes differentiation with respect to $y$. The $G^5_{\ 5}$ equation accounts for the stabilization of the radion and provides a constraint on $T^5_{\ 5}$, not on the metric. Other components vanish at zero order in perturbations. In general, the only solution of Eqs. (\[g00\])–(\[g05\]) such that $n(t,0)=1$ is easily found to be $$\begin{aligned}
a(t,y) &=& \dot{a}(t,0) \left[ y^2 - 2 \pi R y
+ \frac{6 \pi R M_*^3}{\rho(t)} \right]^{1/2} , \nonumber \\
n(t,y) &=& \frac{\dot{a}(t,y)}{\dot{a}(t,0)}.
\label{aandn}\end{aligned}$$ Note that the solution for the background metric is automatically $Z_2$-symmetric. Later, we will assume that this is also the case for metric perturbations.
The expression for $a(t,y)$ leads to the standard Friedmann law on the brane expected with a stabilized radion: $$H^2 \equiv \left[\frac{\dot{a}(t,0)}{a(t,0)}\right]^2 =
\frac{1}{2 \pi R M_*^3} \frac{\rho(t)}{3}.
\label{friedman}$$
In general, the presence of matter on the brane will cause a small readjustment of the radion with respect to its equilibrium value in vacuum. This shift generates corrections to the Friedmann law in Eq. (\[friedman\]) which are quadratic in $\rho$. Under our assumption that the radion is stabilized by a bulk potential characterized by a mass much larger than the other relevant energy scales, we can safely neglect these corrections.
From this expansion law we can define the four-dimensional gravitational constant to be $M_{P}^2 \equiv (8 \pi G)^{-1} \equiv 2 \pi R M_*^3$. Note that when $\pi R H \geq 1$, the solution for the scale factor is singular: it is not consistent to impose radion stability when the size of the extra dimension is larger than the Hubble radius, $H^{-1}$. (If the Hubble radius is interpreted as the causal horizon, this just means that the stabilization mechanism must remain causal.)
The singular nature of the scale factor for $\pi R H \geq 1$ is not an issue since we are only interested in the case in which the size of the extra dimension is smaller than the Hubble radius, $\pi R H \leq 1$, and the cosmological framework is expected to be almost described by four-dimensional physics (up to the correction factors that we wish to calculate).
Therefore, we assume that the energy density during inflation is smaller than $3 M_{P}^2 / (\pi R)^2$, and that deviations from the standard Friedmann law are suppressed up to this scale.
Matching the discontinuity in the components of the Einstein equations (\[g00\]) and (\[gij\]) gives the well known jump conditions for $a'$ and $n'$: $$\left[ \frac{a'}{a} \right]_{0}^{2 \pi R} = \frac{1}{3M_*^3}\ \rho ,
\qquad
\left[ \frac{n'}{n} \right]_{0}^{2 \pi R} = - \frac{1}{3M_*^3} \
(2 \rho + 3 p) ,$$ where, for any function $f$, we define $$[f]^\alpha_\beta \equiv
f(\alpha)-f(\beta)\, .$$ The restriction of the equation for $G^0_{\ 5}$ on the brane yields the usual energy conservation law for a perfect fluid: $\dot{\rho} + 3 H (\rho + p) = 0$.
When the brane only contains a homogeneous inflaton field $\bar{\varphi}(t)$ with potential $V$, the fluid energy conservation law gives the Klein–Gordon equation: $\ddot{\bar{\varphi}} + 3 H \dot{\bar{\varphi}} + \partial V /
\partial \bar{\varphi} = 0$.
If we assume that the density $\rho$ is constant over time, the scale factor is a separable function of time and $y$, and the brane undergoes de Sitter expansion: $$a(t,y) = a_0(t) n(y) , \qquad a_0(t) \propto \exp(H t) , \qquad
n(y) = \left[ H^2 (y^2 - 2 \pi R y) + 1 \right]^{1/2} .
\label{desitter}$$
The Primordial spectrum of tensor perturbations \[tensorsection\]
=================================================================
In this section we compute the present-day power spectrum of tensor modes generated by a primordial period of inflation on our visible brane at $y=0$.
The tensor perturbation of the metric is defined, as usual, in terms of a traceless transverse tensor $h_{ij}$ such that $$ds^2 = n^2(t,y) \, dt^2 - a^2(t,y) (\delta_{ij} + h_{ij})
d x^i \, d x^j - dy^2 .$$ One learns from the perturbed Einstein equations that the two degrees of polarization contained in $h_{ij}$ obey the wave equation $$\ddot{h} + \left(3 \frac{\dot{a}}{a} - \frac{\dot{n}}{n} \right) \dot{h}
- \frac{n^2}{a^2} \Delta h
- n^2 h'' - n^2 \left(3 \frac{a'}{a} + \frac{n'}{n} \right) h' = 0 ,
\label{eq-h}$$ where $h$ is normalized in such a way that $h^{ij} h_{ij} = h^2/2$.
Note that a free scalar field propagating in the bulk would have the same equation of motion as $h$. In the de Sitter background defined in Eqs.(\[desitter\]), and in a Fourier expansion with respect to the three spatial coordinates $x_i$, the equation reads $$\ddot{h}_k + 3 H \dot{h}_k + \frac{k^2}{a_0^2} h_k - n^2 h_k'' - 4 n'n \ h_k'
= 0 .
\label{eq.h}$$ We see from Eq. (\[desitter\]) that $(n^2)'$ is not continuous on the brane, and has a jump $$\left[ 2nn' \right]_0^{2 \pi R} = 4 \pi R H^2 .$$ This implies that $(n^2)''$ contains a $\delta$ function: $$(n^2)'' = 2 H^2 \left[1 - 2 \pi R \ \delta(y) \right] .$$ The solutions of the mode equation, Eq. (\[eq.h\]), are separable in $t$ and $y$, so we can expand $h_k$ in a sum of Kaluza–Klein modes: $$h_k= a_0^{-3/2} n^{-2}
\sum_p \chi_p(t) g_p(y) ,
\label{def.gn}$$ where $\chi_p(t)$ and $g_p(y)$ satisfy the equations $$\begin{aligned}
\ddot{\chi}_p + \left( \frac{k^2}{a_0^2} - \frac{9}{4} H^2 + \omega_p^2 \right)
\chi_p & = & 0 \label{eq-chin} , \\
n^2 g_p'' + \left[ - (n^2)'' + \omega^2_p \right] g_p & = & 0 .
\label{eq-gn}\end{aligned}$$ In the definition of Eq. (\[def.gn\]), the factor $a_0^{-3/2} n^{-2}$ was introduced just for simplicity so that Eqs. (\[eq-chin\]) and (\[eq-gn\]) contain no friction terms. The equation for $g_p$ has to two independent solutions, which are given in terms of the Gauss hypergeometric function ${_2}F_1$, $$g_p = c_1\ {_2}F_1 \left( \frac{-1-b}{4},\frac{-1+b}{4},\frac{1}{2},-x^2
\right) +
c_2\ x\ {_2}F_1 \left( \frac{1-b}{4},\frac{1+b}{4},\frac{3}{2},-x^2
\right) ,$$ where the parameters $b$ and $x$ are defined by $$b \equiv \sqrt{9-4\frac{\omega_p^2}{H^2}}, \qquad
x\equiv \frac{(y-\pi R)H}{\sqrt{1-(\pi RH)^2}}.$$ The metric continuity condition $g_p(0)=g_p(2 \pi R)$ eliminates the solution odd with respect to $(y- \pi R)$, and fixes $c_2=0$. The constant $c_1$ is determined by the normalization condition of $g_p$. We will set the wave function normalization condition to be $$\int_{0}^{2 \pi R} \! \! \! dy \ n^{-2} \left| g_p \right| ^2 = 1 .
\label{wfnorm}$$ Integrating Eq. (\[eq-gn\]) in a neighborhood of the brane leads to the jump condition for $g'$: $$\left[ g_p' \right]_0^{2 \pi R} = 4 \pi R H^2 g_p(0) .
\label{gpdisc}$$ This condition is satisfied only for a discrete set of possible values of $\omega_p$, determined by the equation [@eigenvalues] $${_2}F_1 \left( \frac{-1-b}{4},\frac{-1+b}{4},\frac{1}{2},-a
\right) = \frac{(a+1)(b^2-1)}{8}\
{_2}F_1 \left( \frac{3-b}{4},\frac{3+b}{4},\frac{3}{2},-a
\right) ,
\label{eigenvalue}$$ with $a$ given by $$a \equiv \frac{(\pi R H)^2}{1-(\pi R H)^2}.$$
As far as the time dependence is concerned, the solution of Eq.(\[eq-chin\]) for $\chi_p$ is a Bessel function, and can be normalized to the adiabatic vacuum inside the Hubble radius using as usual the positive frequency condition and the canonical commutation relations. To do so, we start from the five-dimensional action $$S = \frac{1}{8}
\int dt\ dy\ d^3\!x \ M^3_* n a^3 \left( n^{-2} \dot{h}^2
- a^{-2} \delta^{ij} \partial_i h \partial_j h - {h'}^2 \right) .$$ Note that the factor of $na^3$ simply comes from the term $\sqrt{-g}$. In Fourier space and with the de Sitter background the action is $$S = \frac{1}{8} \int dt\ dy\ d^3\!k \ M^3_* n^2 a_0^3
\left[ \dot{h}_k \dot{h}^*_k
+ \frac{k^2}{a_0^2} h_k h^*_k - n^2 h'_k {h^*_k}' \right] .$$ Following Eq. (\[def.gn\]), we can expand each mode along the basis formed by the functions $g_p$. We define $$\begin{aligned}
I_{mp} &=& \int_0^{2 \pi R} \! dy\ n^{-2} g_m \ g_p^* ,
\nonumber \\
J_{mp} &=& \int_0^{2 \pi R} \! dy \left(g_m' - 2 \frac{n'}{n} g_m \right)
\left({g_p^*}' - 2 \frac{n'}{n} g_p^* \right) .\end{aligned}$$ After integration over $y$, the effective action reads $$\begin{aligned}
S & = & \frac{1}{8} \int dt \ d^3\!k \ M_*^3 \sum_{m,p}
\left\{ \left[
\dot{\chi}_m \dot{\chi}^*_p - \frac{3}{2} H
(\chi_m \dot{\chi}^*_p + \dot{\chi}_m \chi^*_p)
+ \left(\frac{k^2}{a_0^2} + \frac{9}{4} H^2 \right) \chi_m \chi^*_p \right]
I_{mp} \right. \nonumber \\
& & \left. \phantom{\left(\frac{k^2}{a_0^2}\right)}
- \chi_m \chi^*_p J_{mp} \right\} .\end{aligned}$$ Even without knowing explicitly the expression for the $g_p$’s, we can find $I_{mp}$ and $J_{mp}$. The first intermediate step is to integrate by parts $$\int_0^{2 \pi R} \! dy \ g_m \ {g_p^*}'' = - \int_0^{2 \pi R} \! dy \
g_m' \ {g_p^*}' + 4 \pi R H^2 g_m(0) \ g_p^*(0) .$$ So, the integral on the left-hand side has the hermitian symmetry. Then, we use Eq. (\[eq-gn\]) and write $$\begin{aligned}
g_m \ {g_p^*}'' + \left[- 2 H^2 \left[1 - 2 \pi R \ \delta(y) \right]
+ \omega^2_p\right] n^{-2} g_m \ g_p^*
&=& 0 , \nonumber \\
g_m'' \ g_p^* + \left[- 2 H^2 \left[1 - 2 \pi R \ \delta(y) \right]
+ \omega^2_m\right] n^{-2} g_m \ g_p^* &=& 0 .\end{aligned}$$ We subtract these two equations and integrate over $y$, taking advantage of the previously found symmetry. We are left with $$(\omega^2_p - \omega^2_m) \int_0^{2 \pi R} \! dy \ n^{-2} g_m \ g_p^* = 0 .$$ So, unless $\omega^2_p=\omega^2_m$, the above integral vanishes. Given the wave function normalization condition of Eq. (\[wfnorm\]), we conclude that $I_{mp} = \delta_{mp}$. We can also integrate by parts $$J_{mp} = \int_0^{2 \pi R} \! dy \ \left[n^{-2} \left(g_m' -
2 \frac{n'}{n} g_m \right) \right] \left( n^2 {g_p^*}' - 2 n' n g_p^* \right) .$$ Using the equation of motion and the jump condition for $g_p$, we find $J_{mp}
= \omega_p^2 \delta_{mp}$. So, the effective four-dimensional action is diagonal: $$S = \sum_{p} \frac{1}{8} \int dt \ dk^3 \ M_*^3
\left[ \dot{\chi}_p \dot{\chi}^*_p + \left( \frac{k^2}{a_0^2}
+ \frac{9}{4} H^2 - \omega^2_p \right) \chi_p \chi^*_p - \frac{3}{2} H
(\chi_p \dot{\chi}^*_p + \dot{\chi}_p \chi^*_p) \right] .$$ Each Kaluza–Klein mode $\chi_p$ has the same action as a free field in four-dimensional de Sitter spacetime, and can be quantized following the standard procedure. Namely, the adiabatic vacuum can be defined in the sub-horizon limit $k/a_0 \gg H$ in which Minkowski spacetime is asymptotically recovered. Then the canonical commutation relation gives $$\hat{p}_{\chi_p} = \frac{1}{8} M_*^3 \dot{\chi}_p , \qquad
\left[ \hat{\chi}_p, \hat{p}_{\chi_p}^{\dagger} \right] = i,$$ which leads to the Wronskian condition $\chi_p \dot{\chi}_p^* - \dot{\chi}_p
\chi_p^* = 8i/M_*^3$.
Let us focus on the zero mode. Equation (\[eq-gn\]) with $\omega_p=0$ has the obvious solution $c_0 n^2$, where $c_0$ is a constant of integration. This solution is automatically continuous and satisfies the jump condition on the brane. The constant of integration is obtained from the condition of Eq.(\[wfnorm\]): $$c_0 = \left( \int_0^{2 \pi R} \! dy \ n^2 \right)^{-1/2}
= \left[ 2 \pi R \left(1 - \frac{2}{3} \pi^2 R^2 H^2\right) \right]^{-1/2} .$$ The function $\chi_0(t)$ is a Bessel function of index $3/2$ and has a simple analytic expression (we retain only the positive frequency solution and we normalize with the Wronskian condition above), $$\chi_0 = \frac{2}{M_*^{3/2}} \sqrt{\frac{a_0}{k}}
\left(i \frac{a_0H}{k} + 1 \right) \exp\left(i \frac{k}{a_0H}\right) .$$
After horizon crossing, $|\chi_0|$ grows like $a_0^{3/2}$. The behavior of the other Kaluza–Klein modes depends on the sign of $\omega_p^2 - 9H^2/4$. As shown in Fig. \[fig\], even $\omega_1^2$ is larger than $9H^2/4$ except for a marginal range when $\pi HR$ is very close to 1. So, all massive Kaluza–Klein modes oscillate at late time with a constant amplitude, and are quickly suppressed with respect to the zero mode by the factor $a_0^{3/2}$.
=4.5in
In addition, the spectrum of the massive Kaluza–Klein modes is extremely blue, as is usually the case for a scalar field with a mass larger than the Hubble parameter during inflation. This means that for astronomical scales of interest the contribution to the tensor spectrum from Kaluza–Klein modes is practically zero. So, we can focus on the asymptotic value of $h_k$ arising from the zero-mode contribution $$h_k(t,0) \rightarrow a_0^{- 3/2} \chi_0(t) g_0(0) \rightarrow i
\left[2 \pi R M_*^3 \left(1 - \frac{2}{3} \pi^2 R^2 H^2\right) \right]^{-1/2}
\frac{2}{k^{3/2}} \ H . \label{ash}$$ Even when the de Sitter stage ends, Eq. (\[eq-h\]) shows that the zero mode remains frozen on wavelengths larger than the horizon, as is the case in four-dimensional physics. Therefore, the primordial spectrum of gravitational waves at horizon re-entry is still given by Eq. (\[ash\]), where $H$ has to be evaluated at the time of the first horizon crossing during inflation: $${\cal P}_T(k) \equiv \frac{k^3}{2\pi^2} \left|h_k(y=0)\right|^2 =
\frac{2}{\pi^2}\left(\frac{H(k)}{M_{P}}\right)^2 \frac{1}{1 - 2\pi^2 R^2
H^2(k)/3}=\frac{\left. {\cal P}_T(k)\right|_{4D}}{1 - 2\pi^2 R^2 H^2(k)/3} ,
\label{psh}$$ where $H(k)$ indicates the value of the Hubble parameter when a given wavelength $\lambda=2\pi/k$ crosses the horizon, [*i.e.*]{}, when $k=a_0H$. The power spectrum of tensor perturbations is normalized such that in a critical density universe the energy density (per octave) in gravitational waves, $\Omega_g(k)$, is related to ${\cal P}_T(k)$ in terms of the transfer function, $T_g^2(k)$, by (the transfer function is discussed in Ref. [@G]) $$\Omega_g(k) = \frac{1}{24} T_g^2(k) {\cal P}_T(k) .$$ The power spectrum of tensor modes is therefore enhanced compared to the four-dimensional result by a factor $(1 - 2\pi^2 R^2 H^2/3)^{-1}$.
This correction factor has a simple explanation: it originates from the zero-mode wave function normalization. However, it can be understood also in terms of the effective gravitational Planck mass $\left.M_{P}^2\right|_{I}$ [*during inflation*]{}, defined by integrating the zero-mode action over $y$. Indeed, the zero mode, $h_0 \propto n^{-2} g_0$, is constant along $y$ (as expected for a free field with no source localized on the brane). So, integrating the zero-mode action gives a factor (setting $a_0=1$ to isolate the gravitational coupling) [@effective] $$\left.M_{P}^2\right|_{I}= M_*^3
\int_0^{2 \pi R} \! \! dy \ \sqrt{- g} \ g^{00} = M_*^3 \int dy \ n^2
= M_{P}^2\left(1 - \frac{2}{3} \pi^2 R^2 H^2\right) .
\label{efpl}$$ This shows that the enhancement of the tensor power spectrum can be rephrased as a shift in the effective gravitational constant during inflation, when it is defined from the effective gravitational action rather than from the expansion law. The tensor perturbation is a purely five-dimensional field, while the expansion law obtains from a density localized on the brane. This implies that the latter depends on the value of $n^2$ on the brane, not on its average. Physically what happens is that the vacuum energy density present on our visible brane during inflation warps the spacetime geometry in the bulk. This effect is manifest in the nontrivial shape of the functions $a^2(y)$ and $n^2(y)$ during the inflationary epoch. As a result, the graviton zero mode, which is free to spread out in the bulk, feels a smaller Planck mass during inflation. Therefore, today we receive a flux of gravitational waves primordially generated during inflation which is larger than its four-dimensional counterpart because during inflation gravity was stronger.
The Primordial spectrum of scalar perturbations
===============================================
In this section we compute the present-day power spectrum of scalar modes generated by a primordial period of inflation on our visible brane at $y=0$.
The first-order scalar perturbations of the metric can be expressed as [@perturb] $$\begin{aligned}
ds^2 & = & n^2 (1 + 2 \phi) \, dt^2
- a^2 [(1 - 2 \psi)\delta_{ij} + 2 \partial_i \partial_j E ] \, d x^i \, d x^j
+ 2 \partial_i B \, d x^i \, d t \nonumber \\
& &
+ 2 \partial_i w \, dx^i \, dy + 2 \delta g_{05} \,dt \, dy
- (1 - \delta g_{55}) \, dy^2 .\end{aligned}$$ The perturbed brane position is specified by another function, $\delta y (t,
x^i)$.
Five-dimensional gauge transformations of the form $x^{\mu} \rightarrow x^{\mu}
+ \xi^{\mu}$, where $\xi^{\mu} = (\xi^0, \partial^i \xi, \xi^5)$, induces the transformations $$\begin{aligned}
\phi &\rightarrow& \phi + \dot{\xi}^0 + \frac{\dot{n}}{n} \xi^0
+ \frac{n'}{n} \xi^5 , \nonumber \\
\psi &\rightarrow& \psi - \frac{\dot{a}}{a} \xi^0 - \frac{a'}{a} \xi^5 ,
\nonumber \\
E &\rightarrow& E + \xi ,\nonumber \\
B &\rightarrow& B + n^2 \xi^0 - a^2 \dot{\xi} , \nonumber \\
w &\rightarrow& w + \xi^5 + a^2 \xi' , \nonumber \\
\delta g_{05} &\rightarrow& \delta g_{05} + \dot{\xi}^5 - n^2 {\xi^0}' ,
\nonumber \\
\delta g_{55} &\rightarrow& \delta g_{55} + 2 {\xi^5}' ,\end{aligned}$$ while the new brane position is $\delta y (t, x^i)
+ \xi^5(t,x^i,\delta y (t, x^i))$.
We will work in a particular gauge, the Gaussian normal gauge. The same gauge choice was made in [*e.g.,*]{} Ref. [@gauge]. In order to eliminate $\delta g_{55}$, we choose $2 {\xi^5}' = - \delta g_{55}$. This fixes the function $\xi^5(t, x^i, y)$ up to a boundary condition, [*i.e.,*]{} up to an arbitrary function of $(t, x^i)$ on one hypersurface (for instance, on the brane). The most convenient boundary condition is $\xi^5(t,x^i,\delta y (t,
x^i)) = - \delta y (t, x^i)$, in order to shift the brane position to $y=0$, even at first order in perturbations. Similarly, in order to eliminate $w$, we may choose $a^2 \xi' = - w - \xi^5$, with the boundary condition $\xi = - E$ on the brane in order to have $E=0$ on the brane. Finally, in order to eliminate $\delta g_{05}$, we choose $n^2 {\xi^0}' = \delta g_{05} + \dot{\xi}^5$, with the boundary condition $n^2 \xi^0 = - B + a^2 \dot{\xi}$, so that $B$ also vanishes on the brane. Of course $E$ and $B$ are still non-zero in the bulk. The perturbed metric reduces to $$ds^2 = n^2 (1 + 2 \phi) \, dt^2
- a^2 [(1 - 2 \psi) \delta_{ij} + 2 \partial_i \partial_j E ] \,
d x^i \, d x^j
+ 2 \partial_i B \, d x^i \, d t - dy^2 .$$ The induced metric on the brane is diagonal, and involves only the perturbations $\phi_0$ and $\psi_0$. It is identical to the four-dimensional perturbed metric in the so-called longitudinal gauge. Since the system is symmetric in $y \longleftrightarrow 2 \pi R - y$, the perturbations are expected to be even functions with respect to $(y - \pi R)$. In the following, the terms “even” and “odd” will be meant always with respect to $(y - \pi
R)$.
We give in Eqs. (\[Einstein.00\])–(\[Einstein.i5\]) the expression of the perturbed Einstein equations in the Gaussian normal gauge. In general, the restrictions of the $G^0_{\ 5}$ and $G^i_{\ 5}$ equations on the brane provide the continuity and Euler equations. When the brane contains only a perturbed scalar field $\varphi (t,x^i) = \bar{\varphi}(t) + \delta \varphi (t,x^i)$, the $\delta G^i_{\ 5}$ equation is trivially satisfied, while the $\delta G^0_{\
5}$ equation gives the standard perturbed Klein–Gordon equation: $$\delta \ddot{\varphi} + 3 \frac{\dot{a}_0}{a_0} \delta \dot{\varphi}
+ \left( \frac{\partial^2 V}{\partial \varphi^2}
- \frac{\Delta}{a_0^2} \right) \delta \varphi
= \dot{\bar{\varphi}} (\dot{\phi}_0 + 3 \dot{\psi}_0) - 2
\frac{\partial V}{\partial \varphi} \phi_0 .
\label{pert-KG}$$ The other components of the perturbed Einstein equations contain some second derivatives with respect to $y$ that have to be matched with source terms on the brane. However, to first order in the perturbations, the scalar field cannot generate anisotropic stress on the brane: $\delta T_{ij}$ is proportional to $\delta_{ij}$. This imposes the continuity of $E''$ across the brane, and therefore, since $E'$ is odd, $E'_0=0$. The other perturbations are sourced on the brane and have to satisfy the jump conditions $$\begin{aligned}
-3 \left[\psi'\right]_0^{2\pi R} & = & M_*^{-3} \delta T^0_{\ 0}
= \dot{\bar{\varphi}} \ \delta \dot{\varphi} - \dot{\bar{\varphi}}^2 \phi
+ \frac{\partial V}{\partial \varphi} \delta \varphi , \nonumber \\
- \left[\phi'\right]_0^{2\pi R} + 2 \left[\psi'\right]_0^{2\pi R} & = &
- M_*^{-3} \delta T^i_i
= \dot{\bar{\varphi}} \ \delta \dot{\varphi} - \dot{\bar{\varphi}}^2 \phi
- \frac{\partial V}{\partial \varphi} \delta \varphi , \nonumber \\
- \frac{1}{2} \left[B'\right]_0^{2\pi R}
& = & M_*^{-3} \dot{\bar{\varphi}} \ \delta \varphi . \end{aligned}$$
A master equation for the scalar perturbations
----------------------------------------------
We would like to find an equation of motion for a single variable that would account for the full scalar perturbation dynamics, as Eq. (\[eq-h\]) did for tensor perturbations. Such a master equation has already been found in the case of a maximally-symmetric background spacetime [@muk], but not in cases where $T_{55}$ accounts for the radion stabilization. The best approach is to work with a set of variables reflecting some gauge-invariant quantities. By studying 5-dimensional gauge transformations, it is straightforward to show that one can build four independent gauge-invariant quantities out of the full set of scalar perturbations of the metric. In our gauge, these quantities reduce to $$\begin{aligned}
\Psi_{\phi} & = & \phi - \frac{\dot{B}}{n^2}
+ a_0^2 [ - \ddot{E} - 2 H \dot{E} + n'n E'] , \nonumber \\
\Psi_{\psi} & = & \psi + H \frac{B}{n^2}
+ a_0^2 [ H \dot{E} - n'n E'] , \nonumber \\
\Psi_{05} & = & B' - 2 \frac{n'}{n} B
+ 2 a_0^2 n^2 [ \dot{E}' + H E' ] , \nonumber \\
\Psi_{55} & = & 2 a_0^2 [ n^2 E'' + 2 n'n E' ] .\end{aligned}$$ Similarly, it is possible to build a gauge-invariant quantity out of the scalar field perturbation $\delta \varphi$ and the metric perturbations $E$ and $B$. However, in our gauge $E$ and $B$ vanish on the brane, so $\delta
\varphi$ directly reflects the gauge-invariant field perturbation. Although the Einstein tensor is not gauge invariant, some of its components can be expressed in terms of $(\Psi_{\phi}, \Psi_{\psi}, \Psi_{05}, \Psi_{55})$. We write the Einstein equations in the de Sitter background, first in terms of $(\phi, \psi, E, B)$ (see Eqs. (\[Einstein.DeSitter.00\])–(\[Einstein.DeSitter.i5\]) of the Appendix), and then in terms of the above variables (some terms in $E$ and $E'$ still remain). The traceless part of $\delta G^i_{\ j}$ just gives $$\Psi_{\psi} - \Psi_{\phi} - \frac{1}{2} \Psi_{55} = 0 ,
\label{traceless}$$ and allows us to eliminate easily one of the four variables: instead of ($\Psi_{\psi}$, $\Psi_{\phi}$, $\Psi_{05}$, $\Psi_{55}$), we can work with ($\Sigma=\Psi_{\psi} + \Psi_{\phi}, \Psi_{05}, \Psi_{55}$). Then, the equation for $\delta G^i_{\ 0}$ provides a simple relation between $\Sigma$ and $\Psi_{05}$: $$2n^2 \left(\dot{\Sigma} + H \Sigma\right) + \left(n^2 \Psi_{05}\right)'
= 2 M_*^{-3} \dot{\bar{\varphi}} \delta \varphi \ \delta(y) .
\label{rel.Sigma.05}$$ The equation for $\delta G^0_{\ 0}$, combined with the previous one, gives a relation between $\Sigma$ and $\Psi_{55}$: $$\begin{aligned}
& & \frac{3}{2} \left( n^2 \Sigma'' + 4 n' n \Sigma' \right)
+ 3 H (H \Sigma + \dot{\Sigma} ) + \frac{\Delta}{a_0^2} \Sigma
+ \frac{3}{2} \left( \frac{n^2}{2} \Psi_{55}'' + 3 n'n \Psi_{55}'
+ 3 H^2 \Psi_{55} \right) \nonumber \\
& & \ \ \ \ \ \ = M_*^{-3} \left[ \delta T^0_{\ 0} + 6 H \dot{\bar{\varphi}}
\delta \varphi + 3 H^2 \ \Psi_{55}(y=0) \right] \delta(y) .
\label{rel.Sigma.55}\end{aligned}$$ Finally, the system is closed, for instance, by the equation for $\delta G^i_{\
5}$. Indeed, the quantity $a_0^2 n^{-2} (n^4 \delta G^i_{\ 5})'=0$ can be combined with the previous constraints in Eqs. (\[rel.Sigma.05\]) and (\[rel.Sigma.55\]), to lead to a master equation for $\Sigma$: $$\begin{aligned}
\ddot{\Sigma} - H \dot{\Sigma} - \frac{\Delta}{a_0^2} \Sigma
- n^2 \Sigma'' - 4 n'n \Sigma' - 2 H^2 \Sigma
& = & M_*^{-3} \left[ \delta T^i_{\ i} - \delta T^0_{\ 0} + 2 a_0
\left( \frac{\dot{\bar{\varphi}} \delta \varphi}{a_0} \right)^. \right]
\delta(y) \nonumber \\
&=& 2 M_*^{-3} \left( - H \dot{\bar{\varphi}} \delta \varphi
+ \ddot{\bar{\varphi}} \delta \varphi + \dot{\bar{\varphi}}^2 \phi_0
\right) \delta(y) .
\label{master}\end{aligned}$$ In the bulk, this equation looks like a five-dimensional wave equation, and in terms of the rescaled variable $a_0^{-2} \Sigma$, it would be identical to that of the tensor perturbations or to that of a canonically normalized bulk scalar field. The first difference with the tensor case is the presence of a source localized on the brane, which imposes a jump condition on the derivative $\Sigma'$: $$- \frac{1}{2} \left[\Sigma'\right]_0^{2\pi R}
= M_*^{-3} \left( H \dot{\bar{\varphi}} \delta \varphi
- \ddot{\bar{\varphi}} \delta \varphi - \dot{\bar{\varphi}}^2 \phi_0 \right) .
\label{Sigma.jump}$$ A similar condition would be found in the case of a bulk scalar field sourced on the brane. However, there is a second difference, reflecting the complicated structure of the Einstein equations, and the integro-differential relations between the various perturbations and boundary conditions. By integrating Eq. (\[rel.Sigma.05\]) over the circle $0 \leq y \leq 2 \pi R$, we get an integrability condition for $\Sigma$: $$\{ H + \partial_t \} \int_0^{2 \pi R} \! \! \! dy \ n^2 \Sigma =
M_*^{-3} \dot{\bar{\varphi}} \delta \varphi .
\label{Sigma.integ}$$
Any even solution of the master equation, Eq. (\[master\]), matching the jump and integrability conditions, Eqs. (\[Sigma.jump\]) and (\[Sigma.integ\]), provides a solution of the full Einstein equations. At any time one can compute the perturbations $(\phi_0, \psi_0)$ on the brane. Indeed, the integral of Eq. (\[rel.Sigma.55\]) over $y$ gives $\Psi_{55}(y=0)$ as a function of $\Sigma$ and of the scalar field (remembering that $\left[\Psi_{55}'\right]_0^{2\pi R}$ is given by the jump conditions, while $\int_0^{2 \pi R} dy \Psi_{55} = 2 a_0^2 \left[n^2 E'\right]_0^{2\pi
R}=0$): $$2 \pi R H^2 \ \Psi_{55} (y=0) =
- \left\{ \partial_t^2 + H \partial_t - \frac{\Delta}{3 a_0^2} \right\}
\int_0^{2 \pi R} \! \! \! \! \! dy \ \Sigma \
+ \frac{2}{3} M_*^{-3} \left( \dot{\bar{\varphi}} \delta \dot{\varphi}
+ 2 \ddot{\bar{\varphi}} \delta \varphi + 2 \dot{\bar{\varphi}}^2
\phi_0 \right) .
\label{Psi.55.0}$$ Finally, the perturbations on the brane may be found from $$\psi_0 = \frac{1}{2} \left[ \Sigma (y=0)
+ \frac{1}{2} \Psi_{55} (y=0) \right] , \qquad
\phi_0 = \frac{1}{2} \left[ \Sigma (y=0)
- \frac{1}{2} \Psi_{55} (y=0) \right] .$$
Comparison to the four-dimensional solution
-------------------------------------------
In four dimensions, the absence of anisotropic stress implies $\phi_0=\psi_0$ and scalar metric perturbations are described by a single variable (matching the Newtonian gravitational potential inside the Hubble radius). The exact computation of the scalar power spectrum for scales leaving the Hubble radius during inflation can be performed by various methods that take into account the coupling between the metric and the scalar field perturbations [@reviewrocky]. For instance, one can integrate the equation of propagation of the Mukhanov variable which is a combination of $\delta
\varphi$ and $\phi_0$ [@brand]. A second possibility is to solve a pair of coupled equations: the perturbed Klein–Gordon equation (\[pert-KG\]), and one of the Einstein equations, for instance that for $\delta G^i_{\ 0}$: $$\dot{\phi}_0 + H \phi_0 = \frac{\dot{\bar{\varphi}} \ \delta \varphi}{2 M_P^2}.
\label{4D.metric.constraint}$$ In the present work, we are interested in the generalization of the scalar power spectrum calculation only at the leading order in slow-roll parameters. In four dimensions, such a calculation is done consistently by writing all the equations of propagation in the exact de Sitter background; simultaneously, in the expression for the perturbations one keeps the leading-order term in the expansion parameter $\dot{H}/H^2$ \[or equivalently $\dot{\bar{\varphi}}^2/(H^2
M_{P}^2)$\], treated as a constant. In that case, the sourcing of the field perturbation by the metric perturbation \[described by the right-hand side in the Klein–Gordon equation Eq. (\[pert-KG\])\] can be safely neglected, and one can solve the homogeneous equation $$\delta \ddot{\varphi} + 3 H \delta \dot{\varphi}
+ \left( \frac{\partial^2 V}{\partial \varphi^2}
+ \frac{k^2}{a_0^2} \right) \delta \varphi = 0 .
\label{free.pert}$$ At leading order in slow-roll parameters we can also neglect $|V''| \ll H^2$ and write the solution, normalized to the adiabatic vacuum inside the Hubble radius, as $$\delta \varphi = \frac{1}{a_0 \sqrt{2 k}} \left( 1 + i \frac{H a_0}{k} \right)
\exp\left(i \frac{k}{a_0 H}\right) .$$ The metric perturbation is just following the field evolution, according to Eq. (\[4D.metric.constraint\]). This gives in the two limits (sub-Hubble-radius and super-Hubble-radius scales): $$\begin{aligned}
\delta \varphi & \simeq & \frac{1}{a_0 \sqrt{2k}} \
\exp\left(i \frac{k}{a_0 H}\right) , \quad
\phi_0 = \frac{i \dot{\bar{\varphi}}}{2 M_P^2 \sqrt{2 k^3}} \
\exp\left(i \frac{k}{a_0 H}\right) \quad
\left( \textrm{for}\,\, \frac{k^2}{a_0^2} \gg H^2 \right) , \label{4d.sw} \\
\delta \varphi & \simeq & \frac{H}{\sqrt{2k^3}} ,
\hspace*{90pt} \phi_0 = \frac{\dot{\bar{\varphi}}}{2 M_P^2 \sqrt{2k^3}}
\hspace*{81pt} \left( \textrm{for}\,\, \frac{k^2}{a_0^2} \ll H^2 \right) .\end{aligned}$$ In both limits, we see that the right-hand side in the Klein–Gordon equation is always very small, and acts as a source for the field perturbations only at a sub-leading order in $\dot{H}/H^2$: $$\left| \dot{\bar{\varphi}} (\dot{\phi}_0 + 3 \dot{\psi}_0) - 2
\frac{\partial V}{\partial \varphi} \phi_0 \right| \sim
\left| \frac{\dot{\bar{\varphi}}^2}{M_P^2} \delta \varphi \right|
\sim | \dot{H} \delta \varphi | \ll H^2 |\delta \varphi| .$$ Another way to reach the same conclusion is to look at the Mukhanov variable $\xi= a_0 ( \delta \varphi + \dot{\bar{\varphi}}\phi_0/H)$. It is straightforward to show that at leading order in slow-roll parameters, $\delta
\varphi$ and $\xi/a_0$ are equal and share the same equations of propagation. The distinction becomes only relevant when slow roll is marginally or temporarily violated, for instance, in inflationary models with phase transitions.
We reviewed this point in the four-dimensional case because it is crucial for the five-dimensional calculation where we will also assume that at leading order $\delta \varphi$ can be treated as a free field, obeying Eq.(\[free.pert\]). We will check [*a posteriori*]{} that the metric perturbations just follow the field and do not alter $\delta \varphi$ in a significant way.
Long-wavelength solution in the five-dimensional model
------------------------------------------------------
If the assumption that the scalar metric perturbations just follow the scalar field perturbations is correct, then it is sufficient to study the coupling between these two degrees of freedom in the long-wavelength regime in order to know the power spectrum $k^3 |\phi_0|^2$ on super-Hubble-radius scales during inflation. This is our purpose in this section. In the next section we will study the short-wavelength solution for consistency.
On super-Hubble-radius scales the scalar field perturbations are approximately constant in time: $\delta \varphi = H / \sqrt{2 k^3}$ (we dropped the arbitrary phase). We first look for a particular solution of the inhomogeneous master equation, Eq. (\[master\]), neglecting the Laplacian term. In slow roll only the first of the three contributions to the source term on the right-hand side is relevant because $|\ddot{\bar{\varphi}} \ \delta \varphi| \ll | H
\dot{\bar{\varphi}} \ \delta \varphi|$ and $ |\dot{\bar{\varphi}}^2
\phi_0 | \sim | \dot{H} \ \phi_0 | \ll H^2 |\phi_0|$. A particular solution is found to be $$\Sigma_i = \frac{1}{2 \pi R M_*^3}
\frac{\dot{\bar{\varphi}} \delta \varphi}{H n^2} .
\label{part}$$ This solution is constant in time, but not in $y$. It matches both the jump and integrability conditions Eqns. (\[Sigma.jump\]) and (\[Sigma.integ\]) (still at leading order in $\dot{H}/{H^2}$). We are free to add to Eq.(\[part\]) a solution of the homogeneous equation, [*i.e.*]{}, any even solution $\Sigma_h$ of the sourceless master equation $$\ddot{\Sigma}_h - H \dot{\Sigma}_h
- n^2 \Sigma_h'' - 4 n'n \Sigma_h' - 2 H^2 \Sigma_h = 0$$ such that $\left[\Sigma_h'\right]_0^{2\pi R}=0$ and $$\{ H + \partial_t\} \int_0^{2 \pi R} \! \! \! \! \! dy \ n^2 \ \Sigma_h = 0 .
\label{Sigma.hom.integ}$$
We almost already have the solutions because the homogeneous master equation written in terms of $(a_0^{-2} \Sigma_h)$ is the same as the equation for the tensor modes. So $\Sigma_h$ is a sum of separable solutions (the Kaluza–Klein modes) with the same values of $\omega_p$ as for the tensors, but with a different normalization condition, Eq. (\[Sigma.hom.integ\]). For the zero mode with $\omega_p=0$, the solution reads $(\Sigma_{h})_{p=0} = C_1 a_0^2 +
C_2 a_0^{-1}$, but the normalization condition imposes $C_1 = 0$. For heavy modes with $\omega_p \geq 3H/2$, one gets on super-Hubble-radius scales $$\begin{aligned}
(\Sigma_{h})_p =C_p a_0^{1/2} \exp\left(\pm i t \ \sqrt{\omega_p^2 -
\frac{9}{4} H^2}
\right) \ n^{-2}(y) g_p(y) , \end{aligned}$$ where $g_p$ is an even solution of Eq. (\[eq-gn\]). But, since the integral of $g_p$ over $y$ does not vanish, the integrability condition imposes $C_p=0$: in the limit under consideration, the Einstein equations are not compatible with any significant contribution of heavy Kaluza–Klein modes. The final solution reads $$\Sigma = \frac{1}{2 \pi R M_*^3}
\frac{\dot{\bar{\varphi}} \delta \varphi}{H n^2} + \frac{C_2}{a_0} ,
\label{sixtyfour}$$ and corresponds to the usual combination of a growing adiabatic mode driven by the scalar field, and a decaying mode that could be normalized only if the full solution was known (from inside the horizon). Since the decaying mode becomes rapidly negligible it will not concern us. Then, we can compute $\Psi_{55}(y=0)$ using Eq. (\[Psi.55.0\]). The leading-order terms are $$2 \pi R H^2 \ \Psi_{55}(y=0)
\simeq - H \partial_t \int_0^{2 \pi R} \! \! dy \ \Sigma +
\frac{2}{3} M_*^{-3} \dot{\bar{\varphi}} \delta \dot{\varphi} \simeq
-\frac{2}{3} \pi R \dot{H} \ \Sigma (y=0) \left[ 1 + {\cal O}(H^2 R^2) \right].
\label{eq:leading}$$ So, to first order in $\epsilon = -\dot{H}/{H^2} = M_{P}^2 (V'/V)^2/2$, we recover $\psi_0=\phi_0 = \Sigma_0/2$, and the metric perturbations on the brane match exactly the four-dimensional result: $$\phi_0 = \psi_0 = \frac{1}{4 \pi R M_*^3}
\frac{\dot{\bar{\varphi}}}{H} \delta \varphi =
\frac{\dot{\bar{\varphi}}}{2 M_P^2 \sqrt{2k^3}}.$$
Since our results indicate that on long wavelengths the perturbations of the gravitational potential coincides with the four-dimensional one, the power spectrum of scalar curvature perturbations will be given by the usual result (we recall that during the de Sitter stage, the large wavelength metric perturbations are related to the curvature perturbations by a factor $\epsilon$) $${\cal P}_S (k) = \frac{k^3}{2 \pi^2} \frac{\phi_0^2}{\epsilon^2} =
\frac{1}{8\pi^2}\frac{1}{\epsilon}\left(\frac{H(k)}{M_{P}}\right)^2 .
\label{cur}$$ Here ${\cal P}_S(k)$ is defined in terms of the observable power spectrum $P(k)$ and the scalar transfer function $T^2(k)$ by $$\frac{k^3}{2\pi^2}P(k) = \left(\frac{k}{aH}\right)^4 T^2(k) {\cal P}_S(k) .$$
In Eq. (\[cur\]), the limit $\epsilon=0$ is singular as in four dimensions. This corresponds to the exact de Sitter limiting case, for which $\dot{\bar{\varphi}}=\ddot{\bar{\varphi}}=0$. Then, the master equation Eq.(\[master\]) and the constraint equations Eqs.(\[Sigma.jump\],\[Sigma.integ\]) have vanishing right-hand sides. The single solution for $\Sigma$ at large wavelength is the decaying mode, $C_2/a_0$ found in Eq. (\[sixtyfour\]). Plugging this mode into Eq.(\[Psi.55.0\]) shows that $\Psi_{55}(y=0)=0$ and $\phi_0 = \psi_0 = C_2 / 2
a_0 $. We reach the same conclusion as in four dimensions: for exact de Sitter expansion, the scalar metric perturbations do not have a non-decaying solution at large wavelength.
It is appropriate to re-emphasize our result that the scalar spectrum is unaltered is only true to lowest order in the slow-roll parameters. For instance, one sees that Eq. (\[eq:leading\]) has corrections of order $(H^2R^2)$, but they are multiplied by $\epsilon$, so to lowest order in the slow-roll parameters we can ignore them.
The absence of any correction factor in $(RH)^2$ can be interpreted in the following way: Unlike the tensor degrees of freedom, which are five-dimensional free fields quantized [*in the bulk,*]{} the scalar metric perturbations only follow the scalar field. The later is quantized [*on the brane*]{} and has the same behavior as in the four-dimensional case in all regimes. So we only need to study the coupling between the field and the metric in the long-wavelength regime when the metric evolves as in the four-dimensional theory. Moreover, the coupling is localized on the brane, so that no signature remains from the non-trivial geometry in the bulk.
Short-wavelength solution in the five-dimensional model
-------------------------------------------------------
In the short-wavelength limit, the scalar field perturbations can be approximated by $\delta \varphi = (a_0 \sqrt{2k})^{-1} \exp[-i(k/a_0)t]$, and $k/a_0$ is a slowly varying parameter ($\delta \dot{\varphi} = - i (k/a_0)
\delta \varphi$). We write the master equation with $H=0$ (and accordingly, with $n=1$): $$\ddot{\Sigma} - \frac{\Delta}{a_0^2} \Sigma - \Sigma'' = 0 .
\label{master.short}$$ The brane source term that was proportional to $H$ has disappeared since the jump in $\Sigma'$ across the brane is found to be negligible is the limit in which the matter perturbations on the brane behave like a fluid with sound speed $c_s^2 = -1$ ([*i.e.*]{}, $\delta T^0_{\ 0} = - \delta T^i_{\ i}$). Since $\Sigma'$ is odd, this implies $[\Sigma']_0^{2\pi R} =0$. The most general even solution of Eq. (\[master.short\]) can be written as a Kaluza–Klein expansion (with $p=0,1,...,\infty$): $$\Sigma = \sum_{p = - \infty}^{+ \infty} \Sigma_p
= \sum_{p = - \infty}^{+ \infty}
\left[ c_p \exp(i \nu_p t) +
d_p \exp(- i \nu_p t) \right] \cos \left[\omega_p (y- \pi R) \right] ,$$ $$\nu_p \equiv \left( \frac{k^2}{a_0^2} + \omega_p^2 \right)^{1/2} ,$$ where $(c_p, d_p)$ are constants of integration and the $\omega_p$’s are imposed by the continuity of $\Sigma'$: $$\omega_p = \frac{p}{R} , \qquad p=0,1,...,\infty .$$ Inserting the general solution into the integrability condition Eq.(\[Sigma.integ\]), which now reduces to $$\int_0^{2 \pi R} \! dy \dot{\Sigma}
= M_*^{-3} \dot{\bar{\varphi}} \delta \varphi ,$$ gives the two constraints $$c_0=0, \quad d_0 = \frac{1}{2 \pi R M_*^3}
\frac{i \dot{\bar{\varphi}}}{\sqrt{2 k^3}}.$$ This implies $$\Sigma_{p=0} = \frac{1}{2 \pi R M_*^3} \
i \frac{a_0}{k} \dot{\bar{\varphi}} \delta \varphi ,
\label{sigma.0}$$ while the constants of integration for $p \geq 1 $ remain arbitrary. In other words, the zero mode is driven by the scalar field, as would be the case for $\phi_0$ in four dimensions, while the heavy Kaluza–Klein modes are independent of the matter on the brane. So, the quantization should be done first for the zero mode and the scalar field together since they only represent one independent degree of freedom, then for each heavy Kaluza–Klein mode separately following the same procedure as for the tensor modes (normalization to the adiabatic vacuum).
Let us focus on the quantization of the zero mode plus the field, since we know that the heavy Kaluza–Klein modes will decouple when $kR/a_0 \ll 1$. We first must determine whether it is consistent to assume that the metric zero-mode contribution to the perturbed Klein–Gordon equation is very small, and that the field perturbation can be quantized as a free field, while the metric zero-mode just follows. In order to discover the answer, we need to compute the contribution to $\left[\Psi_{55}\right]_0^{2\pi R}$ arising from the zero mode only. This is done by integrating Eq. (\[rel.Sigma.55\]), which now simplifies to $$\frac{3}{2} \Sigma'' - \frac{k^2}{a_0^2} \Sigma + \frac{3}{4} \Psi_{55}''
= M_*^{-3} \delta T^0_{\ 0} \ \delta(y) .$$ Replacing $\Sigma$ by $\Sigma_{p=0}$, and using the constraint $\int_0^{2 \pi R} \! dy \, \Psi_{55}=0$, we get $$(\Psi_{55})_{p=0} = \frac{2}{3} \frac{k^2}{a_0^2}
\left[(y - \pi R)^2 - \frac{1}{3} (\pi R)^2 \right]
\Sigma_{p=0}.$$ By evaluating this relation at $y=0$ we find that the relation between $\phi_0$ and $\psi_0$ arising from the zero mode is $$\psi_0 - \phi_0 =
\frac{2 (k \pi R)^2}{9 a_0^2}
(\psi_0 + \phi_0) .
\label{rel.phi0.psi0}$$ Let us first examine the limit $H R \ll k R / a_0 \ll 1$ in which we expect to recover the four-dimensional results. Indeed, in this case we find from Eqs.(\[sigma.0\]) and (\[rel.phi0.psi0\]) that $$\phi_0 = \psi_0 = \frac{1}{4 \pi R M_*^3} \
i \frac{a_0}{k} \dot{\bar{\varphi}} \delta \varphi
= \frac{1}{2 M_P^2} \ i \frac{a_0}{k} \ \dot{\bar{\varphi}} \delta \varphi ,$$ which is exactly the four-dimensional result of Eq. (\[4d.sw\]). On the other hand, in the limit $k R / a_0 \gg 1$ the contribution to $\phi_0$ and $\psi_0$ arising from the zero mode is $$- \phi_0 = \psi_0 = \frac{(k \pi R)^2}{ 9\ a_0^2}
\frac{1}{2 \pi R M_*^3} \ i \frac{a_0}{k} \dot{\bar{\varphi}} \delta \varphi .$$ So, the right-hand side in the perturbed Klein-Gordon equation is of order $$\left| \dot{\bar{\varphi}} ( \dot{\phi}_0 + 3 \dot{\psi}_0 ) \right|
= \left(
\frac{\sqrt{2} k \pi R}{3 \ a_0}
\right)^2
\frac{\dot{\bar{\varphi}}^2}{2 \pi R M_*^3}
\left|
\delta \varphi
\right|
= \left(
\frac{2 k \pi R}{3 \ a_0}
\right)^2
\left| \dot{H} \delta \varphi
\right| \ll
\left(
\frac{2 \pi R H \ k}{3 a_0}
\right)^2
\left|
\delta \varphi
\right|,$$ where we used the slow-roll inequality $\epsilon = -\dot{H}/H^2 \ll 1$. If we remember that $ \pi R H < 1$ and that the leading terms in the homogeneous perturbed Klein-Gordon equation are of order $(k/a_0)^2 | \delta \varphi |$, we see that even in this regime, the dynamics of the scalar field is unaffected by that of the metric perturbations. This justifies the assumption that we made in the long wavelength regime that the field dynamics is the same as in four-dimensional physics (same vacuum normalization and same evolution).
The consistency relation \[consistencysection\]
===============================================
Four-dimensional single-field models of inflation predict a consistency relation [@reviewrocky] relating the amplitude of the scalar perturbations, ${\cal P}_S(k)$, the amplitude of the tensor perturbations, ${\cal P}_T(k)$, and the tensor spectral index, $n_T\equiv d\ln {\cal P}_T(k)/d\ln k$.
Indeed, in four dimensions since ${\cal P}_T(k)\propto H^2(k)$, $n_T$ is given by $n_T = d\ln H^2(k)/d\ln k = -2\epsilon$. The four-dimensional consistency relation is $$\left. \frac{{\cal P}_T(k)}{{\cal P}_S(k)} \right|_{4D}= 16 \epsilon = -8 n_T .$$ In the five-dimensional universe, however, both the amplitude and the tilt of the tensor power spectrum receive corrections which are functions of $(RH)^2$.
If we parameterize the $R$-dependent corrections to the power spectrum of tensor modes as $${\cal P}_T(k) = \frac{2}{\pi^2} \left[\frac{H(k)}{M_{P}}\right]^2
\frac{1}{1 - \alpha R^2 H^2(k)} ,$$ (recall that our result was $\alpha=2\pi^2/3$) we may compute the spectral index of the tensor modes to be $$n_T = \frac{d\ln {\cal P}_T(k)}{d\ln H^2(k)}\frac{d\ln H^2(k)}{d\ln k} =
\frac{1}{1 - \alpha R^2 H^2}
\frac{d\ln H^2(k)}{d\ln k} =-\frac{2\epsilon}{1 - \alpha R^2 H^2} ,
\label{t}$$ where we have used the fact that the change of the Hubble parameter as a function of scale, $d\ln H^2(k)/d\ln k$, is still given by $-2\epsilon$ since the inflaton field is a brane field and the Hubble rate still satisfies the four-dimensional equation $\dot{H} = -\epsilon H^2$.
Using Eqs. (\[psh\]), (\[cur\]), and (\[t\]), we find $$\frac{{\cal P}_T(k)}{{\cal P}_S(k)} =
\frac{16\epsilon}{1 - \alpha R^2 H^2}=-8 n_T .$$ This is a particularly surprising result: the consistency relation remains unaltered at lowest order in the slow-roll parameters. A similar result has been found in Ref. [@hl] for a set-up where the bulk on either side of the brane corresponds to Anti-de Sitter [*AdS*]{} spaces with different cosmological constants. This degeneracy between the usual result in 4-D one-field inflation and in extra-dimensional models will make it more difficult to disentangle the various theoretical possibilities from observations. Our results hold in the case in which only curvature perturbations are generated during the inflationary phase. If isocurvature perturbations are produced, the consistency relation in brane world scenarios is expected to differ from the one obtained here as it happens in the four-dimensional case [@bmr].
Of course this conclusion depends strongly on the particular form of the corrections to the power-spectrum of the tensor perturbations, $(1 - \alpha R^2
H^2)^{-1}$, which holds in our five-dimensional example. One can show that this is actually the only possible functional dependence on $H$ such that the consistency relation remains unaltered.
However, our result seems to be quite robust: even considering more than one extra-dimension or even a Randall–Sundrum like scenario, we show in the next section that the power spectrum of tensor perturbations always gets corrections of the same form, as long as the radius is completely stabilized during inflation.
Of course our calculation of the consistency relation is only to lowest-order in slow-roll parameters. In general, one expects the usual four-dimensional corrections to the lowest-order result, corrections from the five-dimensional background equations if $H$ is not constant, and corrections of order $(H^2R^2)$ to the scalar perturbations as indicated in Eq. (\[eq:leading\]).
Generalization of the results
=============================
We now would like to generalize some of our considerations to the case of more than one extra dimension in the case in which the sizes of the extra dimensions are all equal. If we assume that the compactified geometry of the extra dimensions is stabilized, we can take the background metric in the form of Eq.(\[backgroundmetric\]) with $dy^2\equiv \delta_{\alpha \beta} dy^\alpha
dy^\beta$, $\alpha, \beta =1,\dots ,\delta$. Equations (\[g00\]) and (\[g05\]) for $\delta$ extra dimensions become $$\begin{aligned}
G^0_{\ 0} & = & \frac{3}{a^2}\left( \frac{\dot{a}^2}{n^2}-\frac{1}{2}
\partial_\alpha \partial^\alpha a^2 \right) = M_*^{-(2+\delta)} \rho(t) \
\delta (y_1)\dots \delta (y_\delta),
\label{g00p} \\
G^0_{\ \alpha} & = & -\frac{3}{an} \partial_\alpha
\left( \frac{\dot{a}}{n}\right) =0 .
\label{g05p}\end{aligned}$$
Let us now make the simplifying assumption that because of rotational symmetry in the extra dimensions the scale and lapse functions $a$ and $n$ depend only on the distance from the brane, $r\equiv (\sum_\alpha y_\alpha^2
)^{1/2}$, and on time, but not on the angular variables. Then, Eq. (\[g00p\]) becomes $$G^0_{\ 0} = -\frac{3}{2a^2}\left[ \frac{d^2a^2}{dr^2}+\frac{(\delta -1)}{r}
\frac{da^2}{dr}-2 K^2\right] = M_*^{-(2+\delta)}\rho(t)
\frac{1}{S_\delta r^{\delta -1}} \delta (r) ,
\label{g00pp}$$ where $S_{\delta}$ is the surface of a unit-radius sphere in $\delta$ dimensions. The lapse function is given by $n=\dot{a}/K$, where $K$ is independent of $r$, so we can choose $K= \dot{a}(\bar r )/n(\bar r)$, evaluated at an arbitrary point $r=\bar r$. For $\delta >1$, the solution of Eq.(\[g00pp\]) becomes singular at $r=0$, where the brane is located.
To overcome this difficulty, we define a brane with a finite thickness, and impose our boundary conditions at $r=\epsilon$, keeping only the leading terms in the limit $\epsilon \to 0$. We assume that appropriate density terms within the brane smooth the singularity at the origin. Next, we require compactification conditions that, for simplicity, involve only the variable $r$, and we impose that the values of the scale factor $a$ at $r=\epsilon$ and at $r=2\pi R -\epsilon$ are equal. Then, the solution of Eq. (\[g00pp\]) is $$a=a_\epsilon \sqrt{1-(\pi RH)^2 c(r)},$$ with $c(r)$ given by $$c(r) = \left\{
\begin{array}{ll}
\Frac{2r}{\pi R} - \left( \Frac{r}{\pi R}\right)^2
& \quad \textrm{for } \delta =1 \\ & \\
\Frac{2 \ln r/\epsilon}{\ln 2\pi R/\epsilon} -\Frac{1}{2}
\left( \Frac{r}{\pi R}\right)^2 & \quad \textrm{for } \delta =2
\\ & \\
\Frac{4}{\delta} \left[ 1-\left( \Frac{\epsilon}{r}\right)^{\delta
-2} \right] -\Frac{1}{\delta}\left( \Frac{r}{\pi R}\right)^2 &
\quad \textrm{for } \delta \ge 3.
\end{array}
\right.$$ Here $a_\epsilon =a(r=\epsilon)$, and $H=\dot{a}_\epsilon / (n_\epsilon
a_\epsilon )$ is determined to be proportional to $\sqrt{\rho}$ by the jump condition of $d a/dr$. In the case $\delta =1$ we can safely take the limit $\epsilon \to 0$ and we reproduce Eq. (\[aandn\]).
The correction factor for the tensor perturbations can be obtained following the same argument used in Sec. \[tensorsection\]. The effective gravitational Planck mass during inflation is $$M_P^2|_I =M_*^{\delta -2} \int d^\delta \!y \ n^2= M_P^2\left[ 1- (\pi RH)^2
C\right],$$ with $$C\equiv \frac{\int_\epsilon^{2\pi R -\epsilon} dr \ r^{\delta -1} c(r)}
{\int_\epsilon^{2\pi R -\epsilon} dr \ r^{\delta -1} }.$$ The coefficient $C$ depends on the compactification geometry. In the five-dimensional case studied in Sec. \[tensorsection\], we found $C=2/3$, while in the simplified case of $\delta$ extra dimensions compactified as explained above, we obtain $C=8/[\delta (\delta +2)]$ (for $\delta \ge 2$).
In conclusion, our result that the tensor perturbations are enhanced by a factor $\left[ 1- (\pi RH)^2 C\right]^{-1}$ is a generic consequence of our set up with a stabilized geometry. The factor $C$ depends on the details of the compactification and on the number of extra dimensions, but it is typically a number of order unity. This form of corrections for the tensor perturbations (together with the usual four-dimensional result for the scalar perturbations) have the specific property of preserving the consistency relation, as discussed in Sec. \[consistencysection\].
We can also extend our result to the case of non-factorizable geometries [@rs]. Let us consider a 5-dimensional set up with two branes at $y=0$ and $y=\pi R$, with vacuum densities equal in magnitude but opposite in sign ($V_0=-V_\pi$), and a cosmological constant $\Lambda$ in the bulk. We assume the usual relation between $V_0$ and $\Lambda$ to obtain the warped Randall–Sundrum metric in vacuum, and we include a constant energy density $\rho$ on the negative-tension brane. Imposing the condition that the compactification radius is stabilized during the cosmological evolution, we obtain that the scale and lapse functions are given by [@outstanding] $$a=a_\pi \frac{n}{\Omega}, \qquad \Omega \equiv e^{-\pi KR},$$ $$n^2=e^{-2K|y|}\left[ 1+\frac{(2\Omega^2 -1)H^2}{4\Omega^2K^2}\right]
+\left( \frac{e^{2K|y|}}{2}-1 \right) \frac{\Omega^2H^2}{2K^2},$$ where the 5-dimensional coordinate $y$ is defined in the interval $(-\pi R,
\pi R)$. Here $K\equiv \Lambda /V_\pi$ is the inverse of the [*AdS*]{} radius, and $H \equiv {\dot a}_\pi / (n_\pi a_\pi)$ is the Hubble constant on the visible brane located at $y=\pi R$, with $a_\pi =a(y=\pi R)$ and $n_\pi=
\Omega$. The Hubble constant is related to the energy density $\rho$ by the equation $$H^2=\frac{K\rho}{3M_*^3 (1-\Omega^2)}.$$
We can now compute the Planck mass during inflation, which is given by $$M_P^2|_I \equiv M_*^3 \int_{-\pi R}^{\pi R} dy \ n^2 =
\frac{M_*^3}{K} (1-\Omega^2) \left[ 1+\frac{H^2}{K^2} \left(
\frac{3\Omega^2-1}{4\Omega^2}-\frac{\pi RK \Omega^2}{1-\Omega^2}\right)\right].$$ Recalling that for the Randall–Sundrum model the Planck mass of the 4-dimensional effective theory is given by $M_P^2=(M_*^3/K) (1-\Omega^2)$, we obtain in the limit of small warp factor $\Omega \ll 1$, $$\left.M_P^2\right|_I =M_P^2 \left[ 1- \left(\frac{x_1H}{2m_1}\right)^2\right] .$$ Here $m_1=x_1 K\Omega$ is the mass of the first graviton Kaluza–Klein mode (in the limit $H=0$), and $x_1=3.8$ is the first root of the Bessel function $J_1$. Therefore, the correction factor is quadratic in $H$ also in the case of factorizable geometries, and the typical scale is determined by the Kaluza–Klein mass gap.
We should note that in the case of the Randall–Sundrum model, the energy separation between the Kaluza–Klein graviton mass and the fundamental scale at which the theory becomes strongly interacting is often very small and this is a limitation for the applicability of our result. Uncomputable quantum gravity effects can become important and lead to comparable contributions.
Conclusions
===========
In this paper we have initiated the investigation of the effects of transdimensional physics on the spectrum of the cosmological density perturbations generated during a period of primordial inflation taking place on our visible three-brane. We have shown that the size of the transdimensional effects are of order $(HR)^2$, where $H$ is the Hubble parameter during inflation and $R$ is the typical size of the extra dimensions (or, more precisely, the inverse of the Kaluza–Klein mass gap at zero temperature). The corrections appear in the power spectrum of the tensor modes. The coefficient of the corrections depends upon the compactification geometry, the number of extra dimensions, and if they are flat or warped. As we have already stressed in the Introduction, our treatment should be unaffected by (unknown) quantum effects which might arise at distances below $M_*^{-1}$ as long as the size of extra dimensions is larger than $M_*^{-1}$.
Our results may be generalized in different ways. First of all, our set up is the simplest we could imagine: only one extra dimension and inflation taking place on the brane. One can envisage the possibility of putting the inflaton field responsible for inflation in the bulk. In such a case we expect a different form of corrections. In particular, the power spectrum of scalar perturbations should be modified, thus possibly changing the consistency relation.
One can also relax our working assumption of keeping the radii of extra dimensions fixed. In this case there might be significant corrections to the slope of the power spectra since having a dynamical radion field during inflation amounts to change the Hubble rate during inflation.
In our paper we have also assumed that the energy density $\rho$ on the brane is smaller than about $ M_{P}^2 / R^2$, or equivalently, that the Hubble radius is larger than the radii of compactification. Deviations from the standard four-dimensional Friedmann law are present in the opposite regime and large deviations from the standard results for the power spectra of density perturbations should appear.
Finally, we have assumed that deep in the ultraviolet regime, at distances much smaller than the horizon length, the initial vacuum is the traditional Bunch–Davies vacuum containing no initial particles in the spectrum. This is a reasonable assumption at physical momenta $k/a_0$ much larger than $R^{-1}$, but still smaller than the fundamental scale $M_*$. Of course, for momenta $k/a_0\gg M_*$, unknown quantum effects may take over and change drastically the properties of the vacuum. This would lead to corrections scaling as powers of $H/M_*$ as suggested by the analysis performed in the four-dimensional cases [@tp]. Nevertheless, whenever $M_*R$ is sufficiently large, the computable corrections discussed in this paper dominate over these unknown quantum corrections.
A.R. would like to thank S. Matarrese, D. Lyth, and D. Wands for several discussions and the Theory Group of CERN where part of this work was done. J.L. acknowledges a visit to the University of Padua during which this project was initiated. The work of E.W.K. was supported in part by the Department of Energy and by NASA (NAG5-10842).
[99]{}
For a review, see D. H. Lyth, and A. Riotto, Phys. Rept. [**314**]{}, 1 (1999).
J. Martin and R. H. Brandenberger, Phys. Rev. D [**63**]{}, 123501 (2001); J. C. Niemeyer, Phys. Rev. D [**63**]{}, 123502 (2001); A. Kempf, Phys. Rev. D [**63**]{}, 083514 (2001); J. Martin and R. H. Brandenberger, astro-ph/0012031; J. C. Niemeyer and R. Parentani, Phys. Rev. D [**64**]{}, 101301 (2001); A. Kempf and J. C. Niemeyer, Phys. Rev. D [**64**]{}, 103501 (2001); A. A. Starobinsky, Pisma Zh. Eksp. Teor. Fiz. [**73**]{}, 415 (2001) \[JETP Lett. [**73**]{}, 371 (2001)\]; R. Easther, B. R. Greene, W. H. Kinney, and G. Shiu, Phys. Rev. D [**64**]{}, 103502 (2001); L. Hui and W. H. Kinney, Phys. Rev. D [**65**]{}, 103507 (2002); M. Lemoine, M. Lubo, J. Martin, and J. P. Uzan, Phys. Rev. D [**65**]{}, 023510 (2002); R. Easther, B. R. Greene, W. H. Kinney, and G. Shiu, hep-th/0110226; R. H. Brandenberger, S. E. Joras, and J. Martin, hep-th/0112122; N. Kaloper, M. Kleban, A. E. Lawrence, and S. Shenker, hep-th/0201158; J. Martin and R. H. Brandenberger, Phys. Rev. D [**65**]{}, 103514 (2002); R. H. Brandenberger and J. Martin, hep-th/0202142; G. Shiu and I. Wasserman, Phys. Lett. B [**536**]{}, 1 (2002); H. Danielsson, hep-th/0203198; S. F. Hassan and M. S. Sloth, hep-th/0204110; R. Easther, B. R. Greene, W. H. Kinney, and G. Shiu, hep-th/0204129; U. H. Danielsson, hep-th/0205227; J. C. Niemeyer, R. Parentani, and D. Campo, hep-th/0206149.
A signature of transplanckian physics which does not depend upon nonlinear unknown quantum effects is the one induced by resonant production of the particle during inflation leaving an imprint in the form of sharp features in the primordial power spectrum. This effect has been discussed in D. J. Chung, E. W. Kolb, A. Riotto and I. I. Tkachev, Phys. Rev. D [**62**]{}, 043508 (2000).
N. Arkani-Hamed, S. Dimopoulos, and G. R. Dvali, Phys. Lett. B [**429**]{}, 263 (1998); I. Antoniadis, N. Arkani-Hamed, S. Dimopoulos, and G. R. Dvali, Phys. Lett. B [**436**]{}, 257 (1998).
D. Langlois, R. Maartens, D. Wands, Phys. Lett. B [**489**]{}, 259 (2000).
R. Maartens, D. Wands, B.A. Bassett and I. Heard, Phys. Rev. D 62, 041301 (2000); D. Langlois, R. Maartens, M. Sasaki, D. Wands, Phys. Rev. D 63, 084009 (2001).
Notice that $a_0(t)$ is the value of the scale factor at $y=0$, not its present value.
P. Kanti, K. A. Olive, and M. Pospelov, Phys. Lett. B [**481**]{}, 386 (2000); C. Csaki, M. Graesser, L. Randall, and J. Terning, Phys. Rev. D [**62**]{}, 045015 (2000).
L. Randall and R. Sundrum, Phys. Rev. Lett. [**83**]{}, 3370 (1999); Phys. Rev. Lett. [**83**]{}, 4690 (1999).
P. Binetruy, C. Deffayet, and D. Langlois, Nucl. Phys. B [**565**]{}, 269 (2000); P. Binetruy, C. Deffayet, U. Ellwanger, and D. Langlois, Phys. Lett. B [**477**]{}, 285 (2000).
This equation admits the solution $\omega =0$ for any value of $RH$. In the limit of small $RH$, the solutions of Eq. (\[eigenvalue\]) are $\omega_p =
p/R$, with $p$ ranging over the non-negative integers. Indeed, in the limit $RH\to 0$, the space is not warped ($n \to 1$) and we recover the usual equally-spaced Kaluza–Klein eigenvalues. The solutions for $0<\pi RH<1$ are plotted in Fig. 1.
See [*e.g.,*]{} A. R. Liddle and D. H. Lyth, [*Cosmological Inflation and Large-Scale Structure,*]{} (Cambridge Univ. Press, Cambridge, 2000)
Notice that if we defined the effective four-dimensional Planck mass during inflation by integrating the background gravitational action over $y$, we would obtain a different result, $\left.M_{P}^2\right|_{I}= M_{P}^2
\left(1 - \frac{1}{3} \pi^2 R^2 H^2\right)$. However, we are interested here in the effective gravitational coupling of the zero Kaluza–Klein mode of the tensor perturbation, which is properly defined by Eq. (\[efpl\]).
R. Maartens, Phys. Rev. D [**62**]{}, 084023 (2000); C. van de Bruck, M. Dorca, R. Brandenberger, and A. Lukas, Phys. Rev. D [**62**]{}, 123515 (2000); H. Kodama, A. Ishibashi, and O. Seto, Phys. Rev. D [**62**]{}, 064022 (2000); S. Mukohyama, Phys. Rev. D [**62**]{}, 084015 (2000); Class. Quantum Grav. [**17**]{}, 4777 (2000); Phys. Rev. D [**64**]{}, 064006 (2001); D. Langlois, Phys. Rev. D [**62**]{}, 126012 (2000); , 2212 (2001); A. Neronov and I. Sachs, B [**513**]{}, 173 (2001); C. van de Bruck and M. Dorca, hep-th/0012073; H. Bridgman, K. Malik, and D. Wands, Phys. Rev. D [**65**]{}, 043502 (2002); K. Koyama and J. Soda, Phys. Rev. D [**62**]{}, 123502 (2000); K. Koyama and J. Soda, Phys. Rev. D [**65**]{}, 023514 (2002); V. Bozza, M. Gasperini, and G. Veneziano, Nucl. Phys. [**B619**]{}, 191 (2001); O. Seto and H. Kodama, Phys. Rev. D [**63**]{}, 123506 (2001); J.D. Barrow and R. Maartens, gr-qc/0108073; A. Riazuelo, F. Vernizzi, D. Steer, and R. Durrer, hep-th/0205220.
D. Langlois, in Ref. [@perturb].
S. Mukohyama, in Ref. [@perturb].
For a review, see J. E. Lidsey, A. R. Liddle, E. W. Kolb, E. J. Copeland, T. Barreiro, and M. Abney, Rev. Mod. Phys. [**69**]{}, 373 (1997).
V. F. Mukhanov, H. A. Feldman, and R. H. Brandenberger, Phys. Rept. [**215**]{}, 203 (1992).
G. Huey and J. E. Lidsey, astro-ph/0205236.
N. Bartolo, S. Matarrese, and A. Riotto, Phys. Rev. D [**64**]{}, 083514 (2001); N. Bartolo, S. Matarrese, and A. Riotto, Phys. Rev. D [**64**]{}, 123504 (2001); D. Wands, N. Bartolo, S. Matarrese, and A. Riotto, astro-ph/0205253.
J. Lesgourgues, S. Pastor, M. Peloso, and L. Sorbo, Phys. Lett. B [**489**]{}, 411 (2000).
Background and perturbed Einstein equations
===========================================
We give here the background and perturbed Einstein equations in the Gaussian normal gauge: $$\begin{aligned}
G^0_{\ 0} & = &
\frac{3}{n^2} \left( \frac{\dot{a}}{a} \right)^2
- 3 \left[ \frac{a''}{a}
+ \left( \frac{a'}{a} \right)^2 \right] \nonumber \\
& &
- \frac{6}{n^2} \frac{\dot{a}}{a}
\left( \frac{\dot{a}}{a} \phi + \dot{\psi} \right)
+ \frac{2}{a^2} \Delta \psi + \frac{2}{n^2 a^2} \frac{\dot{a}}{a} \Delta B
+ \frac{2}{n^2} \frac{\dot{a}}{a} \ \Delta \dot{E}
\nonumber \\
& &
+ 3 \left( 4 \frac{a'}{a} \psi' + \psi'' \right)
- 4 \frac{a'}{a} \Delta E' - \Delta E''
\nonumber \\
& = & \delta(y) \left[ \left( \frac{\dot{\bar{\varphi}}^2}{2 n^2}
+ V \right) + \frac{\dot{\bar{\varphi}}
\delta \dot{\varphi} -
\dot{\bar{\varphi}}^2 \phi}{n^2} + V' \delta \varphi
\right] ,
\label{Einstein.00}
\\
G^i_{\ 0} & = & - a^{-2} \partial_i \left\{
2 \frac{\dot{a}}{a} \phi + 2 \dot{\psi}
+ \frac{2}{n^2} \left[ \frac{\ddot{a}}{a}
- \left( \frac{\dot{a}}{a} \right)^2
- \frac{\dot{n}}{n} \frac{\dot{a}}{a} \right] B
\right.
\nonumber \\
& &
\left.
- \left( \frac{n''}{n} + \frac{a'}{a} \frac{n'}{n} \right) B
+ \frac{1}{2} \left( \frac{a'}{a} - \frac{n'}{n} \right) B'
+ \frac{1}{2} B''
\right\}
\nonumber \\
& = & -a^{-2} \delta(y) \partial_i \left(
\dot{\bar{\varphi}} \delta \varphi - \frac{\dot{\bar{\varphi}}^2}{n^2} B
\right)
\\
\textrm{or } G^0_{\ i} &=& n^{-2} \partial_i \left\{
2 \frac{\dot{a}}{a} \phi + 2 \dot{\psi}
\right.
\nonumber \\
& &
\left.
+ \left[ \frac{a'}{a} \frac{n'}{n} - 2 \left( \frac{a'}{a}\right)^2
- \frac{a''}{a} \right] B
+ \frac{1}{2} \left( \frac{a'}{a} - \frac{n'}{n} \right) B'
+ \frac{1}{2} B''
\right\}
\nonumber \\
G^i_{\ j} & = & \left\{
\frac{1}{n^2} \left[2 \frac{\ddot{a}}{a}
+ \frac{\dot{a}}{a} \left(\frac{\dot{a}}{a} - 2 \frac{\dot{n}}{n}\right)\right]
- 2 \frac{a''}{a} - \frac{a'}{a}
\left( \frac{a'}{a} + 2 \frac{n'}{n} \right) - \frac{n''}{n}
\right.
\nonumber \\
& &
\left. - \frac{2}{n^2} \left[ 2 \frac{\ddot{a}}{a}
+ \left( \frac{\dot{a}}{a} \right)^2 - 2 \frac{\dot{a}}{a} \frac{\dot{n}}{n}
\right] \phi - 2 \left( \frac{a'}{a} + \frac{n'}{n} \right) \phi'
- \phi'' - \frac{2}{n^2} \frac{\dot{a}}{a} \dot{\phi}
\right.
\nonumber \\
& &
\left.
+ 2 \left( \frac{n'}{n} + 3 \frac{a'}{a} \right) \psi' + 2 \psi''
- \frac{2}{n^2} \left( 3 \frac{\dot{a}}{a} - \frac{\dot{n}}{n} \right)
\dot{\psi} - \frac{2}{n^2} \ddot{\psi}
\right\} \delta_{ij}
\nonumber \\
& & + a^{-2} (\delta_{ij} \Delta - \partial_i \partial_j)
\left\{ -\phi + \psi
+ \frac{1}{n^2} \left[ \left( \frac{\dot{a}}{a}
- \frac{\dot{n}}{n} \right) B + \dot{B} \right]
+ \frac{a^2}{n^2} \left[
\left( 3 \frac{\dot{a}}{a} - \frac{\dot{n}}{n} \right) \dot{E}
+ \ddot{E} \right]
\right.
\nonumber \\
& &
\left.
- a^2 \left[ \left(3\frac{a'}{a} + \frac{n'}{n} \right) E'
+ E'' \right] \right\}
\nonumber \\
& = & \delta(y) \left( V - \frac{\dot{\bar{\varphi}}^2}{2 n^2}
+ \frac{\dot{\bar{\varphi}}^2 \phi -
\dot{\bar{\varphi}} \delta \dot{\varphi}}
{n^2} + V' \delta \varphi \right) \delta_{ij}
\\
G^0_{\ 5} & = &
\frac{3}{n^2} \left[
\left( \frac{\dot{a}}{a} \frac{n'}{n} - \frac{\dot{a}'}{a} \right)
(1 - 2 \phi) + \frac{\dot{a}}{a} \phi' + \frac{\dot{a}}{a}
\psi' + \left( \frac{a'}{a} - \frac{n'}{n} \right) \dot{\psi}
+ \dot{\psi}' \right] \nonumber \\
& & + \frac{1}{n^2 a^2} \left( \frac{n'}{n} \Delta B
- \frac{1}{2} \Delta B' \right)
\nonumber \\
& &
+ \frac{1}{n^2} \left[
- \Delta \dot{E}' + \left( \frac{n'}{n} - \frac{a'}{a} \right) \Delta \dot{E}
- \frac{\dot{a}}{a} \Delta E'
\right] = 0
\\
G^i_{\ 5} & = & a^{-2} \partial_i\left\{\phi'-2\psi'
+\left({n'\over n}-{a'\over a}\right) \phi
+ \frac{1}{n^2} \left[ \frac{a'}{a} \dot{B} -
\frac{1}{2} \dot{B}' + \frac{1}{2} \left( \frac{\dot{n}}{n} -3
\frac{\dot{a}}{a} \right) B'
\right. \right.
\nonumber \\
& & \left.\left.
+\left( 2\frac{\dot{a}a'}{a^2}-2\frac{\dot{a}'}{a}-\frac{\dot{n}a'}{na}
+3\frac{\dot{a}n'}{an}\right)B
\right] \right\} = 0 .
\label{Einstein.i5}\end{aligned}$$ These equation simplify in the De Sitter background defined in Eq.(\[desitter\]): $$\begin{aligned}
n^2 \delta G^0_{\ 0} & = &
- 6 H \left( H \phi + \dot{\psi} \right) + 2 \frac{\Delta}{a_0^2} \psi
+ 2\frac{ H \Delta}{a_0^2n^2} B + 2 H \Delta \dot{E}
\nonumber \\
& &
+ 3 \left( 4 n'n \psi' + n^2 \psi'' \right) - 4 n'n \Delta E' - n^2\Delta E''
\label{Einstein.DeSitter.00}
\\
- a^2 \delta G^i_{\ 0} &=& \partial_i \left( 2 H \phi + 2 \dot{\psi}
- \frac{1}{n^2} H^2 B + \frac{1}{2} B''
\right)
\\
n^2 \delta G^i_j &=& \left( - 6 H^2 \phi - 4 n'n \phi' - n^2 \phi''
- 2 H \dot{\phi} \right.
\nonumber \\
& &
\left. + 8 n'n \psi' + 2 n^2 \psi'' - 6 H \dot{\psi} - 2 \ddot{\psi}
\right) \delta_{ij}
\nonumber \\
& & + (a_0n)^{-2} (\delta_{ij} \Delta - \partial_i \partial_j)
\left[ -\phi + \psi + n^{-2} \left( H B + \dot{B} \right)
+ a_0^2 \left( 3 H \dot{E} + \ddot{E} \right)
\right.
\nonumber \\
& &
\left.
- a_0^2 \left( 4 n'n E' + n^2 E'' \right) \right]
\\
n^2 \delta G^0_{\ 5} & = &
3 \left( H \phi' + H \psi' + \dot{\psi}' \right)
+ \frac{1}{a_0^2 n^2} \left( \frac{n'}{n} \Delta B - \frac{1}{2} \Delta B'
\right) - \Delta \dot{E}' - H \Delta E'
\\
- a^2 G^i_{\ 5} &=& \partial_i \left[ -\phi'+2\psi' \right.
\nonumber \\
& &
\left.
+ \frac{1}{n^2} \left(
\frac{1}{2} \dot{B}' + \frac{3}{2} H B' - \frac{n'}{n} \dot{B}
- 3 H \frac{n'}{n} B \right) \right] .
\label{Einstein.DeSitter.i5}\end{aligned}$$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We report a ground-state solution for the two-dimensional fermionic Hubbard model, which is obtained via a numerical variational method. The two ingredients in this approach are tensor network states and the time-evolving block decimation. We easily handle the horizontal hopping in the Hamiltonian, and we proceed further to observe the fermion-exchange effect caused by the vertical hopping. By requiring no divergence and no convergence to zero for the ground state, we successively determine the ground-state energy per site as a function of the chemical potential and the lattice length. In addition, we observe saturation in the behavior of the ground-state energy as the lattice length increases.'
address:
- 'College of Science and Technology, Hongik University, Sejong 339-701, Korea'
- 'E-mail: [email protected]'
author:
- 'Myung-Hoon Chung'
title: A Solution of the Hubbard Model
---
Introduction
============
In 1963, to understand the behavior of correlated electrons in solids, a fermion lattice model was proposed independently by three physicists: Martin Gutzwiller [@Gutzwiller], Junjiro Kanamori [@Kanamori], and John Hubbard [@Hubbard1]. This model has become widely known as the Hubbard model [@Hubbard2]. Since this model’s relevance to high $T_c$ superconductors was first suggested [@Dagotto], much attention has been paid to it. Recently, it has become possible to construct experimental implementations of the Hubbard model using an optical lattice for cold atoms [@Greiner; @Jordens], and hence, the research community has refocused on the Hubbard model. Although the model can be represented in a simple form, we encounter notorious difficulties [@Troyer] when we attempt to find a solution even numerically.
One of the main advances in the field of strongly correlated systems is the establishment of the concept of the renormalization group (RG) [@Wilson1]. In fact, Wilson also invented the numerical RG (NRG) [@Wilson2] to solve the Kondo problem [@Kondo]. Inspired by the NRG, White proposed the density-matrix RG (DMRG) [@White], which has proven to be a great success in the simulation of strongly correlated one-dimensional quantum lattice systems. It has been found that the internal structure of the DMRG can be understood with respect to the matrix-product states (MPS) [@Ostlund; @Garcia; @Saberi; @Schollwoeck]. For two-dimensional systems, the projected entangled-pair states (PEPS) [@Verstraete; @Orus] are introduced. More generally, we call all of these states tensor network states (TNS), and they include MPS, PEPS, tree tensor network states [@Murg], the multiscale entanglement renormalization ansatz [@Vidal1], and matrix-product projected states [@Chou]. Beyond the spin-block concept, the tensor network method based on the coarse-grained tensor RG [@Levin] has been applied to a classical spin system. The method was refined to the second RG [@Jiang; @Xie] by globally optimizing the truncation scheme and improving the accuracy.
When a total Hamiltonian is written as a sum of local Hamiltonians, Vidal [@Vidal2; @Vidal3] introduced a powerful method called time-evolving block decimation (TEBD) for finding correlation functions. If the total Hamiltonian also has a type of symmetry such as translational invariance, we can use the so-called infinite TEBD [@Vidal4], in which we assume that the matrices in the TNS have the same form, and we update a few matrices to achieve the ground state. However, because the TNS for the Hubbard model is not an eigenstate of the number operator, the TNS breaks the basic symmetry of particle-number preserving. Furthermore, we do not insist on preserving the translational invariance in the TNS. In consequence, we do not use the infinite TEBD here. We alternatively adopt TEBD and extend it to the case of the Hubbard model using PEPS. If the fermion-exchange effect is involved during the TEBD procedure, a long-range entanglement appears between the tensors of the PEPS. The essence of the Hubbard model is to solve the problem caused by the fermion-exchange effect.
In this paper, we focus on updating the large entangled part in the TNS when we apply TEBD to the Hubbard model. To that end, we first describe the nature of the TNS as an approximate ground state for the Hubbard model. The connections between the tensors in the TNS are represented by three types of bonds: horizontal, vertical, and spin bonds. The set of the TNS is a small subspace of the corresponding huge Hilbert space for the Hubbard model. During the imaginary time evolution in TEBD, we restrict the accessible states to the set of the TNS. Furthermore, in the process of updating bonds, we adjust the proportional factor in front of the state. By requiring no divergence and no convergence to zero for the factor, we determine the form of the TNS for the ground state and the corresponding energy.
This paper is organized as follows. In Sec. 2, a detailed description of the Suzuki-Trotter decomposition is given, and we introduce the tensor network state for the Hubbard model. In Sec. 3, using the Suzuki-Trotter decomposition, we present the framework of the algorithm in the spirit of TEBD. Moreover, in this section, we describe how to update the horizontal, vertical, and spin bonds, and present the method of determining the ground-state energy per site. In Sec. 4, we present consistency checks for the method, and we summarize the numerical results obtained when performing TEBD with small bond dimensions; the bond dimensions should be increased in future works. The results for the ground-state energy show evidence of saturation as the lattice length increases, which indicates that the thermodynamic limit is achieved. We observe the spin-flip symmetry breaking, and present the critical strength of the on-site Coulomb repulsion. In conclusion, we discuss a parallelism for implementation in future work to improve the speed of computing.
Hamiltonian and Tensor Network States
=====================================
We begin by presenting the Hamiltonian for the Hubbard model, which is written as $$\begin{aligned}
H &=& -t\sum_{\langle i j
\rangle}(c^{\dagger}_{i\uparrow}c_{j\uparrow} +
c^{\dagger}_{j\uparrow}c_{i\uparrow} +
c^{\dagger}_{i\downarrow}c_{j\downarrow} +
c^{\dagger}_{j\downarrow}c_{i\downarrow}) \nonumber \\
& &
+U\sum_{i}(n_{i\uparrow}-\frac{1}{2})(n_{i\downarrow}-\frac{1}{2})-\mu
\sum_{i}(n_{i\uparrow}+n_{i\downarrow}) \nonumber \\
&=& H^{\uparrow}_{he} + H^{\uparrow}_{ho} + H^{\uparrow}_{ve} +
H^{\uparrow}_{vo} + H^{\downarrow}_{he} + H^{\downarrow}_{ho} +
H^{\downarrow}_{ve} + H^{\downarrow}_{vo} + H_{d},\end{aligned}$$ where $\langle i j \rangle$ represents nearest-neighbor hopping in a two-dimensional lattice, and $n_{i\uparrow}$ and $n_{i\downarrow}$ are the spin-up and the spin-down number operators, respectively. We let the hopping strength $t$ be 1 and vary the strengths of both the on-site Coulomb repulsion $U$ and the chemical potential $\mu$; the number of fermions is controlled by $\mu$. We divide the hopping term into four parts for each spin, which are denoted by $h$(horizontal), $v$(vertical), $e$(even), and $o$(odd), as shown in Fig. 1. The diagonal Hamiltonian $H_{d}$ for a typical basis contains the last two terms of the on-site repulsion and the chemical potential. The Hubbard model may be the simplest quantum system of interacting fermions on a lattice.
We note that the Hamiltonian has symmetries. First of all, the number operator $N_{op}=\sum_{i}(
c^{\dagger}_{i\uparrow}c_{i\uparrow} +
c^{\dagger}_{i\downarrow}c_{i\downarrow})$ commutes with the Hamiltonian. When we impose the periodic boundary condition, the translational symmetry appears. Furthermore, the Hamiltonian is invariant under the spin-flip operation such as $c_{i\uparrow}\rightarrow c_{i\downarrow}$ and $c_{i\downarrow}\rightarrow c_{i\uparrow}$. We will discuss these symmetries in relation to the TNS later.
For a given Hamiltonian $H$, we introduce an energy shift $E$ and the inverse of the energy $T$, and then, we consider a formal solution to the imaginary time Schrödinger equation: $$|\Psi( T ) \rangle = \exp\{ - ( H - E ) T \} |\Psi( 0 )\rangle.$$ As $T$ goes to infinity, the state $|\Psi( T ) \rangle$ becomes the ground state for properly chosen $E$. In fact, when $E$ is larger or smaller than the ground-state energy, $|\Psi( T
)\rangle$ blows up or shrinks down, respectively, in the limit as $T\rightarrow \infty$. In a numerical approach, we redefine $E$ as a function of $T$ to determine the ground state.
We rewrite the operator using the Suzuki-Trotter decomposition with a given small time step $\tau$ as $$\begin{aligned}
\exp\{ - ( H - E ) T \}&\cong &
\prod^{T/\tau}\exp\{(E-H_{d})\tau\}
\nonumber \\
&\times&\Big[\mbox{The same expression
for spin down}\Big] \nonumber \\
&\times&\exp(-\frac{1}{4}H^{\uparrow}_{he}\tau)
\exp(-\frac{1}{2}H^{\uparrow}_{ho}\tau)
\exp(-\frac{1}{4}H^{\uparrow}_{he}\tau) \nonumber \\
&\times&\exp(-\frac{1}{2}H^{\uparrow}_{ve}\tau)
\exp(-H^{\uparrow}_{vo}\tau)
\exp(-\frac{1}{2}H^{\uparrow}_{ve}\tau) \nonumber \\
&\times&\exp(-\frac{1}{4}H^{\uparrow}_{he}\tau)
\exp(-\frac{1}{2}H^{\uparrow}_{ho}\tau)
\exp(-\frac{1}{4}H^{\uparrow}_{he}\tau).\end{aligned}$$ It is not difficult to employ a higher-order Suzuki-Trotter decomposition to obtain a more accurate calculation. Note that we now decompose the operators in Eq. (3) in terms of elementary operators such as $$\begin{aligned}
\exp(-\frac{1}{4}H^{\uparrow}_{he}\tau) = \prod_{\langle i j
\rangle_{he}} \exp \{\frac{1}{4}t\tau
(c^{\dagger}_{i\uparrow}c_{j\uparrow} +
c^{\dagger}_{j\uparrow}c_{i\uparrow})\}, \nonumber \\
~~~~~~~~\vdots \nonumber \\
\exp\{(E - H_{d})\tau \} = \prod_{i} \exp\{e\tau - U\tau
(n_{i\uparrow}-\frac{1}{2})(n_{i\downarrow}-\frac{1}{2}) +\mu \tau
(n_{i\uparrow}+n_{i\downarrow}) \}, \nonumber\end{aligned}$$ where $e$ is the energy per site, that is, $E/N$, and $N=\sum_{i}1$. Our strategy is to use Vidal’s TEBD with these elementary operators in a small subset of the Hilbert space. This small subset is composed of the TNS characterized by the fixed bond dimension.
For the fermionic Hubbard model, the usual tensor network states should be suitably modified to describe fermions. In previous works, many such attempts have been made; the Jordan-Wigner strings was noticed in relation to fermions [@Barthel], and we find the fermionic projected entangled-pair states [@Kraus; @Corboz1; @Pizorn; @Corboz2; @Corboz3] and the fermionic multiscale entanglement renormalization ansatz [@Pineda; @Corboz4; @Corboz5; @Marti] for the ground states. These fermionic modifications share some similarities, but they do not agree with each other completely; thus, they require further investigation. As a first step, we adopt the scheme of Corboz’s fermionic PEPS for our TNS, however for which we do not insist on preserving the fermionic parity.
We use the one-to-one correspondence between a state of the two-state chain and a state of the Fock space. The state of the chain is represented by $\sigma_{i}$ and $\sigma_{N+i}$ for the spin-up and the spin-down, respectively, and the state of the Fock space is written in terms of the creation operators $c^{\dagger}_{i\uparrow}$ and $c^{\dagger}_{i\downarrow}$ as follows: $$\begin{aligned}
&&| \sigma_{0} \cdots \sigma_{N-1}\sigma_{N} \cdots \sigma_{2N-1}
\rangle \nonumber \\
&&~~~~~~~~~~= (c^{\dagger}_{0\uparrow})^{\sigma_{0}} \cdots
(c^{\dagger}_{N-1\uparrow})^{\sigma_{N-1}}
(c^{\dagger}_{0\downarrow})^{\sigma_{N}} \cdots
(c^{\dagger}_{N-1\downarrow})^{\sigma_{2N-1}}|0 \rangle ,\end{aligned}$$ where $\sigma_{N+i}= 0$ or $1$ means there is a spin-down fermion vacancy or occupancy at the $i$-th site, respectively. It is important to maintain the ordering of the fermions in the state of the Fock space to handle the negative sign caused by the fermion exchange. We adopt the zigzag ordering, which is the approach of numbering sites from left to right and from right to left one by one alternately in horizontal lines. For the example of $N=4
\times 4$, the corresponding ordering of sites is $$\begin{array}{ccccccccc}
0&1&2&3&~~~~~~&16&17&18&19 \\
7&6&5&4&~~~~~~&23&22&21&20 \\
8&9&10&11&~~~~~~&24&25&26&27 \\
15&14&13&12&~~~~~~&31&30&29&28
\end{array}$$ where the numbers from $0$ to $15$ denote the spin-up sites and the numbers from $16$ to $31$ denote the spin-down sites.
When representing the TNS for the Hubbard model, we should appreciate the area law for the entanglement entropy [@Eisert]. Taking into account the square lattice, we include two horizontal bonds and two vertical bonds for each tensor. Because the TNS do not preserve the fermion numbers, it is natural to break the translational symmetry also. Thus, we use different tensors at all sites. In order to consider the general case of the spin-flip symmetry breaking, we introduce different tensors for the spin-down fermions from those for the spin-up fermions. Because the Hubbard model has an on-site interaction, we connect the two tensors of spin-up and spin-down at the same site using the spin bond. In consequence, for each tensor we attach four legs for the right, up, left, and down bonds and one leg for the spin bond as well as the physical index. We assign a Schmidt coefficient vector to each bond. Therefore, for the square system of the length $L$, there are $2L^{2}$ tensors and $2L^{2} \times
2+L^{2}$ Schmidt coefficient vectors in our TNS as shown in Fig. 2. These tensors and vectors will be updated in the process of TEBD with periodic boundary conditions.
A typical one of the $2L^{2}$ tensors, $A_{ruld}^{\sigma s}$, has six indices, among which the physical index $\sigma$ takes a value of $0$ or $1$. For the space-bond degree of freedom, the indices $r$ (right), $u$ (up), $l$ (left), and $d$ (down) run from $0$ to $\chi-1$, where $\chi$ is the bond dimension. For the spin-bond degree of freedom, the index $s$ runs from $0$ to $\kappa -1$, where $\kappa$ is the spin-bond dimension. A state in the space of the tensor network states is written as $$\begin{aligned}
|\mbox{TNS} \rangle =
\sum_{\cdots\sigma\rho\cdots\nu\eta\cdots\alpha \beta \cdots
\delta \gamma \cdots} &&\mbox{Tr} \left[
\begin{array}{cccc}
\ddots & \vdots & \vdots & \\
\cdots & A^{\sigma}_{\uparrow} & B^{\rho}_{\uparrow} & \cdots \\
\cdots & C^{\eta}_{\uparrow} & D^{\nu}_{\uparrow} & \cdots \\
& \vdots & \vdots & \ddots \\
\end{array}
\right] \left[
\begin{array}{cccc}
\ddots & \vdots & \vdots & \\
\cdots & A^{\alpha}_{\downarrow} & B^{\beta}_{\downarrow} & \cdots \\
\cdots & C^{\gamma}_{\downarrow} & D^{\delta}_{\downarrow} & \cdots \\
& \vdots & \vdots & \ddots \\
\end{array}
\right] \nonumber \\
&&\times | \cdots \sigma \rho \cdots \nu \eta \cdots \cdots \alpha
\beta \cdots \delta \gamma \cdots\rangle ,\end{aligned}$$ where we ignore the representation of the internal bond indices on the tensors and all internal bonds are connected by $\mbox{Tr}$, as shown in Fig. 2. Note that the zigzag ordering puts the physical index $\nu$ before $\eta$ in the spin-like chain basis. The thermodynamic limit will be achieved for $L \rightarrow
\infty$ and $\chi,\kappa \rightarrow \infty$.
When we consider the case of preserving the spin-flip symmetry, we duplicate the tensors for spin-down using those for spin-up such as $$\begin{aligned}
|\mbox{TNS} \rangle =
\sum_{\cdots\sigma\rho\cdots\nu\eta\cdots\alpha \beta \cdots
\delta \gamma \cdots} &&\mbox{Tr} \left[
\begin{array}{cccc}
\ddots & \vdots & \vdots & \\
\cdots & A^{\sigma} & B^{\rho} & \cdots \\
\cdots & C^{\eta} & D^{\nu} & \cdots \\
& \vdots & \vdots & \ddots \\
\end{array}
\right] \left[
\begin{array}{cccc}
\ddots & \vdots & \vdots & \\
\cdots & A^{\alpha} & B^{\beta} & \cdots \\
\cdots & C^{\gamma} & D^{\delta} & \cdots \\
& \vdots & \vdots & \ddots \\
\end{array}
\right] \nonumber \\
&&\times | \cdots \sigma \rho \cdots \nu \eta \cdots \cdots \alpha
\beta \cdots \delta \gamma \cdots\rangle .\end{aligned}$$ In the process of TEBD, we update $L^{2}$ tensors and $L^{2}
\times (2+1)$ Schmidt coefficient vectors for the ground state preserving the spin-flip symmetry.
The operator $\exp\{(E - H) \tau \}$ of Eq. (3) acts on the state $|\mbox{TNS}\rangle$ of Eq. (5) consecutively. Thus, the output state $\exp\{(E - H) \tau \}|\mbox{TNS}\rangle$, which is outside the space of the TNS, is approximated into a TNS by updating the tensors and the vectors. The updating procedure is the subject of the next section.
The Updating Procedure
======================
In order to proceed with TEBD, we consider the single elementary hopping term $\exp\{t\tau (c^{\dagger}_{i\uparrow}c_{j\uparrow} +
c^{\dagger}_{j\uparrow}c_{i\uparrow})\}$, where we add a factor of $1/2$ or $1/4$ in front of $t\tau$ if necessary. When the elementary operator acts on the previous TNS, we approximate the output state into our subset of the TNS by updating the tensors and the vectors locally. This procedure is the basic strategy of TEBD.
When the elementary operator $\exp\{t\tau
(c^{\dagger}_{i\uparrow}c_{j\uparrow} +
c^{\dagger}_{j\uparrow}c_{i\uparrow})\}$ acts on a basis vector $|
\sigma_{0}\cdots \sigma_{i} \cdots \sigma_{j} \cdots
\sigma_{N-1}\sigma_{N}\cdots\sigma_{2N-1} \rangle$, we find the following important result [@Chung0], which is written for four cases that correspond to $\sigma_i = 0$ or $1$ and $\sigma_j
= 0$ or $1$: $$\left\{
\begin{array}{l}
\exp\{t\tau (c^{\dagger}_{i\uparrow}c_{j\uparrow} +
c^{\dagger}_{j\uparrow}c_{i\uparrow})\} | \cdots 0 \cdots 0 \cdots
\rangle = |
\cdots 0 \cdots 0 \cdots \rangle \nonumber\\
\exp\{t\tau (c^{\dagger}_{i\uparrow}c_{j\uparrow} +
c^{\dagger}_{j\uparrow}c_{i\uparrow})\} | \cdots 0 \cdots 1 \cdots
\rangle = \cosh (t\tau) |
\cdots 0 \cdots 1 \cdots \rangle \nonumber\\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +\sinh
(t\tau)(-1)^{\sigma_{i+1}+\cdots+\sigma_{j-1}}
| \cdots 1 \cdots 0 \cdots \rangle \nonumber\\
\exp\{t\tau (c^{\dagger}_{i\uparrow}c_{j\uparrow} +
c^{\dagger}_{j\uparrow}c_{i\uparrow})\} | \cdots 1 \cdots 0 \cdots
\rangle = \cosh (t\tau) |
\cdots 1 \cdots 0 \cdots \rangle \nonumber\\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +\sinh
(t\tau)(-1)^{\sigma_{i+1}+\cdots+\sigma_{j-1}}
| \cdots 0 \cdots 1 \cdots \rangle \nonumber\\
\exp\{t\tau (c^{\dagger}_{i\uparrow}c_{j\uparrow} +
c^{\dagger}_{j\uparrow}c_{i\uparrow})\} | \cdots 1 \cdots 1 \cdots
\rangle = | \cdots 1 \cdots 1 \cdots \rangle \nonumber
\end{array} \right .$$ The sign of $(-1)^{\sigma_{i+1}+\cdots+\sigma_{j-1}}$ reflects the fermion-exchange effect. Note that the values of the physical indices at the sites numbered from $i+1$ to $j-1$ are related to the sign, which makes it difficult to handle the vertical hopping. The above equations play a key role in updating the TNS.
Updates to the Horizontal Bonds
-------------------------------
Because a horizontal bond connects a site to the next site, that is, to $j=i+1$ in our ordering of sites, it is straightforward to update horizontal bonds. As we can see in the Suzuki-Trotter decomposition of Eq. (3), we first handle $H_{he}$, then $H_{ho}$, and then $H_{he}$ again. Here, we present the typical procedure for updating a horizontal bond.
For example, to update $A$, $B$, and $\lambda^{AB}$ in Fig. 2 with the periodic boundary condition, we consider the tensor product that is represented symbolically as follows: $$(A \otimes B )^{\sigma s \rho
\tilde{s}}_{uld\tilde{r}\tilde{u}\tilde{d}} \equiv
\sum_{x=0}^{\chi -1} A^{\sigma s}_{xuld} \lambda^{AB}_x B^{\rho
\tilde{s}}_{\tilde{r}\tilde{u}x\tilde{d}} \lambda^{AA}_s
\lambda^{MA}_u \lambda^{DA}_l \lambda^{AE}_d
\lambda^{BB}_{\tilde{s}}
\lambda^{BC}_{\tilde{r}}\lambda^{NB}_{\tilde{u}}\lambda^{BF}_{\tilde{d}},$$ where the eight Schmidt coefficients are attached to $A$ and $B$. Using the result of Eq. (7), we find the ten-index tensor $\Theta^{\sigma s \rho\tilde{s}}_{uld\tilde{r}\tilde{u}\tilde{d}}$ to update $A$, $B$, and $\lambda^{AB}$: $$\left\{
\begin{array}{l}
\Theta^{0s0\tilde{s}}_{uld\tilde{r}\tilde{u}\tilde{d}} = (A
\otimes B )^{0 s 0
\tilde{s}}_{uld\tilde{r}\tilde{u}\tilde{d}} \nonumber\\
\Theta^{0s1\tilde{s}}_{uld\tilde{r}\tilde{u}\tilde{d}} = \cosh
(t\tau)\times(A \otimes B )^{0 s 1
\tilde{s}}_{uld\tilde{r}\tilde{u}\tilde{d}}
+ \sinh (t\tau)\times(A \otimes B )^{1 s 0
\tilde{s}}_{uld\tilde{r}\tilde{u}\tilde{d}} \nonumber\\
\Theta^{1s0\tilde{s}}_{uld\tilde{r}\tilde{u}\tilde{d}} = \cosh
(t\tau)\times(A \otimes B )^{1 s 0
\tilde{s}}_{uld\tilde{r}\tilde{u}\tilde{d}}
+ \sinh (t\tau)\times (A \otimes B )^{0 s 1
\tilde{s}}_{uld\tilde{r}\tilde{u}\tilde{d}} \nonumber\\
\Theta^{1s1\tilde{s}}_{uld\tilde{r}\tilde{u}\tilde{d}}=(A \otimes
B )^{1 s 1 \tilde{s}}_{uld\tilde{r}\tilde{u}\tilde{d}} \nonumber
\end{array} \right .$$ We emphasize the physical-index exchange between $0$ and $1$ in the tensor product multiplied by $\sinh(t\tau)$. By employing singular value decompositions (SVD), we obtain the updated $\tilde{\lambda}^{AB}_{x}$ by keeping the $\chi$ largest weights: $$\begin{aligned}
\Theta^{\sigma s \rho \tilde{s}}_{uld\tilde{r}\tilde{u}\tilde{d}}
&\rightarrow& \sum_{x=0}^{\chi-1} \bar{A}^{\sigma s}_{xuld}
\tilde{\lambda}^{AB}_x \bar{B}^{\rho
\tilde{s}}_{\tilde{r}\tilde{u}x\tilde{d}}\nonumber \\
&=&\sum_{x=0}^{\chi-1} \tilde{A}^{\sigma s}_{xuld}
\tilde{\lambda}^{AB}_x \tilde{B}^{\rho
\tilde{s}}_{\tilde{r}\tilde{u}x\tilde{d}}\lambda^{AA}_s
\lambda^{MA}_u \lambda^{DA}_l \lambda^{AE}_d
\lambda^{BB}_{\tilde{s}}
\lambda^{BC}_{\tilde{r}}\lambda^{NB}_{\tilde{u}}\lambda^{BF}_{\tilde{d}}.\end{aligned}$$ By dividing and attaching the eight weights, we find $\tilde{A}$ and $\tilde{B}$ in the above. We denote this process graphically as follows: $$\begin{array}{ccccccccccc}
&|& &|& & & & |& & |& \\
-& &\Theta & &-& \rightarrow &-&\tilde{A}&-&\tilde{B}&- \\
&|& &|& & & & |& & |&
\end{array}$$ where we omit both the spin bonds and the physical indices.
A similar procedure is performed for other tensors and other vectors. By updating all $2L^{2}$ tensors, we finish the horizontal-bond update. We note that it is possible to update all $2L^{2}$ tensors simultaneously if we use multi-core computers. Thus, we can easily parallelize the horizontal-bond update.
Updates to the Vertical Bonds
-----------------------------
The vertical bonds exhibit a striking difference from the horizontal bonds during the update process: the notorious fermion-exchange effect appears when fermions are hopping vertically. For the vertical bonds, we should consider $\exp\{t\tau (c^{\dagger}_{i\uparrow}c_{j\uparrow} +
c^{\dagger}_{j\uparrow}c_{i\uparrow})\}|\cdots\sigma_i
\cdots\sigma_j \cdots \rangle$, where the number of sites between $i$ and $j$ is given by a value from $0$ to $2L-2$ in the zigzag ordering. Therefore, all tensors at the sites between $i$ and $j$ should be updated. Here, we introduce the method for updating the tensors between $i$ and $j$ one by one.
For example, to update $A$, $E$, and $\lambda^{AE}$ on the vertical bond in Fig. 2 with the periodic boundary condition, we begin by writing the tensor product of $A$ and $E$ as follows: $$\left(
\begin{array}{l}
A \\
\otimes \\
E
\end{array}
\right){}^{\sigma s \eta
\tilde{s}}_{rul\tilde{r}\tilde{l}\tilde{d}} \equiv
\sum_{x=0}^{\chi-1} A^{\sigma s}_{rulx} \lambda^{AE}_x E^{\eta
\tilde{s}}_{\tilde{r}x\tilde{l}\tilde{d}} \lambda^{AA}_s
\lambda^{AB}_r \lambda^{MA}_u \lambda^{DA}_l
\lambda^{EE}_{\tilde{s}}
\lambda^{EF}_{\tilde{r}}\lambda^{HE}_{\tilde{l}}\lambda^{EI}_{\tilde{d}}.$$ From the result of Eq. (7), when the vertical-hopping term acts on the TNS we find the portion that should be updated into a single tensor network: $$\begin{array}{cccccccccccc}
& & | & &|& &|& &|& \\
&-& &-&B&-&C&-&D&- \\
& & \Phi & &|& &|& &|& \\
&-& &-&F&-&G&-&H&- \\
& & | & &|& &|& &|& \\ \\
& & | & &|& &|& &|& \\
&-& &-&(-1)^{\rho}B&-&(-1)^{\alpha} C&-&(-1)^{\gamma} D&- \\
+& &\Psi & & |& & |& & |& \\
&-& &-&(-1)^{\nu} F&-&(-1)^{\beta} G&-&(-1)^{\delta} H&- \\
& & | & &|& &|& &|& \\
\end{array}$$ where we omit the spin bonds and the legs for the physical indices. Because of the fermion exchange, the many signs appear in front of the tensors in the second term. Each power of $(-1)$, such as $\rho$, $\nu$, $\alpha$, $\beta$, $\gamma$, or $\delta$, is the physical index of the corresponding tensor. Just as we introduce a ten-index tensor for the horizontal-bond update, we similarly find two ten-index tensors for the vertical-bond update, namely, $\Phi^{\sigma s \eta
\tilde{s}}_{rul\tilde{r}\tilde{l}\tilde{d}}$ and $\Psi^{\sigma s
\eta \tilde{s}}_{rul\tilde{r}\tilde{l}\tilde{d}}$, which are written in terms of the tensor product as follows: $$\begin{array}{ll}
\Phi^{0 s 0 \tilde{s}}_{rul\tilde{r}\tilde{l}\tilde{d}} = \left(
\begin{array}{l}
A \\
\otimes \\
E
\end{array}
\right){}^{0 s 0 \tilde{s}}_{rul\tilde{r}\tilde{l}\tilde{d}} &
\Psi^{0 s 0 \tilde{s}}_{rul\tilde{r}\tilde{l}\tilde{d}}= 0
\\
\Phi^{0 s 1 \tilde{s}}_{rul\tilde{r}\tilde{l}\tilde{d}} =
\cosh(t\tau)\times\left(
\begin{array}{l}
A \\
\otimes \\
E
\end{array}
\right){}^{0 s 1 \tilde{s}}_{rul\tilde{r}\tilde{l}\tilde{d}} &
\Psi^{0 s 1
\tilde{s}}_{rul\tilde{r}\tilde{l}\tilde{d}}=\sinh(t\tau)\times\left(
\begin{array}{l}
A \\
\otimes \\
E
\end{array}
\right){}^{1 s 0 \tilde{s}}_{rul\tilde{r}\tilde{l}\tilde{d}}
\\
\Phi^{1 s 0 \tilde{s}}_{rul\tilde{r}\tilde{l}\tilde{d}} =
\cosh(t\tau)\times\left(
\begin{array}{l}
A \\
\otimes \\
E
\end{array}
\right){}^{1 s 0 \tilde{s}}_{rul\tilde{r}\tilde{l}\tilde{d}} &
\Psi^{1 s 0
\tilde{s}}_{rul\tilde{r}\tilde{l}\tilde{d}}=\sinh(t\tau)\times\left(
\begin{array}{l}
A \\
\otimes \\
E
\end{array}
\right){}^{0 s 1 \tilde{s}}_{rul\tilde{r}\tilde{l}\tilde{d}}
\\
\Phi^{1 s 1 \tilde{s}}_{rul\tilde{r}\tilde{l}\tilde{d}} = \left(
\begin{array}{l}
A \\
\otimes \\
E
\end{array}
\right){}^{1 s 1 \tilde{s}}_{rul\tilde{r}\tilde{l}\tilde{d}} &
\Psi^{1 s 1 \tilde{s}}_{rul\tilde{r}\tilde{l}\tilde{d}}= 0
\end{array}$$
At this point, we propose a crucial idea to update the long tensor chain given above. We call this idea [*doubling*]{}. Doubling means that we enlarge the bond dimension for the indices $r$(right) and $l$(left) by a factor of two such that they now run from $0$ to $2\chi -1$. Graphically, doubling is represented by changing from $-B-$ to $=B=$, and similarly for $=C=$ and other tensors. Explicitly, we let $$(=B=)^{\rho s}_{ruld} \equiv \left\{
\begin{array}{ll}
B^{\rho s}_{ruld}&~~\mbox{for}~~r < \chi~\mbox{and}~l<\chi \nonumber\\
(-1)^{\rho}B^{\rho s}_{(r-\chi)u(l-\chi)d}&~~\mbox{for}~~r \ge \chi~\mbox{and}~l \ge \chi \nonumber\\
0&~~\mbox{otherwise} \nonumber
\end{array} \right .$$ Correspondingly, the ten-index tensors $\Phi$ and $\Psi$ are combined into $\Theta$ as follows: $$({}^{-}_{-}\Theta^{=}_{=})^{\sigma s
\eta\tilde{s}}_{rul\tilde{r}\tilde{l}\tilde{d}} \equiv \left\{
\begin{array}{ll}
\Phi^{\sigma s
\eta\tilde{s}}_{rul\tilde{r}\tilde{l}\tilde{d}}&~~\mbox{for}~~r < \chi~\mbox{and}~\tilde{r}<\chi \nonumber\\
\Psi^{\sigma s
\eta\tilde{s}}_{(r-\chi)ul(\tilde{r}-\chi)\tilde{l}\tilde{d}}&~~\mbox{for}~~r \ge \chi~\mbox{and}~\tilde{r} \ge \chi \nonumber\\
0&~~\mbox{otherwise} \nonumber
\end{array} \right .$$ Obviously, the vectors with the enlarged bond dimensions are defined as $$\lambda^{AB}_{r} \equiv \left\{
\begin{array}{ll}
\lambda^{AB}_{r}&~~\mbox{for}~~r < \chi \nonumber\\
\lambda^{AB}_{r-\chi}&~~\mbox{for}~~r \ge \chi \nonumber
\end{array} \right .$$ As a result of doubling, the addition of two tensor networks can be written as a single tensor network but with the increased bond dimensions for the $r$ and $l$ indices. In consequence, we can write the chain as $$\begin{array}{cccccccccccc}
& & | & &|& &|& &|& \\
&-& &=&B&=&C&=&D&- \\
& & \Theta & &|& &|& &\parallel& \\
&-& &=&F&=&G&=&H&- \\
& & | & &|& &|& &|&
\end{array}$$ where the rightmost tensors $D$ and $H$ have the doubled indices $d$ and $u$, respectively, and are as follows: $$(=D-)^{\gamma s}_{ruld} \equiv \left\{
\begin{array}{ll}
D^{\gamma s}_{ruld}&~~\mbox{for}~~d < \chi~\mbox{and}~l<\chi \nonumber\\
(-1)^{\gamma}D^{\gamma s}_{ru(l-\chi)(d-\chi)}&~~\mbox{for}~~d \ge \chi~\mbox{and}~l \ge \chi \nonumber\\
0&~~\mbox{otherwise} \nonumber
\end{array} \right .$$ $$(=H-)^{\delta s}_{ruld} \equiv \left\{
\begin{array}{ll}
H^{\delta s}_{ruld}&~~\mbox{for}~~u < \chi~\mbox{and}~l<\chi \nonumber\\
(-1)^{\delta}H^{\delta s}_{r(u-\chi)(l-\chi)d}&~~\mbox{for}~~u \ge \chi~\mbox{and}~l \ge \chi \nonumber\\
0&~~\mbox{otherwise} \nonumber
\end{array} \right .$$ It is useful to see the matrix forms of $\Theta$ and $B$ written in the following way: $${}^{-}_{-}\Theta^{=}_{=} \equiv \left(
\begin{array}{c|c}
{}^{-}_{-}\Phi^{-}_{-} & 0 \\
\hline 0 & {}^{-}_{-}\Psi^{-}_{-}
\end{array}
\right) ~~~\mbox{and}~~~ =B= \equiv \left(
\begin{array}{c|c}
-B- & 0 \\
\hline 0 & -(-1)^{\rho}B-
\end{array}
\right)$$ and the similar forms for $D$ and $H$.
To maintain the bond dimension, we make an approximation using SVD. We perform SVD for $\Theta$ first; then, we obtain $\acute{A}$, $\acute{E}$, and simultaneously, we obtain the vector $\tilde{\lambda}^{AE}$. Next, we perform SVD again from $-\acute{A}=B=$ to $-\tilde{A}-\acute{B}=$ as follows: $$\begin{array}{cccccccccccccccccc}
& | & &|& & & & | & &|& & & & | & & | & \\
-& &=&B&=& &-&\acute{A}&=&B&=& &-&\tilde{A}&-&\acute{B}&= \\
& \Theta & &|& &\rightarrow & & | & &|& &\rightarrow & & | & & | & \\
-& &=&F&=& &-&\acute{E}&=&F&=& &-&\tilde{E}&-&\acute{F}&= \\
& | & &|& & & & | & &|& & & & | & & | &
\end{array}$$ For $B$ and $C$ in Fig. 2, we change from $-\acute{B}=C=$ to $-\tilde{B}-\acute{C}=$ by using SVD. We continue performing SVD tensor by tensor until we reach the rightmost $D$. We also do the same thing for the lower half-chain from $E$ to $H$. Finally, we obtain $\acute{D}$ and $\acute{H}$, and we perform SVD as follows: $$\begin{array}{ccccccccc}
& |& & & & |& \\
-&\acute{D}&-& &-&\tilde{D}&- \\
&\parallel& & \rightarrow & & |& \\
-&\acute{H}&-& &-&\tilde{H}&- \\
& |& & & & |&
\end{array}$$ where the right-hand bonds of $D$ and $H$ are connected to the left-hand bonds of the leftmost tensors by the periodic boundary condition.
It is worth noting that we can approximate the tensor chain in different orderings. For example, we perform SVD first for the bond between $B$ and $C$ or $A$ and $B$ as shown below: $$\begin{array}{cccccccccc}
&-&A&=&B&=&C&=&D&-\\
\rightarrow&-&A&=&B&-&C&=&D&-\\
\rightarrow&-&A&-&B&-&C&=&D&-\\
\rightarrow&-&A&-&B&-&C&-&D&-
\end{array}$$ or $$\begin{array}{cccccccccc}
&-&A&=&B&=&C&=&D&-\\
\rightarrow&-&A&-&B&=&C&=&D&-\\
\rightarrow&-&A&-&B&-&C&=&D&-\\
\rightarrow&-&A&-&B&-&C&-&D&-
\end{array}$$ We note that there are twice as many singular values for $=B=C=\rightarrow =B-C=$ as those for $-B=C=\rightarrow -B-C=$ in the approximation by SVD. Because we keep only a fixed number of singular values, we lose more for $=B=C=\rightarrow =B-C=$ than for $-B=C=\rightarrow -B-C=$. It is reasonable to perform SVD one by one from the end as our scheme above.
When we apply $\exp(-\frac{1}{2}H_{ve}^{\uparrow}\tau)$ to the tensor network state, we assume that the elementary operators act on the tensors one by one from right to left. Thus, the vectors on the horizontal bonds are updated repeatedly. This convention is different from the case of the horizontal hopping $\exp(-\frac{1}{4}H_{he}^{\uparrow}\tau)$, which updates the vectors on the horizontal bonds only once. As a result, after we perform SVD repeatedly for $\exp(-\frac{1}{2}H_{ve}^{\uparrow}\tau)$, we return to the same form of the tensor network state with modified tensors and vectors: $$\begin{array}{ccccccccccccccccc}
& |& & |& & |& & |& \\
-&\tilde{A}&-&\tilde{B}&-&\tilde{C}&-&\tilde{D}&-\\
& |& & |& & |& & |& \\
-&\tilde{E}&-&\tilde{F}&-&\tilde{G}&-&\tilde{H}&-\\
& |& & |& & |& & |&
\end{array}$$
For $\exp(-H_{vo}^{\uparrow}\tau)$, we follow a similar procedure for the odd vertical bonds for example between $H$ and $L$ in Fig. 2. After doubling, we represent the portion that should be updated as $$\begin{array}{ccccccccccc}
&| & &|& &|& & |& \\
-&E &=&F&=&G&=& &- \\
&\parallel & &|& &|& & \Theta& \\
-&I &=&J&=&K&=& &- \\
&| & &|& &|& & |& \\
\end{array}$$ We assume that the elementary operators in $\exp(-H_{vo}^{\uparrow}\tau)$ act from left to right. Because doubling should be performed in the left-hand part of our zigzag ordering, $\Theta$ is in the right-hand part of this network. We follow the same procedure for approximation: via the SVD of ${}^{=}_{=}\Theta^{-}_{-}$, we obtain the tensors of $=\acute{H}-$ and $=\acute{L}-$, and simultaneously, we obtain the updated vector $\tilde{\lambda}^{HL}$. Again, we repeat the SVD process to reduce the doubled bond dimensions to the original dimensions. In the end, we obtain updated tensors and vectors.
Updates to the Spin Bonds
-------------------------
Because the elementary operator for the spin bond $\exp\{e\tau -
U\tau (n_{i\uparrow}-\frac{1}{2})(n_{i\downarrow}-\frac{1}{2})+\mu
\tau (n_{i\uparrow}+n_{i\downarrow}) \}$ is diagonal with respect to our typical base vectors, we obtain $$\begin{aligned}
&&\exp\{e\tau - U\tau
(n_{i\uparrow}-\frac{1}{2})(n_{i\downarrow}-\frac{1}{2})+\mu \tau
(n_{i\uparrow}+n_{i\downarrow}) \}|\cdots\sigma_i
\cdots\sigma_{N+i}\cdots \rangle
\nonumber \\
&&~= \exp\{e\tau -U\tau(\sigma_i
-\frac{1}{2})(\sigma_{N+i}-\frac{1}{2}) + \mu\tau (\sigma_i +
\sigma_{N+i} ) \}|\cdots\sigma_i \cdots\sigma_{N+i}\cdots \rangle
. \nonumber\end{aligned}$$ From this equation, we determine the ten-index tensor $\Theta^{\sigma
\tilde{\sigma}}_{ruld\tilde{r}\tilde{u}\tilde{l}\tilde{d}}$ to update the vectors on the spin bonds as follows: $$\Theta^{\sigma
\tilde{\sigma}}_{ruld\tilde{r}\tilde{u}\tilde{l}\tilde{d}}=
\exp\{e\tau -U\tau(\sigma-\frac{1}{2})(\tilde{\sigma}-\frac{1}{2})
+ \mu\tau (\sigma + \tilde{\sigma}) \} \times \left[
\begin{array}{c}
A_{\uparrow} \\
\otimes \\
A_{\downarrow}
\end{array}
\right]{}^{\sigma
\tilde{\sigma}}_{ruld\tilde{r}\tilde{u}\tilde{l}\tilde{d}}$$ where the tensor product is given by $$\left[
\begin{array}{c}
A_{\uparrow} \\
\otimes \\
A_{\downarrow}
\end{array}
\right]{}^{\sigma
\tilde{\sigma}}_{ruld\tilde{r}\tilde{u}\tilde{l}\tilde{d}} \equiv
\sum_{x=0}^{\kappa -1} A^{\sigma x}_{\uparrow ruld} \lambda^{AA}_x
A^{\tilde{\sigma} x}_{\downarrow
\tilde{r}\tilde{u}\tilde{l}\tilde{d}} \lambda^{AB}_{\uparrow r}
\lambda^{MA}_{\uparrow u} \lambda^{DA}_{\uparrow l}
\lambda^{AE}_{\uparrow d} \lambda^{AB}_{\downarrow\tilde{r}}
\lambda^{MA}_{\downarrow\tilde{u}}\lambda^{DA}_{\downarrow\tilde{l}}\lambda^{AE}_{\downarrow\tilde{d}}$$ with the periodic boundary condition in Fig. 2.
As in the horizontal-bond update, we perform SVD for $\Theta$ to find $\tilde{\lambda}^{AA}$. By dividing and attaching the eight vectors, we obtain the tensor update $\tilde{A}$: $$\begin{aligned}
\Theta^{\sigma \tilde{\sigma}
}_{ruld\tilde{r}\tilde{u}\tilde{l}\tilde{d}} &\rightarrow&
\sum_{x=0}^{\kappa-1} \bar{A}^{\sigma x}_{\uparrow ruld}
\tilde{\lambda}^{AA}_x \bar{A}^{\tilde{\sigma}
x}_{\downarrow\tilde{r}\tilde{u}\tilde{l}\tilde{d}}\nonumber \\
&=&\sum_{x=0}^{\kappa -1} \tilde{A}^{\sigma x}_{\uparrow ruld}
\tilde{\lambda}^{AA}_x \tilde{A}^{\tilde{\sigma} x}_{\downarrow
\tilde{r}\tilde{u}\tilde{l}\tilde{d}} \lambda^{AB}_{\uparrow r}
\lambda^{MA}_{\uparrow u} \lambda^{DA}_{\uparrow l}
\lambda^{AE}_{\uparrow d} \lambda^{AB}_{\downarrow\tilde{r}}
\lambda^{MA}_{\downarrow\tilde{u}}\lambda^{DA}_{\downarrow\tilde{l}}\lambda^{AE}_{\downarrow\tilde{d}}.\end{aligned}$$
We follow the same procedure for all $2L^{2}$ tensors. It is easy to parallelize this process using multi-core computers.
Energy Updates
--------------
Being inspired by the diffusion Monte Carlo [@Ceperley; @Chung1], we introduced the energy per site $e$ in the operator of the Suzuki-Trotter decomposition. While the energy in the diffusion Monte Carlo is adjusted by controlling the number of replicas, here we determine $e$ by managing the factor in front of the wave function. The algorithm is as follows: when a typical operator $\exp(h\tau)$ acts on a tensor network state $|\mbox{TNS}
\rangle$, we perform SVD and obtain $\chi$ singular values of $\lambda_{0} \ge \lambda_{1} \ge \cdots \ge \lambda_{\chi -1}$. We take $\lambda_{0}$ and place it in front of the wave function, and we modify the singular values as follows: $1 \ge
\lambda_{1}/\lambda_{0} \ge \cdots \ge \lambda_{\chi
-1}/\lambda_{0}$. In this way, we normalize $|\mbox{TNS} \rangle$ such that all $5L^{2}$ vectors on each bond have the maximum value of 1. Thus, whenever the weights are modified by $\exp(E_{k}\tau-H\tau)$ acting on the $k$-th time step state $|\mbox{TNS}_{k} \rangle$, we take out the maximum weight to obtain the factor $F$ in front of the state $$\exp(E_{k}\tau - H\tau)|\mbox{TNS}_{k} \rangle =
F|\mbox{TNS}_{k+1} \rangle ,$$ where $|\mbox{TNS}_{k+1} \rangle$ is a normalized TNS. We obtain the factor $F$ such that $F~*=\lambda_{0}$ whenever any bond is modified. Because we require no divergence and no convergence to zero for the state, as in the diffusion Monte Carlo, we adjust the next energy value $e_{k+1}$ for $F$ to approach 1 in this way: $$e_{k+1} = e_{k} - \xi \log F ,$$ where the value of the feedback parameter $\xi$ is not sensitive in this algorithm. After we find $e_{k+1}$, we set $F=1$ again for the next iteration in the computer simulation. We note that during the time evolution, $e_{k}$ is stable and approaches the ground-state energy per site in the limit of $k \rightarrow
\infty$. The solution of $|\mbox{TNS}_{\infty} \rangle$ is also stable.
Numerical Results
=================
It is instructive to summarize the parameters that are involved in our task of calculating the ground state of the Hubbard model. The model Hamiltonian itself contains three parameters. The tensor network states are defined by the internal-bond dimension, the spin-bond dimension, and the lattice length. We need the Trotter parameter, the feedback parameter for energy adjustment, and the seed for the random number generator we used to set the tensors and vectors for an initial state in TEBD. Thus, we should set nine values initially in the simulation: $$\begin{aligned}
t, ~~U, ~~\mu &~~\mbox{in}~~H,\\
\chi, ~~\kappa, ~~L &~~\mbox{in}~~|\mbox{TNS}\rangle,\\
\tau, ~~\xi,~~\mbox{seed} &~~\mbox{in}~~\mbox{TEBD}.\end{aligned}$$ There are several alternative methods for creating initial states; for instance, all components of tensors and vectors are fixed intentionally without using the random number generator. In this case, we need no seed.
Our goal is to find the stable $e$ and $|\mbox{TNS}\rangle$. From the Suzuki-Trotter decomposition of Eq. (3), we describe the procedure for the computational simulation:
1. For a given seed number, all $2\kappa\chi^{4}$ components of the $2L^{2}$ tensors are given by random numbers between $-0.5$ and $0.5$, and all components of the $5L^{2}$ vectors on the bonds are given by random numbers between $0$ and $1$. Another option is that, with no seed numbers, all components of all of the tensors and vectors are given by $1$. After choosing an initial tensor network state, let $e=0$ and $F=1$.
2. Update the horizontal bonds, the vertical bonds, and then the horizontal bonds in the spin-up layer.
3. Update the horizontal bonds, the vertical bonds, and then the horizontal bonds in the spin-down layer.
4. Update the spin bonds.
5. Update $e$, and set $F=1$. Repeat from step (ii) until $F$ remains stably $1$.
First, we present the typical behavior of the converging energy $e$ in Fig. 3. We find that, regardless of which nine parameters are used in our calculations, we obtain similar behavior for $e$ to what is shown in Fig. 3 for all of the other cases. We find that $\xi$ has no effect on the converging energy value as long as it is small enough. Furthermore, we have varied the Trotter parameter $\tau$, and we find that there are no significant variations in the converging energy $e$ as a function of $\tau$. Hence, in the main simulations, we fix $\xi=0.03$ and $\tau =
0.02$ for TEBD.
Because it is reasonable that the converging energy value is independent of any initial state, we should obtain the same ground-state energy up to the Suzuki-Trotter uncertainty $\tau^{2}$ as long as $t$, $U$ and $\mu$ are fixed. However, it seems that there are some barriers in the Hilbert space that prevent the evolving state from accessing the true ground state. In other words, if the initial state begins from a topologically different sector, it will never approach the true ground state in the process of TEBD. For example, in Fig. 3, we find the difference between the two converging energy values for the initial random tensors and the initial tensors whose components are fixed as 1. Thus, in further calculations, we should repeat simulations with several seed numbers to study the ground-state degeneracy and the disjoint space of tensor network states.
By changing the other parameters $t$, $U$, $\mu$, $\chi$, $\kappa$, and $L$, we can further verify the consistency. It is obvious that the ground-state energy $e$ should become twice as large when we simultaneously double $t$, $U$, and $\mu$. We have checked this consistency so that we can fix the value of $t$ as usual as $1$. Because the exact ground-state energy for the non-interacting infinite system [@Valenti] is known as $-1.6211$, we can compare the exact value to our value of $-1.3422$ for $\chi=\kappa=2$ in the system of $t=1$, $U=0$, $\mu=0$, and $L=10$. Furthermore, there is another exact result of the ground-state energy $-1.8514$ at $t=1$, $U=4$, $\mu=0$, and $L=4$ [@Parola]. We compare it to our result of $-1.6508$ for $\chi=\kappa=2$ and $-1.6539$ for $\chi=\kappa=3$. Because of the finite values of $\chi$ and $\kappa$, there are some differences between the exact and ours. We find a tendency for our value to more closely approach the exact value as we increase $\chi$. However, the difference of $0.2$ is not small, and increasing $\chi$ may not improve the ground-state energy significantly. It indicates that the tensor network state in Fig. 2 may be incorrect.
In order to find the finite size effect related to $L$, we calculate the ground-state energy for the Hamiltonian of $t=1$, $U=4$ and $\mu=0$ by changing $L$. To save computing time, we perform the calculation for only the easy case of $\chi=\kappa=2$. We summarize the numerical results for various values of $L$ in Fig. 4 where we observe the saturation at large $L$.
Because the spin-flip symmetry preserving states in Eq. (6) are living in the subset of the Hilbert space for the spin-flip symmetry breaking states in Eq. (5), the converging energy for the state of Eq. (6) should be greater than or equal to the energy for the state of Eq. (5). For the spin-flip symmetry preserving states, in the process of TEBD, we perform updating the tensors and vectors in the spin-up layer, and then we duplicate the tensors and vectors in the spin-down layer from those in the spin-up layer. Because we duplicate the bonds in the spin-down layer, we should modify the factor $F$ such as $F~*=\lambda_{0}^{2}$ in the process of the horizontal and vertical bonds updating. We present the numerical results at $\mu
= 0$ in Fig. 5, comparing the ground-state energy of the spin-flip symmetry preserving state with that of the spin-flip symmetry breaking state. We note that the ground-state energy of the spin-flip symmetry preserving state is slightly lower than that of the spin-flip symmetry breaking state at small $U$. This is caused by numerical uncertainties, and it is understood as equality. This means that the symmetry breaking does not take place yet. We find from Fig. 5 that there is a transition at $U=0.39(1)$ for $\mu=0$. At large $U$, the ground-state energy of the spin-flip symmetry breaking state depends heavily on $U$. We note that the ground-state energy of the symmetry preserving state is almost independent of $U$. This independence means that the expectation value of the number operator is given by $\langle n_{i\uparrow}
\rangle = \langle n_{i\downarrow} \rangle \approx \frac{1}{2}$ for any $i$.
We conclude that our method is effective in searching for the ground state of the Hubbard model. We emphasize that it is possible to determine the energy and the ground state for any chemical potential.
Conclusion
==========
In summary, we have presented a method for obtaining the ground-state energy and the wave function for two-dimensional quantum many-fermion systems, especially the Hubbard model. We may call this method diffusive TEBD. Because there is a certain discrepancy between the exact ground-state energy and our value for the TNS, it is still questionable whether or not the TNS is correct and the diffusive TEBD is useful. We suggest that the diffusive TEBD is an effective method.
Although we built a user-friendly library in the framework of previous computer code [@Chung2], we obtain only preliminary numerical results because we use the full SVD, which is very inefficient. In future work, we will implement an SVD package based on the Lanczos algorithm with partial reorthogonalization [@Simon] to find only a few eigenvectors and their corresponding singular values, which are sufficient for our truncation scheme.
When we use multi-core computers, it is possible to parallelize the local updates of the horizontal bonds and the spin bonds. For the vertical-bond update, we may apply the concept of a pipeline to optimize the roles of the multiple cores. We anticipate progress in this parallel computing scheme.
In future work, for a fixed $U$, we need to investigate whether any phase transitions happen as we change the chemical potential $\mu$ in the Hamiltonian. If there are any transitions in simulations, the phase transitions may be related to topological orders [@Wen] or the topological entanglement entropy [@Kitaev]. In connection with topological orders, we should give a definitive answer to the ground-state degeneracies. Furthermore, it is necessary to perform the same simulation by changing periodic or open boundary conditions.
It is of interest to extend our method to the case of two-body interactions. A typical topic of interest for two-body interactions may be the fractional quantum Hall effect, for which MPS can be used as an accessible subset of the huge Hilbert space. In the fractional quantum Hall effect, the energy gap between the ground state and the first excited state provides a lesser entanglement entropy, which makes it possible to use MPS with a relatively small bond dimension.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was partially supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education, Science and Technology (Grant No. 2011-0023395) and by the Supercomputing Center at Korea Institute of Science and Technology Information with their supercomputing resources, including technical support (Grant No. KSC-2012-C1-09). The author would like to thank Michelle Ebbs for reading the manuscript.
References {#references .unnumbered}
==========
[32]{}
Gutzwiller M C 1963 Effect of Correlation on the Ferromagnetism of Transition Metals [*Phys. Rev. Lett.*]{} [**10**]{} 159
Kanamori J 1963 Electron Correlation and Ferromagnetism of Transition Metals [*Prog. Theor. Phys.*]{} [**30**]{} 275
Hubbard J 1963 Electron Correlations in Narrow Energy Bands [*Proc. R. Soc. A*]{} [**276**]{} 238
Hubbard J 1964 Electron Correlations in Narrow Energy Bands. III. An Improved Solution [*Proc. R. Soc. A*]{} [**281**]{} 401
Dagotto E 1994 Correlated electrons in high-temperature superconductors [*Rev. Mod. Phys.*]{} [**66**]{} 763
Greiner M, Mandel O, Esslinger T, Hansch T W and Bloch I 2002 Quantum phase transition from a superfluid to a Mott insulator in a gas of ultracold atoms [*Nature*]{} [**415**]{} 39
Jördens R, Strohmaier N, Günter K, Moritz H and Esslinger T 2008 A Mott insulator of fermionic atoms in an optical lattice [*Nature*]{} [**455**]{} 204
Troyer M and Wiese U J 2005 Computational Complexity and Fundamental Limitations to Fermionic Quantum Monte Carlo Simulations [*Phys. Rev. Lett.*]{} [**94**]{} 170201
Wilson K G and Kogut J 1974 The renormalization group and the $\epsilon$ expansion [*Phys. Rep.*]{} [**12**]{} 75
Wilson K G 1975 The renormalization group: Critical phenomena and the Kondo problem [*Rev. Mod. Phys.*]{} [**47**]{} 773
Kondo J 1964 Resistance Minimum in Dilute Magnetic Alloys [*Prog. Theor, Phys.*]{} [**32**]{} 37
White S R 1992 Density matrix formulation for quantum renormalization groups [*Phys. Rev. Lett.*]{} [**69**]{} 2863
Östlund S and Rommer S 1995 Thermodynamic Limit of Density Matrix Renormalization [*Phys. Rev. Lett.*]{} [**75**]{} 3537
García-Ripoll J J 2006 Time evolution of Matrix Product States [*New J. Phys.*]{} [**8**]{} 305
Saberi H, Weichselbaum A and von Delft J 2008 Matrix-product-state comparison of the numerical renormalization group and the variational formulation of the density-matrix renormalization group [*Phys. Rev. B*]{} [**78**]{} 035124
Schollwöck U 2011 The density-matrix renormalization group in the age of matrix product states [*Ann. Phys.*]{} [**326**]{} 96
Pérez-García D, Verstraete F, Wolf M M and Cirac J I 2008 PEPS as unique ground states of local Hamiltonians [*Quant. Inf. Comp.*]{} [**8**]{} 0650
Orús R 2012 Exploring corner transfer matrices and corner tensors for the classical simulation of quantum lattice systems [*Phys. Rev. B*]{} [**85**]{} 205117
Murg V, Verstraete F, Legeza Ó and Noack R M 2010 Simulating strongly correlated quantum systems with tree tensor networks [*Phys. Rev. B*]{} [**82**]{} 205105
Vidal G 2007 Entanglement Renormalization [*Phys. Rev. Lett.*]{} [**99**]{} 220405
Chou C P, Pollmann F and Lee T K 2012 Matrix-product-based projected wave functions ansatz for quantum many-body ground states [*Phys. Rev. B*]{} [**86**]{} 041105
Levin M and Nave C P 2007 Tensor Renormalization Group Approach to Two-Dimensional Classical Lattice Models [*Phys. Rev. Lett.*]{} [**99**]{} 120601
Jiang H C, Weng Z Y and Xiang T 2008 Accurate Determination of Tensor Network State of Quantum Lattice Models in Two Dimensions [*Phys. Rev. Lett.*]{} [**101**]{} 090603
Xie Z Y, Chen J, Qin M P, Zhu J W, Yang L P and Xiang T 2012 Coarse-graining renormalization by higher-order singular value decomposition [*Phys. Rev. B*]{} [**86**]{} 045139
Vidal G 2003 Efficient Classical Simulation of Slightly Entangled Quantum Computations [*Phys. Rev. Lett.*]{} [**91**]{} 147902
Vidal G 2004 Efficient simulation of one-dimensional quantum many-body systems [*Phys. Rev. Lett.*]{} [**93**]{} 040502
Vidal G 2007 Classical Simulation of Infinite-Size Quantum Lattice Systems in One Spatial Dimension [*Phys. Rev. Lett.*]{} [**98**]{} 070201
Barthel T, Pineda C and Eisert J 2009 Contraction of fermionic operator circuits and the simulation of strongly correlated fermions [*Phys. Rev. A*]{} [**80**]{} 042333
Kraus C V, Schuch N, Verstraete F and Cirac J I 2010 Fermionic projected entangled pair states [*Phys. Rev. A*]{} [**81**]{} 052338
Corboz P, Orús R, Bauer B and Vidal G 2010 Simulation of strongly correlated fermions in two spatial dimensions with fermionic projected entangled-pair states [*Phys. Rev. B*]{} [**81**]{} 165104
Pižorn I and Verstraete F 2010 Fermionic implementation of projected entangled pair states algorithm [*Phys. Rev. B*]{} [**81**]{} 245110
Corboz P, Jordan J and Vidal G 2010 Simulation of fermionic lattice models in two dimensions with projected entangled-pair states: Next-nearest neighbor Hamiltonians [*Phys. Rev. B*]{} [**82**]{} 245119
Corboz P, White S R, Vidal G and Troyer M 2011 Stripes in the two-dimensional t-J model with infinite projected entangled-pair states [*Phys. Rev. B*]{} [**84**]{} 041108
Pineda C, Barthel T and Eisert J 2010 Unitary circuits for strongly correlated fermions [*Phys. Rev. A*]{} [**81**]{} 050303
Corboz P and Vidal G 2009 Fermionic multiscale entanglement renormalization ansatz [*Phys. Rev. B*]{} [**80**]{} 165129
Corboz P, Evenbly G, Verstraete F and Vidal G 2010 Simulation of interacting fermions with entanglement renormalization [*Phys. Rev. A*]{} [**81**]{} 010303
Marti K H, Bauer B, Reiher M, Troyer M and Verstraete F 2010 Complete-graph tensor network states: a new fermionic wave function ansatz for molecules [*New J. Phys.*]{} [**12**]{} 103008
Eisert J, Cramer M and Plenio M B 2010 Area laws for the entanglement entropy [*Rev. Mod. Phys.*]{} [**82**]{} 277
Chung M H 2014 Diffusive time-evolving block decimation method for matrix product states in one-dimensional spinless fermion systems [*J. Korean Phys. Soc.*]{} [**64**]{} 999
Ceperley D and Alder B 1986 Quantum Monte Carlo [*Science*]{} [**231**]{} 555
Chung M H and Landau D P 2012 Diffusion Monte Carlo for fermions with replica reduction [*Phys. Rev. B*]{} [**85**]{} 115115
Valentí R, Stolze J and Hirschfeld P J 1991 Lower bounds for the ground-state energies of the two-dimensional Hubbard and $t$-$J$ models [*Phys. Rev. B*]{} [**43**]{} 13743
Parola A, Sorella S, Baroni S, Parrinello M and Tosatti E 1989 Static properties of the 2D Hubbard model on a $4 \times 4$ cluster [*Int. J. Mod. Phys. B*]{} [**3**]{} 1865
Chung M H 2008 Science Code .Net: Object-oriented programming for science [*Sci. Comput. Program.*]{} [**71**]{} 242
Simon H D 1984 The Lanczos algorithm with partial reorthogonalization [*Math. Comp.*]{} [**42**]{} 115
Wen X G 1990 Topological Orders in Rigid States [*Int. J. Mod. Phys. B*]{} [**4**]{} 239
Kitaev A and Preskill J 2006 Topological Entanglement Entropy [*Phys. Rev. Lett.*]{} [**96**]{}, 110404
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Measurement incompatibility is the most basic resource that distinguishes quantum from classical physics. Contextuality is the critical resource behind the power of some models of quantum computation and is also a necessary ingredient for many applications in quantum information. A fundamental problem is thus identifying when incompatibility produces contextuality. Here, we show that, given a structure of incompatibility characterized by a graph in which nonadjacent vertices represent incompatible ideal measurements, the necessary and sufficient condition for the existence of a quantum realization producing contextuality is that this graph contains induced cycles of size larger than three.'
author:
- 'Zhen-Peng Xu'
- Adán Cabello
title: Necessary and sufficient condition for contextuality from incompatibility
---
[*Incompatibility versus contextuality.*]{} Measurement incompatibility is arguably the most basic resource that distinguishes quantum and classical physics. Incompatibility is ubiquitous in protocols with a quantum-over-classical advantage and has been proven to be necessary for no-cloning [@HMZ16] and nonlocality [@KC85; @WPF09; @QBHB16]. On the other hand, contextuality (a concept resulting from the Kochen-Specker theorem [@Specker60; @Bell66; @KS67], but here used in the exact sense used in Refs. [@KCBS08; @Cabello08; @YO12; @KBLGC12; @AQB13; @CSW14]) is the critical resource behind the quantum advantage of some models of quantum computation [@AB09; @HB11; @Raussendorf13; @HWVE14; @DGBR15; @ASM17; @BDBOR17; @RBDOB17] and a necessary ingredient for many quantum protocols (e.g., device-independent quantum key distribution [@Ekert91; @BHK05], quantum advantage in zero-error classical communication [@CLMW10], and some cryptographic protocols [@CDNS11]). Therefore, a fundamental question is what is the relation between incompatibility and contextuality. This is the problem we address in this Rapid Communication
The definition of measurement incompatibility is independent of any physical theory. Two measurements, $A$, with outcome set $\{a_x\}_{x \in X}$, and $B$, with outcome set $\{b_y\}_{y \in Y}$, are incompatible (or not jointly measurable) if there is no measurement $C$ with outcome set $\{c_{x,y}\}_{x \in X, y \in Y}$ such that, for all initial states $\rho$, the probability $P(a_x|\rho)= \sum_{y \in Y} P(c_{x,y}|\rho)$, for all outcomes $a_x$, and the probability $P(b_y|\rho)=\sum_{x \in X} P(c_{x,y}|\rho)$, for all outcomes $b_y$. If such a $C$ exists, then $A$ and $b$ are compatible (or jointly measurable). In other words, two measurements $A$ and $B$ are incompatible if there does not exist a measurement $C$ such that both $A$ and $B$ are coarse grainings of $C$.
A measurement scenario is characterized by a set ${\cal M}$ of measurements, their respective outcomes, and the subsets of ${\cal M}$ that are compatible. The relations of compatibility between the measurements in a scenario are usually represented by a hypergraph in which each vertex represents a measurement and vertices in the same hyperedge are mutually compatible (see, e.g., Refs. [@HRS08; @HFR14; @AT18]).
In general, contextuality indicates that the outcome statistics of an experiment involving several contexts (i.e., sets of compatible measurements) cannot be explained assuming that the outcomes reveal preexisting values that are independent of the context. However, there are several definitions of contextuality in the literature. The one for which a crucial connection with quantum computation has been established [@HWVE14] is the one used in Refs. [@KCBS08; @Cabello08; @YO12; @KBLGC12; @AQB13; @CSW14]. Given a measurement scenario where all measurements are ideal, a behavior (i.e., a set of probability distributions, one for each context) is contextual if it does not belong to the polytope whose vertices are all possible deterministic assignments of outcomes to the measurements in that scenario. A measurement is ideal (or sharp) [@CY14; @CY16] if (i) it yields the same outcome when performed consecutive times, (ii) it only disturbs measurements that are incompatible with it, and (iii) all its coarse grainings have realizations satisfying (i) and (ii). In quantum theory, an ideal measurements is represented by a self-adjoint operator $A$ on a Hilbert space or, equivalently, by the set of orthogonal projectors (onto distinct, possibly degenerate, eigenspaces of $A$) summing to the identity in the spectral decomposition of $A$. On the other hand, compatible measurements are represented in quantum theory by commuting operators.
The restriction of the definition of contextuality to scenarios involving only ideal measurements obeys three main reasons: (I) It assures that compatible measurements do not disturb each other (which is what naturally happens in Bell scenarios due to the fact that measurements are spatially separated). A measurement $A$ disturbs a measurement $B$ if, for some initial state, from the outcome statistics of $B$, one can detect whether $A$ was performed. Recall that, for nonideal measurements, compatibility does not imply nondisturbance [@HRS08; @HW10; @HFR14]. (II) It assures that the contextuality of a behavior can be taken as a signature of nonclassicality. On the one hand, as pointed out in Ref. [@Spekkens14], the assumption that the outcome of a measurement depends deterministically on the ontic state (which is the assumption satisfied by the extreme points of the set of noncontextual behaviors) is reasonable if and only if the measurement is ideal. In particular, it is not a physically plausible assumption when applied to a noisy measurement (even a classical one), since, in this case, the outcome may have an indeterministic dependence on the ontic state of the measured system. On the other hand, the classical simulation of quantum contextuality for ideal measurements has a quantifiable memory [@KGPLC11; @CGGX18] and thermodynamical overcosts [@CGGLK16]. (III) It assures that the classical and quantum sets of behaviors are direct generalizations of the corresponding sets for Bell scenarios. In particular, for contextuality scenarios that, by spacelike separating the measurements, can be converted into Bell scenarios, the sets of behaviors are identical regardless of whether there is timelike or spacelike separation.
For ideal measurements, if in a set of measurements every two of them are compatible, then all of them are compatible [@CY14] (this is not true for nonideal measurements [@Specker60; @HRS08]). As a consequence, the relations of compatibility between ideal measurements can be represented by a simple graph, called compatibility graph, in which any clique of vertices represents a set of compatible measurements (see, e.g., Refs. [@AT18; @KRK12; @CDLP13]). A clique of a graph is a set of vertices every pair of which are adjacent.
The fundamental problem is what is the relation between incompatibility and contextuality. Clearly, incompatibility is necessary for contextuality. Otherwise, if all measurements are compatible, then there is only one context. However, not every set of measurements that includes incompatible measurements produces contextuality. Therefore, the crucial question is what incompatibility structures can produce quantum contextuality and which ones cannot. Surprisingly, we have not found the answer to this question in the literature.
A first step towards solving this problem is a theorem introduced by Vorob’yev [@Vorob'yev63; @Vorob'yev67] that has been used in connection to quantum theory in Refs. [@BM10; @BC12; @RSKK12; @Barbosa14; @Barbosa15; @BMC16]. The theorem states that, for any set of measurements whose corresponding compatibility graph is chordal (i.e., does not contain induced cycles of size larger than three), there is always a joint probability distribution for every behavior (see the Appendix). Therefore, in this case, all quantum behaviors can be simulated by a noncontextual hidden variable model. Recall that an induced subgraph of a graph $G(V,E)$, with vertex set $V$ and edge set $E$, is a graph with vertex set $S \subseteq V$ and edge set comprising all the edges of $G$ with both ends in $S$. An $n$-vertex cycle, denoted $C_n$, is a graph with $n$ vertices connected in a closed chain, e.g., $C_4$ is a square and $C_5$ is a pentagon. Therefore, a necessary condition for quantum contextuality is that the compatibility graph is not chordal.
[*Main result.*]{} The aim of this Rapid Communication is to prove and explore the consequences of the following result.
[*Theorem.*]{} For a given compatibility graph $G(V,E)$, with vertex set $V$ and edge set $E$, there is a set of quantum ideal measurements $\mathcal{M} = \{M_i\}_{i\in V}$ satisfying the incompatibility/compatibility structure given by $G(V,E)$ and producing contextuality if and only if $G(V,E)$ is not a chordal graph.
[*Proof.*]{} That the nonchordality of the compatibility graph is a necessary condition for contextuality follows from the proof of Vorob’yev’s theorem (see the Appendix). That nonchordality of the compatibility graph is a sufficient condition for contextuality can be proven as follows. Let $G_{1}(V_1, E_1)$ be the compatibility graph of ${\cal M}_1 = \{M_i\}_{i \in V_1}$. Let $G_{2}(V_2, E_2)$ be a compatibility graph such that $G_{1}(V_1, E_1)$ is an induced subgraph of $G_{2}(V_2, E_2)$ and $V_2 = V_1 \ \bigcup \{v_0\}$, where $v_0$ is a vertex that is not in $V_1$. The following set of measurements, ${\cal M}_2 = \{\overline{M}_i\}_{i \in V_2}$, has $G_{2}(V_2, E_2)$ as its compatibility graph, $$\overline{M}_i = M_i \bigotimes_{j\in V_1} \Pi_{i,j},\,\forall i \in V_1\;\;\text{and}\;\;\overline{M}_{v_0} = \mathbb{I}_d \bigotimes_{j\in V_1} P_{j},$$ where $$\Pi_{i,j} = \begin{cases}
\mathbb{I}_2 , & j\neq i,\\
|0\rangle\langle 0|, & j=i,
\end{cases}$$ $$P_j = \begin{cases}
\mathbb{I}_2, & (v_0,j) \in E_2,\\
|\psi\rangle\langle \psi|, & (v_0,j)\not\in E_2,
\end{cases},$$ $d$ is the dimension of each of the elements in $\{M_i\}_{i \in V_1}$, $\mathbb{I}_k$ is the identity operator in dimension $k$, and $|\psi\rangle = (|0\rangle + |1\rangle)/\sqrt{2}$. By construction, the state $\overline{\rho} = \rho \otimes_{j\in V_1} |0\rangle\langle 0|$ and the measurements $\overline{\mathcal{M}}_1 = \{\overline{M}_i\}_{i\in V_1}$ produce the same probabilities as the ones produced by $\rho$ and $\mathcal{M}_1$. That is, for every outcome $\overline{m_i}$ of $\overline{M}_i \in \overline{\mathcal{M}}_1$ and every outcome $m_i$ of $M_i \in \mathcal{M}_1$, $$P( \overline{m}_i | \overline{\rho} ) = P ( m_i | \rho).$$
This implies that, if there is an induced subgraph of a given compatibility graph $G$ that can produce contextuality, then the graph $G$ can produce at least the same amount of contextuality. Now, notice that a compatibility graph $G$ is not chordal if and only if it has induced cycles of size $k \ge 4$. Let us suppose that $C_k$ is one of them. If we could find a set of measurements ${\cal M}_C = \{M_i\}_{i \in V_C}$ whose compatibility graph were isomorphic to $C_k$ and that would produce contextuality, then, by the previous result, $G$ would also produce contextuality (at least the same amount as the induced $C_k$), thus proving our claim. For any $k$, explicit examples of sets of measurements satisfying all these requirements can be found in Ref. [@AQB13].
[*Classification of the scenarios with quantum contextuality.*]{} An interesting consequence of the previous theorem is that it allows us to identify and classify all measurement scenarios in which incompatibility can produce contextuality and tells us how to use quantum theory to produce contextuality in each of them. Given a fixed number $k$ of ideal measurements, to identify all scenarios that can produce contextuality, it is enough to compute all nonchordal graphs with $k$ vertices and avoid the cases in which one of the measurements is not needed for contextuality by removing those graphs in which one of the vertices does not belong to any cycle of length four or more. For $k$ up to $6$, the complete list of compatibility graphs corresponding to scenarios in which contextuality can occur is shown in Fig. \[Fig1\]. All these compatibility graphs can be realized in experiments with sequential measurements on single systems, such as the experiments of Refs. [@MWZ00; @YLB03; @KZG09; @LMZNACH17]. In addition, some of the compatibility graphs can be realized in multipartite scenarios, since their sets of vertices can be divided into disjoint subsets, each subset corresponding to the measurements of one party and containing some nonadjacent vertices (i.e., incompatible measurements), and such that each vertex in a subset is adjacent to all vertices in the other subsets. According to this criterion, the graphs of compatibility that can produce quantum contextuality can be classified in three types:
\(a) Nonchordal compatibility graphs that are complete $n$-partite, with $n\ge2$ (i.e., whose sets of vertices can be divided into $n$ disjoint and independent subsets such that each vertex in a set is adjacent to all vertices in the other subsets), as the graphs in Figs. \[Fig1\](a1)–\[Fig1\](a5). If $n=2$, then the graphs have realizations as bipartite Bell scenarios. For example, the graphs in Figs. \[Fig1\](a1)–\[Fig1\](a4). If $n=3$, then the graphs have realizations as tripartite Bell scenarios. For example, the graph in Fig. \[Fig1\](a5). The sets of classical and quantum behaviors for these scenarios have been studied extensively, since the boundaries of the classical (noncontextual) sets are tight Bell inequalities. Specifically, for measurements with two outcomes, the exhaustive list of tight Bell inequalities that bound the set of classical behaviors for the scenario whose compatibility graph is in Fig. \[Fig1\](a1) is in Refs. [@CHSH69; @Tsirelson80; @SZ81; @Froissart81; @Fine82], while the corresponding set of quantum behaviors is exhaustively characterized in Refs. [@Tsirelson80; @Tsirelson85; @KC85; @Tsirelson93; @NPA07; @NPA08]. Similarly, for the scenarios in Figs. \[Fig1\](a2)–\[Fig1\](a3), the tight Bell inequalities and their quantum violations are presented in Ref. [@BG08], and in Ref. [@CG04] for the scenario in Fig. \[Fig1\](a4). Finally, the full set of tight Bell inequalities for the scenario in Fig. \[Fig1\](a5) is in Refs. [@PS01; @Sliwa03] and the corresponding quantum violations in Ref. [@LXC16].
\(b) Nonchordal compatibility graphs that have realizations as multipartite scenarios (since their vertices can be divided into disjoint sets, each of them containing some nonadjacent vertices, and such that each vertex in a subset is adjacent to all vertices in the other subsets), but in which at least one party has at least two measurements that are compatible (i.e., at least one of the subsets is not an independent set). These graphs are shown in Figs. \[Fig1\](b1)–\[Fig1\](b8). So far, to our knowledge, these types of compatibility graphs have been considered only in relation with scenarios of nonlocality via local contextuality [@Cabello10; @LHC16] and monogamy between nonlocality and local contextuality [@KCK14; @ZZL16]. However, unlike in all these cases, in the scenarios in Figs. \[Fig1\](b1)–\[Fig1\](b6), none of the parties has a set of measurements capable to locally produce contextuality, thus our result reveals a different form of quantum contextuality that is worth closer examination. Specifically, it would be interesting to compare the classical and quantum sets of behaviors with those of the scenario in Fig. \[Fig1\](a1), since it seems that there are quantum behaviors that are contextual in the scenarios of Figs. \[Fig1\](b1)–\[Fig1\](b6) but that are noncontextual when we ignore the measurements in Bob’s side that are not in the scenario of Fig. \[Fig1\](a1). However, this contextuality is not merely local (as occurs in Refs. [@Cabello10; @LHC16; @KCK14; @ZZL16]).
\(c) Nonchordal compatibility graphs that do not admit realizations as multipartite scenarios (since their sets of vertices cannot be divided into disjoint subsets containing some nonadjacent vertices and such that each vertex in a subset is adjacent to all vertices in the other subsets). These graphs are shown in Figs. \[Fig1\](c1)–\[Fig1\](c11). The most famous of them is the pentagon of compatibility shown in Fig. \[Fig1\](c1), which corresponds to the scenario studied by Klyachko, Can, Binicioğlu, and Shumovsky (KCBS) [@KCBS08]. To our knowledge, so far, the classical and quantum sets of behaviors have been exhaustively characterized only for the scenarios corresponding to this graph and the hexagon in Fig. \[Fig1\](c2) [@AQB13]. Our result allows us to identify new simple scenarios that can produce quantum contextuality. Curiously, the first Bell inequality different than the Clauser-Horne-Shimony-Holt inequality proposed in the literature, the two-party three-setting chained Bell inequality, proposed in Ref. [@Pearle70] and rediscovered in Ref. [@BC90], is a tight noncontextuality inequality for a scenario corresponding to the compatibility graph in Fig. \[Fig1\](c2) [@AQB13], but is not a tight Bell inequality for the two-party three-setting Bell scenario corresponding to the compatibility graph in Fig. \[Fig1\](a4).
{width="0.93\linewidth"}
[*Conclusion.*]{} Here, we have investigated the connection between the most basic form of nonclassicality—incompatibility—and the resource that has been proven to be necessary to explain the power of some leading models of quantum computation and many quantum information protocols—contextuality. We have proven that a necessary and sufficient condition for the existence of a quantum behavior that is contextual in the sense of Refs. [@KCBS08; @Cabello08; @YO12; @KBLGC12; @AQB13; @CSW14] is that the compatibility graph that encodes the relations of incompatibility between the measurements is nonchordal. Since being nonchordal implies containing induced cycles of size larger than three, our result points out the crucial role for quantum contextuality of the $n$-cycle compatibility scenarios with $n\ge 4$ (whose complete list of tight noncontextuality inequalities and their maximal quantum violation are presented in Ref. [@AQB13]).
The scenarios in which contextuality can happen can be classified in three types: Bell scenarios, KCBS-type scenarios, and a third type in between them that worth closer examination. This classification holds not only for quantum theory but for general probabilistic theories, as it is based on the observation that contextuality can only occur in scenarios whose compatibility graph is nonchordal, and nonlocality can only occur if, in addition, the vertices of the compatibility graph can be divided into disjoint sets, each of them containing only nonadjacent vertices. In fact, one of the interesting consequences of our result is the observation that what is special about quantum theory is that contextuality and nonlocality occur in all scenarios in which they can, respectively, occur.
We thank Teiko Heinosaari, Andrei Khrennikov, and an anonymous referee of another paper for repeatedly asking the question addressed in this Rapid Communication, Costantino Budroni for bringing to our attention Vorob’yev’s theorem, and Antonio J. López-Tarrida for comments. This work was supported by Project No. FIS2017-89609-P, “Quantum Tools for Information, Computation and Research” (MINECO-MICINN, Spain) with FEDER funds, the FQXi Large Grant “The Observer Observed: A Bayesian Route to the Reconstruction of Quantum Theory,” and the project “Photonic Quantum Information” (Knut and Alice Wallenberg Foundation, Sweden). Z.-P.X. is supported by the Natural Science Foundation of China (Grant No. 11475089) and the China Scholarship Council.
Vorob’yev’s theorem
===================
Here, we restate in the language of graph theory and prove a theorem introduced, without a proof, by Vorob’yev in 1963 [@Vorob'yev63] and then proven independently by Kellerer [@Kellerer64a; @Kellerer64b], Vorob’yev [@Vorob'yev67], and others [@Malvestuto88]. Vorob’yev’s theorem is the basis of a fundamental result in the field of expert systems [@LS88].
Recall that a perfect elimination ordering in a graph $G$ is an ordering of the vertices of $G$ such that, for each vertex $v_i$, $v_i$ and the vertices of $G$ that are adjacent to $v_i$ and occur after $v_i$ in the order form a clique. A graph is chordal if and only if it has a perfect elimination ordering [@FG65].
[*Theorem.*]{} Any set of probabilities for the outcomes of a set of measurements whose compatibility relations are represented by a chordal graph admits a global extension to a joint probability distribution.
[*Proof.*]{} Suppose an $n$-vertex chordal graph $G$. Since $G$ is chordal, $G$ has a perfect elimination order $(v_n,v_{n-1},\ldots,v_1)$. Let $A_k$ be the set of vertices of $G$ that are adjacent to $v_k$ and occur after $v_k$ in that order. By definition of perfect elimination ordering, $A_k$ is a clique. Therefore, $\{M_v\}_{v\in A_k}$ is a set of mutually compatible measurements. Let $\{M_{v_i} = m_{v_i}\}_{i=1}^k$, for $i=1,2,\ldots,k$, be the set of events in which the output of measurement $M_{v_i}$ is $m_{v_i}$. Let us define
$$\begin{aligned}
P_1(M_{v_1} = m_{v_1}) &:= & \operatorname{Prob}(M_{v_1} = m_{v_1}),\\
P_k(\{M_{v_i} = m_{v_i}\}_{i=1}^k) & := &
\frac{\operatorname{Prob}(\{M_v = m_v\}_{v\in A_k})P_{k-1}(\{M_{v_i} = m_{v_i}\}_{i=1}^{k-1})}{\operatorname{Prob}(\{M_v = m_v\}_{v \in A_k\backslash v_k})},
\end{aligned}$$
where $\operatorname{Prob}(\{M_v = m_v\}_{v\in A})$ denotes the probability distribution for a set of compatible measurements $\{M_v\}_{v\in A}$.
We have to prove that $P_n(\{M_{v_i} = m_{v_i}\}_{i=1}^n)$ is a joint probability distribution which coincides with any $\operatorname{Prob}(\{M_v = m_v\}_{v\in A})$, where $A$ is a clique in $G$. We will prove it by induction. By definition, $P_1(M_{v_1} = m_{v_1})$ coincides with any $\operatorname{Prob}(\{M_v = m_v\}_{v\in A})$. Let us assume that $P_{t}(\{M_{v_i} = m_{v_i}\}_{i=1}^{t})$ also coincides with $\operatorname{Prob}(\{M_v = m_v\}_{v\in A})$, for $1 \leq t \leq k-1$ and any clique $A$ in $G$, that is,
$$\begin{aligned}
\sum_{m_v, v\in \{v_1,\ldots,v_{t}\}\backslash A} P_{t}(\{M_{v_i} = m_{v_i}\}_{i=1}^{t}) & = & \sum_{m_v, v\in A\backslash \{v_1,\ldots,v_{t}\}} \operatorname{Prob}(\{M_v = m_v\}_{v\in A}).
\end{aligned}$$
Then, for any clique which does not contain $v_k$,
$$\begin{aligned}
\sum_{m_v, v\in \{v_1,\ldots,v_k\}\backslash A} P_k(\{M_{v_i} = m_{v_i}\}_{i=1}^k) &=& \sum_{m_v, v\in \{v_1,\ldots,v_{k-1}\}\backslash A} \frac{\left(\sum_{m_{v_k}}\operatorname{Prob}(\{M_v = m_v\}_{v\in A_k})\right)P_{k-1}(\{M_{v_i} = m_{v_i}\}_{i=1}^{k-1})}{\operatorname{Prob}(\{M_v = m_v\}_{v \in A_k\backslash v_k})}\\
&=& \sum_{m_v, v\in \{v_1,\ldots,v_{k-1}\}\backslash A} P_{k-1}(\{M_{v_i} = m_{v_i}\}_{i=1}^{k-1})\\
&=& \sum_{m_v, v\in A\backslash \{v_1,\ldots,v_{k-1}\}} \operatorname{Prob}(\{M_v = m_v\}_{v\in A})\\
&=& \sum_{m_v, v\in A\backslash \{v_1,\ldots,v_k\}} \operatorname{Prob}(\{M_v = m_v\}_{v\in A}).
\end{aligned}$$
If $A$ contains $v_k$, then $A \cap \{v_1,\ldots,v_k\} \subseteq A_k$ by definition. Therefore,
$$\begin{aligned}
\sum_{m_v, v\in \{v_1,\ldots,v_k\}\backslash A} P_k(\{M_{v_i} = m_{v_i}\}_{i=1}^k) &=& \sum_{m_v, v\in A_k\backslash A} \frac{\operatorname{Prob}(\{M_v = m_v\}_{v\in A_k})
\left(\sum_{m_v, v\in \{v_1,\ldots,v_k\}\backslash A_k}P_{k-1}(\{M_{v_i} = m_{v_i}\}_{i=1}^{k-1})\right)}{\operatorname{Prob}(\{M_v = m_v\}_{v \in A_k\backslash v_k})}\\
&=& \sum_{m_v, v\in A_k\backslash A} \operatorname{Prob}(\{M_v = m_v\}_{v\in A_k})\\
&=& \sum_{m_v, v\in A\backslash \{v_1,\ldots,v_k\}} \operatorname{Prob}(\{M_v = m_v\}_{v\in A}).
\end{aligned}$$
So $P_k(\{M_{v_i} = m_{v_i}\}_{i=1}^k)$ also coincides with any $\operatorname{Prob}(\{M_v = m_v\}_{v\in A})$, where $A$ is a clique in $G$. By induction, so does $P_n(\{M_{v_i} = m_{v_i}\}_{i=1}^n)$.
[99]{}
T. Heinosaari, T. Miyadera, and M. Ziman, An invitation to quantum incompatibility, [J. Phys. A: Math. Theor. **49**, 123001 (2016).](http://dx.doi.org/10.1088/1751-8113/49/12/123001)
L. A. Khalfin and B. S. Cirel’son, Quantum and quasi-classical analogs of Bell inequalities, in [*Symposium on the Foundations of Modern Physics: 50 Years of the Einstein-Podolsky-Rosen Experiment*]{}, edited by P. J. Lahti and P. Mittelstaedt (World Scientific, Singapore, 1985), p. 441.
M. M. Wolf, D. Pérez-García, and C. Fernández, Measurements Incompatible in Quantum Theory Cannot Be Measured Jointly in Any Other No-Signaling Theory, [Phys. Rev. Lett. **103**, 230402 (2009).](https://doi.org/10.1103/PhysRevLett.103.230402)
M. T. Quintino, J. Bowles, F. Hirsch, and N. Brunner, Incompatible quantum measurements admitting a local-hidden-variable model, [Phys. Rev. A **93**, 052115 (2016).](https://doi.org/10.1103/PhysRevA.93.052115)
E. P. Specker, Die Logik nicht gleichzeitig entscheidbarer Aussagen, [Dialectica **14**, 239 (1960)](https://doi.org/10.1111/j.1746-8361.1960.tb00422.x) (English translation: The logic of non-simultaneously decidable propositions, [.](http://arxiv.org/abs/1103.4537))
J. S. Bell, On the problem of hidden variables in quantum mechanics, [Rev. Mod. Phys. **38**, 447 (1966).](https://doi.org/10.1103/RevModPhys.38.447)
S. Kochen and E. P. Specker, The problem of hidden variables in quantum mechanics, J. Math. Mech. **17**, 59 (1967).
A. A. Klyachko, M. A. Can, S. Binicioğlu, and A. S. Shumovsky, Simple Test for Hidden Variables in Spin-1 Systems, [Phys. Rev. Lett. **101**, 020403 (2008).](http://dx.doi.org/10.1103/PhysRevLett.101.020403)
A. Cabello, Experimentally Testable State-Independent Quantum Contextuality, [Phys. Rev. Lett. **101**, 210401 (2008).](http://dx.doi.org/10.1103/PhysRevLett.101.210401)
S. Yu and C. H. Oh, State-Independent Proof of Kochen-Specker Theorem with 13 Rays, [Phys. Rev. Lett. **108**, 030402 (2012).](https://doi.org/10.1103/PhysRevLett.108.030402)
M. Kleinmann, C. Budroni, J.-Å. Larsson, O. G[ü]{}hne, and A. Cabello, Optimal Inequalities for State-Independent Contextuality, [Phys. Rev. Lett. **109**, 250402 (2012).](https://doi.org/10.1103/PhysRevLett.109.250402)
M. Araújo, M. T. Quintino, C. Budroni, M. Terra Cunha, and A. Cabello, All noncontextuality inequalities for the $n$-cycle scenario, [Phys. Rev. A **88**, 022118 (2013).](https://doi.org/10.1103/PhysRevA.88.022118)
A. Cabello, S. Severini, and A. Winter, Graph-Theoretic Approach to Quantum Correlations, [Phys. Rev. Lett. **112**, 040401 (2014).](https://doi.org/10.1103/PhysRevLett.112.040401)
J. Anders and D. E. Browne, Computational Power of Correlations, [Phys. Rev. Lett. **102**, 050502 (2009).](https://doi.org/10.1103/PhysRevLett.102.050502)
M. J. Hoban and D. E. Browne, Stronger Quantum Correlations with Loophole-Free Postselection, [Phys. Rev. Lett. **107**, 120402 (2011).](https://doi.org/10.1103/PhysRevLett.107.120402)
R. Raussendorf, Contextuality in measurement-based quantum computation, [Phys. Rev. A **88**, 022322 (2013).](http://dx.doi.org/10.1103/PhysRevA.88.022322)
M. Howard, J. Wallman, V. Veitch, and J. Emerson, Contextuality supplies the ‘magic’ for quantum computation, [Nature (London) **510**, 351 (2014).](http://dx.doi.org/10.1038/nature13460)
N. Delfosse, P. A. Guerin, J. Bian, and R. Raussendorf, Wigner Function Negativity and Contextuality in Quantum Computation on Rebits, [Phys. Rev. X **5**, 021003 (2015).](http://dx.doi.org/10.1103/PhysRevX.5.021003)
S. Abramsky, R. Soares Barbosa, and S. Mansfield, Contextual Fraction as a Measure of Contextuality, [Phys. Rev. Lett. **119**, 050504 (2017).](https://doi.org/10.1103/PhysRevLett.119.050504)
J. Bermejo-Vega, N. Delfosse, D. E. Browne, C. Okay, and R. Raussendorf, Contextuality as a Resource for Models of Quantum Computation with Qubits, [Phys. Rev. Lett. **119**, 120505 (2017).](https://doi.org/10.1103/PhysRevLett.119.120505)
R. Raussendorf, D. E. Browne, N. Delfosse, C. Okay, and J. Bermejo-Vega, Contextuality and Wigner-function negativity in qubit quantum computation, [Phys. Rev. A **95**, 052334 (2017).](https://doi.org/10.1103/PhysRevA.95.052334)
A. K. Ekert, Quantum Cryptography Based on Bell’s Theorem, [Phys. Rev. Lett. **67**, 661 (1991).](https://doi.org/10.1103/PhysRevA.95.052334)
J. Barrett, L. Hardy, and A. Kent, No Signaling and Quantum Key Distribution, [Phys. Rev. Lett. **95**, 010503 (2005).](http://dx.doi.org/10.1103/PhysRevLett.95.010503)
T. S. Cubitt, D. Leung, W. Matthews, and A. Winter, Improving Zero-Error Classical Communication with Entanglement, [Phys. Rev. Lett. **104**, 230503 (2010).](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.104.230503)
A. Cabello, V. D’Ambrosio, E. Nagali, and F. Sciarrino, Hybrid ququart-encoded quantum cryptography protected by Kochen-Specker contextuality, [Phys. Rev. A **84**, 030302(R) (2011).](http://doi.org/10.1103/PhysRevA.84.030302)
T. Heinosaari, D. Reitzner, and P. Stano, Notes on joint measurability of quantum observables, [Found. Phys. **38**, 1133 (2008).](https://doi.org/10.1007/s10701-008-9256-7)
C. Heunen, T. Fritz, and M. L. Reyes, Quantum theory realises all joint measurability graphs, [Phys. Rev. A **89**, 032121 (2014).](https://doi.org/10.1103/PhysRevA.89.032121)
B. Amaral and M. Terra Cunha, [[*On Graph Approaches to Contextuality and their Role in Quantum Theory*]{} (Springer, Cham, Switzerland, 2018).](https://doi.org/10.1007/978-3-319-93827-1)
G. Chiribella and X. Yuan, Measurement sharpness cuts nonlocality and contextuality in every physical theory, [.](http://arxiv.org/abs/1404.3348)
G. Chiribella and X. Yuan, Bridging the gap between general probabilistic theories and the device-independent framework for nonlocality and contextuality, [Information and Computation **250**, 15 (2016).](https://doi.org/10.1016/j.ic.2016.02.006)
T. Heinosaari and M. M. Wolf, Nondisturbing quantum measurements, [J. Math. Phys. **51**, 092201 (2010).](https://doi.org/10.1063/1.3480658) R. W. Spekkens, The status of determinism in proofs of the impossibility of a noncontextual model of quantum theory, [Found. Phys. **44**, 1125 (2014).](https://doi.org/10.1007/s10701-014-9833-x)
M. Kleinmann, O. G[ü]{}hne, J. R. Portillo, J.-Å. Larsson, and A. Cabello, Memory cost of quantum contextuality, [New J. Phys. **13**, 113011 (2011).](https://doi.org/10.1088/1367-2630/13/11/113011)
A. Cabello, M. Gu, O. G[ü]{}hne, and Z.-P. Xu, Optimal Classical Simulation of State-Independent Quantum Contextuality, [Phys. Rev. Lett. **120**, 130401 (2018).](https://doi.org/10.1103/PhysRevLett.120.130401)
A. Cabello, M. Gu, O. G[ü]{}hne, J.-Å. Larsson, and K. Wiesner, Thermodynamical cost of some interpretations of quantum theory, [Phys. Rev. A **94**, 052127 (2016).](https://doi.org/10.1103/PhysRevA.94.052127)
P. Kurzyński, R. Ramanathan, and D. Kaszlikowski, Entropic Test of Quantum Contextuality, [Phys. Rev. Lett. **109**, 020404 (2012).](https://doi.org/10.1103/PhysRevLett.109.020404)
A. Cabello, L. E. Danielsen, A. J. López-Tarrida, and J. R. Portillo, Basic exclusivity graphs in quantum correlations, [Phys. Rev. A **88**, 032104 (2013).](https://doi.org/10.1103/PhysRevA.88.032104)
N. N. Vorob’yev, Markov measures and Markov extensions, [Theory Probab. Appl. **8**, 420 (1963).](https://doi.org/10.1137/1108047)
N. N. Vorob’yev, Coalition games, [Theory Probab. Appl. **12**, 251 (1967).](https://doi.org/10.1137/1112028)
C. Budroni and G. Morchio, The extension problem for partial Boolean structures in quantum mechanics, [J. Math. Phys. **51**, 122205 (2010).](https://doi.org/10.1063/1.3523478)
C. Budroni and A. Cabello, Bell inequalities from variable elimination methods, [J. Phys. A: Math. Theor. **45**, 385304 (2012).](https://doi.org/10.1088/1751-8113/45/38/385304)
R. Ramanathan, A. Soeda, P. Kurzyński, and D. Kaszlikowski, Generalized Monogamy of Contextual Inequalities from the No-Disturbance Principle, [Phys. Rev. Lett. **109**, 050404 (2012).](https://doi.org/10.1103/PhysRevLett.109.050404)
R. Soares Barbosa, On monogamy of non-locality and macroscopic averages: examples and preliminary results, in [[*Proceedings of the 11th Workshop on Quantum Physics and Logic*]{}, edited by B. Coecke, I. Hasuo, and P. Panangaden, EPTCS 172, 36 (2014).](https://doi.org/10.4204/EPTCS.172.4)
R. Soares Barbosa, Contextuality in quantum mechanics and beyond, Ph.D. thesis, University of Oxford, 2015.
C. Budroni, N. Miklin, and R. Chaves, Indistinguishability of causal relations from limited marginals, [Phys. Rev. A **94**, 042127 (2016).](https://doi.org/10.1103/PhysRevA.94.042127)
M. Michler, H. Weinfurter, and M. Żukowski, Experiments towards Falsification of Noncontextual Hidden Variable Theories, [Phys. Rev. Lett. **84**, 5457 (2000).](https://doi.org/10.1103/PhysRevLett.84.5457)
Y. Hasegawa, R. Loidl, G. Badurek, M. Baron, and H. Rauch, Violation of a Bell-like inequality in single-neutron interferometry, [Nature (London) **425**, 45 (2003).](https://doi.org/10.1038/nature0188)
G. Kirchmair, F. Zähringer, R. Gerritsma, M. Kleinmann, O. G[ü]{}hne, A. Cabello, R. Blatt, and C. F. Roos, State-independent experimental test of quantum contextuality, [Nature (London) **460**, 494 (2009).](http://doi:10.1038/nature08172)
F. M. Leupold, M. Malinowski, C. Zhang, V. Negnevitsky, A. Cabello, J. Alonso, and J. P. Home, Sustained state-independent quantum contextual correlations from a single ion, [Phys. Rev. Lett. **120**, 180401 (2018).](https://doi.org/10.1103/PhysRevLett.120.180401)
J. F. Clauser, M. A. Horne, A. Shimony, and R. A. Holt, Proposed Experiment to Test Local Hidden-Variable Theories, [Phys. Rev. Lett. **23**, 880 (1969).](http://dx.doi.org/10.1103/PhysRevLett.23.880)
B. S. Cirel’son, Quantum generalizations of Bell’s inequality, [Lett. Math. Phys. **4**, 93 (1980).](https://doi.org/10.1007/BF00417500)
P. Suppes and M. Zanotti, When are probabilistic explanations possible?, [Synthese **48**, 191 (1981).](https://doi.org/10.1007/BF01063886)
M. Froissart, Constructive generalization of Bell’s inequalities, [Nuovo Cimento Soc. Ital. Fis. B **64**, 241 (1981).](https://doi.org/10.1007/BF02903286)
A. Fine, Hidden Variables, Joint Probability, and the Bell Inequalities, [Phys. Rev. Lett. **48**, 291 (1982).](https://doi.org/10.1103/PhysRevLett.48.291)
B. S. Cirel’son, Zap. Nauchn. Semin LOMI **142**, 174 (1985) \[Quantum analogues of the Bell inequalities. The case of two spatially separated domains, [J. Sov. Math. **36**, 557 (1987)](https://doi.org/10.1007/BF01663472)\].
B. S. Cirel’son, Some results and problems on quantum Bell-type inequalities, Hadronic Journal Supplement **8**, 329 (1993).
M. Navascués, S. Pironio, and A. Acín, Bounding the Set of Quantum Correlations, [Phys. Rev. Lett. **98**, 010401 (2007).](https://doi.org/10.1103/PhysRevLett.98.010401)
M. Navascués, S. Pironio, and A. Acín, A convergent hierarchy of semidefinite programs characterizing the set of quantum correlations, [New J. Phys. **10**, 073013 (2008).](https://doi.org/10.1088/1367-2630/10/7/073013)
N. Brunner and N. Gisin, Partial list of bipartite Bell inequalities with four binary settings, [Phys. Lett. A **372**, 3162 (2008).](https://doi.org/10.1016/j.physleta.2008.01.052)
D. Collins and N. Gisin, A relevant two qubit Bell inequality inequivalent to the CHSH inequality, [J. Phys. A: Math. Gen. **37**, 1775 (2004).](http://dx.doi.org/10.1088/0305-4470/37/5/021)
I. Pitowsky and K. Svozil, Optimal tests of quantum nonlocality, [Phys. Rev. A **64**, 014102 (2001).](https://doi.org/10.1103/PhysRevA.64.014102)
C. Śliwa, Symmetries of the Bell correlation inequalities, [Phys. Lett. A **317**, 165 (2003).](http://dx.doi.org/10.1016/S0375-9601(03)01115-0)
S. López-Rosa, Z.-P. Xu, and A. Cabello, Maximum nonlocality in the $(3,2,2)$ scenario, [Phys. Rev. A **94**, 062121 (2016).](https://doi.org/10.1103/PhysRevA.94.062121)
A. Cabello, Proposal for Revealing Quantum Nonlocality via Local Contextuality, [Phys. Rev. Lett. **104**, 220401 (2010).](http://dx.doi.org/10.1103/PhysRevLett.104.220401)
B.-H. Liu, X.-M. Hu, J.-S. Chen, Y.-F. Huang, Y.-J. Han, C.-F. Li, G.-C. Guo, and A. Cabello, Nonlocality from Local Contextuality, [Phys. Rev. Lett. **117**, 220402 (2016).](https://doi.org/10.1103/PhysRevLett.117.220402)
P. Kurzyński, A. Cabello, and D. Kaszlikowski, Fundamental Monogamy Relation between Contextuality and Nonlocality, [Phys. Rev. Lett. **112**, 100401 (2014).](http://dx.doi.org/10.1103/PhysRevLett.112.100401)
X. Zhan, X. Zhang, J. Li, Y. Zhang, B. C. Sanders, and P. Xue, Realization of the Contextuality-Nonlocality Tradeoff with a Qubit-Qutrit Photon Pair, [Phys. Rev. Lett. **116**, 090401 (2016).](http://dx.doi.org/10.1103/PhysRevLett.116.090401)
P. M. Pearle, Hidden-variable example based upon data rejection, [Phys. Rev. D **2**, 1418 (1970).](https://doi.org/10.1103/PhysRevD.2.1418)
S. L. Braunstein and C. M. Caves, Wringing out better Bell inequalities, [Annals of Physics **202**, 22 (1990).](https://doi.org/10.1016/0003-4916(90)90339-P)
H. G. Kellerer, Ma[ß]{}theoretische Marginalprobleme, [Mathematische Annalen **153**, 168 (1964).](https://doi.org/10.1007/BF01360315)
H. G. Kellerer, Verteilungsfunktionen mit gegebenen Marginalverteilungen, [Zeitschrift für Wahrscheinlichkeitstheorie und Verwandte Gebiete **3**, 247 (1964).](https://doi.org/10.1007/BF00534912)
F. Malvestuto, Existence of extensions and product extensions for discrete probability distributions, [Discrete Mathematics **69**, 61 (1988).](https://doi.org/10.1016/0012-365X(88)90178-1)
S. L. Lauritzen and D. J. Spiegelhalter, Local computations with probabilities on graphical structures and their application to expert systems, [Journal of the Royal Statistical Society. Series B **50**, 157 (1988).](https://www.jstor.org/stable/2345762)
D. Fulkerson and O. Gross, Incidence matrices and interval graphs, [Pac. J. Math. **15**, 835 (1965).](https://msp.org/pjm/1965/15-3/p11.xhtml)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper briefly reports the GeoMFree$^{3D}$, a meshfree / meshless software package designed for analyzing the problems of large deformations and crack propagations of rock and soil masses in geotechnics. The GeoMFree$^{3D}$ is developed based on the meshfree RPIM, and accelerated by exploiting the parallel computing on multi-core CPU and many-core GPU. The GeoMFree$^{3D}$ is currently being under intensive developments. To demonstrate the correctness and effectiveness of the GeoMFree$^{3D}$, several simple verification examples are presented in this paper. Moreover, future work on the development of the GeoMFree$^{3D}$ is introduced.'
author:
- Gang Mei
- Nengxiong Xu
- Liangliang Xu
- Yazhe Li
bibliography:
- 'MyRef.bib'
title: 'GeoMFree$^{3D}$: An Under-Development Meshfree Software Package for Geomechanics'
---
Introduction
============
In geotechnics, the large deformations of rock and soil masses commonly occur in various geo-disasters such as landslides, debris flows, rock collapses, and ground subsidence. Moreover, when analyzing the stability of rock or soil slopes, the distribution and propagation of natural cracks is one of the most crucial issues that needs to be considered. To understand the mechanisms behind the above-mentioned geo-disasters, physical experiments and numerical investigations are commonly employed in practice.
The large-deformations and crack propagations of rock and soil masses have been examined using various mesh-based or meshfree numerical methods [@01; @02; @03; @04; @05; @06]. When employing those mesh-dependent numerical analysis methods such as Finite Element Method (FEM), Finite Volume Method (FVM), Finite Difference Method (FDM) to analyze the large deformation or crack propagation in geotechnics, the mesh element would generally be distorted or need to be broken. Mainly motivated by addressing those problems mentioned above occurring in mesh-based methods, the meshfree / meshless methods such as SPH, MLPG, LBIM, EFG, RPIM are proposed; see several excellent reviews [@07; @08; @09].
Recently, there are several meshfree software packages that have been developed or are being under development. For example, Hsieh and Pan described the Essential Software Framework for Meshfree methods (ESFM) [@10]. Cercos-Pita [@11] introduced the AQUAgpusph, a new free 3D SPH solver accelerated with OpenCL. Sinaie et *al* [@12] presents the implementation of the material point method (MPM) using Julia. Winkler et *al* [@13] introduced gpuSPHASE, a shared memory caching implementation for 2D SPH using CUDA. Vanaverbeke et *al* [@14] presented GRADSPMHD, a completely Lagrangian parallel magnetohydrodynamics code based on the SPH formalism. Zhang et *al* [@15; @16; @17] developed the 3D explicit parallel MPM code, MPM3D.
This short paper briefly reports the GeoMFree$^{3D}$, a meshfree / meshless software package for Geomechanics. The objective for developing the GeoMFree$^{3D}$ is to numerically analyze large deformations [@18; @19; @20], and crack propagations [@21; @22] of rock and soil masses in geomechanics. The package GeoMFree$^{3D}$ is currently under intensive developments. The underlying algorithm behind the GeoMFree3D is the Radial Point Interpolation Method (RPIM) proposed by Liu G.R. [@23; @24], . In addition, to improve the computational efficiency when analyzing large-scale problems [@25], the GeoMFree3D is parallelized on multi-core CPU and many-core GPU using the OpenMP [@26] and CUDA [@27], respectively.
GeoMFree$^{3D}$
===============
The GeoMFree$^{3D}$ is a meshfree software package designed for numerically analyzing the large deformations and crack propagations of rock and soil masses in geomechanics. The GeoMFree$^{3D}$ is currently capable of analyzing linear and nonlinear static problems, and is being developed for addressing dynamic problems. The GeoMFree$^{3D}$ is written in C/C++ and accelerated by exploiting the parallel computing on multi-core CPU and many-core GPU.
The process of numerical modeling using the GeoMFree$^{3D}$ is illustrated in Figure \[fig1\]. There are three major stages in the GeoMFree$^{3D}$. The first stage is to assemble the global stiffness matrix by looping over all field nodes to create the element stiffness matrix of each field node. The second is to enforce the boundary conditions. And the third is to solve the system equation to obtain displacements and then the stresses, etc.
To improve the computational efficiency, the first stage of assembling the global stiffness matrix is parallelized on multi-core CPU using the API OpenMP [@26]. The meshfree RPIM is inherently suitable to be parallelized since there is no data dependencies between the forming of any two element stiffness matrices for any pair of field nodes. That is, the assembly of the element stiffness matrix of one field node is completely independent of that for another field node. Therefore, we can allocate $n$ threads on the multi-core CPU; and each thread is responsible for assembling the element stiffness matrix for one field node. In this case, the assembly of element stiffness matrices for $n$ field nodes can be conducted concurrently. This is the essential idea behind parallelizing the assembly of global / system stiffness matrix on multi-core CPU.
Similarly, to enhance the computational efficiency, the second stage of enforcing the boundary condition can also be parallelized. More specifically, we adopt the penalty function method to enforce the displacement boundary conditions. This procedure is performed in parallel on the many-core GPU. Assuming there are $m$ field nodes on the displacement boundary, and we can allocate $m$ GPU threads to enforce the displacement boundary conditions for the $m$ field nodes concurrently, where each thread takes responsibility for enforcing the displacement boundary condition for one boundary field node.
The final stage is the solving of system equations to obtain nodal displacements and then the stresses. In meshfree RPIM, the assembled global stiffness matrix is large, sparse, and asymmetric. When analyzing large-scale problems and requiring a large number of field nodes, the global system matrix could be very large. To improve the computational efficiency in solving system equations, we have employed the `cuSparse` and `cuSolver` library integrated in CUDA [@27] to solve the system of equations.
![Process of our meshfree software package GeoMFree$^{3D}$[]{data-label="fig1"}](Figure1.pdf){width="0.78\linewidth"}
Verification
============
This section will present several computational examples to verify the validation and features of the reported meshfree software package GeoMFree$^{3D}$.
Example 1: Stresses of a Cubic Domain
-------------------------------------
First, to verify the correctness of the GeoMFree$^{3D}$, we specifically calculate the distribution of displacements and stresses of a cubic domain; see Figure \[fig2\]. In this quite simple verification example, only the force of gravity is considered and there are no other forces. The density of the cube is set as 2600 kg / m$^{3}$. The stress on the bottom of the cube can be theoretically calculated, and is noted as the *theoretical* result. In contrast, we can also numerically calculate the nodal stress on the bottom using our meshfree package GeoMFree$^{3D}$ which is noted as the *numerical* result. Then, by comparing the theoretical result to the numerical counterpart, we could validate the correctness of the GeoMFree$^{3D}$.
The theoretical nodal stress on the bottom of the cube is 2.548 MPa, while the numerically calculated one is 2.376 MPa. There is a slight difference between the theoretical and numerical results. And thus, we can conclude that the correctness of the GeoMFree$^{3D}$ has been verified, although the employed verification example is extremely simple.
Example 2: Displacements of a Cantilever Beam with Crack
--------------------------------------------------------
To further verify the effectiveness of our GeoMFree$^{3D}$, we have employed it to calculate the displacement and stress field of a Cantilever beam with a crack; see Figure \[fig3\]. Moreover, we have also computed the displacements and stresses of the beam using a FDM numerical software FLAC$^{3D}$; see Figure \[fig3b\]. The numerical results calculated by our package GeoMFree$^{3D}$ and the commercial numerical software FLAC$^{3D}$ are almost the same; see Figure \[fig3b\] and \[fig3c\]. And this indicates that currently our package GeoMFree$^{3D}$ is capable of analyzing the very simple cases of crack propagation. In the near future, we hope that the GeoMFree$^{3D}$ can be used to model and simulate dynamic crack propagations in three-dimensions.
Example 3: Displacements of a Simplified Slope
----------------------------------------------
As having been introduced several times, the motivation why we are developing our meshfree package GeoMFree$^{3D}$ is that: we hope to employ one of the meshfree methods, i.e., the RPIM, to well model and simulate the large deformations and crack propagations of rock and soil masses. Currently, the GeoMFree$^{3D}$ cannot be used to model the continuously developed large deformations of rock or soil masses. But it can be used to calculate the displacements of a simplified slope; see Figure \[fig4\] and Figure \[fig5\]. And we are working on analyzing the stability of slopes using the GeoMFree$^{3D}$ based upon the Strength Reduction Method (SRM).
In meshfree methods, the study domain is discretized with a set of field nodes. These field nodes could be (1) *regularly* or (2) *irregularly* distributed in the domain. The patterns of nodal distributions are of strong influence on both the computational accuracy and efficiency. To verify the flexibility of the GeoMFree3D for addressing problems with regular or irregular discretization, we have decomposed a simplified slope model with: (1) regular nodes (Figure \[fig4b\]) and (2) irregular nodes (Figure \[fig5a\]). We then calculated the displacements of the above two models using our package GeoMFree$^{3D}$, and also compared the results calculated by the GeoMFree$^{3D}$ to those by the commercial software FLAC$^{3D}$.
The numerical results illustrated in Figure 4 and Figure 5 indicate that: (1) our package GeoMFree$^{3D}$ is capable of analyzing the problems with regular or irregular nodal distributions; (2) our package GeoMFree$^{3D}$ can be used to address the problems with relatively complex geometric domains and boundaries.
Conclusion and Future Work
==========================
A meshfree software package, GeoMFree$^{3D}$, has been briefly introduced in this paper. The package GeoMFree$^{3D}$ is designed for the numerical investigation of large-deformations and crack propagations of rock and soil masses in geotechnics. The GeoMFree$^{3D}$ is developed based on the RPIM, and is currently under intensive developments. To validate the effectiveness of the introduced GeoMFree$^{3D}$, several verifications have been conducted. The verification examples have demonstrated that the current version of GeoMFree$^{3D}$ is capable of analyzing the deformation of simple study domains.
The GeoMFree$^{3D}$ is currently under intensive developments. We are focusing on improving the computational efficiency by developing accurate and efficiency meshfree shape functions [@28; @29; @30], for example, the parallel RBF [@31], MLS [@32], and Shepard [@33; @34] interpolations. Currently, we are also aiming at numerically modeling the crack propagation of multiple tensile and shear cracks of rock masses. In future, we hope that: the GeoMFree$^{3D}$ can be used to (1) model the large-deformations of strata induced by underground mining and (2) analyze the stability of jointed rock slopes via modeling the very complex crack propagations of rock masses.
Acknowledgements {#acknowledgements .unnumbered}
================
This research was supported by the Natural Science Foundation of China (Grant Numbers 41772326 and 11602235), and the Fundamental Research Funds for the Central Universities. The authors would like to thank the editor and reviewers for their contributions on the paper.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In systems where one coordinate undergoes periodic oscillation, the net displacement in any other coordinate over a single period is shown to be given by differentiation of the action integral associated with the oscillating coordinate. This result is then used to demonstrate that the action integral acts as a Hamiltonian for slow coordinates providing time is scaled to the “tick-time”of the oscillating coordinate. Numerous examples, including charged particle drifts and relativistic motion, are supplied to illustrate the varied application of these results.'
author:
- 'Rory J. Perkins and Paul M. Bellan'
bibliography:
- 'Perkins-Bellan-2010-PRL-bib-refs.bib'
date: date
title: 'Wheels within wheels: Hamiltonian dynamics as a hierarchy of action variables'
---
Hamiltonian dynamics is almost ubiquitous in physics and describes such varied phenomena as celestial mechanics, optics, fluid dynamics, quantum mechanics, and charged particle motion in electromagnetic fields. Guiding center theory, an approximation of Hamiltonian dynamics for charged particle motion in magnetic fields, describes the motion of the particle’s cyclotron-orbit averaged position, or guiding center [@Alfven63]. The guiding center can be thought of as a quasi-particle subject to new types of forces and manifesting various drifts. We develop a general model, not restricted to charged particle motion, of multi-dimensional systems with a periodic variable and find drifts that cannot be calculated using guiding center theory which becomes a limited example of the more general model. The model shows that the action integral associated with the oscillatory coordinate acts as an effective Hamiltonian for the remaining, slow coordinates providing time is measured in clock cycles of the oscillations. We note that Hamiltonian-type aspects of action integrals have been previously discussed in specific situations [@Kadomtsev1959; @*Northrop1960; @*White1984] but without developing a general demonstration and relying on the detailed equations of motion in their proofs. The model presented here generates a hierarchy of wheels-within-wheels Hamiltonian systems such that the action integral associated with periodic motion at any level in the hierarchy acts as the Hamiltonian for the next slower periodic motion.
Consider a 2D time-independent Hamiltonian system $H(\xi,P_{\xi},P_{\eta})$ with an ignorable coordinate $\eta$ and where the $\xi$-motion is periodic, i.e., $\xi(t+\Delta t)=\xi(t)$, with no limit on the amplitude of $\xi.$ $P_{\eta}$ evolves trivially: $\dot{P}_{\eta}=0$, but the $\eta$ evolution is in general nontrivial. The period $\Delta t$ can be imagined as a clock tick over which $\eta$ undergoes a net change $\Delta\eta$. We claim that $$\Delta\eta= - \partial J / \partial P_{\eta}, \label{result}$$ where $$J(H,P_{\eta})= \oint P_{\xi}(H,\xi,P_{\eta})d\xi\label{J}$$ is the action integral associated with $\xi.$ Equation (\[result\]) means that if $J(H,P_{\eta})$ is known, the net change of $\eta$ during one period of $\xi$ can be calculated without having to consider the potentially complicated form of $\dot{\eta}.$
To prove Eq. (\[result\]), first note that there is no contribution from differentiating the integral’s bounds, so $$\frac{\partial J}{\partial P_{\eta}} = \oint\frac{\partial P_{\xi
}}{\partial P_{\eta}}d\xi.\label{dJ-dPeta}$$ The differential of $H$ is $$dH=\frac{\partial H}{\partial\xi}d\xi\ +\frac{\partial H}{\partial P_{\xi}}dP_{\xi}+\frac{\partial H}{\partial P_{\eta}}dP_{\eta}\ \label{dH},$$ so $$\frac{\partial P_{\xi}}{\partial P_{\eta}}=-\frac{\partial H/\partial P_{\eta
}}{\partial H/\partial P_{\xi}}.\label{dPxi by dPeta}$$ Using Eq. (\[dPxi by dPeta\]) and Hamilton’s equations in Eq. (\[dJ-dPeta\]) gives $$\begin{aligned}
\frac{\partial J}{\partial P_{\eta}} & = - \oint\frac{\partial
H/\partial P_{\eta}}{\partial H/\partial P_{\xi}}d\xi\nonumber\\
& = - \oint\frac{d\eta/dt\ }{d\xi/dt}d\xi=- \Delta\eta.\label{delta eta}$$ If there are other ignorable coordinates in the system, then suitably adjusted versions of Eq. (\[delta eta\]) apply to each of them. Equation (\[result\]) generalizes the well-known theorem [@Goldstein02] that the period of motion is given by a partial derivative of $J$ with respect to $H,$ namely $$\Delta t = \partial J / \partial H.\label{delta t}$$ Equation (\[delta t\]) resembles Eq. (\[result\]) because $(t,-H)$ form a pair of canonical coordinates in extended phase space, so Eq. (\[delta t\]) is a special case of the theorem presented here. The drift, or net time evolution, of $\eta$ is clearly $$\frac{\Delta\eta}{\Delta t}=-\frac{\partial J/\partial P_{\eta}}{\partial
J/\partial H\ },\ \label{drift}$$ which generalizes the particle drifts associated with guiding center theory.
We now relax the requirement that $\eta$ is ignorable and allow the oscillations to evolve adiabatically. We do so by coupling the original Hamiltonian, now denoted as $H_\mathrm{loc}$, to an external system $H_\mathrm{ext}$ that is otherwise isolated. This gives a total Hamiltonian $$H(\xi,P_{\xi},\eta,P_{\eta}) = H_{\mathrm{loc}}(\xi,P_{\xi},\eta,P_{\eta
})+H_{\mathrm{ext}}(\eta,P_{\eta}). \label{Htot}$$ We presume the system behaves as follows. First, the $\xi$-oscillation is described by the *local* Hamiltonian $H_\mathrm{loc}$ in which $\eta$ and $P_\eta$ play the role of slowly varying parameters: $d \xi / dt = \partial H_\mathrm{loc} / \partial P_\xi$ and $dP_\xi / d t = - \partial H_\mathrm{loc} / \partial \xi.$ Second, the “parametric” coordinates $\eta$ and $P_\eta$ are described by the *total* Hamiltonian $H$, so $d \eta / d t = \partial H / \partial P_\eta$ and $d P_\eta / dt = - \partial H / \partial \eta$. The local and external systems exchange energy, but the total energy $E=E_{\mathrm{loc}}(t) + E_{\mathrm{ext}}(t)$ is conserved since the entire system is isolated. $J$ is defined as in Eq. (\[J\]) but is now also a function of $\eta$. As in Ref. [@Landau1969], we assume it is a good approximation to hold the parametric coordinates $\eta$ and $P_\eta$ fixed while evaluating the $\xi$ action integral. $J$ is an adibatic invariant and is thus conserved. Furthermore, $J$ depends only on $H_{\mathrm{loc}},$ i.e. $J = J(H_\mathrm{loc}, \eta, P_\eta) = J( H - H_\mathrm{ext}(\eta, P_\eta), \eta, P_\eta),$ because $H_{\mathrm{loc}}$ is sufficient to prescribe the $\xi$ dynamics. A proof analogous to that of Eq. (\[result\]) shows $$\frac{\partial J}{\partial\eta}= \Delta P_{\eta},\qquad
\frac{\partial J}{\partial P_{\eta}} = - \Delta\eta. \label{delta P eta}$$ Note that $J = J( H - H_\mathrm{ext}(\eta, P_\eta), \eta, P_\eta)$ depends on $\eta$ and $P_\eta$ both implicitly through $H_\mathrm{ext}$ and also explicitly. Accordingly, $\Delta \eta$ and $\Delta P_\eta$ have two terms: one term comes from the explicit dependence and is the drift of the system; the second term comes from the implicit dependence and is the slow change of $H_\mathrm{ext}$.
Equations (\[delta P eta\]) have the makings of a Hamiltonian system with $-J$ serving as the Hamiltonian. They are *precisely* Hamiltonian as follows. We define discretized derivatives $d\eta/dt=\Delta\eta/\Delta t$ and $dP_{\eta}/dt=$ $\Delta P_{\eta}/\Delta t$ that capture the net rates of change of $\eta$ and $P_{\eta}.$ Upon invocation of a rescaled time $\tau$ normalized by the $\xi$-period: $$d\tau = dt / \Delta t,\ \label{def tau}$$ Eqs. (\[delta P eta\]) become $$\frac{d\eta}{d\tau} = \frac{\partial}{\partial P_{\eta}}\left( -J\right), \qquad \frac{dP_{\eta}}{d\tau} = -\frac{\partial}{\partial\eta}\left(
-J\right). \label{dPeta-dtau}$$ Thus, $-J$ is the Hamiltonian for the averaged system provided time is measured in units of $\Delta t.$ It should be noted that $\tau$ is the angle variable conjugate to $J$.
Alternatively, we can obtain a Hamiltonian for the $\xi$-averaged system by solving $J=J(H_{loc},\eta,P_{\eta})$ for $H_{loc}=H_{loc}(J,\eta,P_{\eta})$ which upon inserting in Eq. (9) gives$$H=H_{loc}(J,\eta,P_{\eta})+H_{ext}(\eta,P_{\eta}).\label{H averaged}$$ Solution of Eq. (13) for $J$ gives $J=J(H,\eta,P_{\eta}).$ The differential of $J$ using this latter form is$$dJ=\frac{\partial J}{\partial H}dH+\frac{\partial J}{\partial\eta}d\eta
+\frac{\partial J}{\partial P_{\eta}}dP_{\eta}.\label{dJ}$$ Equation (14) determines $\partial H/\partial P_{\eta}=-\left( \partial
J/\partial P_{\eta}\right) /\left( \partial J/\partial H\right) $ etc., so using Eqs. (10)$$\begin{aligned}
\frac{\partial H}{\partial P_{\eta}} & =-\frac{\partial J/\partial P_{\eta}}{\partial J/\partial H}=-\frac{-\Delta\eta}{\Delta t}=\frac{d\eta}{dt}\label{eta dot regular time}\\
\frac{\partial H}{\partial\eta} & =-\frac{\partial J/\partial\eta}{\partial
J/\partial H}=-\frac{\Delta P_{\eta}}{\Delta t}=-\frac{dP_{\eta}}{dt}.\label{P eta dot regular time}$$ Thus, $H$, written as Eq. (\[H averaged\]), generates the discretized derivatives. The term $H_\mathrm{loc}(J, \eta, P_\eta)$ is an adiabatic potential [@Stefanski1985], the residue of averaging the periodic $\xi$-motion. For systems approximating a harmonic oscillator, $J = 2 \pi H_\mathrm{loc} / \omega(\eta, P_\eta)$, so the adiabatic potential is $H_\mathrm{loc} = J \omega / 2\pi$, showing that $J$ acts like an electrostatic charge and $\omega(\eta,P_{\eta})$ acts like an electrostatic potential. The magnitude of this $J$-charge depends on the amplitude of the $\xi$-oscillation. The use of $-J(H, \eta, P_\eta)$ as a Hamiltonian with normalized time $\tau$ and the use of $H(J, \eta, P_\eta)$ with regular time are entirely equivalent. Practically, though, there are techniques to evaluate $J$ directly [@Born1960], so using $-J$ as the Hamiltonian spares one from inverting $J$ for $H$, which might not be analytically feasible.
\[h\]
[sample\_orbit]{}
We now provide examples illustrating applications. Fig. \[sample\_orbit.eps\] shows an electron moving in the $rz$-plane and subjected to the magnetic field of a current-carrying wire aligned along the $z$-axis. The $z$ coordinate corresponds to $\eta$ and is ignorable; the radial motion is periodic and not ignorable because of the magnetic field gradient. The electron displaces itself an axial distance $\Delta z$ with every gyration as shown in Fig. \[sample\_orbit.eps\]. Using the characteristic velocity $\beta=\mu_{\circ}Ie/2\pi m$ [@Neuberger1982], the Hamiltonian is $$H=\frac{1}{2}mv^{2}=\frac{P_{r}^{2}}{2m}+\frac{\left( P_{z}-m\beta\ln \left(
r / R \right) \right) ^{2}}{2m}, \label{H-1/r}$$ where $R$ is an arbitrary constant of integration. $J$ can be evaluated exactly [@Wouters1995] using the substitution $\cos
\theta=\left( P_{z}-m\beta\ln\left( r/R\right) \right) /mv$ and the integral representation of the modified Bessel function $I_{n}(x)=\pi^{-1}\int_{0}^{\pi}e^{x\cos\theta}\cos(n\theta)d\theta$ [@Abramowitz1970] so that $$J = \oint P_{r}dr = 2 \pi mv r_\textrm{gc} I_{1}\left( \frac{v}{\beta}\right) , \label{Jplanar-1/r}$$ where $r_\textrm{gc} = R\exp(P_{z}/m\beta)$. $r_\textrm{gc}$, plotted as a dashed line in Fig. \[sample\_orbit.eps\], is the radial position at which the $z$-velocity vanishes, as can be checked from $P_{z}=mv_{z}+m\beta\ln r/R.$ $J$ generalizes the first adiabatic invariant $\mu=mv_{\perp}^{2}/2B$ [@Alfven63] and reduces to $2 \pi (m/e) \mu$ when $v\ll\beta$, which for this system is the condition for the guiding center approximation to hold [@Neuberger1982]. $\Delta z$ and $\Delta t$ can be computed using Eqs. (\[result\]) and (\[delta t\]) and noting that $v=\sqrt{2H/m}.$ The exact drift velocity, computed without appealing to the guiding center approximation, is $$\mathbf{v}_{d}=\frac{\Delta z}{\Delta t}\hat{z}=-v\frac{I_{1}(v/\beta)}{I_{0}(v/\beta)}\hat{z}.\label{wire-drift}$$ Equation (\[wire-drift\]) holds for orbits of all energies even when the guiding center approximation fails. The $v\ll\beta$ limit of Eq. (\[wire-drift\]) reduces to the grad B drift [@Bellan96] of the guiding center approximation.
Next we show how $J$ can be used as a Hamiltonian to give the magnetic mirror force [@Bellan96]. For a magnetic field mainly in the $z$ direction, the cyclotron motion is essentially harmonic oscillation at the gyrofrequency $\omega = q B / m$ in the perpendicular direction, so we identify $\eta$ with $z$ and $H_{\mathrm{ext}}$ with $P_{z}^{2}/2m.$ Then $$J = 2\pi \frac{H_\textrm{loc}}{\omega} = 2 \pi \frac{H-P_{z}^{2}/2m}{qB/m} = 2\pi \frac{mv_{\perp}^{2}/2}{qB/m} \label{J mu},$$ which equals $\mu$ except for the factor $2 \pi m/q.$ If $B$ depends on $z$ then Eqs. (\[eta dot regular time\]) and (\[P eta dot regular time\]) become $$\begin{aligned}
\frac{dz}{dt} & =-\frac{\partial J/\partial P_{z}}{\partial J/\partial
H\ }=\ \frac{P_{z}/m}{qB/m}\frac{qB}{m}=\frac{P_{z}}{m}\label{dz by dt mirror}\\
\frac{dP_{z}}{dt} & =\frac{\partial J/\partial z}{\partial J/\partial
H\ }=-\frac{qJ}{2\pi m}\frac{\partial B}{\partial z}=-\mu\frac{\partial B}{\partial z},
\label{dPz by dt mirror}$$ establishing the magnetic mirror force without considering the microscopic motion.
A slightly different approach retrieves the grad B drift. Suppose $\mathbf{B=}B_{z}(x)\hat{z},$ so $\mathbf{A}=A_{y}(x) \hat{y}$ with $B_{z}=\partial A_{y}/\partial x$. We define the $x$ component of the guiding center as the position $x_\textrm{gc}$ where $\ v_{y}$ vanishes: $P_y = q A_y (x_\textrm{gc})$. Setting $\eta=y$, Eq. (\[eta dot regular time\]) applied to Eq. (\[J mu\]) gives $$\begin{aligned}
\frac{dy}{dt} & =-\frac{\partial J/\partial P_{y}}{\partial J/\partial
H\ }= \mu \left( \frac{\partial B_{z}}{\partial
x}\right)_{x_\textrm{gc}} \frac{\partial x_{gc}}{\partial P_{y}}\label{y dot}.\end{aligned}$$ We then use $\partial x_\textrm{gc} / \partial P_y = (q B(x_\textrm{gc}))^{-1}$, obtained by differentiating $P_y = q A_y (x_\textrm{gc})$ with respect to $P_{y}$. Equation (\[y dot\]) thus becomes$$\frac{dy}{dt} = \frac{\mu}{qB_{z}} \left( \frac{\partial B_{z}}{\partial x}\right)_{x_\textrm{gc}} \label{grad B}$$ which is the grad $B$ drift.
A surprising application arises in relativistic mechanics, where it is found that in crossed electric and magnetic fields $\mathbf{E}=E\hat{x}$ and $\mathbf{B}=B\hat{z}$ with $E<Bc$ a particle’s $z$-velocity is modulated by the cyclotron motion, in contrast to the non-relativistic situation where $v_{z}$ is constant and independent of the cyclotron motion. The modulation arises from the periodic addition and subtraction of the $\mathbf{E}\times\mathbf{B}$ drift to the cyclotron velocity, which modulates $\gamma=\left( 1-v^{2}/c^{2}\right)^{-1/2}$ and hence the particle’s effective mass; $v_{z}$ then varies because $v_{z}=P_{z}/\gamma m$ and $P_{z}$ is invariant as $z$ is ignorable. Using the relativistic canonical momenta $\mathbf{P}=m\gamma\mathbf{v}+q\mathbf{A}$ with $\mathbf{A=}Bx\hat{y}$, the electrostatic potential $\phi=-Ex,$ and the relativistic Hamiltonian $H=c\sqrt{(\mathbf{P}-q\mathbf{A})^{2}+m^{2}c^{2}}+q\phi,$ it is found that the relativistic $x$-action $J=\oint P_{x}dx$ evaluates to $$\frac{J}{2\pi} = \frac{(BP_{y}+HE/c^{2})^{2}}{2q(B^{2}-E^{2}/c^{2})^{3/2}} + \frac{(H^{2}/c^{2}-P_{z}^{2}-P_{y}^{2}-m^{2}c^{2})}{2q(B^{2}-E^{2}/c^{2})^{1/2}}.\label{J-relativistic}$$ Calculating $\Delta z$ and $\Delta t$ using Eqs. (\[result\]) and (\[delta t\]) gives the $E$-dependent $z$-drift velocity $$v_d = \frac{\Delta z}{\Delta t}=\frac{B^{2}-E^{2}/c^{2}}{BEP_{y}/c^{2}+B^{2}H/c^{2}}P_{z}.\label{rel-drift}$$ As shown in Fig. (\[fig:relativistic\]), this $v_{z}$ drift has been verified by direct numerical integration of the relativistic equation of motion $d(\gamma m\mathbf{v})/dt=q\left( \mathbf{E}+\mathbf{v\times B}\right) \ $which shows that the modulation of $v_{z}$ is typically spiky as $\gamma \approx 1$ for a short interval during each cyclotron period and then $\gamma \gg 1$ for the remaining fraction of the cyclotron period. Clearly, this analysis generalizes to force-drifts by replacing $\mathbf{E}$ with $\mathbf{F} / q$.
\[h\]
![A particle undergoing relativistic $E \times B$ motion in the $xy$-plane with $E = 0.95 B c$ and initial momentum $P_x = 0$, $P_y = 0.7 m c$, and $P_z = 0.3 m c$. (a) The $z$-velocity (from numerical integration of the relativistic equation of motion) is non-constant, spiking when the particle’s $xy$-motion slows down so that $\gamma \approx 1$. (b) Solid line is the numerically integrated $z$-position; dashed line, calculated using Eq. (\[rel-drift\]), captures the $z$-drift motion.[]{data-label="fig:relativistic"}](relativistic){width=".5\textwidth"}
Kepler motion provides a non-relativistic and non-electromagnetic example. The radial action is [@Goldstein02] $$J_{r} = -2\pi \left\vert P_{\phi}\right\vert + 2 \pi \sqrt{m} mMG/ \sqrt{2 \left\vert H \right\vert },\label{J Kepler}$$ where $P_{\phi}$ is the conserved angular momentum. Equation (\[result\]) gives $\Delta\phi=\pm 2\pi$ depending on the sign of $P_{\phi}$, immediately proving that bounded Kepler orbits are always closed.
We now present a purely mechanical system which exhibits the equivalent of magneticmirroring. Consider a non-relativistic particle in a long groove where the width of the groove varies with position. The Hamiltonian is $$H=\frac{P_{x}^{2}}{2m}+\frac{P_{y}^{2}}{2m}+\frac{1}{2}\kappa x^{2}\left(
1+\alpha y^{2}\right) +\frac{\lambda}{2}y^{2};\label{H groove}$$ where $y$ is the distance along the groove and $x$ is the distance across the groove. Presuming that the $y$-position changes slowly relative to the oscillations across the groove (i.e., $\left\vert \alpha\right\vert $ and $\left\vert \lambda\right\vert $ are small compared to $\kappa)$ the $y$-dependent frequency of $x$-oscillation is $$\omega(y)=\sqrt{\frac{\kappa}{m}}\sqrt{1+\alpha y^{2}}.\label{omega groove}$$ We identify $H_{\mathrm{loc}} = P_{x}^{2}/2m + m\omega(y)^{2}x^{2}/2,$ so the $x$-action is $J= 2 \pi H_{\mathrm{loc}}/\omega(y),$ and Eq. (\[H averaged\]) becomes $$H=\frac{P_{y}^{2}}{2m}+\frac{\lambda}{2}y^{2} + \frac{\omega(y)}{2\pi} J.\label{H reduced groove}$$ Equation (\[P eta dot regular time\]) gives an average force $\ -(J/2\pi)\,\partial\omega/\partial y=-yJ\kappa\alpha/2\pi m\omega$ in the $y$-direction due to the increase in $x$-oscillation energy where the groove narrows. This is a restoring force and, if sufficiently strong, can overwhelm the contribution from $\lambda.$ A negative $\lambda$ corresponds to a potential hill, and if $J=0$ the particle falls down the hill. However, if $J$ is sufficiently large and $\alpha$ is positive, the particle does not fall down but instead oscillates about the top of the hill! This mechanical analog of a magnetic mirror has been verified by direct numerical integration as shown in Fig. \[groove\_sample1.jpg\].
\[h\]
![A particle in a thin saddle-like groove can undergo oscillatory motion due to narrowing of the groove. $x$ is the direction across the groove, $y$ along the groove, and $z$ the vertical. $H$ is given by Eq. (\[H groove\]) with $m=\kappa=\alpha=1$ and $\lambda = -.01$, and the particle starts at $x=y=0$ with $v_x=.25$ and $v_y = 1$.[]{data-label="groove_sample1.jpg"}](groove){width=".5\textwidth"}
For oscillatory $y$-motion, Eq. (\[H reduced groove\]) admits an action integral in the $y$-direction, which we denote by $K$, that acts as a Hamiltonian for the $x$-averaged system. This is a two-tier heirachy of action variables, or a wheel within a wheel. For the reduced system, $J$ is a conserved quantity, so we develop an analog of Eq. (\[result\]): $$\begin{aligned}
\frac{\partial K}{\partial J} &=& \oint \frac{ \partial P_y (H, J, y) }{ \partial J} dy = \oint \frac{1}{\partial J / \partial P_y} dy \\
&=& \oint \frac{- 1 }{d y / d \tau} dy = - \Delta \tau, \end{aligned}$$ where we use Eq. (\[dJ\]) to evaluate $\partial P_y / \partial J$ and Eq. (\[dPeta-dtau\]) to evaluate $\partial J / \partial P_y$. Since $\tau$ counts $x$-cycles, $-\partial K / \partial J$ gives the number of $x$-cycles per $y$-cycle. If this quantity is a rational number, the trajectory is closed. This is of interest when quantizing the system, as there is sometimes a one-to-one correspondence between periodic classical trajectories and quantum energy levels [@Gutzwiller1970].
*Acknowledgments:* Supported by USDOE and NSF.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Frequent supernova explosions in compact starburst regions are a main shaper of these regions’ interstellar media (ISM). In most starbursts, the supernova remnants blast open a hot phase that fills the regions and launches a superwind. Denser starbursts are too overpressured for hot wind formation, but supernovae still stir up the ISM. I argue that supernovae power ubiquitous turbulence through each of the starburst ISM phases, including the hot wind, and that a fluctuation dynamo amplifies magnetic fields until they are in equipartition with the turbulence. Supernovae can drive turbulence with speeds of $\sim 1000\ \kms$ in the hot wind and $\sim 20\ \kms$ in the cold molecular gas, depending on the outer scale. I predict magnetic field strengths of $70\ \muGauss$ in the Galactic Center starburst, $300\ \muGauss$ in M82 and NGC 253, and $2\ \mGauss$ in Arp 220’s nuclei. The mean magnetic field strengths are a few times stronger in molecular gas than in hot winds, but do not vary strongly with density within a starburst. I explain how the dominance of supernova-driven turbulence leads to near equipartition between the components of starburst ISM. I also consider implications for cosmic ray (CR) diffusion in starbursts. The high amounts of power cascading to small scales could confine CRs very effectively in starbursts, so much that CR transport is dominated by the flow of gas rather than diffusion through the gas. In addition, I discuss the effects of turbulence on X-ray line width, the far-infrared–radio correlation, observed radio polarization, and Faraday rotation measures. Finally, I discuss the many questions raised regarding the physics of turbulence in starbursts.'
author:
- 'Brian C. Lacki'
title: |
*Sturm und Drang*:\
Supernova-Driven Turbulence, Magnetic Fields, and Cosmic Rays in the Chaotic Starburst Interstellar Medium
---
Introduction {#sec:Introduction}
============
Starbursts are regions of galaxies that convert gas to stars quickly ($\tau_{\rm gas} \approx 10$ – 100 Myr), resulting in high star-formation surface densities ($\Sigma_{\rm SFR} \ga 0.1\ \Msun\ \yr^{-1}\ \kpc^{-2}$). They host up to 10% of the cosmic star formation rate at all redshifts [@Hopkins10; @Sargent12]. In the present day Universe, they are the most extreme conditions for star formation. They also serve as models for the conditions of star formation in high redshift main sequence galaxies, the birthplaces of most of the Universe’s stars [e.g., @Chary01; @LeFloch05; @Noeske07; @Daddi07; @Magnelli09; @Sargent12]. Those galaxies typically had much larger $\Sigma_{\rm SFR}$ and gas densities and somewhat smaller $\tau_{\rm gas}$ than present-day normal galaxies [@Daddi10-BzK; @Tacconi10; @Genzel10].
The intense conditions of a starburst drive mass and energy transfer through the interstellar medium (ISM), providing feedback that regulates star-formation. The starburst ISM is extremely dynamic and chaotic, with several phases. These phases include relatively cold ($\sim 100\ \Kelv$) and dense molecular gas that makes up most of the mass of starbursts, and a rarefied supernova-heated ($\sim 10^8\ \Kelv$) superwind phase that is thought to fill most of the volume of starbursts. The molecular gas sets the physical conditions of how gas is converted into stars, while the superwind expels gas before it forms stars and enriches the metallicity of the surrounding intergalactic medium. An important feature of the starburst ISM is turbulence, a phenomenon spanning many decades in eddy size. Strong supersonic turbulence (with Mach number ${\cal M} \approx 100$) is present in the molecular gas of starbursts, creating huge density fluctuations and providing high pressures [@Downes98; @MacLow04].
But aside from the thermalized gas of the starburst, star formation also propels the nonthermal phases of starbursts, cosmic rays (CRs) and magnetic fields, which thread through the other phases. The presence of cosmic rays in starbursts is demonstrated by recent gamma-ray detections of M82 and NGC 253 at GeV and TeV energies [@Acero09; @Acciari09; @Abdo10-Starburst]. The presence of CR electrons and positrons ($e^{\pm}$) and magnetic fields is demonstrated by the synchrotron radio emission of starburst galaxies [@Condon92]. The existence of the FIR-radio correlation (FRC), a constant ratio of star-formation rate and continuum GHz radio emission for star-forming galaxies [@Helou85; @Yun01], indicates that the amount of star-formation in a galaxy or starburst is tied to the strength of the magnetic field and amount of cosmic rays [@Lacki10-FRC1]. The nonthermal phases are not just passive ingredients of starburst ISM, but are active participants. Magnetic fields provide strong pressure in starbursts [e.g., @Thompson06] and may halt gas collapse into stars [@Mouschovias76], while CRs are another source of pressure [@Socrates08] and produce ionization in the molecular gas for the magnetic fields to hold onto [@Suchkov93; @Papadopoulos10-CRDRs].
Turbulence links the energy flows in the nonthermal ISM with those in the nonrelativistic gas. Enough energy is stored in the turbulent random motions of the ISM to dominate the fluid pressure and support the ISM from gravitational collapse. On the other hand, some gas in a supersonic turbulent flow is compressed enough to collapse gravitationally and form stars and stellar clusters [e.g., @Krumholz05; @Padoan11; @Hopkins13-Frag]. Turbulence mixes the ISM, distributing metals quickly [@Scalo04]. A turbulent dynamo can amplify a magnetic field until it approaches equipartition with the turbulent energy density (@Cho00 [@Groves03]; I will discuss its efficiency in detail in Section \[sec:Dynamoes\]). Turbulence in the magnetic fields in turn deflects CRs, confining them within a galaxy or starburst [@Schlickeiser02]. More speculatively, strong turbulence can reaccelerate CRs to high energy [@Amano11; @Melia11]. To understand the gas of a starburst, and how magnetic fields and cosmic rays interact with it, we must therefore understand the role of turbulence in starburst galaxies.
Fluctuation dynamoes and equipartition between turbulence and magnetic fields {#sec:Dynamoes}
-----------------------------------------------------------------------------
One of the basic premises of this work is that turbulence is roughly in equipartition with magnetic fields in starbursts [c.f., @Thompson06], because of the fluctuation dynamo. The fluctuation dynamo is thought to be a ubiquitous phenomenon. It relies on the freezing of magnetic fields in most astrophysical plasmas. In plasmas with random motions, the field lines get tangled, and the magnetic field strength increases.
The characteristic residence time of turbulence is an eddy-crossing time, $\ell_{\rm outer} / \sigma$, where $\ell_{\rm outer}$ is the outer scale and $\sigma$ is the turbulent speed. $\ell_{\rm outer}$ is the scale at which turbulence is driven; turbulent energy cascades to smaller scales. The key result of the energy cascade is a spectrum of velocity fluctuations reaching from $\ell_{\rm outer}$ down to an inner scale $\ell_{\rm inner}$ where the energy dissipates. Typically, the velocity fluctuation spectrum is thought to be some kind of power law [@Tennekes72].
As the velocity spectrum develops, the chaotic motions of the plasma deforms the frozen-in magnetic fields, resulting in magnetic fluctuations. A magnetic fluctuation spectrum grows in parallel to the velocity fluctuation spectrum, possibly from small scales to large scales. On some timescale of order the eddy-crossing time, the magnetic fields approach equipartition with turbulence and alter the dynamics of the fluid motions. At this point, the fluctuation dynamo enters a nonlinear regime; it ultimately “saturates” and the magnetic field growth stops [@Cho00; @Beresnyak12]. The fluctuation dynamo does not by itself generate regular fields, as observed in spiral galaxies – these must arise from some other magnetic dynamo [@Beck96].
At least this is the basic picture, but there are many disagreements on the details. The nature of the spectrum of magnetic fluctuations remains unclear. For hydrodynamical turbulence, most power is at large scales (see the discussion in section \[sec:TurbSpectrum\]). But whether the magnetic fluctuations are in equipartition with these velocity fluctuations at each scale is a matter of debate. According to @Schekochihin04, in a flucatuation dynamo, the magnetic fluctuation spectrum only traces the velocity fluctuation spectrum at scales below the resistive scale ($\ell_{\eta} = \ell_{\rm outer} / \sqrt{{\cal R}_m}$, where ${\cal R}_m$ is the magnetic Reynolds number; @Subramanian06), which is tiny in astrophysical systems. As a result, the ratio of saturated magnetic field energy density to turbulent energy density, $\epsilon_B$, is very small. But @Haugen04 instead argued that $\epsilon_B \approx 1/3$, with most of the power on large scales.
Galaxy clusters are a proving ground for fluctuation dynamo theories; they are filled with hot, turbulent plasmas and are known to have magnetic fields from Faraday rotation and synchrotron emission measurements [@Ferrari08]. It’s been known for some time that the magnetic fluctuations cannot peak at the resistive scale, because then there would be no Faraday rotation signal [@Goldshmidt93]. @Subramanian06 and @Ensslin06 argued that the magnetic power peaks at a scale $\ell_{\rm peak} \approx \ell_{\rm outer} / \sqrt{{\cal R}_m^{\rm crit}}$, where ${\cal R}_m^{\rm crit} \approx 35$ is the minimum magnetic Reynolds number for driving magnetic turbulence [see also @Brandenburg05]. They then find that $\epsilon_B \approx 1/4$. This gives results that are compatible with galaxy cluster rotation measures. The simulations by @Beresnyak12 imply that only $\sim 5\%$ of the turbulent energy is transferred into magnetic fields per eddy-crossing time; in starburst galaxies where the wind is advected away in a few eddy-crossing times, this suggests that $\epsilon_B$ is not much greater than $\sim 0.1$.
Even if $\epsilon_B \approx 1/6$ in galaxy clusters, it is not clear whether this applies throughout the starburst ISM. @Stone98 simulated magnetohydrodynamical (MHD) turbulence in molecular clouds and found $\epsilon_B \approx 1/2$. In actual Galactic molecular clouds, magnetic fields and turbulence are in rough equipartition, implying $\epsilon_B \approx 1/3$–$1$ [@Crutcher99; @Troland08; @Falgarone08]. Likewise, magnetic fields in the Galactic cold neutral medium are in equipartition with turbulence [@Heiles05]. @Balsara04 argued that this indicated that the magnetic fields arose from a fast magnetic dynamo, powered by supernova driven turbulence, although in their own simulations, $\epsilon_B$ only reached 0.01. For that matter, the saturated state of the fluctuation dynamo may also depend on the kind of turbulent forcing. @Federrath11 found that solenoidal forcing led to much higher $\epsilon_B$ than compressive forcing. Supernova explosions push the ISM in a compressive manner, yet they can amplify solenoidal motions dramatically in cold and warm gas [@Balsara04]. A “shear dynamo” operating in sheared, turbulent flows leads to not only turbulent magnetic field growth, but a field stronger along one direction than others (@Yousef08; c.f., @Laing80).
For the rest of the paper, I simply scale all of my results to $\epsilon_B = 1$. Remember, though, that the actual value of $\epsilon_B$ is plausibly as low as 0.1 or smaller. I also ignore the time for the turbulent dynamo to reach saturation. The growth time could be important in starbursts if the wind removes the turbulent gas from the star-forming region before the dynamo saturates.
Finally, magnetic fields driven by a turbulent dynamo are unlikely to be homogeneous. Turbulence is “intermittent”, with large fluctuations in strength from place to place and time to time. Instead, the magnetic field structure is thought to include a low-strength volume-filling field, with much higher magnetic field strengths in sheets, ribbons, and filaments [e.g., @Schekochihin04; @Subramanian06; @Bhat13]. These features are roughly $\ell_{\rm outer}$ in length, but only $\ell_{\rm peak}$ in thickness [@Subramanian06]. I generally ignore this complication, but it may prove important in understanding starburst magnetic fields.
Outline of the Paper
--------------------
This paper explores the role of turbulence in starbursts and how it determines the characteristics of the pervasive magnetic fields in these environments. I begin in Section \[sec:BroadView\] by considering the basic structure of starburst ISM as it is shaped by supernova explosions. The bulk characteristics of supernova-driven turbulence and the equipartition magnetic fields in the hot superwind are calculated in Section \[sec:HotTurbulence\]. Similar calculations are performed for turbulence and magnetic fields in cold molecular gas in Section \[sec:ColdTurbulence\]. I demonstrate that energy input from supernovae naturally leads to equipartition between many energy densities in the ISM of starbursts in Section \[sec:Equipartition\]. Section \[sec:Cascade\] explores the cascade of turbulence to small scales, and its implications for CR propagation. The observable effects of turbulence and the magnetic fields, including X-ray line widths, Faraday rotation measures, supernova remnant magnetic fields, and the infrared-radio correlation, are considered in Section \[sec:Implications\]. Finally, I summarize these results in Section \[sec:Conclusion\].
Throughout the paper, I use a Salpeter initial mass function (IMF) from 0.1 to 100$\ \Msun$.
A broad view of the starburst ISM: Physical conditions and phases {#sec:BroadView}
=================================================================
Throughout this work, I compare my results to four prototypical starburst regions: the Galactic Center Central Molecular Zone (GCCMZ), the starburst cores of NGC 253 and M82, and the starburst nuclei of Arp 220. The conditions in these regions span those of $z = 0$ starburst regions. In Table \[table:BasicProperties\], I list the basic properties that I use in computations for these regions. These regions are relatively well studied. Models of the nonthermal emission from each has constrained the magnetic field strength and CR energy densities, and all but Arp 220 have gamma-ray detections. All have star-formation surface densities above the “Heckman threshold" of $0.1\ \Msun\ \yr^{-1}\ \kpc^{-2}$ [c.f., @Heckman90]. The gas consumption times of NGC 253, M82, and Arp 220 are all of the order 10 – 20 Myr, values typical of true starbursts throughout the Universe [@Tacconi06; @Daddi09; @Genzel10; @Riechers11]. The GCCMZ, by contrast, has a far slower gas consumption, more typical of $z \approx 2$ main sequence galaxies [@Genzel10], but since it is much more easily resolved and displays several of the same phenomena as true starbursts, I include it.
[llccccc]{} SFR & $\Msun\ \yr^{-1}$ & 0.07 & 3 & 10 & 100 & a\
R & pc & 100 & 150 & 300 & 100 & b\
h & pc & 50 & 50 & 50 & 50 & c\
$\Sigma_{\rm SFR}$ & $\Msun\ \yr^{-1}\ \kpc^{-2}$ & 2.2 & 42 & 35 & 3200 & d\
$\Gamma_{\rm SN}$ & $\yr^{-1}$ & 0.00052 & 0.022 & 0.074 & 0.74 & e\
$M_{\rm gas}$ & $\Msun$ & $3 \times 10^7$ & $3 \times 10^7$ & $2 \times 10^8$ & $10^9$ & f\
$\tau_{\rm gas}$ & $\Myr$ & 430 & 10 & 20 & 10 & g\
$\Sigma_g$ & $\gcm2$ & 0.20 & 0.087 & 0.15 & 6.7 & h\
${\ensuremath{\langle n_H \rangle}}_{\rm SB}$ & $\cm^{-3}$ & 390 & 170 & 300 & 13000 & i\
$B_{\rm true}$ & $\muGauss$ & 50 - 100 & 100 - 300 & 100 - 300 & 2000 - 6000 & j\
$L_{\rm bol}$ & $\Lsun$ & $3.9 \times 10^8$ & $1.7 \times 10^{10}$ & $5.6 \times 10^{10}$ & $5.6 \times 10^{11}$ & k\
$\dot{E}_{\rm mech}$ & $\Lsun$ & $4.6 \times 10^6$ & $1.9 \times 10^8$ & $6.6 \times 10^8$ & $6.6 \times 10^9$ & l\
$\dot{\varepsilon}_{\rm mech}$ & $\erg\ \cm^{-3}\ \sec^{-1}$ & $1.9 \times 10^{-22}$ & $3.7 \times 10^{-21}$ & $3.1 \times 10^{-21}$ & $2.7 \times 10^{-19}$ & m\
$\dot{\varepsilon}_{\rm mech} / n_H$ & $\erg\ {\rm H}^{-1}\ \sec^{-1}$ & $4.9 \times 10^{-25}$ & $2.2 \times 10^{-23}$ & $1.0 \times 10^{-23}$ & $2.1 \times 10^{-23}$ & \[table:BasicProperties\]
Energy densities in starburst ISM {#sec:UStarburst}
---------------------------------
The flows of energy within galaxies are reflected in the energy densities present in the ISM. The fact of equipartition between the Galactic ISM’s energy fields (thermal, turbulent, radiation, magnetic, CRs) is well known. It is often attributed to couplings between the components that drive them to equal pressures. Equipartition between various ISM components in external galaxies is a frequent assumption, generally to estimate the strength of relatively unknown quantities like the magnetic field or CR pressure. Appeals to pressure balance are also a common way to make a statement about the plausible existence of an ISM phase, such as the hot superwind, or a volume-filling WIM [@Murray10; @Lacki13-LowNu]. In this paper, I argue that the magnetic field and turbulent energy densities are roughly equal. But what actually is our present knowledge of the energy densities in the starburst ISM?
While the pressures have large uncertainties, it is clear from Table \[table:Pressures\] that many of the energy densities in starburst regions are of the same order for each starburst region. The typical pressure $P/k_B$ in the GCCMZ is a few $\times 10^6\ \cm^{-3}$ [see also @Law10], a few $\times 10^7\ \cm^{-3}$ in M82 and NGC 253, and a few $\times 10^9\ \cm^{-3}$ in the intense conditions of Arp 220’s nuclei. Equipartition seems to hold between turbulent energy density in molecular gas, the thermal pressure in superwinds, radiation pressure, and magnetic energy density.
*Molecular gas turbulence* – Although they are in line with the other energy densities, the molecular gas turbulent pressures are actually poorly known. I compute these as ${\ensuremath{\langle \rho \rangle}}_{\rm SB} \sigma^2 / 2$, where $\sigma$ is the turbulent speed. The turbulent speeds of NGC 253 and M82’s molecular gas are not even given in the literature. For M82, $\sigma$ has an upper limit of $70 - 110\ \kms$ from the molecular line widths [@Muehle07]. I assume $\sigma = 50\ \kms$, between the values in the GCCMZ ($\sim 15-30\ \kms$; @Shetty12) and Arp 220 ($\sim 80\ \kms$; @Downes98). That speed also matches the turbulent speed in the H II region surrounding one of M82’s super star clusters [@Smith06]. However, measured turbulent speeds may overestimate the actual level of turbulence, in studies where the observing beam cannot resolve out larger-scale laminar flows [@Ostriker11].
The other missing data are the density of the gas in the molecular clouds. While the mean densities averaged over the entire starburst volume ${\ensuremath{\langle \rho \rangle}}_{\rm SB}$ are fairly well measured, the molecular clouds may only fill a small fraction of that volume, at least in the three weaker starbursts (in ULIRGs, the entire volume may be filled with molecular gas). A filling fraction $\sim 0.1$, entirely plausible given our current state of knowledge, would raise the average density and turbulent energy densities of molecular clouds by $\sim 10$. Thus the molecular gas turbulent pressures should be considered uncertain at the order of magnitude level.
At this point, I should differentiate the true molecular filling factor with the filling factor of the dense molecular clumps inferred by various photodissociation region models of (for example) M82. These clumps are described as clouds with densities $\sim 10^3$ – $10^5\ \cm^{-3}$, sometimes even denser, filling $\sim 0.1 - 1\%$ of the starburst volume [e.g., @Wolfire90; @Wild92; @Lord96; @Mao00; @Weiss01; @Ward03; @Naylor10]. In an isothermal turbulent medium with Mach factor ${\cal M}$, most of the mass is concentrated into regions overdense by a factor $\sim \sqrt{1 + b^2 {\cal M}^2}$, where $b \sim 0.3 - 1$. In the canonical lognormal density distribution for a medium with supersonic turbulence, most of the volume is underdense by the same factor $\sqrt{1 + b^2 {\cal M}^2}$ [@Padoan97; @Ostriker01], although the actual volume-filling density may in fact be near the volume-average density [@Hopkins13-rhoDist]. If M82’s molecular gas has a volume-averaged density $300 f_{\rm fill}^{-1}\ \cm^{-3}$ and ${\cal M} \approx 30$ turbulence, most of the mass is predicted to be in clumps of density $\ga 9000 f_{\rm fill}^{-1}\ \cm^{-3}$ for a lognormal distribution. When the isothermal assumption is relaxed, though, the density distribution is not lognormal anymore, but tends to a power-law form at high densities [@Scalo98].
Therefore the observed “clouds” may not be permanent objects separated by another phase, but rather, temporary structures of a much larger volume molecular medium. The true filling factor of the molecular medium is a factor $\sim {\cal M}$ greater than the inferred filling factor of the clumps – perhaps of the order $\sim 10\%$.
But the thermal pressure in the molecular medium is far below equipartition with the turbulent energy density. The sound speed is a mere $\sqrt{(5/6) k_B T/m_H} = 0.8\ \kms (T/100\ \Kelv)^{1/2}$, so the Mach factors in starburst molecular media are ${\cal M} \approx 10$ – $100$. The ratio of turbulent to thermal pressure scales as ${\cal M}^2$, and is of order $\sim 1000$ in starbursts. Even in the clumps that contain most of the molecular mass, the density is enhanced by only a factor $\sim {\cal M}$, so the thermal pressure inferred for these clumps is still a factor $\sim {\cal M}$ below the mean turbulent pressure.
[lrrrrc]{} Molecular gas turbulence & $\sim 14 \times 10^6$ & $\sim 3 \times 10^7$ & $\sim 5 \times 10^7$ & $\sim 5 \times 10^9$ & (1)\
Superwind thermal (CC85) & $1.1 \times 10^6$ & $2.4 \times 10^7$ & $2.5 \times 10^7$ & $1.5 \times 10^9$ & (2)\
Superwind turbulence (this paper) & $(1 - 2) \times 10^6$ & $(1 - 3) \times 10^7$ & $(1 - 3) \times 10^7$ & $(0.4 - 2) \times 10^9$ & (3)\
H II region thermal & $(3 - 24) \times 10^6$ & $(0.2 - 2) \times 10^7$ & $(0.3 - 2) \times 10^7$ & $(0.02 - 4) \times 10^9$ & (4)\
H II region turbulence & & & $30 \times 10^7$ & & (5)\
& $0.7 \times 10^6$ & $1.4 \times 10^7$ & $1.5 \times 10^7$ & $0.9 \times 10^9$ &\
& $0.8 \times 10^6$ & $2.1 \times 10^7$ & $2.8 \times 10^7$ & $44 \times 10^9$ &\
Magnetic fields & $\sim (0.7 - 3) \times 10^6$ & $\sim (0.3 - 3) \times 10^7$ & $\sim (0.3 - 3) \times 10^7$ & $\sim (1.2 - 10) \times 10^9$ & (7)\
Cosmic rays & $\sim 0.2 \times 10^6$ & $\sim (0.3 - 0.4) \times 10^7$ & $\sim (0.3 - 0.4) \times 10^7$ & $\sim (0.03 - 1) \times 10^9$ & (8)\
& $60 \times 10^6$ & $1 \times 10^7$ & $3 \times 10^7$ & $70 \times 10^9$ &\
& $1700 \times 10^6$ & $6 \times 10^7$ & $17 \times 10^7$ & $130 \times 10^9$ &\
“Orbital” & $150 \times 10^6$ & $10 \times 10^7$ & $18 \times 10^7$ & $80 \times 10^9$ & (10) \[table:Pressures\]
*Cosmic rays* – The most glaring outlier in Table \[table:BasicProperties\] is the CR energy density, which is admittedly poorly constrained but appears to be an order of magnitude lower than the other densities. In the GCCMZ, M82, and NGC 253, the CR energy density is constrained by gamma-ray detections. But deriving these constraints requires knowledge of the gas density the CRs interact with; while typically the mean gas density is used, whether the CR lifetime is actually determined by the mean gas density is not clear. Imposing equipartition between magnetic and CR energy density gives energy densities that are a bit lower than ISM pressure in M82 and NGC 253, $\sim 10^7\ \Kelv\ \cm^{-3}$ [@Lacki13-Equip].
While the CR energy density is perhaps an order of magnitude too low in M82 and NGC 253, the disparity is much more spectacular in Arp 220’s nuclei. The equipartition estimate between magnetic and CR energy densities is a mere $10^8\ \Kelv\ \cm^{-3}$ – roughly 30 – 100 times lower than the pressure in the molecular phase and H II regions [@Thompson06; @Lacki13-Equip]. But even this may be too high. Models of the CR population in Arp 220’s nuclei imply higher magnetic field strengths (roughly $\sim 2$ – $6\ \muGauss$) but lower CR energy densities (a few times $10^7\ \Kelv\ \cm^{-3}$; @Torres04 [@Lacki10-FRC1; @Lacki13-XRay]). In that case, the CRs are a factor of $\sim 100$ below equipartition with the ISM pressure, while the magnetic field energy density is comparable. As with the gamma-ray estimates, these models require knowledge of a density.
*H II region pressure* – Measured H II region thermal pressures show a wide spread, but on average are somewhat high in the GCCMZ and somewhat low in NGC 253, M82, and Arp 220. Either way, it is a very different situation than in the Milky Way, where H II regions are massively overpressured bubbles expanding into the ISM (see section \[sec:WarmGas\]). In addition to the thermal pressures, there is probably turbulent pressure in the H II regions [@Smith06].
*The hydrostatic pressure, the radiation pressure, and rotation* – A starburst’s gas must be in hydrostatic equilibrium, otherwise it will either collapse or be blown apart. For a thin, homogeneous disk with surface density $\Sigma_g$ and gas fraction $f_{\rm gas}$, the necessary midplane hydrostatic pressure is $\pi G \Sigma_g \Sigma_{\rm tot} = \pi G \Sigma_g^2 / f_{\rm gas}$. When compared to the other pressures in Table \[table:Pressures\], $\pi G \Sigma_g^2$ seems roughly in line with the other pressures in NGC 253, and M82, especially when considering that $f_{\rm gas}$ is probably of order a few tenths. But it fails spectacularly at predicting the characteristic pressures in Arp 220.
What is the cause of this discrepancy? The radiation pressure in Arp 220 could be high enough to reach $\pi G \Sigma_g^2$, but only if it is an extreme scattering atmosphere with $\tau_{\rm IR} \approx 100$, so that the energy density accumulates as photons diffuse out of the starburst. Those high levels motivate radiation-pressure supported models of starbursts [e.g., @Thompson05; @Thompson06; @Murray10]. Yet an odd fact is clear from Table \[table:Pressures\]: $\pi G \Sigma_g^2$ is much greater than the mean turbulent energy density. Thus, if radiation pressure pressurizes ULIRGs that amount, it apparently does so *without driving turbulence to equipartition* (Table \[table:Pressures\]), in defiance of the arguments of @Thompson06. An additional problem is that actual starbursts are inhomogeneous, and may have low density chimneys where IR photons can escape. A recent simulation by @Krumholz12 indicates that instabilities reduce $\tau_{\rm IR}$ by a factor $\sim 5$ in actual starbursts. That implies a radiation energy density $\sim 10^{10}\ \Kelv\ \cm^{-3}$ in Arp 220, comparable to the turbulent energy density.
If the radiation pressure does not overwhelmingly dominate the pressures in Arp 220, why does $\pi G \Sigma_g^2$ fail as an estimate? One possibility is that the surface density in Arp 220 is overestimated by a factor $\sim 10$, reducing both $\pi G \Sigma_g^2$ and $U_{\rm turb}$ to $\sim 10^9\ \Kelv\ \cm^{-3}$. That would require substantial errors in our ability to measure gas masses, though. A second possibility is that the “nuclei” of Arp 220 are in fact transient features and will collapse in a free-fall time $\sim 1 /\sqrt{G\rho} \approx 10^6\ \yr$. Indeed, the two nuclei appear to be concentrations within a larger gas disk with ${\ensuremath{\langle \Sigma_g \rangle}} \approx 1\ \gcm2$ [@Downes98]. Yet the concentration of radio supernovae in the nuclei requires that the nuclei have been stable for at least a few million years [e.g., @Smith98; @HerreroIllana12].
The discrepancy could indicate that Arp 220’s nuclei are very inhomogeneous. This is supported by the relatively low Toomre ${\cal Q} \approx \sqrt{2} \sigma v_{\rm circ} / (\pi G \Sigma_g R)$ values of $\sim 0.3$ – 1 for the nuclei [c.f., @Kruijssen12]. Gas under such conditions fragments into small clumps. The clumps’ gravity holds down any interclump gas, but they do not themselves need to be supported by interclump gas. The clumps orbit around the nucleus; the nucleus as a whole is supported by the orbital motions of these clumps rather than the turbulent motions of gas within the clumps (just as the orbital speeds of stars in the Milky Way’s disk are much greater than their local velocity dispersion). Indeed, I find that the orbital kinetic energy density in Arp 220’s nuclei is comparable to $\pi G \Sigma_g^2$ (see Table \[table:Pressures\]), when I use the orbital speed of $325\ \kms$ [@Downes98].
Similar considerations seem to apply to the GCCMZ, where the clumpiness of the gas can be directly observed [e.g., @Molinari11]. Note that if $f_{\rm gas}$ is very small, as suggested by @Launhardt02, then the hydrostatic pressure apparently is much greater than even the orbital kinetic energy (Table \[table:Pressures\]). This may mean that $f_{\rm gas}$ is actually of order a few tenths in this region, as in M82 and NGC 253.
*Basic conclusions* – Loose equipartition, at least, seems to hold between the various gas phases of the ISM in starbursts. The major outliers are the CR energy density and the hydrostatic pressure estimate. The hot superwind phase, if it exists, is in equipartition with the cold molecular gas; thus pressure arguments alone do not rule it out. The magnetic energy density is comparable to the turbulent energy density.
Cold clouds in a hot wind or hot bubbles in a cold sea: Which phase is volume-filling? {#sec:VolumeFillingPhase}
--------------------------------------------------------------------------------------
With these estimates of the central pressure of starburst regions in mind, which phase fills most of the volume in starbursts? Do supernova remnants (SNRs) expand until they largely overlap, or are they stopped and form little bubbles in the ISM? The volume overlap factor is essentially given by: $$Q_{\rm SNR} = \rho_{\rm SN} \times \frac{4}{3} \pi R_{\rm max}^3 \times t_{\rm SNR},$$ the product of the volumetric rate of supernova explosions ($\rho_{\rm SN}$), the maximum volume of the SNR (set by the maximum radius of the remnant, $R_{\rm max}$), and the time the SNR survives before being washed away by the surrounding ISM ($t_{\rm SNR}$). The filling fraction of the SNRs is then $f_{\rm fill} = 1 - \exp(-Q_{\rm SNR})$ [@McKee77; @Heckman90]. The survival time is generally assumed to be of order the ISM sound-crossing time of the SNR, $R_{\rm max} / \sqrt{P_{\rm ISM} / \rho_{\rm ISM}}$ [@McKee77].
Suppose the ISM originally consists of cold molecular gas. Then, the evolution of SNRs is most likely to be set by radiative losses in the dense ISM. The maximum radius of a radiative SNR is $$R_{\rm max} = 6.6\ \pc\ E_{51}^{0.32} \left(\frac{n_{\rm ext}}{100\ \cm^{-3}}\right)^{-0.16} \left(\frac{P/k_B}{10^7\ \Kelv\ \cm^{-3}}\right)^{-0.2}$$ from @McKee77. In these conditions, and according to @McKee77 and @Heckman90, the overlap factor is $$\begin{aligned}
\nonumber Q_{\rm SNR} & = & 3.4\ E_{51}^{1.28} \left(\frac{\rho_{\rm SN}}{10\ \kpc^{-3}\ \yr^{-1}}\right) \left(\frac{n_{\rm ext}}{100\ \cm^{-3}}\right)^{-0.14} \\
& & \times \left(\frac{P/k_B}{10^7\ \Kelv\ \cm^{-3}}\right)^{-1.3},\end{aligned}$$ In this equation, $E_{51}$ is the mechanical energy of the supernova in units of $10^{51}\ \erg$, $\rho_{\rm SN}$ is the volumetric supernova rate, $n_{\rm ext}$ is the density of the surrounding ISM, and $P_{\rm ISM}$ is the ISM pressure.
We need two ingredients: the density of the surrounding ISM, and the pressure of the surrounding ISM. It is possible to relate the density to $\Sigma_{\rm SFR}$ with the Schmidt law [@Kennicutt98], but the recent literature suggests that the gas masses in some starbursts (e.g., M82 and NGC 253) are lower than used in @Kennicutt98 [@Weiss01]; furthermore, true merger-driven starbursts appear to lie on a different Schmidt law than main sequence galaxies [@Daddi10-Schmidt]. I instead relate $\Sigma_g$ and $\Sigma_{\rm SFR}$ through the gas consumption time $\tau_{\rm gas}$: $$\left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right) = 240\ \left(\frac{\Sigma_g}{\gcm2}\right) \left(\frac{\tau_{\rm gas}}{20\ \Myr}\right)^{-1}.$$ I then set $n_{\rm ext} = \delta (\Sigma_g / (2 h m_H)) = \delta {\ensuremath{\langle \rho \rangle}}_{\rm SB} / m_H$, where $\delta$ accounts for the possibility that SNe generally go off in an underdense medium.
As noted in the previous subsection, there are several different estimates we can use for the ISM pressure. The first is the turbulent energy density of the gas, ${\ensuremath{\langle \rho \rangle}}_{\rm ext} \sigma^2 / 2$. In this case, the SNR overlap fraction is: $$\begin{aligned}
\nonumber Q_{\rm SNR}^{\rm turb} & = & 0.55\ \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right)^{-0.44} \left(\frac{\tau_{\rm gas}}{20\ \Myr}\right)^{-1.44}\\
& & \times \left[E_{51}^{1.28} h_{50}^{0.44} \delta^{-0.14} \sigma_{50}^{-2.6}\right].\end{aligned}$$ The auxiliary variables parameterize the mechanical energy per supernova as $10^{51} E_{51}\ \erg$, the starburst scale height as $50 h_{50}\ \pc$, and the turbulent dispersion as $50 \sigma_{50}\ \kms$. The second option is to suppose the pressure everywhere in the ISM is equal to the CC85 wind pressure (given in Appendix \[sec:CC85Properties\]), giving: $$\begin{aligned}
\nonumber Q_{\rm SNR}^{\rm CC85} & = 0.77\ \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right)^{-0.44} \left(\frac{\tau_{\rm gas}}{20\ \Myr}\right)^{-0.14}\\
& \times \left[E_{51}^{1.28} f_{\rm geom}^{2.6} \delta^{-0.14} \zeta^{-1.3} h_{50}^{-0.86} \left(\frac{\beta}{2}\right)^{-0.65} \left(\frac{\epsilon_{\rm therm}}{0.75}\right)^{-0.65}\right].\end{aligned}$$ Finally, we can use the homogeneous hydrostatic pressure, $\pi G \Sigma_g^2 / f_{\rm gas}$: $$\begin{aligned}
\nonumber Q_{\rm SNR}^{\rm hydro} & = & 80\ \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right)^{-1.74} \left(\frac{\tau_{\rm gas}}{20\ \Myr}\right)^{-2.74}\\
& & \times \left[E_{51}^{1.28} f_{\rm gas}^{1.3} \delta^{-0.14} h_{50}^{-0.86}\right].\end{aligned}$$
Because of its strong dependence on pressure, the overlap fractions differ greatly, but there is a universal trend towards smaller $Q_{\rm SNR}$ as $\Sigma_{\rm SFR}$ increases (Figure \[fig:QSNR\]). In the weakest starbursts, $Q_{\rm SNR} \approx 1$, while $Q_{\rm SNR}$ drops to a few percent in powerful ULIRGs. This trend is repeated in Table \[table:BHotPredictions\], where I calculate the different $Q_{\rm SNR}$ for the GCCMZ, NGC 253, M82, and Arp 220. The overlap factors are of order half in the NGC 253, and M82, and only a few percent in Arp 220. There is a very wide disagreement in the $Q_{\rm SNR}$ for the GCCMZ, because the pressures in Table \[table:Pressures\] are themselves disparate. If I use $P \approx 10^6\ \Kelv\ \cm^{-3}$ (the superwind, magnetic, radiation pressure), then $Q_{\rm SNR} \approx 1$. Using the much greater molecular turbulence pressure gives values of a few percent, and the hydrostatic pressure implies that $Q_{\rm SNR} \approx 0$.
{width="8cm"}
In practice, a few effects raise the overlap fractions from my fiducial parameters. Star formation in starbursts occurs in clusters [@OConnell95; @Kruijssen12]. The combined effect of the stellar winds in the clusters digs out superbubbles. In the low density superbubbles, the mechanical power of supernovae is not as easily lost to radiative losses. In addition, some hot bubbles are heated by multiple supernovae. Clustering increases the overlap fraction, since the overlap fraction depends more strongly on the number of supernovae in the bubble ($\propto E_{51}^{1.28}$) than on the number of bubbles ($\propto N$). These effects may raise the overlap fraction a few times. Evidently, $Q_{\rm SNR}$ *is* enhanced in M82 enough for the hot wind to fill the volume, as demonstrated by the presence of hot X-ray gas [@Strickland07] and the rapid expansions of SNRs (see Section \[sec:SNREvolution\]). In Arp 220, however, the overlap fraction would have to be raised by a factor $\ga 20$ for the hot phase to dominate. @Melioli04 found that the supernova heating efficiency is bistable in starbursts, tending either towards high radiative losses if dense molecular gas fills the starburst volume or efficient supernova heating once the hot phase grows enough.
High-redshift galaxies, both main sequence and true starburst, have greater scale heights. At a given $\Sigma_{\rm SFR}$ and $\Sigma_g$, these galaxies should have smaller ${\ensuremath{\langle \rho \rangle}}$ and ${\ensuremath{\langle \rho_{\rm SN} \rangle}}$. If the mean pressure in these galaxies is given by ${\ensuremath{\langle \rho \rangle}} \sigma^2 / 2$, then the decreased volumetric density implies that the overlap fraction is larger in these galaxies. But both the CC85 wind pressure and the (homogeneous) hydrostatic pressure depend on column rather than volumetric density, so using these pressures implies that the overlap fraction decreases, as the larger scale height dilutes the SNRs.
If the hot ionized medium is established, and the SNe go off in it, then what is the new overlap fraction? The SNRs fade into the hot ISM before they suffer radiative losses [compare with the cooling radii listed in Table 2 of @Draine91]. In this case, the SNR ceases to be distinct from the background ISM when its internal energy density is comparable to the surrounding energy density: $$R_{\rm max}^{\rm therm} \approx \left[\frac{E_{\rm SN}}{(4/3) \pi U_{\rm ISM}}\right]^{1/3} \approx \left[\frac{E_{\rm SN}}{2 \pi n_{\rm wind} k T_c}\right]^{1/3}$$ Taking the CC85 wind pressure and density, I find $$\begin{aligned}
\nonumber R_{\rm max}^{\rm therm} & \approx 35\ \pc\ \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right)^{-1/3}\\
& \times \left[E_{51}^{1/3} f_{\rm geom}^{2/3} \zeta^{-1/3} \left(\frac{\beta}{2}\right)^{-1/6} \left(\frac{\epsilon_{\rm therm}}{0.75}\right)^{-1/6} \right].\end{aligned}$$ The overlap fraction is then $$\begin{aligned}
\nonumber Q_{\rm SNR}^{\rm hot} & \approx 0.52\ \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right)^{-1/3}\\
& \times \left[E_{51}^{4/3} f_{\rm geom}^{8/3} \zeta^{-4/3} \left(\frac{\beta}{2}\right)^{-1/6} \left(\frac{\epsilon_{\rm therm}}{0.75}\right)^{-7/6} \right].\end{aligned}$$ with a SNR survival time $$\begin{aligned}
\nonumber t_{\rm SNR}^{\rm hot} & \approx 47\ \kyr\ \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right)^{-1/3}\\
& \times \left[E_{51}^{1/3} f_{\rm geom}^{2/3} \zeta^{-1/3} \left(\frac{\beta}{2}\right)^{1/3} \left(\frac{\epsilon_{\rm therm}}{0.75}\right)^{-2/3} \right].\end{aligned}$$
The overlap fraction is no bigger than in the cold phase, despite the lack of radiative losses to stop SNR growth, because the sound speed in the hot ISM is much larger, shortening the SNR survival time. In all starbursts with $\Sigma_{\rm SFR} > 1\ \Msun\ \yr\ \kpc^{-2}$, $Q_{\rm hot} < 1$ (Figure \[fig:QSNR\]). While $t_{\rm SNR}^{\rm hot}$ is short, the SNR is simply merging into a background of hot plasma, not being filled in by cold gas. Instead, the low value of $Q_{\rm hot}$ means that while the hot phase fills some starbursts, the hot phase is truly diffuse as opposed to being a bunch of mostly-overlapping SNR bubbles. Most of the hot plasma in weaker starbursts cannot be attributed to individual SNRs, but only to the collective action of supernovae over hundreds of kyr.
For convenience, I group starbursts into two varieties through this work. I call the lower density starbursts, including the GCCMZ, NGC 253, and M82, “hot” starbursts because the hot phase fills most of their ISMs. Higher density starbursts, including Arp 220 and submillimeter galaxies, are grouped as “cold” starbursts. Both groups stand in contrast to the Milky Way and other $z \approx 0$ normal galaxies, where warm neutral and warm ionized gas fills most of the ISM, and thus might be called “warm” galaxies. Figure \[fig:ISMSketches\] depicts the different ISM phase structure in warm, hot, and cold starbursts.
I address other objections to the existence of hot starbursts in Appendix \[sec:HotObjections\].
SNR evolution observed in starbursts {#sec:SNREvolution}
------------------------------------
I motivate the above picture by appealing to long-term evolution of SNRs in starbursts. The properties of SNRs in nearby starbursts can be directly determined from the properties of compact radio and X-ray sources. With Very Long Baseline Interferometry (VLBI), it is actually possible to directly determine both the size and expansion speeds of SNRs and radio SNe. Since the late evolution of SNRs is strongly affected by which phase they are in, these observations are powerful tools for understanding the ISM phase structure of starbursts.
The basic idea, that supernovae going off in molecular clouds have highly radiative remnants whereas those that go off in the hot wind are in the Sedov phase when they fade away, was described by @Chevalier01. Under the assumption that molecular clouds encase supernovae, and assuming an average density of ${\ensuremath{\langle n_H \rangle}} \approx 1000\ \cm^{-3}$, @Chevalier01 predicted that most SNRs in M82 progress through the Sedov and enter the radiative phase rapidly. By the time they reach the observed radii of most SNRs in M82, a few parsecs, they should be slowed down to a speed $\sim 500\ \kms$. @Chevalier01 argued that these SNRs would evolve slowly, explaining why the SNR’s radio fluxes do not appear to vary with time.
Yet VLBI observations repeatedly show this not to be the case for most SNRs. Instead, the typical expansion speeds of the SNRs are roughly $10^4\ \kms$ [@Pedlar99; @Beswick06; @Fenech10]. The speeds are, in fact, about the same as for new supernovae [@Brunthaler10]. If the SNRs are embedded in the molecular medium, it would be difficult to explain these observations. However, if the SNRs are in the hot wind, they only reach the Sedov phase when they have expanded to a radius $R_{\rm Sedov} = [3 M_{\rm ej} / (4 \pi \rho_{\rm ext})]^{1/3}$: $$\label{eqn:RSedov}
R_{\rm Sedov} \approx 2.1\ \pc \left(\frac{M_{\rm ej}}{1\ \Msun}\right)^{1/3} \left(\frac{n_{\rm ext}}{1\ \cm^{-3}}\right)^{-1/3},$$ where $M_{\rm ej}$ is the mass of the SN ejecta.[^1]
As for why the SNRs in M82 are not embedded within molecular clouds as originally supposed by @Chevalier01, note that the typical time after a stellar population forms until the first SNe go off is several Myr. On the other hand, the turbulent gas within M82’s molecular medium rearranges itself within one eddy time, $\ell_{\rm outer} / \sigma \approx 200\ \kyr\ (\ell_{\rm outer} / 10\ \pc) (\sigma / 50\ \kms)^{-1}$. Thus, even ignoring the fact that stellar clusters may disrupt the molecular clouds they were born in [e.g., @Murray10] and the “cruel cradle effect” [e.g., @Kruijssen12], the transient clumps of molecular gas that swathe a stellar nursery can be long gone by the time the first SNe go off.
What about Arp 220, where I argue that the molecular medium fills the entire starburst? In Arp 220, VLBI observations paint a far different picture [e.g., @Smith98; @Rovilos05; @Lonsdale06; @Parra07]. Aside from an extremely large number of radio supernovae, the radio sources observed in Arp 220 appear to be very small SNRs with radii $\sim 0.27 - 0.38\ \pc$ [@Batejat11]. By noting the presence of the SNRs in observations from the prior decade, @Batejat11 estimate that the expansion speed is at most $5000\ \kms$, implying that the SNe are well into the Sedov phase. Comparing to equation \[eqn:RSedov\], this is consistent with the SNe going off in a density $\sim 10^4\ \cm^{-3}$ [@Batejat11]. By contrast, the Sedov radius for a CC85 wind with $n_e \approx 25\ \cm^{-3}$ is $0.7\ \pc\ (M_{\rm ej} / \Msun)^{1/3}$. Assuming that SNe go off in random places in Arp 220’s ISM, the small, slowly expanding SNRs observed by @Batejat11 imply that the cold molecular gas is volume-filling.
It would be interesting to see where the demarcation between hot wind and cold gas dominated starburst occurs by conducting VLBI observations of starbursts with properties between those of M82 and Arp 220. One possible target is Arp 299; there have been several VLBI observations but no SNR expansion speeds reported yet [@Neff04; @Ulvestad09; @PerezTorres09].
What drives the turbulence in starbursts? {#sec:TurbulenceSources}
-----------------------------------------
Many drivers of turbulence in star-forming galaxies, tapping different energy sources, have been proposed. These include gravitationally-powered disk-scale instabilities [@Fleck81; @Wada02; @Agertz09], galactic accretion [@Klessen10], the expansion of superbubbles and SNRs [@Norman96; @Korpi99; @Dib06; @Joung06; @Joung09], H II region expansion [@Matzner02], protostellar winds and jets [@McKee89; @Li06], and radiation pressure [@Murray10]. A combination of any of these may operate in different classes of galaxy and in different ISM phases. For example, the primary power source for Milky Way turbulence is probably superbubbles and supernovae, but with additional contributions from the other sources depending on the scale [@Miesch94; @Norman96; @MacLow04; @Haverkorn08; @Brunt09; @Lee12].
The hot superwind of starbursts is transparent to radiation and is not bound to the starburst. The only possible drivers of turbulence are those with mechanical forcing: the expansion of superbubbles and SNRs, in particular. Supernovae and stellar winds are the power source for these; their mechanical luminosity is well known (see Appendix \[sec:CC85Properties\]).
The situation is less clear in molecular clouds and in cold starbursts. Supernovae go off in these regions, and might drive turbulence there as well [@Ostriker11]. Supernova remnants may be highly radiative in these regions, with $\sim 10\%$ of their mechanical energy going into turbulent forcing [@Thornton98]. On the other hand, radiation is potentially a powerful source of turbulence in the extreme infrared radiation pressure and optical depths in these environments (Table \[table:Pressures\]) (@Murray10 [@Ostriker11]; but see @Krumholz12). The Toomre ${\cal Q}$ is observed to be $\la 1$ in Arp 220 (see Section \[sec:UStarburst\]; @Kruijssen12), a condition which promotes disk instabilities [@Ostriker11].
The debate over the source of turbulence in nearby starbursts is related to the similar debate over turbulence driving in high-redshift galaxies. High redshift disks are observed to host clumps and large velocity dispersions. Although @Green10 argued that a correlation between turbulence and star-formation rate proves star formation powers turbulence, other authors note that Toomre ${\cal Q}$ is near 1 in the clumps and propose that disk instabilities are the source of turbulence [@Genzel08; @Burkert10].
In any case, though, stellar mechanical luminosity sets a lower bound on the turbulent energy injection. I therefore focus my discussion on turbulence powered by SNRs and other mechanical sources. The addition of other turbulent drivers would lead to faster turbulent velocity dispersions, and stronger turbulent magnetic fields.
What about the cool and warm gas? {#sec:WarmGas}
---------------------------------
Throughout my discussion, I have largely ignored the diffuse warm phases, the Warm Neutral Medium (WNM) and the Warm Ionized Medium (WIM), as well as the Cool Neutral Medium (CNM). HI gas makes up the great majority of the Milky Way’s gas mass, and the WIM and WNM fill roughly half of its volume (see Figure \[fig:ISMSketches\]). However, the molecular fraction of gas increases with increasing $\Sigma_g$ and $\Sigma_{\rm SFR}$; for starbursts, the majority of the gas mass is molecular [@Wong02; @Ferriere07; @Krumholz09].
This still leaves a minority of the gas mass in neutral phases in weaker starbursts (e.g., @Ferriere07 and references therein for the GCCMZ; @Yun93 and @Wills98 for M82). Yet its density should tend to be near the mean molecular gas density. The neutral gas’ pressure support is either thermal or neutral. If the neutral gas were very underdense, it would need to have much higher turbulent speeds than the molecular gas ($\gg 50\ \kms$) to resist collapse. Yet such speeds would eject the neutral gas. On the other hand, the sound speed of warm gas, $\sim 16\ \kms$ is only a few times smaller than the turbulent speed of the molecular gas. Thus, warm gas cannot remain in equilibrium if it is in very overdense, because the thermal pressure alone would exceed the surrounding turbulent pressure. Cool neutral gas might be overdense if it had a small turbulent velocity, but as I show in Section \[sec:ColdTurbulence\], the expected turbulent energy density depends weakly on the gas density.
A WIM with a substantial filling fraction of starbursts is unlikely to exist in starburst regions [@Lacki13-LowNu]. Briefly, the high densities of starburst gas leads to rapid recombination. Therefore, the ionizing photons produced by the starburst can only ionize a few percent of the gas mass. In addition, not all ionizing photons may be available, if some are destroyed by dust absorption or escape through the hot wind. A low density WIM is inconsistent with the pressure constraints. Instead, the only WIM that can survive is in H II regions that fill a small fraction of the volume and have densities comparable to the mean. And indeed, such H II regions are known to exist through radio recombination line studies [@Zhao93; @RodriguezRico05; @RodriguezRico06; @Law09]: the H II regions most likely *are* the WIM in starbursts [@Lacki13-LowNu].
In short, the WIM, WNM, and CNM, to the extent they exist in starbursts, have mean densities comparable to the molecular gas. Since the calculation of turbulent speeds and magnetic fields does not depend on temperature, these quantities should be similar to those of the molecular gas as derived in Section \[sec:ColdTurbulence\]. The one difference is the Mach number, which is much smaller in the WIM and WNM than in molecular gas because of the higher temperatures. Smaller Mach numbers lead to smaller density contrasts in the turbulence [@Ostriker01].
It therefore appears that starburst ISMs are, in some sense, simpler than the Galactic ISM. The famous three phases (cold, warm, and hot) of the Milky Way and other warm galaxies [@McKee77] are reduced to two in hot starbursts (cold and hot) and finally only one in cold starbursts, as depicted in Figure \[fig:ISMSketches\].
{width="7cm"}
Turbulence and Magnetic Fields in Superwinds {#sec:HotTurbulence}
============================================
Motivation for magnetic fields in the superwind {#sec:HotWindMotivation}
-----------------------------------------------
What evidence do we even have that there is turbulence and magnetic fields in the hot wind? Although SNRs no doubt push into the wind, such motions do not necessarily fill the whole medium. Nor is it obvious that a fluctuation dynamo operates in the hot wind.
Although we have little direct evidence regarding the magnetic fields and turbulence in superwinds, there is circumstantial evidence from the synchrotron emission. VLBI observations resolve out most of a starburst’s radio emission, indicating it is truly diffuse and not confined to small clumps [e.g., @Lonsdale06]. In the GCCMZ, one of the few starburst regions we can resolve in low frequency radio, 74 MHz observations reveal synchrotron emission shining from beyond H II regions, implying the CRs are not confined to these regions [@Brogan03; @Nord06]. A MERLIN image of M82 also shows synchrotron emission at 400 MHz outside of a large H II region [@Wills97]. I propose the simplest explanation for this diffuse radio emission is that CR $e^{\pm}$ actually do permeate the volume of the starburst and are radiating synchrotron throughout it. In hot starbursts, this means that the CRs spend most of their time in the volume-filling phase, the superwind. The diffuse synchrotron emission traces magnetic fields in the superwind.
In addition, the gamma-ray luminosities of M82 and NGC 253 imply that CR protons are being advected out by the starburst wind. If they instead remained in the starburst, they would eventually be destroyed by pionic losses as they passed through dense gas clouds. In this “proton calorimeter” limit [c.f., @Pohl94], the energy that went into accelerating CR protons would then be converted into pions, about a third of it ending up as gamma-rays. The actual gamma-ray luminosities of M82 and NGC 253 are only $\sim 40\%$ of the calorimetric limit, implying that CRs are escaping [@Lacki11-Obs; @Abramowski12; @Ackermann12]. Furthermore, the hard GeV to TeV gamma-ray spectra of these starbursts imply that CRs escape through an energy-independent process, which is contrary to our normal understanding of CR diffusion [@Abramowski12]. The GCCMZ also displays underluminous but hard gamma-ray emission and even synchrotron radio emission for its star-formation rate, which suggests rapid energy-independent CR escape [@Crocker11-Wind]. Advection by the starburst wind naturally accounts for these observations. But whereas the hot wind is consistently expelled out of the starburst, only $\sim 1\%$ of the cold and warm gas is ejected per advection time ($\sim 300\ \kyr$). Otherwise, the gas mass would be depleted within a Myr, quenching star formation. Unless the CRs “know” which warm and cold clouds are going to be expelled, advective escape implies they are primarily in the hot superwind.
The supernova observations I describe in \[sec:SNREvolution\] support CR injection directly into the superwind too. The rapidly expanding supernova remnants of M82 can only exist if they are in a low density phase, like the superwind. Yet they are clearly glowing in synchrotron, implying CR $e^{\pm}$ acceleration. Additional support comes from 400 MHz MERLIN observations of supernova remnants in M82. These also emit synchrotron radiation, but a significant fraction display no free-free absorption [@Wills97], again implying they are not in H II regions.
Finally, radio haloes are observed around starbursts, indicating CR $e^{\pm}$ and magnetic fields are present in their winds [@Seaquist91]. A similar kiloparsec-scale radio halo is observed around the GCCMZ [@Crocker11-Wind]. Since there is warm and cold gas mixed with the wind, this by itself does not prove the CR $e^{\pm}$ and magnetic fields are in the hot phase per se. Strong adiabatic losses may also cool the wind and the CR $e^{\pm}$ on kiloparsec scales from the starburst (CC85), so the physical conditions may not be equivalent to the hot ISM inside the starburst. But the radio haloes are consistent with magnetic fields and CR $e^{\pm}$ being present in the hot wind itself.
The existence of turbulence (or at least volume-filling random motions), magnetic fields, and CRs throughout hot superwinds is therefore plausible. Clearly this cannot be the whole picture: the (presumably pionic) gamma-ray emission from the GCCMZ, NGC 253, and M82 implies that CRs must at some point enter denser gas. In fact, models of the nonthermal emission of M82 and NGC 253 are consistent with CRs experiencing the average density of gas in starbursts over their lifetimes (e.g., @Domingo05 [@Persic08; @deCeaDelPozo09-M82; @Rephaeli10; @Lacki10-FRC1; @Lacki13-XRay; @YoastHull13]; higher than average densities are possible as shown by @Paglione12), suggesting that they traverse all of the phases of the ISM.[^2] The situation would then be analogous to the Milky Way, where CRs diffuse through the low density, warm/hot Galactic halo outside of the gas disk most of the time, but experience most of their gas interactions in the neutral and molecular Galactic disk. As in the Milky Way, most of the synchrotron emission of starbursts comes from underdense hot plasma, but occasionally the CRs penetrate into molecular clouds and/or H II regions and experience energy losses from pion production, bremsstrahlung, and ionization. This could happen if there are a sufficiently large number of molecular clouds, so that most magnetic field lines (and, thus, most CR paths) traverse a molecular cloud, as described in detail by @YoastHull13.
I note that this scenario, which motivates my calculation of the turbulence and magnetic fields in superwinds, bears some similarities to that presented by @Becker09. @Becker09 argued that the thermal, magnetic, CR, and turbulent energy densities in the hot ISM are in equipartition in actively star-forming galaxies. They argued this meant that the timescales regulating each are equal, and derived magnetic field strengths by relating the CR diffusive escape time to the radiative cooling time. My hypothesis is weaker, simply that turbulence and magnetic fields are in equipartition. As it turns out, the four energy densities in the wind end up being roughly equal in weaker starbursts, but this is not explicitly required in this paper and the defining timescale is the flow crossing time (see Section \[sec:Equipartition\]). In denser starbursts, equipartition between CRs and magnetic fields fails entirely.
My discussion is guided by the CC85 theory of starburst winds, which gives approximate values for the wind temperature, central density, and pressure.
Are superwinds collisional? {#sec:CollisonlessWinds}
---------------------------
Many processes can shape the wind and alter the CC85 values, such as nonthermal pressure from magnetic fields and CRs, but one deserves special mentioning, the low collisionality of hot dilute plasmas. The CC85 theory assumes that the wind is a hydrodynamic fluid, which has a locally defined pressure and temperature. Yet, when the plasma has a small enough density, it becomes more collisionless, and its electrons and ions may have different temperatures ($T_e$ for electrons and $T_i$ for ions), or even no temperature at all. Although @Strickland07 briefly investigated the possibility that M82’s wind plasma is collisionless, the issue bears greater examination when considering the very different conditions in wind fluids for starburst regions ranging from the GCCMZ to ULIRGs.
Collisions are most effective at transferring energy when the two particles have the same mass. While the effective cross sections for collisions between particles with the same charge are the same, lighter particles have a higher microscopic speed at a given energy, so their collision rate is faster. Thus, the electron-electron collision frequency is highest, so electrons are the most collisional particles with the best-defined temperature. The electron-electron collision frequency is $\nu_{ee} = 4 \pi n_e e^4 (\ln \Lambda)/ \sqrt{m_e (k_B T_e)^3}$. The typical value of the Coulomb logarithm ($\ln \Lambda$) in a superwind plasma is $\sim 30$ [@Huba11]. In a CC85 superwind, the electron-electron collision frequency is: $$\begin{aligned}
\nonumber \nu_{ee}^{-1} & = 1.3\ \kyr\ \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right)^{-1} \left(\frac{T_e}{T_c}\right)^{3/2}\\
& \times \left[f_{\rm geom}^2 \zeta^{-1} \left(\frac{\ln \Lambda}{30}\right)^{-1} \left(\frac{\beta}{2}\right)^{-3} \left(\frac{\epsilon_{\rm therm}}{0.75}\right)^2 \right].\end{aligned}$$ Note that the time for the wind to escape the starburst region is typically a few hundred kyr. In all starburst superwinds, the electrons are collisional and have established a Maxwellian pressure and temperature distribution.
Since ions are heavier, they are slower and collide more rarely, with $\nu_{ii} = 4 \pi n_H e^4 / \sqrt{m_H (k_B T_e)^3}$ (for hydrogen ions): $$\begin{aligned}
\nonumber \nu_{ii}^{-1} & = 66\ \kyr\ \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right)^{-1} \left(\frac{T_i}{T_c}\right)^{3/2}\\
& \times \left[f_{\rm geom}^2 \zeta^{-1} \left(\frac{\ln \Lambda}{30}\right)^{-1} \left(\frac{\beta}{2}\right)^{-3} \left(\frac{\epsilon_{\rm therm}}{0.75}\right)^2 \right].\end{aligned}$$ In the weakest starbursts, like the GCCMZ, the ion-ion collision approaches the advection time. Thus, the ions in these regions will only approximately establish a Maxwellian distribution.
Finally, electrons and ions are poorly coupled in wind plasmas because they have such different masses. The light electrons are easily deflected by the much heavier ions with an electron-ion collision frequency that is just $\nu_{ei} = \nu_{ee} / (2 \sqrt{2})$. Yet very little energy transfer occurs in these deflections, only about $m_e / m_H \sim 1/2000$ of the particle kinetic energy [@Kulsrud05]. The timescale for energy transfer between the ions and electrons is set by the ion-electron collison frequency, which is much slower even than $\nu_{ii}$: $\nu_{ie} = \nu_{ee} m_e / (2\sqrt{2} m_i)$ or $$\begin{aligned}
\nonumber \nu_{ie}^{-1} & = 6.9\ \Myr\ \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right)^{-1} \left(\frac{T_{e,i}}{T_c}\right)^{3/2}\\
& \times \left[f_{\rm geom}^2 \zeta^{-1} \left(\frac{\ln \Lambda}{30}\right)^{-1} \left(\frac{\beta}{2}\right)^{-3} \left(\frac{\epsilon_{\rm therm}}{0.75}\right)^2 \right].\end{aligned}$$ Over a broad range of starburst conditions – including the GCCMZ, NGC 253, and M82 – the ion-electron collision time is comparable to or longer than the advection time. The electrons and ions do not necessarily have the same temperature in these starbursts.
Collisions only matter for the dynamics on the largest scales, those bigger than the mean free path. The mean free path is given by $\lambda \approx v / \nu$, where $v$ is the thermal speed of the particle. The mean free paths are similar for electrons and hydrogen ions. For the specific case of electrons, with a thermal speed $v_e \approx \sqrt{k T_e / m_e}$, the mean free path is : $$\begin{aligned}
\nonumber \lambda_{ee} & = 32\ \pc\ \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right)^{-1}\\
& \times \left[f_{\rm geom}^2 \zeta^{-1} \left(\frac{\ln \Lambda}{30}\right)^{-1} \left(\frac{\beta}{2}\right)^{-7/2} \left(\frac{\epsilon_{\rm therm}}{0.75}\right)^{5/2} \right].\end{aligned}$$ The GCCMZ’s scale height is actually not much bigger than the particle mean-free path – collisionless dynamics play a role in all scales in that region. In M82 and NGC 253, the superwind plasma is collisionless on scales smaller than a parsec.
Therefore, unlike molecular clouds or the warm ionized medium of the Milky Way, a starburst superwind is *not* a standard MHD fluid characterized by a single pressure and temperature. Instead, it is governed by kinetic processes, with plasma waves and magnetic fields shaping the particle distributions on parsec and smaller scales. Hydrodynamical models might leave out important physics, and quantities like the central temperature in the superwind are potentially inaccurate. For the remainder of this paper, I assume that the CC85 temperature and density (equations \[eqn:nC\] and \[eqn:SuperwindTc\]) are correct, at least well enough to estimate the properties of turbulence in the superwind, but a kinetic treatment would be very useful.
Finally, what might be the signs of collisionless plasma? First, the electron and ion temperatures can be different. Fortunately, the different populations emit hard X-rays through distinct mechanisms. X-ray line emission comes from ions, while bremsstrahlung continuum is mostly emitted by the easily-deflected electrons. @Strickland07 actually found that the diffuse hard continuum and iron line emission of M82 appeared to be tracing plasma of different properties. Non-equilibrium between electrons and ions may account for this. Second, as kinetic effects involving magnetic fields regulate the plasma, the magnetic field might be expected to introduce a preferred direction and *anisotropy* to the wind plasma.
### The Reynolds number: Can true turbulence exist in hot winds? {#sec:Reynolds}
Given the low collisionality of wind plasma, can turbulence exist at all in the superwind? In these conditions, the Braginskii viscosity is very high, preventing hydrodynamical turbulence from appearing even if collisions are effective. The viscosity is dominated by ions moving along the magnetic field lines. This characteristic Braginskii viscosity is given by $$\eta_0^i = 5.11 n_i k_B T_i \nu_{\rm ii}^{-1},$$ which is typically $\sim 600\ \gram\ \sec^{-1}\ \cm^{-1}$ for the superwind phase. The Reynolds number of a flow in the superwind is then ${\cal R} = \rho v \ell / \eta_0^i$: $$\begin{aligned}
\nonumber {\cal R} & = 0.63 \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right) \left(\frac{v}{1000\ \kms}\right) \left(\frac{\ell}{50\ \pc}\right)\\
& \times \left[\zeta f_{\rm geom}^{-2} \left(\frac{\ln \Lambda}{30}\right) \left(\frac{\beta}{2}\right)^4 \left(\frac{\epsilon_{\rm therm}}{0.75}\right)^{-3}\right].\end{aligned}$$ For standard hydrodynamical flows, turbulence appears when ${\cal R} \ga 10^3 - 10^4$, implying starburst superwinds are laminar flows: while random motions may be present, the energy in them does not cascade down to small scales and dissipate as in true turbulence [see similar considerations for the Galactic coronal phase in @McIvor77]. Motions driven on yet smaller scales will have smaller velocities, and still smaller Reynolds numbers.
The problem with this argument is that collisionless processes can reduce the thermal ion mean free path and, with it, the viscosity. First, as discussed in § \[sec:CollisonlessWinds\], starburst regions themselves can be smaller than the collisional mean free path, which means that collisionless processes must set the actual ion mean free path. At most the effective kinematic viscosity is $c_s \ell_{\rm outer} / 3$, so that the effective Reynolds number is ${\cal R} \ga 3 v / c_S$ [@Subramanian06]. The Reynolds number of the wind is therefore unlikely to be less than $\sim 1$. While this lower bound is still not large enough to guarantee hydrodynamical turbulence, this is a conservative estimate that assumes that collisionless processes can deflect ions only on length-scales comparable to the starburst size itself.
A very similar situation occurs in galaxy clusters, which are filled with hot ($\sim 10^8\ \Kelv$) and rarefied gas, just like starburst winds. Even though the formal Reynolds number is $\la 1000$, turbulence appears to be present. But the effective Reynolds number in galaxy clusters is thought to be raised by collisionless processes such as plasma instabilities [e.g., @Schekochihin05; @Schekochihin06; @Lazarian06; @Subramanian06; @Brunetti07]. Similar processes may operate in starburst winds.
What is the outer scale of turbulence in the hot wind? {#sec:HotOuterScale}
------------------------------------------------------
Turbulence dissipates more slowly on large scales than on small scales. The final energy density of the turbulence, and its strength at each scale in the turbulent cascade, therefore depends on the scale most of the energy is injected. In the Milky Way, this outer scale $\ell_{\rm outer}$ of the turbulence seems to depend on the location and ISM phase, with values ranging from a few parsecs for ionized gas in the Galactic spiral arms [@Haverkorn08] to about a hundred parsecs [@Haverkorn08; @Chepurnov10]. However, such large scales are not attainable in starbursts, simply because that is larger than the physical size of the starburst itself. Instead, the largest possible scale of the turbulence is the scale height of the starburst, $h \approx 50\ \pc$. In the GCCMZ, there is evidence that turbulence is injected on fairly large scales, $\ga 30\ \pc$ [@Shetty12], which would be consistent with $\ell_{\rm outer} \approx h$.
For a lower bound, the outer scale of turbulence is plausibly the scale on which individual supernova remnants fade into the background ISM. One specific possibility is that $\ell_{\rm outer}$ is the size of a supernova remnant whose energy density is equal to the pressure of the external medium: $\ell_1 \equiv [3 E_{\rm SN} / (4 \pi P_{\rm ext})]^{1/3}$. This is consistent with the turbulent outer scale in the Milky Way: for an average supernova kinetic energy of $10^{51}\ \erg$ and Milky Way-like pressures, the outer scale is $\sim 180\ \pc\ [(P_{\rm ext}/ k_B) / (10^4\ \Kelv\ \cm^{-3})]^{1/3}$. Assuming the external pressure is the thermal energy density in the hot superwind, the outer scale is then found as $$\ell_1 = \left[\frac{3 E_{\rm SN}}{4 \pi \times (3/2) n k T}\right]^{1/3} = \left[\frac{E_{\rm SN}}{2 \pi n K T}\right]^{1/3}.$$
Another possible outer scale is that it is the radius $\ell_2$ of a supernova remnant whose shock speed $v$ is equal to the velocity dispersion of the surrounding ISM velocity field. At this point, the supernova remnant ceases to be a distinct region in the ISM. I calculate this scale by supposing that the supernova remnant is in the non-radiating Sedov phase, since the hot ISM cools inefficiently. Then the kinetic energy of the supernova remnant is conserved with time, and assuming a uniform density, $v = \sqrt{3 E_{\rm SN} / (2 \pi r^3 \rho)}$. The outer scale is then $$\ell_2 = \left[\frac{3 E_{\rm SN}}{2 \pi \sigma^2 \rho}\right]^{1/3}.$$ Note this scenario is the same as supposing that the energy density of the supernova is equal to the external turbulent pressure.
Finally, most star-formation in starbursts occurs in large stellar clusters, some with masses far exceeding $10^6\ \Msun$ [@OConnell95; @McCrady07; @Kruijssen12]. The combined energy input from the stars and supernovae in these clusters create large superbubbles. The superbubbles also drive the ISM, but can do so on larger scales than individual supernovae [e.g., @Castor75; @Silich07]. If these superbubbles dominate energy injection, the outer scale of turbulence may plausibly be as large as $h$. Turbulent driving can even occur over a wide range of scales; this is thought to be the case in the Milky Way [@Norman96; @Joung06].
The energetics of superwind turbulence {#sec:TurbDensity}
--------------------------------------
The turbulent energy injection rate into the superwind is $\epsilon_{\rm turb} \dot{E}_{\rm mech}$, where $\epsilon_{\rm turb}$ is the efficiency of conversion of mechanical power into turbulence. If the turbulence survives for a timescale $t_{\rm turb}$, the turbulent energy density is $$\label{eqn:UTurbBasic}
U_{\rm turb} = \frac{\epsilon_{\rm turb} \dot{E}_{\rm mech} t_{\rm turb}}{V}$$ where $V = 2 \pi R^2 h$ is the volume of the starburst.
If the turbulence has a characteristic speed $\sigma$, it is expected to dissipate into heat over one crossing time, $\ell_{\rm outer} / \sigma$, where $\ell_{\rm outer}$ is the outer scale [e.g., @Tennekes72]. We can solve for $\sigma$ by using $U_{\rm turb} = \rho \sigma^2 / 2$ and equation \[eqn:UTurbBasic\]: $$\label{eqn:sigmaAnalytic}
\sigma = \left[\frac{\epsilon_{\rm turb} \dot{E}_{\rm mech} \ell_{\rm outer}}{\rho_c \pi R^2 h}\right]^{1/3} = \left[\frac{\epsilon_{\rm turb} \dot{E}_{\rm mech}}{n_c \mu m_H (\pi R^2)} \left(\frac{\ell_{\rm outer}}{h}\right)\right]^{1/3}.$$ Using the energy injection rate in eqn. \[eqn:EDot\] and superwind density in eqn. \[eqn:nC\], I find a turbulent velocity dispersion of $$\begin{aligned}
\label{eqn:sigmaTurb}
\nonumber \sigma & = 1020\ \kms\ \left(\frac{\ell_{\rm outer}}{h}\right)^{1/3}\\
& \times \left[\epsilon_{\rm turb} \zeta^{-1} f_{\rm geom}^{2} \left(\frac{\epsilon_{\rm therm}}{0.75}\right)^{1/2} \left(\frac{\beta}{2}\right)^{-3/2}\right]^{1/3}\end{aligned}$$ Since both $\dot{E}_{\rm mech} / (\pi R^2)$ and $n_c$ increase linearly with $\Sigma_{\rm SFR}$, the star-formation rate cancels out. If the $\ell_{\rm outer} \approx h$, the turbulence is weakly supersonic (${\cal M} \approx 1.1$).
The turbulent energy density is $$\label{eqn:Uturb}
U_{\rm turb} = \frac{\rho_c^{1/3}}{2} \left[\frac{\epsilon_{\rm turb} \dot{E}_{\rm mech}}{\pi R^2} \left(\frac{\ell_{\rm outer}}{h}\right)\right]^{2/3}$$ Note that since $\rho \propto \Sigma_{\rm SFR}$, we have $U_{\rm turb} \propto \Sigma_{\rm SFR}$. The energy dissipated in the turbulence goes into heating the superwind. Indeed, turbulent dissipation may be *the* way that supernova mechanical energy is thermalized.
Smaller outer scales lead to quicker dissipation and smaller turbulent energy densities. If the outer scale is set by the time it takes a SNR overpressure to fade into the ISM (Section \[sec:HotOuterScale\]), then: $$\begin{aligned}
\nonumber \ell_1 & = 35\ \pc \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right)^{-1/3}\\
& \times \left[E_{51} f_{\rm geom}^{2} \zeta^{-1} \left(\frac{\beta}{2}\right)^{-1/2} \left(\frac{\epsilon_{\rm therm}}{0.75}\right)^{-1/2} \right]^{1/3}\end{aligned}$$ The outer scale decreases from several tens of parsecs in the Galactic Center to only a few parsecs in extreme starbursts like Arp 220 (Table \[table:BHotPredictions\]). If the outer scale is $\ell_1$, then the turbulent velocity dispersion is only $$\begin{aligned}
\nonumber \sigma_1 & = 900\ \kms\ \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right)^{-1/9}\\
& \times \left[h_{50}^{-3} E_{51} \epsilon_{\rm turb}^{3} f_{\rm geom}^{8} \zeta^{-4} \left(\frac{\epsilon_{\rm therm}}{0.75}\right) \left(\frac{\beta}{2}\right)^{-5} \right]^{1/9}.\end{aligned}$$
If $\ell_{\rm outer}$ is set by the scale where SNRs merge with the ISM velocity field, $$\begin{aligned}
\nonumber \ell_2 & = 41\ \pc\ \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right)^{-3/11} \\
& \times \left[h_{50}^{2} E_{51}^{3} f_{\rm geom}^{2} \epsilon_{\rm turb}^{-2} \zeta^{-1}\ \left(\frac{\beta}{2}\right)^{-3/2} \left(\frac{\epsilon_{\rm therm}}{0.75}\right)^{1/2} \right]^{1/11}.\end{aligned}$$ I then find $$\begin{aligned}
\nonumber \sigma_2 & = 950\ \kms\ \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right)^{-1/11}\\
& \times \left[h_{50}^{-3} E_{51} \epsilon_{\rm turb}^{3} f_{\rm geom}^{8} \zeta^{-4} \left(\frac{\epsilon_{\rm therm}}{0.75}\right)^{2} \left(\frac{\beta}{2}\right)^{-6}\right]^{1/11}.\end{aligned}$$ Note that these turbulent velocities are slightly subsonic, and in fact decrease to only a few hundred kilometers per second for Arp 220 (Table \[table:BHotPredictions\]).
[llcccc]{} $f_{\rm geom}$ & & 1.41 & 1.29 & 1.15 & 1.41\
$n_e$ & $\cm^{-3}$ & 0.015 & 0.33 & 0.35 & 21\
$P_{\rm wind} / k_B$ & $\Kelv\ \cm^{-3}$ & $1.1 \times 10^6$ & $2.4 \times 10^7$ & $2.5 \times 10^7$ & $1.5 \times 10^9$\
$\nu_{ee}^{-1}$ & $\kyr$ & 1.2 & 0.052 & 0.050 & 0.00083\
$\nu_{ii}^{-1}$ & $\kyr$ & 59 & 2.6 & 2.5 & 0.041\
$\nu_{ie}^{-1}$ & $\kyr$ & 6200 & 270 & 260 & 4.3\
$\lambda_{ee}$ & $\pc$ & 29 & 1.3 & 1.0 & 0.020\
$Q_{\rm SNR}^{\rm turb}$ & & 0.028 & 0.29 & 0.11 & 0.013\
$Q_{\rm SNR}^{\rm CC85}$ & & 0.86 & 0.31 & 0.23 & 0.060\
$Q_{\rm SNR}^{\rm hydro}$ & & $5.9 \times 10^{-5}$ & 0.097 & 0.020 & 0.00012\
$Q_{\rm SNR}^{\rm hot}$ & & 1.2 & 0.34 & 0.27 & 0.10\
$\sigma$ & $\kms$ & 1300 & 1200 & 1100 & 1300\
$U_{\rm turb} / k_B$ & $\Kelv\ \cm^{-3}$ & $1.7 \times 10^6$ & $3.4 \times 10^7$ & $3.0 \times 10^7$ & $2.4 \times 10^9$\
$B_{\rm turb}$ & $\muGauss$ & 76 & 340 & 330 & 2900\
$\ell_{\rm in}$ & $\km$ & 1800 & 370 & 360 & 46\
$\ell_{\rm outer}$ & pc & 33 & 12 & 12 & 3.0\
$U_{\rm turb} / k_B$ & $\Kelv\ \cm^{-3}$ & $1.3 \times 10^6$ & $1.3 \times 10^7$ & $1.2 \times 10^7$ & $0.36 \times 10^9$\
$\sigma$ & $\kms$ & 1100 & 740 & 690 & 500\
$B_{\rm turb}$ & $\muGauss$ & 68 & 210 & 200 & 1100\
$\ell_{\rm in}$ & $\km$ & 1800 & 460 & 480 & 86\
$\ell_{\rm outer}$ & pc & 35 & 16 & 16 & 4.9\
$\sigma$ & $\kms$ & 1100 & 820 & 770 & 590\
$U_{\rm turb} / k_B$ & $\Kelv\ \cm^{-3}$ & $1.3 \times 10^6$ & $1.6 \times 10^7$ & $1.4 \times 10^7$ & $0.51 \times 10^9$\
$B_{\rm turb}$ & $\muGauss$ & 68 & 230 & 220 & 1300\
$\ell_{\rm in}$ & $\km$ & 1800 & 420 & 440 & 73 \[table:BHotPredictions\]
It is interesting that the predicted outer scales in the GCCMZ are of order the same length as the mysterious radio filaments, whose origin remain unknown. Turbulent may naturally concentrate magnetic fields into ribbons (see section \[sec:Dynamoes\]), perhaps providing an explanation for the filaments’ existence. This was also the conclusion of @Boldyrev06, who first suggested the filaments are turbulent structures; although they estimated $\ell_{\rm outer}$ as the radius of a SNR when it first runs into another SNR’s shell.
The Magnetic Field Strength {#sec:BHot}
---------------------------
If a fluctuation dynamo operates in the starburst wind (Section \[sec:Dynamoes\]), then $B = \sqrt{8 \pi \epsilon_B U_{\rm turb}}$: $$B_{\rm turb} = \sqrt{4 \pi \epsilon_B} \rho_c^{1/6} \left[\frac{\epsilon_{\rm turb} \dot{E}}{\pi R^2} \left(\frac{\ell_{\rm outer}}{h}\right) \right]^{1/3}.$$
Substituting in the numerical values, I find $$\begin{aligned}
\label{eqn:BAsSigmaSFR}
\nonumber B_{\rm turb} & = 57\ \muGauss \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right)^{1/2} \left(\frac{\ell_{\rm outer}}{h}\right)^{1/3}\\
& \times \sqrt{\epsilon_B} \left[\epsilon_{\rm turb}^{2} \zeta f_{\rm geom}^{-2} \left(\frac{\epsilon_{\rm therm}}{0.75}\right)^{-1/2} \left(\frac{\beta}{2}\right)^{3/2} \right]^{1/6}.\end{aligned}$$
A smaller outer scale leads to smaller turbulent energy densities, and thus smaller magnetic fields. The magnetic field strength for an outer scale of $\ell_1$ is $$\begin{aligned}
\nonumber B_1 & = 51\ \muGauss\ \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right)^{7/18}\\
& \times \sqrt{\epsilon_B} \left[h_{50}^{-3} E_{51} \epsilon_{\rm turb}^{3} f_{\rm geom}^{-1} \zeta^{1/2} \left(\frac{\epsilon_{\rm therm}}{0.75}\right)^{-5/4} \left(\frac{\beta}{2}\right)^{7/4} \right]^{1/9}\end{aligned}$$ and for an outer scale of $\ell_2$ is $$\begin{aligned}
\nonumber B_2 & = 54\ \muGauss\ \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right)^{9/22}\\
& \times \sqrt{\epsilon_B} \left[h_{50}^{-3} E_{51} \epsilon_{\rm turb}^{3} \zeta^{3/2} f_{\rm geom}^{-3} \left(\frac{\epsilon_{\rm therm}}{0.75}\right)^{-3/4} \left(\frac{\beta}{2}\right)^{9/4} \right]^{1/11}.\end{aligned}$$
{width="8cm"}
### Comparison with estimated magnetic field strengths {#sec:ModelBComparison}
Using one-zone models of the multiwavelength (particularly radio and gamma-ray) emission of some nearby starburst regions, many papers have constrained the magnetic field strength of these individual starbursts under certain assumptions. Examples of starbursts modelled this way include the Galactic Center Central Molecular Zone [@Crocker11-Wild; @Crocker12; @Lacki13-XRay], NGC 253 [@Paglione96; @Domingo05; @Rephaeli10; @Paglione12], M82 [@Persic08; @deCeaDelPozo09-M82; @Paglione12; @YoastHull13], and the starburst nuclei of Arp 220 [@Torres04; @Lacki13-XRay]. These models often use a gas density that is equal to the average gas density and assume that CRs are accelerated with a similar efficiency as in the Milky Way. Upper limits on the amount of Inverse Compton and bremsstrahlung emission constrain the density of synchrotron-emitting CR $e^{\pm}$, and therefore set a lower limit on $B$ [e.g., @Condon91; @Thompson06; @Crocker10]. If the density is known, the amount of synchrotron emission sets a upper limit on $B$; if it is too high, too much power is radiated through the synchrotron mechanism, whereas with a lower $B$ most of the power goes into bremsstrahlung, ionization, and Inverse Compton losses. But beware: it is possible that CRs experience a non-average density in the inhomogeneous ISM, leading to different allowed $B$ values [@Paglione12].
It is worthwhile to compare my calculations of $B_{\rm turb}$ (Table \[table:BHotPredictions\]) to these model-constrained magnetic field strengths (Table \[table:BasicProperties\]). The calculated magnetic field strengths are within the ranges found by modeling. Thus, supernova-driven turbulence can plausibly generate pervasive magnetic fields in the starburst region that are responsible for the synchrotron radio emission.
Turbulence and magnetic fields in cold and warm gas {#sec:ColdTurbulence}
===================================================
I showed in Section \[sec:VolumeFillingPhase\] that while the hot superwind should fill much of M82’s volume, the high pressures in ULIRGs confines the hot gas too much for it to fill those starbursts. Instead, most of the volume is in turbulent molecular gas. Furthermore, even in hot starbursts, typical CRs probably enter the molecular medium at some point in their lives (see the discussion in section \[sec:HotWindMotivation\]).
But what are the magnetic fields in the molecular medium? If the molecular gas has some free charges, the turbulent dynamo pushes the magnetic energy density close to equipartition with the turbulence [@Stone98], is implied by observations of Galactic molecular clouds that are near equipartition [@Crutcher99]. A turbulent dynamo naturally explains this fact [@Balsara04]. Typical turbulent velocity dispersions in starburst molecular gas are $25 - 150\ \kms$, with the lower end representative of motions in the GCCMZ [@Shetty12] and the high end found in $z \approx 2$ galaxies [@Green10]. This implies a mean turbulent energy density of $$\frac{P}{k_B} = \frac{{\ensuremath{\langle \rho \rangle}} \sigma^2}{2 k_B} = 6.1 \times 10^8\ \Kelv\ \cm^{-3} \left(\frac{{\ensuremath{\langle n_H \rangle}}}{1000\ \cm^{-3}}\right) \left(\frac{\sigma}{100\ \kms}\right)^2,$$ and equipartition magnetic field strength: $$B_{\rm turb} = 1.45\ \mGauss\ \epsilon_B \left(\frac{{\ensuremath{\langle n_H \rangle}}}{1000\ \cm^{-3}}\right)^{1/2} \left(\frac{\sigma}{100\ \kms}\right).$$ Essentially, this is the hypothesis of @Groves03 to explain the existence of the far-infrared–radio correlation, except that $\sigma$ is $\sim 100\ \kms$ instead of $10\ \kms$.
As in Section \[sec:HotTurbulence\], I assume that supernovae drive turbulence in the cold and warm gas (see section \[sec:TurbulenceSources\]). What complicates the calculation is that molecular gas is neither homogeneous nor necessarily volume-filling. There are at least three relevant densities for the problem. The first is the density averaged over the volume of the entire starburst, ${\ensuremath{\langle \rho \rangle}}_{\rm SB} = M(H_2) / V_{\rm SB}$. It is the density that is most closely related to the mass, which sets the *total* kinetic energy of the turbulence and is what is actually measured. The second relevant density is the density averaged only over the volume of the molecular clouds, ${\ensuremath{\langle \rho \rangle}}_{\rm MC} = M(H_2) / (V_{\rm SB} f_{\rm fill})$, where $f_{\rm fill}$ is the fraction of the starburst volume occupied by molecular clouds. That is what sets the typical turbulent *energy density* in the molecular clouds. Finally, there is $\rho_{\rm ext}$, which is the density *exterior* to a SNR expanding into a molecular cloud. It sets the maximum size of the SNR, and presumably the outer scale of turbulence. Roughly, $\rho_{\rm ext}$ is the median density in the volume of molecular clouds, and is somewhere in the range ${\ensuremath{\langle \rho \rangle}}_{\rm MC} / \sqrt{1 + b^2 {\cal M}^2}$ to ${\ensuremath{\langle \rho \rangle}}_{\rm MC}$ for isothermal turbulence [@Hopkins13-rhoDist].
As in the superwind phase, the turbulent kinetic energy is roughly the power injected multiplied by the eddy-crossing time. For the molecular gas, we have $M(H_2) \sigma^2/2 = \dot{E}_{\rm mol} \ell_{\rm outer} / \sigma$, where $\dot{E}_{\rm mol}$ is the power input into the molecular phase only. We can solve for $\sigma$, getting $$\sigma = \left[\frac{\dot{E}_{\rm mech} \epsilon_{\rm turb} \epsilon_{\rm mol}}{A {\ensuremath{\langle \rho \rangle}}_{\rm SB}} \left(\frac{\ell_{\rm outer}}{h}\right)\right]^{1/3}.$$ The factor $\epsilon_{\rm mol} \equiv \dot{E}_{\rm mol} / \dot{E}_{\rm mech}$ represents the fraction of mechanical luminosity that is available for the cold gas because the sources are within that phase. For example, if the molecular phase occupies a small fraction of the starburst volume, few of the supernovae go off in it; most do not stir up the molecular gas. If sources are located randomly in the starburst, $\epsilon_{\rm mol} \approx f_{\rm fill}$; if they are instead clustered inside molecular clouds, $\epsilon_{\rm mol} \approx 1$.
If the outer scale of turbulence is some constant ratio of the starburst scale height, then the characteristic turbulent speed is $$\begin{aligned}
\nonumber \sigma & = 25\ \kms\ \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right)^{1/3} \left(\frac{{\ensuremath{\langle n \rangle}}_{\rm SB}}{1000\ \cm^{-3}}\right)^{-1/3}\\
& \times \left[\epsilon_{\rm turb} \epsilon_{\rm mol} \left(\frac{\ell_{\rm outer}}{h}\right)\right]^{1/3}.\end{aligned}$$ The equipartition magnetic field strength is $$\begin{aligned}
\nonumber B_{\rm turb} & = 360\ \muGauss\ \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right)^{1/3} \left(\frac{{\ensuremath{\langle n \rangle}}_{\rm SB}}{1000\ \cm^{-3}}\right)^{1/6} \\
& \times \sqrt{\frac{\epsilon_B}{f_{\rm fill}}}\left[\epsilon_{\rm turb} \epsilon_{\rm mol} \left(\frac{\ell_{\rm outer}}{h}\right)\right]^{1/3}.\end{aligned}$$ See Table \[table:BColdPredictions\] for these magnetic field strengths for the prototypical starbursts.
However, as with the superwind phase, turbulence can be much slower if $\ell_{\rm outer}$ is small. The outer scales can easily be tiny in starbursts, where the high pressures confine supernova remnants to small radii. Supernova remnants are highly radiative in dense molecular environments, unlike in the hot superwinds where the size evolution of the remnants could be understood by energy conservation. At their maximum radius, the supernova remnants will then be in the momentum-conserving phase [@Chevalier74; @McKee77]. To do a more realistic calculation of the properties of the turbulence, I therefore set $\ell_{\rm outer} \approx \ell_3 \equiv R_{\rm max}$. In addition, the surrounding ISM pressure is mostly from the turbulence itself, so I set $P_{\rm ext} = {\ensuremath{\langle \rho \rangle}}_{\rm MC} \sigma^2 / 2$. Under these conditions, the outer scale is $$\ell_3 = 4.8\ \pc\ \Psi_{\ell}\ \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right)^{-2/17} \left(\frac{{\ensuremath{\langle n \rangle}}_{\rm SB}}{1000\ \cm^{-3}}\right)^{-1/5},$$ where the nuisance parameter $\Psi_{\ell}$ contains the information on the efficiency of the turbulence driving, geometry, filling factor, and density contrasts: $$\Psi_{\ell} = \frac{E_{51}^{24/85} f_{\rm fill}^{27/85} h_{50}^{2/17}}{\epsilon_{\rm turb}^{2/17} \epsilon_{\rm mol}^{2/17} \Delta^{12/85}}$$ with $\Delta = \rho_{\rm ext} / {\ensuremath{\langle \rho \rangle}}_{\rm MC}$. The length scale is about an order of magnitude smaller than the scale height of the starburst, implying smaller turbulent speeds and magnetic fields by a factor of $\sim 2$. The turbulent speed is $$\sigma = 11\ \kms\ \Psi_{\sigma}\ \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right)^{5/17} \left(\frac{{\ensuremath{\langle n \rangle}}_{\rm SB}}{1000\ \cm^{-3}}\right)^{-2/5},$$ with $$\Psi_{\sigma} = \frac{E_{51}^{8/85} \epsilon_{\rm turb}^{5/17} \epsilon_{\rm mol}^{5/17} f_{\rm fill}^{9/85}}{h_{50}^{5/17} \Delta^{4/85}}.$$ Finally, the equipartition magnetic field strength is $$B_{\rm turb} = 170\ \muGauss\ \Psi_B\ \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right)^{5/17} \left(\frac{{\ensuremath{\langle n \rangle}}_{\rm SB}}{1000\ \cm^{-3}}\right)^{1/10},$$ with $$\Psi_B = \frac{E_{51}^{8/85} \epsilon_B^{1/2} \epsilon_{\rm turb}^{5/17} \epsilon_{\rm mol}^{5/17}}{f_{\rm fill}^{67/170} h_{50}^{5/17} \Delta^{4/85}}.$$
The big question is, what are the values of $\Psi_{\ell}$, $\Psi_{\sigma}$, and $\Psi_B$, and the factors that comprise them? It is generally thought that supernova remnants expanding in dense molecular media lose energy by radiation; according to the simulations of @Thornton98, only $\sim 10\%$ of the supernova mechanical energy is available for turbulence. While $f_{\rm fill}$ is probably near 1 for cold starbursts, it may be much smaller in hot starbursts. Presumably, $\epsilon_{\rm mol}$ is equal to $f_{\rm fill}$, but this is not necessarily the case if supernovae only go off in molecular clouds (although there is evidence against this for the weaker starbursts; Section \[sec:SNREvolution\]). Finally, in the Mach 10 – 100 turbulence of starburst molecular gas, $\Delta \approx 0.01$ – $1$ [e.g., @Krumholz05; @Hopkins13-rhoDist].
To answer these issues, I consider some “natural” values of $\Psi_{\ell}$, $\Psi_{\sigma}$, and $\Psi_B$. In all cases, I set $\epsilon_{\rm turb} = 0.1$. For the filling factor $f_{\rm fill}$, I use the somewhat arbitrary value of $25\%$ for the GCCMZ, NGC 253, and M82, but 100% for Arp 220’s nuclei. Likewise, the $\epsilon_{\rm mol}$ geometrical factor is set equal to $f_{\rm fill}$. Finally, I set $\Delta = 0.01$ for Arp 220’s nuclei and $\Delta = 0.03$ for the other starbursts. The resultant “natural” values for $\Psi_L$, $\Psi_{\sigma}$, and $\Psi_B$ are listed in Table \[table:BColdPredictions\], as well as the “corrected” outer scales, velocity dispersions, and magnetic field strengths. With these values, the turbulent driving is much weaker, and the available energy density must push more gas per unit volume (the smaller filling factors mean higher densities in those regions that *do* have molecular gas). The turbulent velocity dispersions are now smaller by a factor of $\sim 2$, or $\sim 4$ for Arp 220. Likewise, the turbulent pressures are a factor 4–10 times smaller than those inferred from observations. The resulting magnetic field strengths are lower, more in line with those inferred by modeling, and roughly equal to those in the turbulent superwind.
[llcccc]{} $\sigma_{\rm obs}$ & $\kms$ & 25 & $\sim 50$ & $\sim 50$ & 80\
$U_{\rm turb}$ & $\Kelv\ \cm^{-3}$ & $4.5 \times 10^6$ & $2.6 \times 10^7$ & $4.5 \times 10^7$ & $5.0 \times 10^9$\
$B_{\rm turb}$ & $\muGauss$ & 130 & 300 & 400 & 4200\
$\sigma$ & $\kms$ & 45 & 160 & 120 & 160\
$U_{\rm turb}$ & $\Kelv\ \cm^{-3}$ & $48 \times 10^6$ & $26 \times 10^7$ & $28 \times 10^7$ & $20 \times 10^9$\
$B_{\rm turb}$ & $\muGauss$ & 410 & 950 & 980 & 8200\
$\ell_{\rm outer}$ & pc & 5.2 & 4.4 & 4.0 & 1.1\
$\sigma$ & $\kms$ & 21 & 70 & 53 & 44\
$U_{\rm turb}$ & $\Kelv\ \cm^{-3}$ & $1.1 \times 10^7$ & $5.1 \times 10^7$ & $5.1 \times 10^7$ & $1.6 \times 10^9$\
$B_{\rm turb}$ & $\muGauss$ & 190 & 420 & 420 & 2300\
Natural $\Psi_{\ell}$ & & 1.6 & 1.6 & 1.6 & 2.5\
Natural $\Psi_{\sigma}$ & & 0.34 & 0.34 & 0.34 & 0.63\
Natural $\Psi_B$ & & 0.69 & 0.69 & 0.69 & 0.63\
$\ell_{\rm outer}$ & pc & 8.6 & 7.1 & 6.5 & 2.8\
$\sigma$ & $\kms$ & 7.3 & 24 & 18 & 28\
$U_{\rm turb}$ & $\Kelv\ \cm^{-3}$ & $5.0 \times 10^6$ & $2.4 \times 10^7$ & $2.4 \times 10^7$ & $0.61 \times 10^9$\
$B_{\rm turb}$ & $\muGauss$ & 130 & 290 & 290 & 1500 \[table:BColdPredictions\]
*Why aren’t the observed $\sigma$ this low?* – If the observed turbulent speeds are correct [c.f., @Ostriker11], there must be more turbulent power than provided by highly radiative supernova remnants. First, there could be other sources of turbulence within molecular clouds. As discussed in § \[sec:TurbulenceSources\], protostellar outflows, radiation pressure, and starburst disk instabilities may all help drive turbulence in cold starbursts.
In addition, many of the young stars in the nuclei are likely in clusters [@Kruijssen12]. By aggregating the energy input into a few, more powerful sources, the outer scale of the clusters’ superbubbles will increase as should the turbulent energy density.
On a broader level, turbulent power does not necessarily have to be directly injected into the molecular clouds. Turbulent power on large scales can actually cascade down into molecular clouds on smaller scales. Supernovae going off in the superwind near the molecular cloud could effectively stir the medium on 10 pc scales with minimal radiative losses. The mechanical pushing and pulling and stirring of the superwind then drives turbulence on smaller scales in molecular clouds stuck within the eddies, bypassing the need for small-scale stirring by individual supernova remnants. The driving of molecular cloud turbulence by external processes on larger scales likely happens in the Milky Way [@Brunt03; @Brunt09], as demonstrated by Larsons’ relation where the fastest turbulent occurs on the largest scales [@Larson81; @Miesch94; @Heyer04]. Indeed, Larsons’ relation holds in the GCCMZ starburst too [@Shetty12].
*Why aren’t the calculated $\sigma$ and $B$ much lower?* – Yet what is just as striking is that the order of magnitude variation in $f_{\rm fill}$ and $\epsilon_{\rm turb}$ has a relatively small effect on $\ell_{\rm outer}$, $\sigma$, and $B$. To take this even further, let $\epsilon_{\rm turb}$ vary over the entire range of 0.01 to 1.0, and independently vary $f_{\rm fill} = \epsilon_{\rm mol}$ over 0.1 to 1, and let $\Delta$ range from 0.01 to 1 too. Even then, $\Phi_{\ell}$ varies at most by a factor of 5 (from 0.72 to 3.29), $\Phi_{\sigma}$ varies by at most a factor of 12 (from 0.10 to 1.24), and $\Phi_B$ varies at most a factor of 6 (from 0.26 to 1.6). The turbulent energy density is only uncertain by 1.5 orders of magnitude.
The weak dependencies of the nuisance parameters on these basic factors results from the physical factors having several counteracting effects. For example, suppose we lowered $\epsilon_{\rm turb}$. The main effect is to lessen the amount of turbulence, which overall pushes $\sigma$ down. But with slower turbulent flows, the eddy-crossing time grows. And, since the turbulent pressure is lower, SNRs are not confined to such small radii, the outer scale becomes longer, and the eddy-crossing time grows further. Hence, $\sigma$ decreases, but quite slowly, as $\epsilon_{\rm turb}$ decreases. Similar considerations apply for the filling factor.
*How does $B$ compare between the phases?* – The great uncertainty in relevant quantities make it impossible to exactly predict $\ell_{\rm outer}$, $B$, or $\sigma$ from first principles. But it is clear that the magnetic fields in the molecular material are at least as strong as those in the hot superwind. If we push the turbulence to maximal values, with $\ell_{\rm outer} = h$ and $\epsilon_{\rm turb} = 1$, then $B$ may be as high as 3 to 7 times higher in the cold molecular gas than in the hot superwind. However, those high levels of turbulence would manifest in turbulent speeds ($> 100\ \kms$; Table \[table:BColdPredictions\]) that are ruled out by observations. In a more likely scenario with $\ell_{\rm outer} \approx R_{\rm max}$, the magnetic field strengths are roughly equal or somewhat greater than those in the hot superwind, but generally within a factor of $\sim 2$.
As a result, the value of $B$ appears to be roughly constant as we move from one phase to the next in the starburst ISM. This is good news for models of synchrotron emission from starburst galaxies. Although the magnetic fields are probably stronger in molecular clouds, there is not orders of magnitude of variation between phases. Probably the more important factor is the intermittent nature of turbulence, which in and of itself causes fluctuations of $B$ from one place to another.
Equipartition in starbursts revisited {#sec:Equipartition}
=====================================
On some level, pressure balance between the phases must occur; otherwise, the overpressured phases will crush the others until they are in pressure balance. In this sense, it’s not a surprise that rough equipartition holds between the phases in starbursts (Table \[table:Pressures\]). Of course, I also assume equipartition between turbulence and magnetic fields from the start.
But why are most of the “natural” values of the various energy densities in starbursts so close to each other, *even when considered independently of each other*? The CC85 wind solution does not include any squeezing by an overpressured molecular medium; it is as if the molecular medium is not there at all – yet the theoretical pressure is in rough equipartition with the molecular turbulence. Likewise, my calculations of the turbulent molecular pressure did not depend on the pressure of the superwind pushing it into equipartition. Furthermore, the pressure balance argument only applies to the *total* pressure. There is no a priori reason the thermal and turbulent pressures have to be anywhere near each other, and they are not in the molecular gas. But they are in the superwind phase, and they are within an order of magnitude of each other in the H II regions. Why is radiation in equipartition with the gas even in weak starbursts like the GCCMZ that are transparent to infrared photons? And why is the CR energy density an order of magnitude below the turbulent pressure – except for the ULIRGs, where the ratio plummets? It’s as if the phases *started out* in pressure balance even before any mechanical coupling between them.
I explore these questions in this section. Ultimately, the energy density is a combination of a power density and a timescale. In several cases (turbulence, magnetic fields, CRs, superwind thermal pressure), the power source is the same – supernovae mechanical energy, and the timescales – roughly the superwind sound crossing time – are near each other, explaining equipartition. In other cases (radiation and H II thermal pressure), there is a fortuitous canceling between a large power and a short timescale.
Why are the turbulent pressures between the phases nearly equal?
----------------------------------------------------------------
From equation \[eqn:Uturb\], the turbulent energy density can be written as $U_{\rm turb} = [2 \rho \dot{\varepsilon}_{\rm turb}^2 \ell_{\rm outer}^2]^{1/3}$. Therefore the ratio of the turbulent energy density in two phases is equal to $$\frac{U_a}{U_b} = \left(\frac{\dot{\varepsilon}_a}{\dot{\varepsilon}_b}\right)^{2/3} \left(\frac{\rho_a}{\rho_b}\right)^{1/3} \left(\frac{\ell_a}{\ell_b}\right)^{2/3}.$$
It is clear that the turbulent energy density depends weakly on density, but relatively strongly on energy input. Thus, even though the density in starburst molecular clouds is 100 to $10^4$ (depending on filling factor) times higher than that in the superwind, that alone would only boost the turbulent energy density by an order of magnitude. Put another way, in the molecular clouds, the supernovae have to push more mass. On the one hand, the turbulence is slower but on the other the eddy-crossing time is longer, and these effects partly cancel out.
But in molecular clouds and H II regions, the supernova remnants are highly radiative, injecting only 10% of the input mechanical energy [@Thornton98], reducing the molecular turbulent energy density by a factor $\sim 5$. The outer scale lengths are fairly near each other, about twice as long in the superwind as in the molecular ISM and H II regions, if we consider the maximum radius of a SNR before fading into the ISM (see Tables \[table:BHotPredictions\] and \[table:BColdPredictions\]). Therefore, we have $U_{\rm cold}/U_{\rm hot} \approx 10 \times (1/5) \times (1/2) \approx 1$.
Why are the thermal and turbulent pressures nearly equal in the superwind?
--------------------------------------------------------------------------
The superwind is heated by supernovae in its starburst region. The plasma in the superwind is hot enough that it is not affected by the starburst’s gravity, and therefore explodes out of the starburst in roughly a dynamical (sound-crossing) time. Thus, the superwind has a thermostat that regulates its internal thermal pressure, one directly related to the flow properties. If the wind is too cool, it will stay in the starburst longer and heat up more; if it heats too much, it will expand quickly and prevent further heating.
Note the similarity with the turbulent energy density: the thermal energy density is a power density times a flow-crossing time. The only substantial difference is that here we want the time it takes the wind to cross the starburst rather than just the outer scale. We can write $$U_{\rm therm} = \frac{\dot{\varepsilon} h \psi}{c_S} = \frac{3}{2\gamma} \rho c_S^2,$$ The $\psi$ factor corrects the thermal energy density to the CC85 solution and is equal to $3.08 \epsilon_{\rm therm} \mu \zeta / f_{\rm geom}^2 \approx 1$. The solution for the sound speed is $c_S = [2\gamma \dot{\varepsilon}_{\rm mech} h \psi / (3\rho)]^{1/3}$. Then we have $$U_{\rm therm} = \left(\frac{3}{2\gamma}\right)^{1/3} (\dot{\varepsilon}_{\rm mech} h \psi)^{2/3} \rho^{1/3}.$$ Once we compare to the turbulent energy density, it is clear that the thermal and turbulent energy densities must be in near equipartition in superwinds. Their ratio is $$\frac{U_{\rm therm}}{U_{\rm turb}} = \left(\frac{12}{\gamma}\right)^{1/3} \left(\frac{h}{\ell_{\rm outer}}\right)^{2/3} \left(\frac{\psi}{\epsilon_{\rm turb}}\right)^{2/3}.$$ Since $\ell_{\rm outer}/h \approx 0.1 - 1$, we have $U_{\rm therm} / U_{\rm turb} \approx 0.3 - 2$.
By contrast, the starburst’s molecular medium has a much different thermostat mechanism than the superwind. Molecular gas cools by radiating line emission and by collisions with dust grains. Even in regions where it has the same power source of supernova-driven turbulence, the time it retains that heat is much smaller than the time it retains turbulent kinetic energy. Therefore, the turbulence in the molecular gas is extremely supersonic. But in the $10^7 - 10^8\ \Kelv$ superwind, radiative cooling is completely negligible so that the dynamical thermostat operates.[^3]
Why is the thermal pressure in H II regions within an order of magnitude of the turbulent pressure?
---------------------------------------------------------------------------------------------------
The thermal pressures measured for H II regions within starburst regions (Table \[table:Pressures\]) tends to be a few times lower than the turbulent pressure I predict. But why should this be true, when it is not true in the other dense phase, the molecular gas? Neither the heating nor cooling of H II regions has anything to do with supernova-driven turbulence. Put another way, why is the natural turbulent speed of a few tens $\kms$ for dense starburst gas just a bit higher than the sound speed of $\sim 15\ \kms$?
H II regions are heated by the photoelectric absorption of ionizing radiation of their OB stars. We can estimate the heating rate per unit volume by dividing the ionizing luminosity by the Strömgren volume: $L_{\star} n_{\rm H\,II}^2 \alpha_B / Q_{\rm ion}^{\star}$, where $\alpha_B$ is the recombination coefficient, $L_{\star}$ is the ionizing power of the OB stars, and $Q_{\rm ion}^{\star}$ is the rate ionizing photons are emitted by those stars [e.g., @Draine11]. Note the ratio $L_{\star} / Q_{\rm ion}^{\star}$ equals ${\ensuremath{\langle h \nu_{\star} \rangle}}$, the mean energy of an ionizing photon. I find that the average volumetric heating rate in H II regions with a temperature $10^4\ \Kelv$ is $$\begin{aligned}
\nonumber \dot{\varepsilon}_{\rm therm} & = 1.2 \times 10^{-17}\ \erg\ \cm^{-3}\ \sec^{-1} \left(\frac{n_{\rm H\,II}}{1000\ \cm^{-3}}\right)^2 \\
& \times \left(\frac{{\ensuremath{\langle h \nu_{\star} \rangle}}}{30\ \eV}\right)\end{aligned}$$
The thermal energy within the H II regions is retained for one cooling time, $t_{\rm cool} = 3 n_{\rm H\,II} k T / \Lambda$, before being radiated away as line and free-free emission. Given that the cooling coefficient has an approximate value $2 \times 10^{-24}\ \erg\ \cm^{-3}\ \sec^{-1} \times n_e n_H$ for Solar metallicity [@Osterbrock89], the cooling time is $$t_{\rm cool} = 33\ \yr\ \left(\frac{n_{\rm H\,II}}{1000\ \cm^{-3}}\right)^{-1}$$ for a totally ionized plasma at $10^4\ \Kelv$.
By contrast, the turbulent volumetric power from supernova mechanical energy is much smaller: $$\begin{aligned}
\nonumber \dot{\varepsilon}_{\rm turb} & = 1.1 \times 10^{-20} \erg\ \cm^{-3}\ \sec^{-1}\ \epsilon_{\rm turb} \left(\frac{{\ensuremath{\langle n_H \rangle}}_{\rm SB}}{1000\ \cm^{-3}}\right)\\
& \times \left(\frac{\tau_{\rm gas}}{20\ \Myr}\right)^{-1}.\end{aligned}$$ At the same time, the flow crossing time is much longer than the cooling time: $$t_{\rm eddy} = 200\ \kyr \left(\frac{\ell_{\rm outer}}{5\ \pc}\right) \left(\frac{\sigma}{25\ \kms}\right)^{-1}.$$
We can now see why equipartition holds: although the heating rate is over ten thousand times the turbulent power, the heat is retained for only a ten-thousandth of the time. The ratio of thermal to turbulent energy density in H II regions is $$\begin{aligned}
\nonumber \frac{U_{\rm therm}}{U_{\rm turb}} & = \frac{\dot{\varepsilon}_{\rm therm}}{\dot{\varepsilon}_{\rm turb}} \frac{t_{\rm cool}}{t_{\rm eddy}}\\
\nonumber & \approx 0.18 \epsilon_{\rm turb} \left(\frac{n_{\rm H\,II}}{{\ensuremath{\langle n_H \rangle}}_{\rm SB}}\right) \left(\frac{\tau_{\rm gas}}{20\ \Myr}\right) \left(\frac{{\ensuremath{\langle h\nu_{\star} \rangle}}}{30\ \eV}\right)\\
& \times \left(\frac{\ell_{\rm outer}}{5\ \pc}\right)^{-1} \left(\frac{\sigma}{25\ \kms}\right).
\label{eqn:HIITurbThermRatio}\end{aligned}$$ Thus, thermal H II region pressure is naturally $\sim 5$ times smaller in H II regions of average density, at least in nuclear starburst regions where $\tau_{\rm gas}$ and $\ell_{\rm outer}$ are small.
This coincidence may not apply to high redshift main sequence galaxies or other environments with long $\tau_{\rm gas}$. On the one hand, $\tau_{\rm gas}$ is much longer in these regions, $\sim 0.5$ Gyr. That means the galaxies have lower SFR for their gas masses, and hence, less turbulent power per unit mass. Turbulence might naturally be weaker. There might be a sign of this in Table \[table:Pressures\]: in the GCCMZ where $\tau_{\rm gas} \approx 400\ \Myr$, the H II region thermal energy densities are higher than the other energy densities (equation \[eqn:HIITurbThermRatio\] implies a ratio of 3.6). If the thermal energy density is much greater than the surrounding ISM pressure, the H II regions will expand until they are in equilibrium: thus $n_H / {\ensuremath{\langle n_H \rangle}}_{\rm SB}$ would be lower. On the other hand, in the lower gas densities of main sequence galaxies, the turbulent outer scale might be larger, increasing $t_{\rm eddy}$.
Why is the radiation pressure near equipartition with turbulence?
-----------------------------------------------------------------
The near equipartition between radiation and turbulent energy densities in starburst regions is another coincidence. The amount of power released as light by a starburst, $L_{\rm bol} \approx 2.15 \times 10^{43} \erg\ \sec^{-1} \times ({\rm SFR} / \Msun\ \yr^{-1})$ [@Kennicutt98], is much greater than the mechanical power. On the other hand, the light escapes very quickly, in a light crossing time if the starburst is not optically thick: $t_{\rm rad} \approx \tau_{\rm IR} h / c$, where the IR optical depth $\tau_{\rm IR}$ is typically a few or less [@Krumholz12]. These two factors cancel each other out.
From the timescale argument, we again have $$\begin{aligned}
\nonumber \frac{U_{\rm turb}}{U_{\rm rad}} & \approx \frac{\dot{E}_{\rm mech} \epsilon_{\rm turb} \ell_{\rm outer} / \sigma}{L_{\rm bol} \tau_{\rm IR} h / c}\\
& \approx 3.5 \frac{\epsilon_{\rm turb}}{\tau_{\rm IR}} \frac{\ell_{\rm outer}}{h} \left(\frac{\sigma}{1000\ \kms}\right)^{-1}.\end{aligned}$$ Thus, by coincidence, the turbulent pressure in hot superwinds nearly equals the radiation pressure. In the densest cold starbursts, $\epsilon_{\rm turb}$ is only $\sim 0.1$, $\ell_{\rm outer} \approx 0.1 h$ and $\tau_{\rm IR} \ga 1$. On the other hand, the turbulent speeds in the cold ISM are $\la 100\ \kms$, so the turbulence resides for longer. In these regions, $U_{\rm rad}$ goes from being a few times smaller than $U_{\rm turb}$ to a few times larger than $U_{\rm turb}$.
Why is $U_{\rm CR}$ near $U_B$ only in weaker starbursts? {#sec:CREquipartition}
---------------------------------------------------------
Whether equipartition between CRs and magnetic fields holds in starburst regions is a matter of debate [e.g., @Thompson06]. Generally, the equipartition assumption gives reasonable results for hot starbursts but too small $B$ estimates for cold starbursts. For example, for M82 and NGC 253, the equipartition formula gives results of $\sim 190\ \muGauss$ [@Beck12], even after being corrected for secondaries and strong losses [@Lacki13-Equip]. This is in the range derived from detailed modeling and this paper’s estimate for $B$ (see section \[sec:BHot\]), so equipartition appears plausible. However, the magnetic field strengths derived for ULIRGs like Arp 220 are typically only $\sim 500\ \muGauss$ [@Thompson06; @Lacki13-Equip]. This is inconsistent with modeling (section \[sec:ModelBComparison\]) and my own estimate for $B$, which indicate magnetic field strengths of several mG.
The CR energy density is $$\label{eqn:UCRBasic}
U_{\rm CR} = \frac{\epsilon_{\rm CR} \dot{E}_{\rm mech} t_{\rm CR}}{V}$$ where $t_{\rm CR}$ is the residence time for CRs in the superwind and $\epsilon_{\rm CR} \approx 0.1$ is the CR acceleration efficiency. The bulk of CR energy is in protons. Protons in starbursts either escape through advection if they traverse low density ISM, or lose energy to pionic losses if they experience a high enough density. The gamma-ray observations of M82 and NGC 253 indicate that $F_{\rm cal} \approx 20 - 40\% < 1$ in these galaxies. Thus, it appears that advection sets the lifetimes of CRs in these starbursts.
We therefore have $$\label{eqn:UCRAdvection}
U_{\rm CR}^{\rm adv} = \frac{\epsilon_{\rm CR} \dot{E}_{\rm mech}}{2 \pi R^2 v_{\rm wind}}$$ where $v_{\rm wind}$ is the average speed that CRs are transported out of the starburst. Note that the outflow speed rapidly accelerates at the sonic point [@Chevalier85]; within the starburst, CRs flow outwards at a slower speed, $v_{\rm wind} \approx 300$ – $600 \kms$, until they reach the starburst edge. Comparing with $U_B = U_{\rm turb}$ in equation \[eqn:UTurbBasic\], I find a ratio of $$\frac{U_{\rm CR}^{\rm adv}}{U_B} = \left(\frac{\epsilon_{\rm CR}}{\epsilon_{\rm turb}}\right) \left(\frac{\ell_{\rm outer}}{h}\right) \left(\frac{\sigma}{v_{\rm wind}}\right)$$ Because only $\sim 10\%$ of the mechanical energy is converted into CRs, $\epsilon_{\rm CR} / \epsilon_{\rm turb} \approx 1/10$. On the other hand, the turbulent outer scale in the hot wind may be a few times smaller than the scale height, and $v_{\rm wind}$ is probably a few times smaller than 1. Overall, we then find that $U_{\rm CR}^{\rm adv}/U_B \sim 1$. This arises not primarily because the CRs and magnetic field drive each other. Instead, equipartition happens because both CRs and magnetic fields draw from the same power source – supernovae – and have similar characteristic times, the sound/turbulent crossing time.[^4]
Likewise, the ratio of CR and thermal pressure in the superwind remains constant as long as advection dominates CR transport: it is roughly $U_{\rm CR} / U_{\rm therm} \approx \psi (\epsilon_{\rm CR} / \epsilon_{\rm therm}) (c_S / v_{\rm wind}) \approx 1/3$. CRs do not provide the majority of the pressure, but they do track the thermal pressure because supernovae power both the superwind and CRs.
Even if the ratio of CR and magnetic energy densities vary by an order of magnitude, the estimated equipartition magnetic field strength is still reasonable. As @Lacki13-Equip showed, for a given observation of a starburst, the dependence of the estimated magnetic field strength depends only as $\xi^{2/7.2}$ where $\xi = U_B / U_{\rm CR}$ if CR $e^{\pm}$ are cooled by bremmstrahlung. Thus, an order of magnitude variation in $\xi$ leads to only a factor 2 change in $B$. Note the estimated $U_{\rm CR}$ has a stronger dependence $\sim \xi^{-3.2/7.2}$, so its estimate is less trustworthy.
However, in dense starbursts, proton calorimetry is unavoidable. Then the CR energy density is *not* set by the sound-crossing time, but by the pionic time, which can be arbitrarily shorter: $$t_{\pi} \approx 5000\ \yr \left(\frac{n_H}{10^4\ \cm^{-3}}\right)^{-1}$$ from @Mannheim94. Thus, once proton calorimetry is achieved, equipartition between magnetic fields and CRs fails. In Arp 220, equipartition fails by nearly two orders of magnitude (assuming the CRs are evenly distributed throughout the starburst; see section \[sec:CRStoch\]). As a result, $\xi$ approaches 50 and the equipartition estimate for $B$ falls short by a factor of $\sim 3$.
The turbulent cascade and CR diffusion {#sec:Cascade}
======================================
Most of the observational probes of starburst turbulence (see Section \[sec:Implications\]) only probe the fluctuation spectrum on large scales, where most of the turbulent power is. But we often want to know what the full spectrum of turbulence is; for example, the microphysics is relevant for the dissipation into heat. In the Milky Way, there are several observational constraints on turbulence, such as radio scintillation [@Scalo04]. Yet most of these probes are impractical for starburst regions other than the GCCMZ, due to their great distances and high extinctions.
The diffusion of CRs is intimately related to the magnetic fluctuation spectrum on small scales, especially on the gyroradius of individual CRs. As a general rule, strong fluctuations slow down CRs. The CR diffusion rate is therefore a powerful probe of the small-scale turbulence, although in practice, interpreting it is difficult. The synchrotron emission of nearby starbursts is resolved, so these studies are in principle possible, and the rate of CR diffusion is a key quantity in predicting this resolved emission [c.f., @Torres12]. Extremely strong turbulence can also energize extant CRs through second-order Fermi acceleration, pushing them towards equipartition.
CRs have an important role in the neutral ISM of starbursts: they provide a low level of ionization, allowing the gas to couple to magnetic fields. Without some form of ionization, the turbulent dynamo described in section \[sec:Dynamoes\] could not operate in the mass-dominant neutral ISM. Ultraviolet photons are unlikely to penetrate through the thick columns of starburst molecular gas. Because of their penetrating power, CRs are a good candidate for ionizing [@Papadopoulos10-CRDRs]. But if the CR diffusion is too slow, the CRs are confined to small bubbles, and we must look to some other source of radiation [@Lacki12-GRDRs; @Lacki12-SLRs].
What is the spectrum of turbulent fluctuations? {#sec:TurbSpectrum}
-----------------------------------------------
Although power is injected at large scales, the turbulent cascade spans orders of magnitude in galaxies. At each scale, eddies decay into smaller eddies in one flow-crossing time. The cascade is finally terminated by microscopic processes like viscosity, converting the power into heat. The differential spectrum of power in modes with wave number between $k$ and $k + dk$ is generally described as a power law with spectral index $-q$: $E(k) \propto k^{-q}$.
A simple but powerful guess for the spectrum of hydrodynamic turbulence is Kolmogorov turbulence. In this theory, turbulent energy cascades through each length scale without significant dissipative losses. That means that the energy at each scale $\ell$ is essentially the amount of power injected at the outer scale multiplied by a residence time, which is taken to be the flow-crossing time at that scale, $\sigma(\ell) / \ell$. Then, as seen from eqn. \[eqn:Uturb\], the turbulent velocity at that scale is just $\sigma \propto \ell^{1/3}$, with a turbulent energy density of $U_{\rm turb} (\ell) \propto \ell^{2/3}$. The power spectrum is the well-known Kolmogorov spectrum $E(k) \propto k^{-5/3}$. While Kolmogorov turbulence was derived under the assumption of nonmagnetic, hydrodynamic turbulence, the same power law is a common feature in simulated MHD turbulence [@Cho02]. @Goldreich95 derived a spectral index of $q = 5/3$ for Alfven waves in MHD turbulence, although these waves are anisotropic. Furthermore, a Kolmogorov-like power law is observed to hold in the Milky Way WIM from parsec scales all the way down to megameter scales (the so-called “Big Power Law in the Sky”; @Armstrong95). Kolmogorov-like turbulence is also observed in Galactic HI gas [e.g., @Roy08], where the presence of a turbulent cascade is supported by the existence of Larson’s laws [@Larson81; @Miesch94; @Shetty12].
Besides the Kolmogorov spectrum, the “Kraichnan” spectrum with $q = 3/2$ is a common hypothesis. There is more power in small-scale magnetic fluctuations in the Kraichnan spectrum than in the Kolmogorov spectrum (Figure \[fig:BSpectra\]).
The inner scale of magnetic fluctuations {#sec:lInner}
----------------------------------------
Aside from the outer scales and energy injection rates calculated in the previous sections, the inner scale is the final ingredient needed to calculate the amount of turbulence at each scale. The physics thought to terminate the turbulent cascades is quite different for ionized and neutral matter. In the Milky Way warm ionized medium, the inner scale is only of order $\sim 10^7\ \cm$ [@Spangler90]. @Spangler90 suggested that the inner scale was the larger of two lengths, both of which are related to the ion cyclotron frequency $\Omega_i = e B / (m_p c)$. The first is ion inertial length, $v_A / \Omega_i$, at which a cyclotron resonance could damp waves; the second is the ion Larmor radius $v_{\rm therm} / \Omega_i$ [@Spangler90]. In starburst superwinds with $U_B = U_{\rm turb} \approx U_{\rm therm}$, both scales are nearly equal. Despite the low densities and high temperatures of the superwind that increase $v_A$ and $v_{\rm therm}$, the strong magnetic fields of starbursts imply that the inner scales of their turbulence is quite small, of order tens to a few thousand kilometers. These values are listed in Table \[table:BHotPredictions\]. Note that, by construction, these inner scales are all smaller than the gyroradius for a GeV ion.
The situation is much different for starburst molecular gas. In this case, the magnetic fields directly act on the free ions and electrons created by occasional ionizing radiation in the otherwise neutral gas. But it is the neutral gas that contains virtually all of the mass and the kinetic energy. The first barrier to the turbulent cascade is ambipolar diffusion, in which the ions drift relative to the neutral gas. On small length scales, ions diffuse out of eddies faster than they are transported by the gas, resulting in friction [@Zweibel83]. The effective Reynolds number for ambipolar diffusion drops below 1 on scales: $$\label{eqn:lADCold}
\ell_{AD} \approx 8.0 \times 10^{16}\ \cm\ {\cal M}_A^{-1} \left(\frac{B}{2\ \mGauss}\right) \left(\frac{x_e}{10^{-5}}\right)^{-1} \left(\frac{n ({\rm H_2})}{10^4\ \cm^{-3}}\right)^{-3/2}$$ which could serve as an inner scale for magnetic fluctuations [@Klessen00]. The Alfven Mach number $\sigma/v_A$ is denoted by ${\cal M}_A$ in this equation. This role of ambipolar diffusion, particularly concerning whether it actually generates small-scale fluctuations itself, is disputed, though [@Tagger95; @Balsara01; @Elmegreen04].
The more fundamental barrier is that the magnetic field couples to the neutrals only through neutral-ion collisions. However, these are relatively infrequent, with the typical mean free path of an ion being $\ell_N \approx v_{\rm therm} / ({\ensuremath{\langle \sigma v \rangle}} n_{H2})$ [@Lithwick01]. The MHD cascade therefore must terminate at $\ell_N$, because the ions have no way of responding to the motions of neutrals without colliding with them. According to @Osterbrock61, the cross section for ion-H$_2$ collisions[^5] is given by ${\ensuremath{\langle \sigma v \rangle}} \approx 1.9 \times 10^{-9}\ \cm^{-3}\ \sec^{-1}$. Then the inner scale of MHD turbulence in starburst molecular gas is $$\label{eqn:lInCold}
\ell_N \approx 1.4 \times 10^{12}\ \cm \left(\frac{T}{100\ \Kelv}\right)^{1/2} \left(\frac{n_(H_2)}{100\ \cm^{-3}}\right)^{-1}.$$ Unlike in the superwinds, the inner scale is much larger than the gyroradius of a GeV ion, which has important implications for CR diffusion. However, the inner scale gets progressively smaller in denser regions.
{width="8cm"}
Calculation of CR diffusion constants {#sec:DCascCalc}
-------------------------------------
[lllccccccccc]{} GCCMZ & Hot & $\ell_{\rm outer} = h$ & $4.4 \times 10^{10}$ & 1.0 & 0.026 & 0.15 & 20 & 1500 & 18 & 0.45 & 2.6\
& & $\ell_{\rm outer} = \ell_1$ & $5.0 \times 10^{10}$ & 0.81 & 0.023 & 0.14 & 12 & 1000 & 18 & 0.52 & 3.2\
& & $\ell_{\rm outer} = \ell_2$ & $4.9 \times 10^{10}$ & 0.84 & 0.023 & 0.14 & 12 & 1100 & 18 & 0.51 & 3.1\
& Cold & $\ell_{\rm outer} = h$; $\Psi = 1$ & $8.2 \times 10^9$ & 0.57 & 0.011 & 0.35 & 0.69 & 1500 & 8200 & 160 & 4900\
& & $\ell_{\rm outer} = R_{\rm max}$; $\Psi = 1$ & $1.7 \times 10^{10}$ & 0.17 & 0.0053 & 0.24 & 0.034 & 160 & 11000 & 340 & 15000\
& & $\ell_{\rm outer} = R_{\rm max}$; “Natural” $\Psi$ & $2.5 \times 10^{10}$ & 0.26 & 0.0082 & 0.63 & 0.019 & 260 & $1.4 \times 10^5$ & 4400 & $3.4 \times 10^5$\
NGC 253 & Hot & $\ell_{\rm outer} = h$ & $9.7 \times 10^9$ & 0.61 & 0.012 & 0.073 & 19 & 1500 & 12 & 0.24 & 1.4\
& & $\ell_{\rm outer} = \ell_1$ & $1.6 \times 10^{10}$ & 0.27 & 0.0076 & 0.057 & 2.7 & 360 & 14 & 0.39 & 3.0\
& & $\ell_{\rm outer} = \ell_2$ & $1.4 \times 10^{10}$ & 0.32 & 0.0083 & 0.060 & 3.9 & 480 & 14 & 0.35 & 2.6\
& Cold & $\ell_{\rm outer} = h$; $\Psi = 1$ & $3.5 \times 10^9$ & 0.44 & 0.0074 & 0.12 & 2.4 & 1500 & 500 & 8.4 & 140\
& & $\ell_{\rm outer} = R_{\rm max}$; $\Psi = 1$ & $7.9 \times 10^9$ & 0.11 & 0.0033 & 0.081 & 0.095 & 140 & 650 & 19 & 470\
& & $\ell_{\rm outer} = R_{\rm max}$; “Natural” $\Psi$ & $1.2 \times 10^{10}$ & 0.18 & 0.0050 & 0.21 & 0.053 & 220 & 8600 & 250 & 10000\
M82 & Hot & $\ell_{\rm outer} = h$ & $1.0 \times 10^{10}$ & 0.62 & 0.013 & 0.078 & 17 & 1500 & 14 & 0.29 & 1.8\
& & $\ell_{\rm outer} = \ell_1$ & $1.7 \times 10^{10}$ & 0.28 & 0.0077 & 0.061 & 2.5 & 360 & 17 & 0.47 & 3.7\
& & $\ell_{\rm outer} = \ell_2$ & $1.5 \times 10^{10}$ & 0.33 & 0.0086 & 0.064 & 3.8 & 490 & 16 & 0.42 & 3.1\
& Cold & $\ell_{\rm outer} = h$; $\Psi = 1$ & $3.4 \times 10^9$ & 0.43 & 0.0072 & 0.14 & 1.9 & 1500 & 810 & 14 & 250\
& & $\ell_{\rm outer} = R_{\rm max}$; $\Psi = 1$ & $7.9 \times 10^9$ & 0.11 & 0.0031 & 0.089 & 0.065 & 120 & 1100 & 32 & 900\
& & $\ell_{\rm outer} = R_{\rm max}$; “Natural” $\Psi$ & $1.1 \times 10^{10}$ & 0.17 & 0.0048 & 0.23 & 0.037 & 200 & 14000 & 410 & 20000\
Arp 220 Nuclei & Hot & $\ell_{\rm outer} = h$ & $1.2 \times 10^9$ & 0.30 & 0.0042 & 0.024 & 20 & 1500 & 5.3 & 0.073 & 0.43\
& & $\ell_{\rm outer} = \ell_1$ & $3.0 \times 10^9$ & 0.063 & 0.0016 & 0.015 & 0.46 & 91 & 7.2 & 0.19 & 1.7\
& & $\ell_{\rm outer} = \ell_2$ & $2.5 \times 10^9$ & 0.083 & 0.0019 & 0.017 & 0.89 & 150 & 6.8 & 0.16 & 0.14\
& Cold & $\ell_{\rm outer} = h$; $\Psi = 1$ & $4.1 \times 10^8$ & 0.21 & 0.0025 & 0.041 & 2.4 & 1500 & 250 & 2.9 & 48\
& & $\ell_{\rm outer} = R_{\rm max}$; $\Psi = 1$ & $1.4 \times 10^9$ & 0.026 & 0.00070 & 0.022 & 0.015 & 34 & 370 & 10 & 320\
& & $\ell_{\rm outer} = R_{\rm max}$; “Natural” $\Psi$ & $2.3 \times 10^9$ & 0.055 & 0.0014 & 0.055 & 0.024 & 86 & 2000 & 51 & 2000 \[table:DPredictions\]
A CR is largely unaffected by magnetic inhomogeneities smaller than its Larmor radius $r_L$, and it basically follows a magnetic field line if it curves on scales larger than its Larmor radius. Magnetic fluctuations of the same size as the Larmor radius can interact with CRs resonantly and deflect it. This is the gyroresonance. Since the inhomogenieties $\Delta B (r_L)$ at the Larmor radius are much typically smaller than the large-scale $B$ setting the size of the Larmor radius, the CRs are deflected by only a small angle $\sim \Delta B (r_L) / B$ with each Larmor orbit. If the fluctuations are random, it takes $\sim (B / \Delta B (r_L) )^2$ orbits to be isotropized; from this the mean free path of a CR can be calculated [@Kulsrud05].
*Kolomogorov spectrum* – If isotropic Kolmogorov spectrum turbulence extends down to the gyroradius of the CR, it provides a minimum level of magnetic fluctuations that deflect the CR. The mean free path of a CR is no more than $$\lambda_{\rm CR} \approx r_L^{1/3} \ell_{\rm outer}^{2/3}$$ assuming the magnetic field is entirely turbulent [@Schlickeiser02]. The spatial diffusion constant is $D_{\rm casc} \approx \lambda v_{\rm CR} / 3$, or $$D_{\rm casc} \approx r_L^{1/3} \ell_{\rm outer}^{2/3} \beta_{\rm CR} c / 3,$$ where $\beta_{\rm CR} c$ is the CR’s speed.
Using the outer scales and magnetic fields I found in sections \[sec:HotTurbulence\] and \[sec:ColdTurbulence\], I calculate CR diffusion constants and list them in Table \[table:DPredictions\]. These are listed for CR protons with energies of 1 GeV, the peak of the CR energy density spectrum. The diffusion constants are very small compared to the values of $3 \times 10^{28}\ \cm^2\ \sec^{-1}$ in the Solar neighborhood: the high volumetric rate of supernovae drives vast amounts of turbulent power through each scale of the cascade, and the high magnetic field strengths lead to small Larmor radii and smaller mean free paths. The time for CRs to escape the starburst through diffusion is then $t_{\rm diff} \approx h^2 / D$, about $3\ \Myr$. Then CR diffusion is much slower than advection, as supported by the hard gamma-ray spectra of M82 and NGC 253, which indicate that the dominant transport process is energy-independent [@Abramowski12].
*Fast modes in MHD turbulence and the Kraichnan spectrum* – In the modern theory of MHD turbulence developed by @Goldreich95, most of the turbulent power is in the form of Alfven waves. The waves are extremely anisotropic at small scales, with the wavelength parallel to the magnetic field much longer than the wavelength perpendicular to the magnetic field ($\lambda_{\|} / \lambda_{\bot}) \propto (\ell_{\rm outer} / \lambda_{\bot})^{1/3}$. The power spectrum of these waves is in fact Kolmogorov, $k_{\bot}^{-5/3}$ [@Goldreich95]. The problem for CR confinement is that CRs are thought to scatter on fluctuations moving parallel to the magnetic field. Thus, by the point that $\lambda_{\|}$ equals $r_L$, $\lambda_{\bot}$ is tiny, and the fluctuations are too weak to confine CRs.
Fortunately, Alfven waves are not the only MHD modes present in magnetic turbulence. Simulations of MHD turbulence reveal that a fraction of the turbulent power goes into producing fast magnetosonic waves [@Cho02; @Cho03], which are isotropic and thus can scatter CRs [@Yan02]. The normalization of the cascade spectrum is set by this fraction. Although this fraction is small when the magnetic field is ordered or when $v_A \ll c_S$ [@Cho02; @Cho03], it may be of order unity for starbursts where the magnetic field is essentially all turbulent and is in equipartition with or far exceeds the thermal pressure.
Even better for the confinement of CRs, the fast magnetosonic wave spectrum decays has a Kraichnan spectrum at small scales [@Cho02] (see Figure \[fig:BSpectra\]).
If the magnetic fluctuation strength due to fast modes is $B_{\rm fast}$, and if a fraction $\epsilon_{\rm fast}$ of turbulent power is in these modes, then $(\Delta B_{\rm fast} / B) \approx (r_L / \ell_{\rm outer})^{1/2} \epsilon_{\rm fast}$. We then have, by a similar argument to the Kolomogorov spectrum, $\lambda_{\rm CR} \approx \epsilon_{\rm fast}^{-1} \sqrt{r_L \ell_{\rm outer}}$. The diffusion constant for the Kraichnan spectrum is $$\begin{aligned}
\label{eqn:DCascKraich}
D_{\rm casc} & \approx \frac{\beta_{\rm CR} c}{3 \epsilon_{\rm fast}} \sqrt{r_L \ell_{\rm outer}}.\end{aligned}$$
@Schlickeiser98 calculated the diffusion constant of CRs scattering on fast modes using quasilinear theory (see also @Schlickeiser02) in which they included both the resonance and a “transit time damping” contribution (described in section \[sec:OtherD\]). They found that the diffusion constant is limited by CRs with small pitch angles, resulting in a diffusion constant that is roughly $\sim \beta_{\rm CR} c / v_A$ times larger than in equation \[eqn:DCascKraich\]. The exact value of the diffusion constant for fast modes with $q = 3/2$ in @Schlickeiser98 is $$\begin{aligned}
D_{\rm fast} & \approx \frac{4}{3\pi {\cal Z}(5/2)} \epsilon_{\rm fast}^{-1} \sqrt{\frac{1}{2\pi} \frac{\beta_{\rm CR}^3 c^3}{v_A} r_L \ell_{\rm outer}},\end{aligned}$$ where ${\cal Z}$ is the zeta function.
I list $D_{\rm casc} (q = 3/2)$ and $D_{\rm fast}$ for supernova-driven turbulence within starbursts in Table \[table:DPredictions\]. We see that the diffusion constants are very small, and tiny compared to those inferred for the Milky Way ($\sim 10^{28}$ – $10^{29}\ \cm^2\ \sec^{-1}$). If these diffusion constants apply, then standard CR diffusion plays no role in the large scale motion of CRs through and out of starbursts.
*Diffusive reacceleration* – The scattering of CRs off magnetic turbulence also results in momentum diffusion, with CRs gaining roughly $\sim (v_A/c)^2$ in energy per mean free path through second-order Fermi acceleration. While this process is slow in the Milky Way, it could be much faster in starbursts because the mean free paths are shorter and the turbulent speeds are faster. The effective momentum diffusion constant is $D_{pp} \approx p^2 v_A^2 / D_{\rm casc}$, and the associated timescale for momentum gain is $t_{\rm reacc} \approx 9 D_{\rm casc} / v_A^2$ [@Strong07].[^6]
If the diffusion constants calculated above are correct, diffusive reacceleration is extremely fast in starburst winds. The reacceleration timescale is only $\sim 10\ \kyr$ for Kolmogorov turbulence and just a few centuries for Kraichnan turbulence (Table \[table:DPredictions\]). As this is much shorter than the escape time, at face value, this implies the CR energy density grows exponentially. Of course, the turbulent energy would be damped by this process, so what must happen is that CRs, turbulence, and magnetic fields approach equipartition. If this is true, CR acceleration in starburst winds is not primarily a process happening at discrete supernova remnants, but is a collective process throughout the hot ISM [c.f., @Amano11; @Melia11]. The efficiency for converting supernova mechanical power into CR kinetic energy could then be higher than the canonical value of 10%. Note also that some of the energy would go into MeV suprathermal particles, not just the GeV-TeV particles observed by gamma-ray telescopes.
Diffusive reacceleration is much slower in molecular regions, where the turbulent speeds are $\la 100\ \kms$. The reacceleration times are $\ga \Myr$ for Kolmogorov timescales and $\sim 10\ \kyr$ for Kraichnan turbulence (see Table \[table:DPredictions\]). Note that the radiative lifetime of CRs in ULIRGs is just a few kyr; therefore, reacceleration probably does not affect the overall energetics of CRs in these environments.
Turbulent mixing {#sec:TurbulentMixing}
----------------
The transport of CRs involves not just the flow of CRs relative to matter but the flow of the matter containing the magnetic fluctuations. For small enough scattering diffusion constants, like those found in Section \[sec:DCascCalc\], the CRs are effectively frozen into the fluid. Thus the motion of the ISM itself becomes increasingly important. While the large-scale bulk advection is well known to carry away CRs, the small-scale chaotic advection has not been explored as much.
Turbulence is well known for its ability to mix fluids. The chaotic flows knead regions with different properties into one another, smoothing out the coarse-grained distribution. While it does not directly even out the fine-grained distribution, the inhomogeneities are on such tiny scales that even very low levels of microphysical diffusion smooth out the fine-grained distribution [see the review by @Shraiman00]. Turbulence is suspected to be a prime mover behind the transport of metals in the ISM of star-forming regions [e.g., @Roy95; @deAvillez02; @Scalo04; @Pan10].
CRs are another “pollutant” injected by star-forming regions that turbulence will spread. We can define an effective diffusion constant for the turbulent mixing as $$D_{\rm mix} \approx \ell_{\rm outer} \sigma$$ [@Tennekes72]. Turbulent mixing is most effective in regions with high $\sigma$, particularly the hot superwind phase. Another noteworthy feature of turbulent mixing is that it is *energy independent*. Thus, observations of the gamma-ray spectral shape do not constrain turbulent mixing.
But, while we can define a “diffusion constant” for the process, there are several important differences with standard diffusion. Turbulence creates inhomogeneities in the flow; for transonic or supersonic turbulence, the fluctuations are extremely strong. If CRs are frozen into the flow, turbulent mixing squeezes and stretches them too. The resulting strong fluctuations in the density of CRs are purely random and not directly related to the locations of CR sources. On top of that, adiabatic expansion and compression of the CRs enhances density gradients, as $P \propto V^{-4/3}$ for relativistic particles. Furthermore, turbulent mixing is generally intermittent on large scales, so that even if the underlying fluid were homogeneous, there would be fluctuations in the density of CRs simply because the mixing cannot even out the CR abundance perfectly [@Shraiman00]. By contrast, standard diffusion reduces inhomogeneities, simply because there are more CRs in a high density fluctuation to flow out than CRs outside the fluctuation to flow in.
The essential difference is that standard diffusion works to even out the *density* of CRs, the number of CRs per volume, whereas turbulent mixing works to even out the *abundance* of CRs, the number of CRs per mass [c.f., @Ensslin11]. The former can be thought of as a “Eulerian” process, where the CRs follow the volume. This is reflected by the Eulerian methods used to treat diffusive propagation in models of galactic CR populations like GalProp [as described in @Strong98]. But turbulent mixing is a “Lagrangian” process, where the CRs follow the mass, suggesting a possible need for Lagrangian methods when modeling starburst CR populations. An example of these methods are provided by MHD simulations of CR populations in galaxy clusters, where Lagrangian CR transport is included (often as the only form of CR transport) [e.g., @Hanasz03; @Pfrommer08].
The ${\cal M} \approx 100$ turbulence in the molecular medium has especially extreme density contrasts and extreme intermittency. Current CR models assume that CRs experience essentially steady losses in a uniform medium. But actual losses in a supersonic turbulent medium may be highly irregular. Suppose CRs are injected into such a medium at random locations. At any given time, the median gas density in the volume is $\sim {\ensuremath{\langle \rho \rangle}} / {\cal M}$ to ${\ensuremath{\langle \rho \rangle}}$, but most of the mass is in clumps with density $\sim {\cal M} {\ensuremath{\langle \rho \rangle}}$. Thus, most CRs could be injected into underdense regions, where they are presumably frozen into the flow. In these regions, losses are much slower than supposed by one-zone models. Yet the flow rearranges itself on an eddy crossing time, $\sim \ell_{\rm outer} / \sigma \approx 10\ \kyr$, and most of the fluid containing the CRs ends up in high density clumps. As the flow is squeezed into these clumps, the radiative losses become orders of magnitude stronger – up to a factor ${\cal M}^2 \approx 10^4$ faster. At some point, the radiative losses become faster than the advection times, and the CRs are destroyed. Thus, CR losses could occur in short bursts.
Although proton calorimetry implies the total gamma-ray luminosity does not depend on the proton cooling time, adiabatic compression of the CRs converts mechanical energy into CR kinetic energy. When the CRs are destroyed, that energy is mostly radiated away as gamma rays and neutrinos. Turbulent mixing in supersonic turbulence can therefore decouple the power injected in CRs from the power radiated as gamma rays. The intermittent destruction of CRs and the implications for proton calorimetry need to be studied.
Comparison with Observations
----------------------------
*The CR diffusion constant around 30 Doradus* – By comparing the resolved synchrotron and Inverse Compton emission of the super star cluster 30 Doradus in the Large Magellanic Cloud, @Murphy12 derived the CR diffusion constant and its energy dependence at GeV energies in that region. The star-formation rate, $0.15\ \Msun\ \yr^{-1}$, and size of the region the CRs diffused through, $\sim 100\ \pc$, correspond to a star-formation surface density of $5\ \Msun\ \kpc^{-2}\ \yr^{-1}$, of the same order as in the GCCMZ.
The measured CR diffusion constant is $10^{27}\ \cm^2\ \sec^{-1} (E/\GeV)^{0.7}$ for relativistic particles [@Murphy12]. They conclude that the smallness of the diffusion constant is due to the high level of turbulence surrounding 30 Doradus. At 1 GeV, the measured value is indeed comparable to the expected diffusion constant in the GCCMZ for a Kolmogorov spectrum of turbulence. The steep energy dependence is, however, incompatible with the standard Kolmogorov cascade. But according to @Blasi12, diffusion constants with $D \propto E^{0.7}$ arise when CRs amplify waves from a previously existing Kolmogorov spectrum as they self-confine. This process decreases $D$.
Yet the 100 pc surrounding 30 Doradus may not be typical of starburst regions as a whole. There is only one super star cluster, whereas starburst regions contain many (dozens of similar mass in M82, for example; @OConnell95). Furthermore, most of the star formation responsible for accelerating the CRs occurs in R136, with a diameter of a few pc, at the center of 30 Doradus [e.g., @Hunter95], so the estimates of turbulence I derived above do not apply to the CRs 100 pc away. For example, perhaps the CRs are amplifying relatively weak turbulence in these distant regions to produce a small but steeply increasing diffusion constant. In contrast, star formation occurs on 100 pc scales throughout starbursts. Observations of the GCCMZ, M82, and NGC 253 exclude such steep energy dependencies within their starbursts, suggesting different physics.
*The TeV emission of the GCCMZ and NGC 253* – If the TeV emission from the GCCMZ region is in fact powered by star formation, it indicates that CR diffusion is quite slow at TeV energies [@Crocker11-Wild]. The spectrum observed by HESS remains relatively hard, with a power law index of $\sim 2.3$, until the highest observed energies of 10 TeV [@Aharonian06]. This implies energy-independent escape for CRs with energies of up to 100 TeV, or $D \la 4 \times 10^{27}\ \cm^2\ \sec^{-1}$
Likewise, the hard GeV-TeV gamma-ray spectra of M82 and NGC 253 set stringent constraints on the rate of energy-dependent diffusion in these starbursts. According to @Abramowski12, the lack of a spectral break in the gamma-ray spectrum of NGC 253 as observed with *Fermi* and HESS indicates that $D \la 3 \times 10^{27}\ \cm^2\ \sec^{-1}$ for protons with energies up to 30 TeV. We can compare this number with the predicted cascade diffusion constants $D_{\rm casc}$ listed in Table \[table:DPredictions\]. I find that the 30 TeV diffusion constants are $\sim 10^{28}\ \cm^2\ \sec^{-1}$ for the Kolomogorov spectrum and $\sim 10^{27}\ \cm^2\ \sec^{-1}$ for the Kraichnan spectrum.
While there is mild tension between the $D_{\rm casc}$ and the observed TeV spectrum for the Kolmogorov case, the Kraichnan case is in line with observations. It is interesting that above 100 TeV, the predicted $D_{\rm casc}$ for $k = 3/2$ finally exceeds the limits set by HESS, and a break in the proton spectrum should appear. Unfortunately, at such high energies (corresponding to $\ga 10\ \TeV$ gamma rays), NGC 253 becomes opaque as photons are destroyed by the $\gamma\gamma$ absorption process [@Inoue11; @Lacki13-XRay]. The GCCMZ, however, is transparent at these energies [@Lacki13-XRay], and observable up to a few hundred TeV [@Moskalenko06].
What truly confines CRs in starbursts?
--------------------------------------
### Problems with confinement by turbulent fluctuations
*Does the turbulence cascade in hot winds?* – As noted in Section \[sec:lInner\], the estimates of $D_{\rm casc}$ in hot winds only apply if the turbulence can cascade to small scales. Many authors have argued that the turbulence is halted at large scales in the hot medium of the Milky Way (and by extension, in hot starbursts). A common objection is that the small Reynolds number prevents any true cascade of magnetic energy to small scales [@Hall80]. But, as I have argued, the collisionless nature of the wind implies the hydrodynamical Reynolds number is not necessarily the relevant quantity (Section \[sec:Reynolds\]).
A more subtle problem raised by detailed theories of MHD turbulence is the possibility that the fast modes most likely to confine CRs are damped in hot ISM. According to @Yan04, fast modes are collisionally damped on scales comparable to the mean free path of particles in the medium. Since this length is just a few parsecs, both in the Milky Way hot ISM and in starburst hot winds, this could prevent the cascade from reaching scales relevant for CRs. In the GCCMZ, though, $\ell_{\rm outer}$ is nearly equal to the mean free path, so it is possible the cascade is not damped collisionally.
*Neutrals in cold gas* – The long ambipolar diffusion (equation \[eqn:lADCold\]) and neutral-ion drag scales (equation \[eqn:lInCold\]) in molecular gas pose severe problems for CR confinement in molecular gas in starbursts, especially for ULIRGs where virtually the entire volume may be low density molecular gas. Only CRs with $\ga \TeV$ energies have $r_L \ga \ell_{\rm N}$. This problem is analogous to the situation for MeV positrons propagating in neutral gas within the Milky Way, but at least in the Galaxy, most of the volume is ionized [@Higdon09; @Jean09; @Prantzos11].
### Other confinement processes {#sec:OtherD}
*Self-confinement* – Streaming CRs, with an anisotropic momentum distribution, excite Alfven waves that effectively confine the CRs [@Kulsrud69]. In cold fully-ionized plasmas, these waves force CRs to stream at speeds smaller than $v_A$.
In the hot wind of starbursts, the Alfvenic speed is actually faster than the bulk speed of the wind in the core of the starburst ($\sim 300\ \kms$), and is approximately equal to the asymptotic wind speed ($v_{\infty} \approx 1600\ \kms$). Thus, even if self-confinement did operate, diffusive escape could actually still dominate over advection if no other process intervened. In addition, @Holman79 argued that self-confinement fails to limit CRs to speeds below the sound speed $c_S$, although $c_S \approx v_A$ in hot ISM.
The Alfven speed is much slower in the cold gas, but self-excited Alfven waves may be strongly damped by neutral atoms in these phases. @Higdon09 noted similar problems for confining Galactic MeV positrons in neutral gas. On the other hand, the CR energy densities are much higher than in the Galaxy, providing much stronger driving for the build up of such waves. A calculation of whether the streaming instabilities are enough to overcome drag is needed.
In the warm ionized medium of H II regions, self-confinement should be effective.
*Field line random walk* – Another possibility is that CRs do follow magnetic field lines, but the field lines themselves are so twisted that they confine CRs within the starbursts. This is the Field Line Random Walk (FLRW) mechanism, which is important in the “compound diffusion” discussed by @Lingenfelter71. The field lines are randomized on scales $\ell_{\rm peak}$ where the magnetic fluctuation spectrum peaks. The worst possible case for confinement is when most of the turbulent power is near the outer scale. The effective diffusion constant for FLRW $$D_{\rm FLRW} \approx \ell_{\rm peak} c / 3,$$ as long as $r_L < \ell_{\rm peak}$. These values when $\ell_{\rm peak} = \ell_{\rm outer}$ are listed in Table \[table:DPredictions\].
While this mechanism does not work if the outer scale is nearly the size of the starburst, confinement is effective if the outer scale is just a few parsecs, because the starburst would have a scattering optical depth to CRs of order $\sim 10$ and the typical number of scatterings before escape would be $\sim 10^2$. This effective diffusion constant, $D_{\rm FLRW} \approx 1.5 \times 10^{29}\ \cm^2\ \sec^{-1}\ (\ell_{\rm peak} / 5\ \pc)$, is comparable to $D$ in the Milky Way: in the Milky Way, the mean free path is also of order a few parsecs.
*Firehose and mirror plasma instabilities* – The need for some kind of CR confinement in the Milky Way’s hot ISM prompted suggestions that magnetic fluctuations could be generated directly on small scales. These fluctuations could be generated by the firehose and mirror instabilities, which operate when the pressure is anisotropic enough in the plasma [@Hall80]. It’s not clear these could work in starburst wind, though. A key point in @Hall80’s argument is that the magnetic pressure is much smaller than the thermal pressure; this allows the instabilities grow. But in the starburst wind, the magnetic pressure approaches the thermal pressure, inhibiting the growth of these instabilities. @Schekochihin05 came to this conclusion for galaxy cluster plasma too: the plasma instabilities driven by anisotropic pressure work when magnetic pressure is much smaller than the thermal pressure, but die away as magnetic fields become strong.
*Transit time damping* – Although the gyroresonance is the most obvious way to scatter CRs, compressible turbulent MHD waves can also deflect CRs through transit time damping (TTD) [@Schlickeiser98; @Yan04]. TTD results from the “mirror” force as magnetic field strengths fluctuate. TTD naturally occurs in the fast modes, and in fact is accounted for in the $D_{\rm fast}$ given by @Schlickeiser98. Unlike the gyroresonance, the cascade of fast modes does not have to reach $r_L$ to scatter CRs. According to @Yan04, TTD is what confines CRs in Galactic neutral clouds and the hot phase, so its role should also be considered for starbursts.
*Confinement by H II regions and SNRs* – CRs that attempt to escape may be trapped by pockets where confinement is effective. Candidate CR traps include H II regions or supernova remnants. While the bubbles have a filling factor much less than 1, their covering factor can be $\ga 1$, so that the typical CR is deflected back inwards as it tries to escape.
### Summary of confinement processes in starbursts
{width="16cm"}
Many processes are conjectured to confine CRs in the Milky Way, and with our far lesser knowledge of starbursts, it is difficult to say which one is most important in those regions. Figure \[fig:CRDiffFlowchart\] lays out the possibilities and the conditions they hold in. It is impossible to even answer basic questions of CR diffusion in starbursts with certainty. The diffusion may be Lagrangian (turbulent mixing) or Eulerian (other kinds of confinement), energy independent (turbulent mixing and FLRW) or dependent (scattering by turbulence), and may be faster (FLRW, self-confinement in the hot phase) or slower (turbulent mixing, scattering by turbulence) than advection.
What we can probably say with high likelihood is that the diffusion is no slower than that given by turbulent mixing ($D_{\rm eddy}$) and no faster than that given by field line random walk ($D_{\rm FLRW}$). The few observations we do have suggest that diffusion is either slow and/or energy-independent, consistent with scattering by a turbulent cascade reaching the gyroscale, turbulent mixing, or a field line random walk.
Do CRs fill all of starbursts? {#sec:CRStoch}
------------------------------
Do CRs in fact reach all of the gas, as is commonly supposed? Suppose that CR accelerators in starbursts inject an instantaneous pulse of CRs. The CRs diffuse away from the accelerator, roughly filling a sphere of radius $R_{\rm diff} = \sqrt{D t}$, where $D$ is a constant diffusion constant that applies throughout the starburst. The CRs reside for a time $t_{\rm CR}$, after which they are lost. Therefore, I consider only CR bubbles that are younger than $t_{\rm CR}$. The maximum size the CR bubbles reach is $R_{\rm diff}^{\rm max} = \sqrt{D t_{\rm CR}}$. If the rate at which CR accelerators appear and inject a burst of CRs is $dN_{\rm acc}/dt$, then the volume filled by CRs is $$V_{\rm CR} = \int_0^{t_{\rm CR}} \frac{dN_{\rm acc}}{dt} \times \frac{4}{3} \pi R_{\rm diff}(t)^3 dt.$$ Now, supposing that SNRs are (short-lived) CR accelerators, we have $dN_{\rm acc}/dt = \Gamma_{\rm SN}$. The estimated volume filled by CRs is $$V_{\rm CR} = \frac{8 \pi}{15} \Gamma_{\rm SN} D^{3/2} t_{\rm CR}^{5/2}.$$ The overlap fraction of these CRs is just $Q_{\rm CR} = V_{\rm CR} / V_{\rm SB}$; if $Q_{\rm CR} \ga 1$, then the CRs are volume-filling.
Although the true CR diffusion constants in starbursts are unknown, the $D_{\rm eddy}$ from turbulent mixing sets a lower limit to the distance CRs can traverse. I compute the $R_{\rm diff}^{\rm max}$ and $Q_{\rm CR}$ for the prototypical starbursts, and list them in Table \[table:CRStoch\].
[lllllcc]{} GCCMZ & $3.1 \times 10^6$ & 160 & Hot & $\ell_{\rm outer} = h$ & 100 & 48\
& & & & $\ell_{\rm outer} = \ell_1$ & 78 & 21\
& & & & $\ell_{\rm outer} = \ell_2$ & 81 & 24\
& & & Cold & $\ell_{\rm outer} = h$; $\Psi = 1$ & 19 & 0.31\
& & & & $\ell_{\rm outer} = R_{\rm max}$; $\Psi = 1$ & 4.3 & 0.0034\
& & & & $\ell_{\rm outer} = R_{\rm max}$; “Natural” $\Psi$ & 3.2 & 0.0015\
NGC 253 & $7.1 \times 10^6$ & 160 & Hot & $\ell_{\rm outer} = h$ & 99 & 830\
& & & & $\ell_{\rm outer} = \ell_1$ & 38 & 46\
& & & & $\ell_{\rm outer} = \ell_2$ & 46 & 80\
& & & Cold & $\ell_{\rm outer} = h$; $\Psi = 1$ & 36 & 39\
& & & & $\ell_{\rm outer} = R_{\rm max}$; $\Psi = 1$ & 7.1 & 0.30\
& & & & $\ell_{\rm outer} = R_{\rm max}$; “Natural” $\Psi$ & 5.3 & 0.13\
M82 & $2.8 \times 10^7$ & 160 & Hot & $\ell_{\rm outer} = h$ & 96 & 620\
& & & & $\ell_{\rm outer} = \ell_1$ & 36 & 33\
& & & & $\ell_{\rm outer} = \ell_2$ & 45 & 63\
& & & Cold & $\ell_{\rm outer} = h$; $\Psi = 1$ & 32 & 23\
& & & & $\ell_{\rm outer} = R_{\rm max}$; $\Psi = 1$ & 5.9 & 0.14\
& & & & $\ell_{\rm outer} = R_{\rm max}$; “Natural” $\Psi$ & 4.4 & 0.061\
Arp 220 Nuclei & $3.1 \times 10^6$ & 5 & Hot & $\ell_{\rm outer} = h$ & 18 & 12\
& & & & $\ell_{\rm outer} = \ell_1$ & 2.8 & 0.041\
& & & & $\ell_{\rm outer} = \ell_2$ & 3.8 & 0.11\
& & & Cold & $\ell_{\rm outer} = h$; $\Psi = 1$ & 6.3 & 0.50\
& & & & $\ell_{\rm outer} = R_{\rm max}$; $\Psi = 1$ & 0.50 & $2.4 \times 10^{-4}$\
& & & & $\ell_{\rm outer} = R_{\rm max}$; “Natural” $\Psi$ & 0.63 & $4.9 \times 10^{-4}$ \[table:CRStoch\]
The first thing to note is that $R_{\rm diff}^{\rm max}$ and $Q_{\rm CR}$ are much larger in the hot winds, because the turbulent speeds are themselves much larger. For the GCCMZ, NGC 253, and M82, CRs should easily fill the hot wind.
Second, the predicted $Q_{\rm CR}$ is typically smaller than 1 if I use the $D_{\rm eddy}$ listed for the cold ISM in Table \[table:DPredictions\]. Of course, in the hot starbursts, only a small fraction of the volume is filled by the molecular clouds and other neutral material, so the use of the cold $D_{\rm eddy}$ is inappropriate. Instead, the relevant question is whether the CRs can diffuse all of the way through a molecular cloud before they are destroyed. From Table \[table:CRStoch\], it appears that CRs that survive for 160 kyr can traverse at least $R_{\rm diff}^{\rm max} \ga 5$ parsecs through starburst molecular clouds, roughly the size of the molecular clouds. On the other hand, in overdense molecular clouds, CRs are not carried by the wind. Rather, they face strong pionic losses that shorten their lifetimes, so it’s still an open question whether CRs make it all the way through the clouds before being destroyed.
Finally, in the Arp 220 nuclei, it appears that CRs may not fill the entire starburst region if they diffuse at only $D_{\rm eddy}$. The turbulent speeds are slow in the molecular gas of these cold starbursts, and the outer scale of turbulence is probably quite small. Instead, CRs are predicted to only fill $0.01 - 0.1\%$ of the volume, residing in about $10^4$ bubbles that are each $\la 1\ \pc$ wide.
If that were true, the CR energy density in these bubbles would be much greater than even the ISM turbulent pressures listed in Table \[table:Pressures\]. Thus, the bubbles would expand until they did reach equipartition, partly because of a greater volume and partly due to adiabatic losses. Indeed, the expansion of such bubbles might themselves be a source of such turbulence in starbursts, on scales smaller than 10 pc.[^7] This just further underscores how little we know about CR diffusion in these environments. Outside of the CR-filled bubbles, some alternate source of ionization is necessary if diffusion is this slow, such as gamma rays [@Lacki12-GRDRs] or short-lived radioactive isotopes [@Lacki12-SLRs].
These lower bounds can be improved by using exact analytic [e.g., @Aharonian96] or numerical solutions for CR diffusion, as done for proton calorimetric homogeneous starbursts in @Torres12.
Observational implications {#sec:Implications}
==========================
Doppler shifts in X-ray lines from superwind turbulence
-------------------------------------------------------
A key prediction of this work is that the volume-filling hot superwind is turbulent, with random motions that are about as fast as the large-scale outflow itself. Ions embedded in the wind are Doppler shifted when they radiate, not just from the bulk flow but from the transonic turbulence. In fact, Fe K line emission is directly observed from M82’s superwind, so a finely-resolved X-ray spectrum should be able to trace these motions [@Strickland07; @Strickland09].
The Doppler shift of the superwind ions from turbulence is $\Delta E \approx E \sigma/c$: $$\Delta E \approx 22\ \eV\ \left(\frac{\sigma}{1000\ \kms}\right) \left(\frac{E}{6.7\ \keV}\right).$$ *Suzaku*’s spectrometer has an energy resolution of 130 eV at 6 keV [@Koyama07]. The soft X-ray spectrometer on Astro-H, scheduled for a 2014 launch, will have an energy resolution of 7 eV in the range 0.3 to 12 keV [@Takahashi12]. Thus, it can easily detect the broadening due to turbulence as well as the bulk outflow.
Unfortunately, while the line widths can indicate the presence of the random motions that are likely present due to supernovae, they cannot so easily indicate whether there is a true turbulent cascade. The small-scale motions in a turbulent cascade necessarily have smaller amplitudes, and are harder to distinguish. Furthermore, the angular resolution of X-ray spectrometers is poor, so the Doppler shifts of many eddies will all be blurred together.
The far-infrared–radio correlation in starbursts
------------------------------------------------
Some of the strongest constraints on magnetic and CR densities in starbursts comes from the existence of the far-infrared–radio correlation (FRC). Star-forming galaxies have a nearly constant ratio between their far-infrared luminosities and their 1.4 GHz synchrotron radio luminosity; this is the FRC [@Helou85; @Condon92; @Yun01]. In denser galaxies, the far-infrared luminosity simply traces the star-formation rate. The reason why the synchrotron emission traces the star-formation rate so well is unclear. In starburst regions, $e^{\pm}$ traced by radio emission generally cool before escaping [@Volk89-Calor]. The radio luminosity apparently is set by a conspiracy between the ratio of pionic secondary $e^{\pm}$ to primary $e^-$ fraction [@Rengarajan05; @Thompson07], and the ratio of synchrotron cooling timescales to $e^{\pm}$ lifetime [@Condon91; @Thompson06; @Lacki10-FRC1]. In normal galaxies, by contrast, primary $e^-$ are overwhelmingly responsible for radio emission, $e^-$ mainly escape, and losses like bremsstrahlung are unimportant [@Chi90; @Lisenfeld96-Normal; @Bell03; @Strong10].
@Lacki10-FRC1 and @Lacki10-FRC2 showed that the effects of bremsstrahlung and ionization cooling could balance the presence of secondary $e^{\pm}$, if the CRs experience average density gas. However, for this balancing to take place, the magnetic field strength must go as $$\label{eqn:BFRC-SigmaG}
B_{\rm FRC1} \approx 400\ \muGauss \left(\frac{\Sigma_g}{\gcm2}\right)^{0.7}$$ or $$\label{eqn:BFRC-rho}
B_{\rm FRC2} \approx 390\ \muGauss \left(\frac{{\ensuremath{\langle n \rangle}}}{1000\ \cm^{-3}}\right)^{0.5},$$ if the @Kennicutt98 Schmidt law held (Figure \[fig:B\]).
A turbulent dynamo powered by supernovae mechanical power can explain why equation \[eqn:BFRC-SigmaG\] holds. The normalization and density dependence of the predicted $B$ for both the hot wind and cold gas are roughly in line with the FRC constraint (sections \[sec:HotTurbulence\] and \[sec:ColdTurbulence\]), as shown in Figure \[sec:ModelBComparison\]. In detail, however, there are discrepancies that might shed light on whether supernovae-driven turbulence sets $B$ in starbursts. First, the magnetic field strength is stronger at low density than equation \[eqn:BFRC-SigmaG\] predicts. Second, the magnetic field strength grows more weakly with $\Sigma_{\rm SFR}$ than in equation \[eqn:BFRC-SigmaG\] if $\ell_{\rm outer}$ also shrinks.
To examine these effects, I run one-zone models of the CR population in starbursts, similar to those described in @Lacki10-FRC1, but using the $B_{\rm turb}$ I derived in sections \[sec:HotTurbulence\] and \[sec:ColdTurbulence\]. I consider both a constant $\tau_{\rm gas} = 20\ \Myr$ and the @Kennicutt98 Schmidt Law. From these, I calculate the 1.4 GHz synchrotron emission and spectral index, as well as the 1-100 GeV gamma-ray emission and spectral index. I ignore free-free emission and free-absorption for the radio, and leptonic emission for the gamma rays. I assume an advection speed of $v = 300\ \kms$. Most importantly, I assume that the CRs experience average density gas (${\ensuremath{\langle \rho \rangle}}_{\rm SB}$), as most other one-zone models do.
{width="8cm"}{width="8cm"}
Figure \[fig:FRCt20\] is a plot of the predicted FRC (left) and gamma-ray–SFR correlation (right) in starbursts with $\tau_{\rm gas} = 20\ \Myr$. There is considerable variation in $\nu L_{\nu} (1.4\ \GHz) / L_{\rm bol}$ for different assumptions about the ISM phase and the outer scale of turbulence. At $\Sigma_{\rm SFR} \ga 25\ \Msun\ \yr^{-1}\ \kpc^{-2}$ both the hot $B$ with $\ell_{\rm outer} = h$ and the cold $B$ with $\ell_{\rm outer} = R_{\rm max}$ predict the FRC within a factor of 2. In starbursts with smaller $\Sigma_{\rm SFR}$, advection carries out $e^{\pm}$ before they radiate much radio emission, so the FRC is predicted to be broken in all models. The gamma-ray emission likewise starts off weak at small $\Sigma_{\rm SFR}$ because of advection but grows to a constant fraction of the bolometric luminosity as proton calorimetry sets in. Another universal feature of the models is that the radio/SFR ratio peaks at $\Sigma_{\rm SFR} \approx 150$ – $500\ \Msun\ \yr^{-1}\ \kpc^{-2}$. This is because the ratio of density to $\Sigma_{\rm SFR}$ remains constant for a constant $\tau_{\rm gas}$, instead of decreasing for the K98 Schmidt law in which $\rho \propto \Sigma_{\rm SFR}^{0.7}$. In addition, when $\ell_{\rm outer}$ shrinks with $\Sigma_{\rm SFR}$, $B$ grows more weakly than $\sqrt{\Sigma_{\rm SFR}}$. Because of these effects, bremsstrahlung and ionization losses grow ever more important with $\Sigma_{\rm SFR}$. Finally, the radio spectra grow flatter with increasing $\Sigma_{\rm SFR}$ [@Thompson06]. This is observed [@Condon91; @Clemens08; @Williams10]; whether it is caused by these cooling processes or free-free absorption is debatable [@Clemens10; @Leroy11; @Lacki13-LowNu].
{width="8cm"}
In contrast, when I assume the K98 Schmidt law applies, it is harder to reproduce the observed FRC with supernova-driven turbulence (see Figure \[fig:FRCK98\]). This is because in the K98 Schmidt law, $\Sigma_{\rm SFR}$ is much lower than found in starbursts like M82 and NGC 253 given their $\Sigma_g$. As a result, there is less turbulent energy density, and very small $B$ for low $\Sigma_{\rm SFR}$. Thus the FRC is broken by advection for the hot wind $B$ until $\Sigma_{\rm SFR} \ga 300\ \Msun\ \yr^{-1}\ \kpc^{-2}$. Indeed, @Lacki10-FRC1 noted this difficulty in reproducing the FRC with the K98 Schmidt Law in starbursts with winds. At higher surface densities, the FRC stabilizes. As in the $\tau_{\rm gas} = 20\ \Myr$ models, $\alpha_{1.4}$ decreases and the GeV emission per star-formation increases with $\Sigma_{\rm SFR}$.
How polarized is the synchrotron emission?
------------------------------------------
Synchrotron emission is notable for being highly polarized in an ordered magnetic field. But if the emission comes from isotropic turbulence, the polarization of radiation emitted varies on a sightline. When a region is not even resolved, the fluctuating polarization on different sightlines cancel even further. Thus, synchrotron radiation from highly turbulent starbursts appears depolarized. Polarization can still be observed if the turbulence has a preferred direction throughout the starburst, perhaps through shearing [@Sokoloff98].
The simplest model of radio polarization is a “cell” model, where magnetic fields are coherent over cells of size $\ell_{\rm cell}$. Different cells have uncorrelated polarizations. The polarization along a sightline can be viewed as a random walk away from zero, with each step being the polarization from one cell. On a sightline of length $s$, each cell contributes $\ell_{\rm cell}/s$ of the emission, and there are $s/\ell_{\rm cell}$ cells, so the degree of polarization is roughly $\sqrt{\ell_{\rm cell}/s}$. Assuming the sightline length is the starburst radius and the cell size is the outer scale of turbulence, the polarization towards any resolved sightline through a starburst is of order $\sim 10$ – $50\%$.
The polarization signal nearly vanishes if the starburst is unresolved, though. In this case, each cell contributes only $\ell_{\rm cell}^3/V_{\rm SB}$ of the total emission and there are $V_{\rm SB}^3/\ell$ cells, leading to a polarization of $\sim \sqrt{\ell_{\rm cell}^3/V_{\rm SB}}$. In M82 and NGC 253, the integrated polarization should be $\la 1\%$, and it essentially vanishes in Arp 220.
As discussed in the next section on Faraday rotation, the observed polarization is higher in the GCCMZ.
Faraday rotation and the magnetic field structure
-------------------------------------------------
Faraday rotation is the twisting of a radio wave’s polarization through an ionized, magnetized medium. It is a frequency-dependent effect, but can be parameterized by the rotation measure (RM) that depends on the electron density, sightline length, and magnetic field structure. The necessary free electrons are present in the superwind and H II region, as well as the molecular medium through the process of CR ionization. The magnetic field structure is not known well (see section \[sec:Dynamoes\]). @Bhat13 finds, from simulations of MHD turbulence, that the rotation measures average to zero but have a dispersion in a turbulent magnetic field. The dispersion is $$\sigma_{\rm RM} = 0.81\ {\rm rad}\ {\rm m}^{-2} \bar{\sigma_{\rm RM}} \left(\frac{{\ensuremath{\langle n_e \rangle}}}{\cm^{-3}}\right) \left(\frac{B_{\rm eq}}{2\sqrt{3}\ \muGauss}\right) \frac{\sqrt{\ell_{\rm outer} s}}{\pc}.$$ Here, $B_{\rm eq}$ is the magnetic field strength if $\epsilon_B = 1$, but note that $\epsilon_B < 1$ in @Bhat13. The factor $\bar{\sigma_{\rm RM}}$ equals 0.4, and $s$ is the sightline length.
I take the starburst radius as the sightline length for the hot and cold ISM. For the molecular medium, I assume that the ionization fraction is $x_e = 10^{-5}$, as expected from cosmic ray ionization [@Lacki12-GRDRs]. Then the free electron density is $n_e = 10^{-5} n_H$. In the Galactic Center, the CR ionization rate may be relatively low, however [@Crocker11-Wild]. I list the expected dispersions in Faraday rotation measures in Table \[table:HII-RMs\].
The $\sigma_{\rm RM}$ grow with $\Sigma_{\rm SFR}$, as both $n_e$ and $B$ increase. The superwind has more free electrons, so its Faraday rotation measure is larger. According to these Faraday measures, the typical sightline should become Faraday thick at frequencies $\nu_F = c \sqrt{\rm RM}$ of $\sim 700\ \MHz$ in the GCCMZ, and $\sim 5\ \GHz$ in NGC 253, M82, and Arp 220’s nuclei. At lower frequencies, the frequency dependence of the polarization angle becomes very complicated for the synchrotron emission occurring within the turbulent magnetic fields. The rotation measure itself then fluctuates with wavelength [e.g., @Chi97; @Sokoloff98].[^8]
Observations of the Faraday rotation measure in starburst are complicated by the large number of H II regions present in these areas. Although the magnetic field structures are probably not much different than in the molecular medium, H II regions are both dense and almost completely ionized. I estimate the rotation measure signals toward typical starburst H II regions in Table \[table:HII-RMs\], using a sightline of 5 pc. The regions become Faraday thick at $\nu \gg 10\ \GHz$. Any weaker Faraday signal from the superwind or molecular medium behind a H II region is completely scrambled.
Worse, the H II regions may have fairly large covering fractions, at least in weaker starbursts, despite filling a very small fraction of the volume. H II regions are directly visible in free-free absorption towards the GCCMZ, and appear to cover roughly $\sim 1/2$ the region [@Brogan03; @Nord06]. The covering fraction of H II regions in M82 and NGC 253 is unknown. According to a very preliminary fit I did in @Lacki13-LowNu, a free-free absorption dip in the low frequency spectrum of the M82 starburst measured by @Adebahr13 is consistent with a covering fraction near 1. However, in Arp 220, radio recombination lines indicate very dense H II regions with very small covering fraction [@RodriguezRico05]. The Faraday signal in Arp 220 may therefore be relatively pure.
[lllcccc]{} GCCMZ & Hot & $\ell_{\rm outer} = h$ & 1 & 100 & 7.5 & 0.82\
& & $\ell_{\rm outer} = \ell_1$ & 1 & 100 & 5.3 & 0.69\
& & $\ell_{\rm outer} = \ell_2$ & 1 & 100 & 5.6 & 0.71\
& Cold & $\ell_{\rm outer} = h$; $\Psi = 1$ & $10^{-5}$ & 100 & 10.5 & 0.97\
& & $\ell_{\rm outer} = R_{\rm max}$; $\Psi = 1$ & $10^{-5}$ & 100 & 1.6 & 0.38\
& & $\ell_{\rm outer} = R_{\rm max}$; “Natural” $\Psi$ & $10^{-5}$ & 100 & 1.4 & 0.36\
& H II & $\ell_{\rm outer} = h$; $\Psi = 1$ & 1 & 5 & $2.3 \times 10^5$ & 150\
& & $\ell_{\rm outer} = R_{\rm max}$; $\Psi = 1$ & 1 & 5 & $3.6 \times 10^4$ & 57\
& & $\ell_{\rm outer} = R_{\rm max}$; “Natural” $\Psi$ & 1 & 5 & $3.2 \times 10^4$ & 53\
NGC 253 & Hot & $\ell_{\rm outer} = h$ & 1 & 150 & 940 & 9.2\
& & $\ell_{\rm outer} = \ell_1$ & 1 & 150 & 280 & 5.0\
& & $\ell_{\rm outer} = \ell_2$ & 1 & 150 & 360 & 5.6\
& Cold & $\ell_{\rm outer} = h$; $\Psi = 1$ & $10^{-5}$ & 150 & 13 & 1.1\
& & $\ell_{\rm outer} = R_{\rm max}$; $\Psi = 1$ & $10^{-5}$ & 150 & 1.7 & 0.39\
& & $\ell_{\rm outer} = R_{\rm max}$; “Natural” $\Psi$ & $10^{-5}$ & 150 & 1.5 & 0.37\
& H II & $\ell_{\rm outer} = h$; $\Psi = 1$ & 1 & 5 & $2.4 \times 10^5$ & 150\
& & $\ell_{\rm outer} = R_{\rm max}$; $\Psi = 1$ & 1 & 5 & $3.1 \times 10^4$ & 53\
& & $\ell_{\rm outer} = R_{\rm max}$; “Natural” $\Psi$ & 1 & 5 & $2.8 \times 10^4$ & 50\
M82 & Hot & $\ell_{\rm outer} = h$ & 1 & 300 & 1300 & 11\
& & $\ell_{\rm outer} = \ell_1$ & 1 & 300 & 390 & 5.9\
& & $\ell_{\rm outer} = \ell_2$ & 1 & 300 & 510 & 6.8\
& Cold & $\ell_{\rm outer} = h$; $\Psi = 1$ & $10^{-5}$ & 300 & 34 & 1.7\
& & $\ell_{\rm outer} = R_{\rm max}$; $\Psi = 1$ & $10^{-5}$ & 300 & 4.1 & 0.61\
& & $\ell_{\rm outer} = R_{\rm max}$; “Natural” $\Psi$ & $10^{-5}$ & 300 & 3.6 & 0.57\
& H II & $\ell_{\rm outer} = h$; $\Psi = 1$ & 1 & 5 & $4.4 \times 10^5$ & 200\
& & $\ell_{\rm outer} = R_{\rm max}$; $\Psi = 1$ & 1 & 5 & $5.3 \times 10^4$ & 69\
& & $\ell_{\rm outer} = R_{\rm max}$; “Natural” $\Psi$ & 1 & 5 & $4.7 \times 10^4$ & 65\
Arp 220 Nuclei & Hot & $\ell_{\rm outer} = h$ & 1 & 100 & $4.0 \times 10^5$ & 190\
& & $\ell_{\rm outer} = \ell_1$ & 1 & 100 & $3.8 \times 10^4$ & 59\
& & $\ell_{\rm outer} = \ell_2$ & 1 & 100 & $5.8 \times 10^4$ & 72\
& Cold & $\ell_{\rm outer} = h$; $\Psi = 1$ & $10^{-5}$ & 100 & 7100 & 25\
& & $\ell_{\rm outer} = R_{\rm max}$; $\Psi = 1$ & $10^{-5}$ & 100 & 300 & 5.2\
& & $\ell_{\rm outer} = R_{\rm max}$; “Natural” $\Psi$ & $10^{-5}$ & 100 & 300 & 5.2\
& H II & $\ell_{\rm outer} = h$; $\Psi = 1$ & 1 & 5 & $1.6 \times 10^8$ & 3800\
& & $\ell_{\rm outer} = R_{\rm max}$; $\Psi = 1$ & 1 & 5 & $6.6 \times 10^6$ & 770\
& & $\ell_{\rm outer} = R_{\rm max}$; “Natural” $\Psi$ & 1 & 5 & $6.6 \times 10^6$ & 770 \[table:HII-RMs\]
We can compare these Faraday rotation dispersions to a few observations, particularly in the well-studied GCCMZ. Sightlines through the GCCMZ have very large RMs, reaching $\sim 3000\ \radm2$ on some sightlines [e.g., @Roy05; @Law11]. These values are completely inconsistent with my predictions for the superwind and molecular medium, but they are smaller than the predicted H II region RMs. However, the observed rotation measures evince a large-scale regular field on scales $\ga 150\ \pc$, with a twisted poloidal structure [@Law11]. The synchrotron emission from the mysterious nonthermal filaments within the GCCMZ is polarized, with both the filaments themselves and their polarization aligned perpendicular to the Galactic plane, again supporting a poloidal magnetic field [@YusefZadeh84; @Tsuboi86].
Other starbursts display evidence for large-scale fields too. Using Faraday rotation measurements, @Heesen11 discovered a helical magnetic field within the wall of NGC 253’s starburst outflow (although not in the starburst itself) using RMs. In M82 itself, polarized submillimeter emission indicates the alignment of dust with a large scale magnetic field [@Greaves00; @Jones00]. Regular magnetic fields are completely unexpected in a generic turbulence model, although the wind introduces a preferred direction. A preferred direction in the density distribution of gas was invoked by @Boldyrev06, who propose the Galactic Center filaments are structures made by supernova-driven turbulence, to explain the alignment of the filaments.
@Robishaw08 observed Faraday rotation towards OH megamasers within Arp 220 and other ULIRGs. While the significance is very weak in Arp 220, the mean RM $\sim 1250\ \radm2$ is roughly expected from the molecular medium. On the other hand, @Robishaw08 found much larger RMs $\gg 10^4\ \radm2$ in III Zw 35; these are hard to explain.
The magnetic fields within starburst SNRs
-----------------------------------------
High resolution radio observations have revealed populations of SNRs in starbursts (Section \[sec:SNREvolution\]). By assuming that equipartition between CRs and magnetic fields holds, one can estimate the magnetic fields within the SNRs from their radio flux. These estimates imply that magnetic field strengths are greater than the ambient ISM [e.g., @Batejat11].
Indeed, we expect for theoretical reasons that the post-shock magnetic field in a SNR is greater than the ambient ISM magnetic field. First, the shock itself compresses the incoming gas and its frozen-in magnetic fields, increasing $B$ by a factor $f_{\rm comp} \sim 3$–7 [@vanDerLaan62; @Reynolds81; @Volk02]: $$B_{\rm SNR}^{\rm comp} = f_{\rm comp} B_{\rm ISM}.$$ Second, plasma instabilities lead to magnetic field amplification in the post-shock region. Through this process, the shock converts some fraction $\epsilon_{\rm amp} \approx 0.01$ of the upstream material’s kinetic energy density (in the shock frame) into magnetic fields [e.g., @Lucek00; @Berezhko04]: $$B_{\rm SNR}^{\rm amp} = \sqrt{8 \pi \epsilon_{\rm amp} \rho_{\rm ext} v_s^2}$$ where $v_s$ is the SNR shock speed and $\rho_{\rm ext}$ is the density of the material the shock is expanding into. Note that $\rho_{\rm ext}$ is not necessarily the mean density of the entire starburst (${\ensuremath{\langle \rho \rangle}}_{\rm SB}$) – the SNR sits in some ISM phase that can be overdense or underdense, and there can be fluctuations in the phase’s density (c.f. Section \[sec:ColdTurbulence\]). Which of these processes sets the SNR magnetic field strength in starbursts is a matter of debate [@Thompson09; @Chomiuk09; @Batejat11].
I address this question using the $B$ estimates developed in this paper. Suppose the SNR goes off in a phase with average density ${\ensuremath{\langle \rho \rangle}}_{\rm phase}$. Then, the ambient magnetic field energy density is $(\epsilon_B/2) {\ensuremath{\langle \rho \rangle}}_{\rm phase} \sigma^2$. By analogy with Section \[sec:ColdTurbulence\], we can define the ratio of the ISM density locally surrounding the SNR with the average density of that ISM phase: $\Delta \equiv \rho_{\rm ext} / {\ensuremath{\langle \rho \rangle}}_{\rm phase}$. For the hot wind plasma, which has subsonic or transonic turbulence, $\Delta \approx 1$. In the cold molecular gas, where the turbulence has high Mach number, the median $\Delta$ could be in the range $0.01$–$1$, depending on the unknown density distribution of gas [@Padoan97; @Ostriker01; @Hopkins13-rhoDist]. Then the ratio of the energy densities in the compressed magnetic field and in the amplified magnetic field are: $$\begin{aligned}
\nonumber \frac{U_B^{\rm comp}}{U_B^{\rm amp}} & = \frac{f_{\rm comp}^2 \epsilon_B}{2 \epsilon_{\rm amp} \Delta} \left(\frac{\sigma}{v_S}\right)^2\\
& = 800 \left(\frac{f_{\rm comp}}{4}\right)^2 \left(\frac{\epsilon_{\rm amp}}{0.01}\right)^{-1} \left(\frac{\epsilon_B}{\Delta}\right) \left(\frac{\sigma}{v_S}\right)^2.\end{aligned}$$
The most important factor controlling whether the SNR magnetic field is dominated by compressed or amplified magnetic fields is the ratio of $\sigma$ and $v_S$. For young SNRs, the shock speed is very large and typically of order $10^4\ \kms$. But the ISM turbulent speed varies by a factor $>10$ between the hot superwind ($\sigma \sim 1000\ \kms$) and the cold molecular gas ($\sigma \sim 50\ \kms$). Thus, which mechanism is most important depends on SNR location within a starburst. In the hot superwind, compression of magnetic fields dominates: $U_B^{\rm comp} / U_B^{\rm amp} \approx 8$. In cold molecular gas, post-shock amplification wins: $U_B^{\rm comp} / U_B^{\rm amp} \approx 0.02$.
In a hot starburst, the typical supernova goes off in the volume-filling superwind. The majority of SNRs therefore have radio fluxes set by the compression of hot wind magnetic fields. However, the *brightest* SNRs with the strongest magnetic fields probably are located within the overdense molecular clouds, where post-shock amplification operates [as suggested by @Thompson09]. The implication is that the radio luminosity function of SNRs in starbursts arises from a complicated distribution of environments; why the observed radio luminosity function seems to be the same between galaxies and starbursts is unclear [@Chomiuk09]. In cold starbursts, amplification alone determines SNR magnetic fields until the SNR shocks have slowed down below $5000\ \kms$, as long as $\Delta \ga 1$.
[lllcccccc]{} GCCMZ & Hot & $\ell_{\rm outer} = h$ & $4 \times 10^{-5}$ & 1 & 310$\ \muGauss$ & 21$\ \muGauss$ & 42$\ \muGauss$ & 84$\ \muGauss$\
& & $\ell_{\rm outer} = \ell_1$ & $4 \times 10^{-5}$ & 1 & 270$\ \muGauss$ & 21$\ \muGauss$ & 42$\ \muGauss$ & 84$\ \muGauss$\
& & $\ell_{\rm outer} = \ell_2$ & $4 \times 10^{-5}$ & 1 & 270$\ \muGauss$ & 21$\ \muGauss$ & 42$\ \muGauss$ & 84$\ \muGauss$\
& Cold & $\ell_{\rm outer} = h$; $\Psi = 1$ & 1 & 1 & 1.6$\ \mGauss$ & 3.2$\ \mGauss$ & 6.4$\ \mGauss$ & 13$\ \mGauss$\
& & $\ell_{\rm outer} = R_{\rm max}$; $\Psi = 1$ & 1 & 1 & 770$\ \muGauss$ & 3.2$\ \mGauss$ & 6.4$\ \mGauss$ & 13$\ \mGauss$\
& & $\ell_{\rm outer} = R_{\rm max}$; “Natural” $\Psi$ & 0.12 & 0.03 & 530$\ \muGauss$ & 1.1$\ \muGauss$ & 2.2$\ \mGauss$ & 4.4$\ \mGauss$\
NGC 253 & Hot & $\ell_{\rm outer} = h$ & 0.002 & 1 & 1.4$\ \mGauss$ & 100$\ \muGauss$ & 200$\ \muGauss$ & 400$\ \muGauss$\
& & $\ell_{\rm outer} = \ell_1$ & 0.002 & 1 & 850$\ \muGauss$ & 100$\ \muGauss$ & 200$\ \muGauss$ & 400$\ \muGauss$\
& & $\ell_{\rm outer} = \ell_2$ & 0.002 & 1 & 930$\ \muGauss$ & 100$\ \muGauss$ & 200$\ \muGauss$ & 400$\ \muGauss$\
& Cold & $\ell_{\rm outer} = h$; $\Psi = 1$ & 4 & 1 & 3.8$\ \mGauss$ & 2.1$\ \mGauss$ & 4.2$\ \mGauss$ & 8.5$\ \mGauss$\
& & $\ell_{\rm outer} = R_{\rm max}$; $\Psi = 1$ & 4 & 1 & 1.7$\ \mGauss$ & 2.1$\ \mGauss$ & 4.2$\ \mGauss$ & 8.5$\ \mGauss$\
& & $\ell_{\rm outer} = R_{\rm max}$; “Natural” $\Psi$ & 0.12 & 0.03 & 1.2$\ \mGauss$ & 730$\ \muGauss$ & 1.5$\ \mGauss$ & 2.9$\ \mGauss$\
M82 & Hot & $\ell_{\rm outer} = h$ & 0.001 & 1 & 1.3$\ \mGauss$ & 100$\ \muGauss$ & 210$\ \muGauss$ & 410$\ \muGauss$\
& & $\ell_{\rm outer} = \ell_1$ & 0.001 & 1 & 800$\ \muGauss$ & 100$\ \muGauss$ & 210$\ \muGauss$ & 410$\ \muGauss$\
& & $\ell_{\rm outer} = \ell_2$ & 0.001 & 1 & 890$\ \muGauss$ & 100$\ \muGauss$ & 210$\ \muGauss$ & 410$\ \muGauss$\
& Cold & $\ell_{\rm outer} = h$; $\Psi = 1$ & 4 & 1 & 3.9$\ \mGauss$ & 2.8$\ \mGauss$ & 5.6$\ \mGauss$ & 11$\ \mGauss$\
& & $\ell_{\rm outer} = R_{\rm max}$; $\Psi = 1$ & 4 & 1 & 1.7$\ \mGauss$ & 2.8$\ \mGauss$ & 5.6$\ \mGauss$ & 11$\ \mGauss$\
& & $\ell_{\rm outer} = R_{\rm max}$; “Natural” $\Psi$ & 0.12 & 0.03 & 1.2$\ \mGauss$ & 970$\ \muGauss$ & 1.9$\ \mGauss$ & 3.9$\ \mGauss$\
Arp 220 Nuclei & Hot & $\ell_{\rm outer} = h$ & 0.002 & 1 & 12$\ \mGauss$ & 800$\ \muGauss$ & 1.6$\ \mGauss$ & 3.2$\ \mGauss$\
& & $\ell_{\rm outer} = \ell_1$ & 0.002 & 1 & 4.5$\ \mGauss$ & 800$\ \muGauss$ & 1.6$\ \mGauss$ & 3.2$\ \mGauss$\
& & $\ell_{\rm outer} = \ell_2$ & 0.002 & 1 & 5.3$\ \mGauss$ & 800$\ \muGauss$ & 1.6$\ \mGauss$ & 3.2$\ \mGauss$\
& Cold & $\ell_{\rm outer} = h$; $\Psi = 1$ & 1 & 1 & 33$\ \mGauss$ & 18$\ \mGauss$ & 37$\ \mGauss$ & 74$\ \mGauss$\
& & $\ell_{\rm outer} = R_{\rm max}$; $\Psi = 1$ & 1 & 1 & 9.2$\ \mGauss$ & 18$\ \mGauss$ & 37$\ \mGauss$ & 74$\ \mGauss$\
& & $\ell_{\rm outer} = R_{\rm max}$; “Natural” $\Psi$ & 0.01 & 0.01 & 5.8$\ \mGauss$ & 1.8$\ \mGauss$ & 3.7$\ \mGauss$ & 7.4$\ \mGauss$ \[table:BSNRPredictions\]
I have listed the predicted $B_{\rm SNR}$ for both magnetic field compression and amplification in Table \[table:BSNRPredictions\]. The SNR magnetic fields are expected to be intense, ranging from a few hundred $\muGauss$ in the GCCMZ to nearly 0.1 G in the youngest SNRs in Arp 220. The magnetic fields are typically stronger in the cold molecular gas by an order of magnitude. The magnetic field strengths in Arp 220’s SNRs, which are observed to have expansion speeds of $\sim 5000\ \kms$ [@Batejat11], are roughly $40\ \mGauss$. This is essentially the equipartition estimate of @Batejat11 from the SNRs’ radio fluxes.
Conclusion {#sec:Conclusion}
==========
The power of supernovae and stellar winds can drive the chaos of the interstellar medium in starbursts. The mechanical luminosity couples efficiently to the gas, stirring it and driving turbulence. The high speed random motions amplify magnetic fields in all phases of the ISM. Supernovae shocks accelerate CRs, which both provide a source of ionization and are observational tracers of the magnetic fields that result from the stirring.
In low density, low pressure (“hot”) starbursts, the expanding supernova remnants successfully sweep up most of the ISM and leave behind rarefied, hot plasma that erupts as a wind. Further supernova explosions both provide needed mass to sustain the wind and, I argue, turbulent driving in the hot plasma. The winds are not thermalized gases as usually pictured, but are typically collisionless on parsec scales. The conditions in the wind are analogous to galaxy clusters, where turbulence and magnetic fields are well studied. Although the hydrodynamical Reynolds number is very low, plasma processes could decrease the viscosity and permit turbulence. I show that the hot wind’s turbulence reaches speeds of $\sim 1000\ \kms$ (${\cal M} \approx 1$). The magnetic field strength grows through the fluctuation dynamo, reaching $\sim 80\ \muGauss$ in the GCCMZ and $\sim 300\ \muGauss$ in M82 and NGC 253. The dissipation of the turbulence plausibly provides the efficient heating that sustains the wind.
High pressure (“cold”) starbursts confine SNRs and are filled with dense molecular gas (Figure \[fig:ISMSketches\]). Although expanding supernova remnants lose much of their mechanical energy through radiation in molecular clouds and in cold starbursts, they still are able to drive turbulence of speeds $\sim 20\ \kms$. The equipartition magnetic fields are a few times higher than those predicted in the superwind. In the extreme conditions of Arp 220’s nuclei, I predict turbulent magnetic fields with strengths of $\sim 2\ \mGauss$. Similar levels of turbulence and magnetic field strengths should hold for warm neutral gas and H II regions.
I argue that CRs in hot starbursts spend most of their time in the hot wind. The observed synchrotron emission largely reflects the magnetic fields in these winds. From time to time, they dive into dense molecular gas where they emit gamma rays [c.f., @YoastHull13].
The importance of supernova mechanical power in starbursts is reflected in the near equipartition of energy densities in starbursts. Turbulence, magnetic fields, CRs, and the thermal energy of the superwind all are powered by supernovae. Since their residence time is typically of order the hot wind sound crossing time of the starburst, they naturally tend to equipartition. In the special case of CRs, equipartition should roughly hold in hot starbursts, but it likely fails in dense starbursts when their residence time is instead set by the pion loss time.
The strong levels of turbulence and the small outer scales possible in starbursts should confine CRs very effectively if the turbulent cascade reaches down to small scales. In fact, the motion of the ISM may be more important than the CR diffusion within the ISM. Thus, CR propagation may be Lagrangian in starbursts instead of Eulerian as in the Milky Way. Even if there is no way of confining CRs, the CRs are forced to follow the tortuous paths of the turbulent magnetic field lines themselves, giving a FLRW diffusion constant. Thus, turbulence rules how CRs propagates in starbursts.
In Section \[sec:Implications\], I considered some observational implications of this scenario. First, X-ray lines from starburst wind should display turbulent broadening roughly comparable to their thermal broadening. Second, I demonstrated that the predicted turbulent magnetic fields are roughly consistent with the FIR-radio correlation. Third, if the magnetic field is indeed mostly turbulent, the synchrotron emission should be depolarized, especially when integrated over the entire starburst. Fourth, I calculate Faraday rotation measures for each of the phases of the starburst ISM. Fifth, there should be milliGauss magnetic fields in starburst SNRs; magnetic field compression sets $B$ in the hot superwind, whereas post-shock amplification sets it in cold molecular gas. Finally, as I noted in section \[sec:TurbulentMixing\], if turbulent advection is the main mode of CR transport, then it may result in fluctuations in the synchrotron emission with no obvious correlate.
Many questions remain about how turbulence, CRs, and magnetic fields interplay in starburst environments.
- First, what is the true ISM structure of starbursts? Does hot superwind plasma fill weaker starbursts like M82 and the GCCMZ? Does cold molecular gas really fill ULIRGs? I argued that CRs fill most of the volume of the starbursts, and their radio synchrotron emission indicates the magnetic field strength in the volume-filling phase. Is that true?
- What is the nature of turbulence in the hot superwind plasma, if it is present? Are there really energy cascades down in spatial scales? Does the kinetic energy end up as heat? If so, the microphysics of turbulence may leave subtle observational signs, such as heating ions and electrons at different rates. For example, maybe turbulently heated electrons explains the anomalous hard X-ray continuum emission of M82 [@Strickland07].
- I evaluated the turbulence and its equipartition magnetic fields in superwinds with properties from CC85 solution. *Chandra* X-ray observations of M82 suggest the solution is adequate for real starbursts [@Strickland09]. But formally speaking, the CC85 solution treats the superwind as an ideal gas and assumes no magnetic fields, cosmic rays, or turbulence. There is need to simulate the superwind plasma self-consistently. Do collisionless effects or strong magnetic fields alter its properties?
- How does the turbulent dynamo operate in starbursts? While there have been numerous MHD simulations of turbulent dynamoes, starbursts differ in that they have large-scale advective winds. Winds twist motions, so even the purely compressive motions from supernovae explosions may be partly transformed into solenoidal motions as they are blasted by the wind. A simulation of that effect should be illuminating.
- What is the role of intermittence in starburst turbulence? What sort of magnetic field structures does it produce? Is the assumption of a single magnetic field strength in each phase adequate for models? Does it lead to unsteady CR destruction and time variability in nonthermal emission?
- What is the spectrum of magnetic fluctuations? Is most of the energy concentrated at large scales (as in the Kolmogorov spectrum)? If so, then Faraday rotation measurements may prove useful. Does the spectrum fall slowly enough to be very good at trapping CRs (as in the Kolmogorov and Kraichnan cases)?
- What is the role of turbulent mixing of gas/plasma in transporting CRs? It would be interesting to construct “Lagrangian” models of CR transport in starbursts, as opposed to the “Eulerian” models we have now like GalProp. Are there order unity fluctuations in the CR energy density that arise simply due to turbulent eddies? These would be entirely uncorrelated with star-formation sites. In supersonic cold gas, if the CRs are frozen into the gas, they essentially live for one eddy-crossing time, before being destroyed when advected into a turbulent clump. How do such transient phenomena square with current steady-state models of CR populations?
- What happens to the turbulent cascade at small scales in cold molecular gas? Are there enough magnetic fluctuations to freeze CRs into the gas? If not, what traps the CRs in ULIRGs – or do they stream away entirely? The situation is analogous to that for MeV positrons in the Milky Way [@Higdon09; @Jean09; @Prantzos11]. Observations of CR propagation in Galactic molecular clouds, especially those in the GCCMZ, could prove helpful.
- Although I discuss turbulence in separate “phases” of the ISM, the reality is probably a continuum of physical conditions [e.g., @Norman96]. How does turbulence work in the chaotically varying conditions of the ISM? Are CRs transported between regions of differing physical conditions, and if so, how does that mixing between “phases” work? @Crocker11-Wild argued that CRs in the GCCMZ do not generally penetrate into dense molecular clouds, based on the gamma-ray faintness of the region. Yet modeling of M82 and NGC 253’s gamma-ray brightness is consistent with CRs experiencing average density material [@Lacki11-Obs]. What microphysics is behind this difference, and how does the relationships between the “phases” come in?
- How analogous is turbulence in $z \approx 0$ starbursts to turbulence in high-redshift disk galaxies? Is turbulence in high-redshift galaxies powered by supernovae? Is the ISM in those galaxies hot or cold? What is the outer scale of turbulence in those galaxies?
Turbulence shapes nearly everything about starburst ISMs, and understanding its role will require synthesizing observations of very different kinds and theories of different processes. Radio observations of synchrotron emission tell us about the turbulent magnetic fields in starbursts, and with high enough resolution, the propagation of CRs. VLBI monitoring of radio supernovae illuminate the sources of the turbulence, the ISM they go off in, and the CRs they accelerate. Observations in molecular lines trace the distribution and kinematics of the stirred molecular gas. With infrared emission, we see the sites of star-formation. The superwind glows in hard X-rays, and X-ray spectroscopy may tell us about its kinematics. Finally gamma rays (and eventually neutrinos) tell us about the CR protons and their interaction with dense gas. On the theory side, a complete theory of starburst turbulence will draw upon fluid dynamics, plasma physics, magnetic dynamo theory, theories of interstellar feedback, and the physics of CRs. Many years of discoveries and further questions await in this area.
I was supported by a Jansky Fellowship for this work from the National Radio Astronomy Observatory. The National Radio Astronomy Observatory is operated by Associated Universities, Inc., under cooperative agreement with the National Science Foundation. The motivation for this paper largely arose from discussions with Todd Thompson while I was at Ohio State University. An early version of this work was delivered at the 2012 Sant Cugat Forum for Astrophysics; I wish to acknowledge the participants. I also thank Jim Condon, Jean Eilek, Josh Marvil, Mark Krumholz, Roger Chevalier, Casey Law, and Chiara Ferrari for discussions and comments.
Bulk properties of the hot superwind in CC85 {#sec:CC85Properties}
============================================
The theory of supernovae-driven superwinds has been thoroughly developed [see @Chevalier85; @Silich04; @Strickland09]. The characteristics of the superwind are given by the rate of hot matter being injected into it ($\dot{M}$) and the energy pumped into the wind ($\dot{E}_{\rm mech}$) by stellar winds and supernovae. @Strickland09 gives the mass injection rate from supernovae and stellar winds as $$\dot{M} = 0.12\ \Msun\ \yr^{-1}\ \left(\frac{\rm SFR}{\Msun\ \yr^{-1}}\right)$$ and the energy injection rate as $$\label{eqn:EDot}
\dot{E}_{\rm mech} = 2.5 \times 10^{41} \ergps\ \left(\frac{\rm SFR}{\Msun\ \yr^{-1}}\right).$$ For these equations, I have converted to a Salpeter IMF from 0.1 to 100 $\Msun$, as used by @Kennicutt98, using ${\rm SFR} (\ge 1\ \Msun) = 0.39 \times {\rm SFR}$. While starbursts may have a non-Salpeter IMF [e.g., @Papadopoulos11-SF; @vanDokkum10], both $\dot{M}$ and $\dot{E}_{\rm mech}$ are set by the formation rate of massive stars, for which the IMF should not vary; supernovae in particular are only generated by stars with masses above $\sim 8\ \Msun$. These massive stars are also the power sources of the radiation used as star-formation indicators (such as the bolometric luminosity, ultraviolet light, radio continuum, and radio recombination lines), so for star-formation rates derived from these tracers, the derived $\dot{E}_{\rm mech}$ and $\dot{M}$ of a starburst should not depend on IMF, as long as it consistent throughout [c.f., @Condon92].
For the physical conditions in the superwind, I use the simple model of CC85, which assumes a spherical geometry with no gravity from the starburst and no radiative cooling from the wind. The star-formation is assumed to be entirely contained and uniform within a radius $R$. Elaborations on this model have been given to include cooling [@Silich04], but the basic arguments should remain the same. After adjusting for geometry [@Strickland09], the central density of a CC85 superwind is parameterized by an effective radius $R_{\rm eff} = f_{\rm geom} R / \sqrt{2}$, where $f_{\rm geom} = \sqrt{1 + 2 h / R}$ for a thin disk (with total surface area $A = 2 \pi R^2 + 4 \pi R h$) and $\sqrt{2}$ for a sphere (with total surface area $4 \pi R^2$). The central density of the superwind in CC85 is $$\label{eqn:rhoC}
\rho_c = 1.860 \frac{\zeta \beta^{3/2}}{f_{\rm geom}^2 \epsilon_{\rm therm}^{1/2}} \frac{\dot{M}^{3/2}}{\dot{E}_{\rm mech}^{1/2} A_{\rm proj}}$$ where $\beta$ is the mass loading, $\zeta$ is the participation fraction, $\epsilon_{\rm therm}$ is the supernova thermalization efficiency, and $A_{\rm proj} = \pi R^2$ is the projected area of the star-forming region. This is for an ideal gas with $\gamma = 5/3$. Note that the density is directly proportional to the mean surface density of star-formation, $\Sigma_{\rm SFR} = {\rm SFR} / A_{\rm proj}$. From observations of diffuse, hard X-ray emitting gas in M82, @Strickland09 find that $\beta \approx 2$, $\zeta \approx 1$, and $\epsilon_{\rm therm} \approx 0.75$. The central number density of electrons, $n_e = n_H + 2 n_{\rm He} = X \rho / m_H + Y \rho / m_{\rm He}$, is $$\label{eqn:nC}
n_e = 0.013\ \cm^{-3} \frac{\zeta}{f_{\rm geom}^2}\ \left(\frac{\beta}{2}\right)^{3/2} \left(\frac{\epsilon_{\rm therm}}{0.75}\right)^{-1/2} \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right),$$ for $X = 0.75$ and $Y = 0.25$. The total particle density is $n = n_e + n_H + n_{\rm He}$: $$\label{eqn:nAll}
n = 0.026\ \cm^{-3} \frac{\zeta}{f_{\rm geom}^2}\ \left(\frac{\beta}{2}\right)^{3/2} \left(\frac{\epsilon_{\rm therm}}{0.75}\right)^{-1/2} \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right).$$
In the CC85 model, the central temperature of the superwind is given by the ratio of the energy and mass injection rates, which should be the same for all starbursts: $$\label{eqn:SuperwindTc}
T_c = \frac{(\gamma - 1) \mu m_H \epsilon_{\rm therm} \dot{E}}{\gamma k_B \beta \dot{M}} = 3.7 \times 10^7\ \Kelv \left(\frac{\epsilon_{\rm therm}}{0.75}\right) \left(\frac{\beta}{2}\right)^{-1}$$ I assume $\mu = 1/(2 X + 3 Y / 4) = 0.59$, appropriate for a completely ionized plasma, since the superwind is extremely hot. The associated sound speed is also universal in the CC85 model: $$\label{eqn:SoundSpeed}
c_s = \sqrt{\frac{(\gamma - 1) \epsilon_{\rm therm} \dot{E}}{\beta \dot{M}}} = 930\ \kms\ \left(\frac{\epsilon_{\rm therm}}{0.75}\right)^{1/2} \left(\frac{\beta}{2}\right)^{-1/2}.$$
Finally, the central thermal pressure in the CC85 model is $P_c = n k_B T$: $$\begin{aligned}
P_c / k_B & = & 9.5 \times 10^{5}\ \Kelv\ \cm^{-3}\ \frac{\zeta}{f_{\rm geom}^2}\ \left(\frac{\epsilon_{\rm therm}}{0.75}\right)^{1/2} \left(\frac{\beta}{2}\right)^{1/2} \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right).\end{aligned}$$
Addressing objections to the existence of a volume-filling hot phase {#sec:HotObjections}
====================================================================
Do X-rays rule out a wind in the GCCMZ?
---------------------------------------
The Galactic Ridge shines in seemingly-diffuse hard X-ray (2 - 10 keV) emission, and is especially bright in the GCCMZ [@Worrall82; @Koyama89]. While a diffuse X-ray hot plasma was considered a possible source, the required to keep the plasma hot enough is implausibly high [@Muno04]. A deep Chandra image presented in @Revnivtsev09 largely settled the matter by demonstrating that much of the X-ray emission, especially at higher energies ($6 - 8\ \keV$) actually comes from discrete sources. But do these observations rule out a superwind from the CMZ [c.f., @Crocker12]?
If the basic theory of CC85 is correct, and the temperature of the superwind is $3.7 \times 10^7\ \Kelv$ (eqn. \[eqn:SuperwindTc\]), the free-free emission per unit volume in 2 - 10 keV X-rays is $$\label{eqn:HardXRayEmissivity}
\varepsilon_{2-10} = 8.5 \times 10^{28} \erg\ \cm^{-3}\ \sec^{-1} f_{\rm geom}^{-4} \left(\frac{\Sigma_{\rm SFR}}{\Msun\ \yr^{-1}\ \kpc^{-2}}\right)^2.$$ The surface density of star-formation in the GCCMZ is $2.2\ \Msun\ \yr^{-1}\ \kpc^{-2}$, and with a radius of $\sim 100\ \pc$, the typical sightline has length $s \sim 100\ \pc$. That means that, if the superwind exists with a CC85-like density, the 2 – 10 keV X-ray intensity observed at Earth from its free-free emission should be $s \varepsilon_{2-10} / (4 \pi) \approx 7.9 \times 10^{-12}\ \erg\ \cm^{-2}\ \sec^{-1}\ \deg^{-2}$. This is just $9\%$ of the total 2 – 10 keV emission from the region [@Revnivtsev09]. For comparison, @Revnivtsev09 calculate that about half of the total emission is resolved at $2 - 3\ \keV$, increasing to $\sim 90 \%$ at $6 - 8\ \keV$. Note that the temperature of the superwind is 3.2 keV, so if the remaining unresolved emission is from the superwind we expect it to fall off at higher energies.
The observations of @Revnivtsev09 therefore are consistent with the existence of a CC85 superwind, if the unresolved emission is partly free-free emission from the superwind: we would not expect the superwind to be as bright as the Galactic X-ray Ridge. The weak constraints on the superwind go back to the energetics issue. The Galactic X-ray Ridge is too bright to be powered by known sources, but a CC85 superwind necessarily can be powered by the supernovae in the CMZ [c.f., @Crocker12].
Do X-ray observations rule out a volume-filling hot phase in Arp 220 and other ULIRGs?
--------------------------------------------------------------------------------------
Although the huge pressures in dense starbursts likely prevent hot winds from filling their volumes, @Murray10 went a step further and argued that the existence of the correlation between hard X-ray emission and star-formation rate ($L_{2-10} \approx 10^{-4} L_{\star}$; @David92 [@Lehmer10]) actually rules out hot superwinds in starbursts with gas surface densities $\Sigma_g \ga 0.15\ \gcm2$ (taking a temperature of $3.7 \times 10^7\ \Kelv$), comparable to M82. However, they assumed that the thermal pressure of the wind was $\pi G \Sigma_g^2$; as I showed in Section \[sec:UStarburst\], that estimate gives problematically high pressures for dense starbursts.
Using equation \[eqn:HardXRayEmissivity\] instead, I calculate a 2 – 10 keV luminosity of $1.0 \times 10^8\ \Lsun$ from free-free emission of both nuclei of Arp 220, or $6 \times 10^{-5}\ L_{\rm TIR}$. Since Arp 220 actually is X-ray faint compared to the X-ray–SFR correlation [@Iwasawa05; @Lehmer10], and since X-ray binaries should contribute an additional $10^{-4}\ L_{\rm TIR}$ [@Grimm03; @Persic04], this is problematic. On the other hand, Arp 220’s nuclei are Compton thick, reducing the X-ray emission by a factor of a few [@Downes98]. I conclude there is some tension between the X-ray emission of Arp 220 and the prediction for a CC85, but not enough to actually rule out the superwind observationally. For M82, where diffuse hard X-ray emission is observed directly, there is no problem with the X-ray emission. I predict its superwind’s diffuse 2 - 10 keV X-ray luminosity is $\sim 1 \times 10^5\ \Lsun$, only $2 \times 10^{-6} L_{\rm TIR}$ and just one tenth of the amount of diffuse, unresolved X-ray emission observed from M82’s starburst [@Strickland07]. (The observed continuum may be brighter because (1) the wind’s transonic turbulence makes it clumpy, and/or (2) the electron temperature is greater than $T_c$ because the ion-electron equilibrium time is greater than the advection time; I argued that both are plausible previously in the paper.)
The X-ray-SFR correlation *does* become constraining for extreme starbursts that are very big, like those in some submillimeter galaxies (SMGs). Whereas the free-free emission from the wind grows as $L_{\rm ff} \propto n_e^2 R^3 \propto \Sigma_{\rm SFR}^2 R^3$, the bolometric emission only grows as $L_{\rm TIR} \propto {\rm SFR} \propto \Sigma_{\rm SFR} R^2$. Thus, free-free emission becomes more dominant for starbursts with high $\Sigma_{\rm SFR}$ and large radius: $L_{\rm ff} / L_{\rm TIR} \propto \Sigma_{\rm SFR} R$. SMGs have typical radii of about two kiloparsecs [@Tacconi06] as opposed to the $\sim 100\ \pc$ scales of compact ULIRGs observed at $z \approx 0$, and SMG surface densities can reach those of Arp 220 [@Walter09]. Furthermore, those without an AGN lie on the X-ray-SFR correlation, demonstrating that they do not have a CC85-like superwind [@Alexander05]. Finally, the winds in such galaxies would emit so much free-free radiation that they cool radiatively before escaping, stalling the wind and invalidating the CC85 theory [@Silich10].
To summarize, the X-ray emission observed from starbursts is perfectly consistent with a superwind existing in M82-like starbursts. In compact ULIRGs observed at $z \approx 0$, the X-ray-SFR correlation is not a conclusive disproof of the superwind phase, although the superwind would have to be a major contributor to the hard X-ray luminosity. Extended ULIRGs as at $z \approx 2$ are indeed too X-ray dim to host a hot superwind.
Abdo, A. A., Ackermann, M., Ajello, M., et al. 2010, , 709, L152
Abramowski, A., Acero, F., Aharonian, F., et al. 2012, , 757, 158
Acciari, V. A., Aliu, E., Arlen, T., et al. 2009, , 462, 770
Acero, F., Aharonian, F., Akhperjanian, A. G., et al. 2009, Science, 326, 1080
Ackermann, M., Ajello, M., Allafort, A., et al. 2012, , 755, 164
Actis, M., Agnetta, G., Aharonian, F., et al. 2011, Experimental Astronomy, 32, 193
Adebahr, B., Krause, M., Klein, U., Weżgowiec, M., Bomans, D. J., Dettmar, R.-J. 2013, , 555, A23
Agertz, O., Lake, G., Teyssier, R., Moore, B., Mayer, L., Romeo, A. B. 2009, , 392, 294
Aharonian, F. A., & Atoyan, A. M. 1996, , 309, 917
Aharonian, F., Akhperjanian, A. G., Bazer-Bachi, A. R., et al. 2006, , 439, 695
Alexander, D. M., Bauer, F. E., Chapman, S. C., Smail, I., Blain, A. W., Brandt, W. N., & Ivison, R. J. 2005, , 632, 736
Amano, T., Torii, K., Hayakawa, T., & Fukui, Y. 2011, , 63, L63
Anantharamaiah, K. R., Viallefond, F., Mohan, N. R., Goss, W. M., & Zhao, J. H. 2000, , 537, 613
Armstrong, J. W., Rickett, B. J., & Spangler, S. R. 1995, , 443, 209
Balsara, D. S., Crutcher, R. M., & Pouquet, A. 2001, , 557, 451
Balsara, D. S., Kim, J., Mac Low, M.-M., & Mathews, G. J. 2004, , 617, 339
Batejat, F., Conway, J. E., Hurley, R., Parra, R., Diamond, P. J., Lonsdale, C. J., Lonsdale, C. J. 2011, , 740, 95
Beck, R., Brandenburg, A., Moss, D., Shukurov, A., & Sokoloff, D. 1996, , 34, 155
Beck, R. 2012, Journal of Physics Conference Series, 372, 012051
Becker, J. K., Biermann, P. L., Dreyer, J., & Kneiske, T. M. 2009, arXiv:0901.1775
Bell, E. F. 2003, , 586, 794
Beresnyak, A. 2012, Physical Review Letters, 108, 035002
Berezhko, E. G., & Völk, H. J. 2004, , 427, 525
Beswick, R. J., Riley, J. D., Marti-Vidal, I., et al. 2006, , 369, 1221
Bhat, P., & Subramanian, K. 2013, , 429, 2469
Blasi, P., Amato, E., & Serpico, P. D. 2012, Physical Review Letters, 109, 061101
Boldyrev, S., & Yusef-Zadeh, F. 2006, , 637, L101
Bradford, C. M., Nikola, T., Stacey, G. J., Bolatto, A. D., Jackson, J. M., Savage, M. L., Davidson, J. A., Higdon, S. J. 2003, , 586, 891
Brandenburg, A., & Subramanian, K. 2005, , 417, 1
Brogan, C. L., Nord, M., Kassim, N., Lazio, J., & Anantharamaiah, K. 2003, Astronomische Nachrichten Supplement, 324, 17
Brunetti, G., & Lazarian, A. 2007, , 378, 245
Brunt, C. M. 2003, , 583, 280
Brunt, C. M., Heyer, M. H., & Mac Low, M.-M. 2009, , 504, 883
Brunthaler, A., Mart[í]{}-Vidal, I., Menten, K. M., et al. 2010, , 516, A27
Burkert, A., Genzel, R., Bouch[é]{}, N., et al. 2010, , 725, 2324
Castor, J., McCray, R., & Weaver, R. 1975, , 200, L107
Chary, R., & Elbaz, D. 2001, , 556, 562
Chepurnov, A., & Lazarian, A. 2010, , 710, 853
Chevalier, R. A. 1974, , 188, 501
Chevalier, R. A., & Clegg, A. W. 1985, , 317, 44
Chevalier, R. A., & Fransson, C. 2001, , 558, L27
Chi, X., & Wolfendale, A. W. 1990, 245, 101
Chi, X., Young, E. C. M., & Beck, R. 1997, , 321, 71
Cho, J., & Vishniac, E. T. 2000, , 538, 217
Cho, J., & Lazarian, A. 2002, Physical Review Letters, 88, 245001
Cho, J., & Lazarian, A. 2003, , 345, 325
Chomiuk, L., & Wilcots, E. M. 2009, , 703, 370
Clemens, M. S., Vega, O., Bressan, A., Granato, G. L., Silva, L., Panuzzo, P. 2008, , 477, 95
Clemens, M. S., Scaife, A., Vega, O., & Bressan, A. 2010, , 405, 887
Condon, J. J., Huang, Z.-P., Yin, Q. F., & Thuan, T. X. 1991, , 378, 65
Condon, J. J. 1992, , 30, 575
Cravens, T. E., & Dalgarno, A. 1978, , 219, 750
Crocker, R. M., Jones, D. I., Melia, F., Ott, J., & Protheroe, R. J. 2010, , 463, 65
Crocker, R. M., Jones, D. I., Aharonian, F., Law, C. J., Melia, F., & Ott, J. 2011a, , 411, L11
Crocker, R. M., Jones, D. I., Aharonian, F., Law, C. J., Melia, F., Oka, T., & Ott, J. 2011b, , 413, 763
Crocker, R. M. 2012, , 423, 3512
Crutcher, R. M. 1999, , 520, 706
David, L. P., Jones, C., & Forman, W. 1992, , 388, 82
Daddi, E., Dickinson, M., Morrison, G., et al. 2007, , 670, 156
Daddi, E., Dannerbauer, H., Stern, D., et al. 2009, , 694, 1517
Daddi, E., Elbaz, D., Walter, F., et al. 2010a, , 714, L118
Daddi, E., Bournaud, F., Walter, F., et al. 2010b, , 713, 686
de Avillez, M. A., & Mac Low, M.-M. 2002, , 581, 1047
de Cea del Pozo, E., Torres, D. F., & Rodriguez Marrero, A. Y. 2009a, , 698, 1054
Dib, S., Bell, E., & Burkert, A. 2006, , 638, 797
Domingo-Santamaría, E. & Torres, D. F. 2005, , 444, 403
Downes, D., & Solomon, P. M. 1998, , 507, 615
Draine, B. T., & Woods, D. T. 1991, , 383, 621
Draine, B. T. 2011, Physics of the Interstellar and Intergalactic Medium (Princeton: Princeton University Press)
Elmegreen, B. G., & Scalo, J. 2004, , 42, 211
En[ß]{}lin, T. A., & Vogt, C. 2006, , 453, 447
En[ß]{}lin, T., Pfrommer, C., Miniati, F., & Subramanian, K. 2011, , 527, A99
Falgarone, E., Troland, T. H., Crutcher, R. M., & Paubert, G. 2008, , 487, 247
Federrath, C., Chabrier, G., Schober, J., Banerjee, R., Klessen, R. S., Schleicher, D. R. G. 2011, Physical Review Letters, 107, 114504
Fenech, D., Beswick, R., Muxlow, T. W. B., Pedlar, A., & Argo, M. K. 2010, , 408, 607
Ferrari, C., Govoni, F., Schindler, S., Bykov, A. M., & Rephaeli, Y. 2008, , 134, 93
Ferri[è]{}re, K., Gillard, W., & Jean, P. 2007, , 467, 611
Fleck, R. C., Jr. 1981, , 246, L151
F[ö]{}rster Schreiber, N. M., Genzel, R., Lutz, D., Kunze, D., & Sternberg, A. 2001, , 552, 544
Genzel, R., Burkert, A., Bouch[é]{}, N., et al. 2008, , 687, 59
Genzel, R., Tacconi, L. J., Gracia-Carpio, J., et al. 2010, , 407, 2091
Goldreich, P., & Sridhar, S. 1995, , 438, 763
Goldshmidt, O., & Rephaeli, Y. 1993, , 411, 518
Greaves, J. S., Holland, W. S., Jenness, T., & Hawarden, T. G. 2000, , 404, 732
Greco, J. P., Martini, P., & Thompson, T. A. 2012, , 757, 24
Green, A. W., Glazebrook, K., McGregor, P. J., et al. 2010, , 467, 684
Grimm, H.-J., Gilfanov, M., & Sunyaev, R. 2003, , 339, 793
Groves, B. A., Cho, J., Dopita, M., & Lazarian, A. 2003, Publ. Astron. Soc. Austaralia 20, 252.
Hall, A. N. 1980, , 190, 353
Hanasz, M., & Lesch, H. 2003, , 412, 331
Harrison, A., Henkel, C., & Russell, A. 1999, , 303, 157
Haugen, N. E., Brandenburg, A., & Dobler, W. 2004, , 70, 016308
Haverkorn, M., Brown, J. C., Gaensler, B. M., & McClure-Griffiths, N. M. 2008, , 680, 362
Heckman, T. M., Armus, L., & Miley, G. K. 1990, , 74, 833
Heesen, V., Beck, R., Krause, M., & Dettmar, R.-J. 2011, , 535, A79
Heiles, C., & Troland, T. H. 2005, , 624, 773
Helou, G., Soifer, B. T., & Rowan-Robinson, M. 1985, , 298, 7
Helou, G. & Bicay, M. D. 1993, , 415, 93
Herrero-Illana, R., P[é]{}rez-Torres, M. [Á]{}., & Alberdi, A. 2012, , 540, L5
Heyer, M. H., & Brunt, C. M. 2004, , 615, L45
Higdon, J. C., Lingenfelter, R. E., & Rothschild, R. E. 2009, , 698, 350
Holman, G. D., Ionson, J. A., & Scott, J. S. 1979, , 228, 576
Hopkins, P. F., & Hernquist, L. 2010, , 402, 985
Hopkins, P. F. 2013a, , 430, 1653
Hopkins, P. F. 2013b, , 430, 1880
Huba, J. D. 2011, *NRL Plasma Formulary*
Hunter, D. A., Shaya, E. J., Holtzman, J. A., Light, R. M., O’Neil, E. J., Lynds, R.. 1995, , 448, 179
Inoue, Y. 2011, , 728, 11
Iwasawa, K., Sanders, D. B., Evans, A. S., Trentham, N., Miniutti, G., Spoon, H. W. W. 2005, , 357, 565
Jean, P., Gillard, W., Marcowith, A., & Ferri[è]{}re, K. 2009, , 508, 1099
Jones, T. J. 2000, , 120, 2920
Jones, D. I., Crocker, R. M., Ott, J., Protheroe, R. J., & Ekers, R. D. 2011, , 141, 82
Joung, M. K. R., & Mac Low, M.-M. 2006, , 653, 1266
Joung, M. R., Mac Low, M.-M., & Bryan, G. L. 2009, , 704, 137
Kennicutt, R. C. 1998, , 498, 541 (K98)
Klessen, R. S., Heitsch, F., & Mac Low, M.-M. 2000, , 535, 887
Klessen, R. S., & Hennebelle, P. 2010, , 520, A17
Korpi, M. J., Brandenburg, A., Shukurov, A., Tuominen, I., & Nordlund, [Å]{}. 1999, , 514, L99
Koyama, K., Awaki, H., Kunieda, H., Takano, S., & Tawara, Y. 1989, , 339, 603
Koyama, K., Tsunemi, H., Dotani, T., et al. 2007, , 59, 23
Kruijssen, J. M. D. 2012, , 426, 3008
Krumholz, M. R., & McKee, C. F. 2005, , 630, 250
Krumholz, M. R., McKee, C. F., & Tumlinson, J. 2009, , 693, 216
Krumholz, M. R., & Thompson, T. A. 2012, , 760, 155
Kulsrud, R., Pearce, W. P. 1969, , 156, 445
Kulsrud, R. M. 2005, *Plasma Physics for Astrophysics*, Princeton, N.J.: Princeton University Press
Lacki, B. C., Thompson, T. A., & Quataert, E. 2010, , 717, 1
Lacki, B. C., & Thompson, T. A. 2010, , 717, 196
Lacki, B. C., Thompson, T. A., Quataert, E., Loeb, A., & Waxman, E. 2011, , 734, 107
Lacki, B. C. 2012a, arXiv:1204.2580
Lacki, B. C. 2012b, arXiv:1204.2584
Lacki, B. C., & Thompson, T. A. 2013, , 762, 29
Lacki, B. C. 2013, , 431, 3003
Lacki, B. C., & Beck, R. 2013, , 430, 3171
Laing, R. A. 1980, , 193, 439
Larson, R. B. 1981, , 194, 809
Launhardt, R., Zylka, R., & Mezger, P. G. 2002, , 384, 112
Law, C. J., Backer, D., Yusef-Zadeh, F., & Maddalena, R. 2009, , 695, 1070
Law, C. J. 2010, , 708, 474
Law, C. J., Brentjens, M. A., & Novak, G. 2011, , 731, 36
Lazarian, A., & Beresnyak, A. 2006, , 373, 1195
Lee, E. J., Murray, N., & Rahman, M. 2012, , 752, 146
Le Floc’h, E., Papovich, C., Dole, H., et al. 2005, , 632, 169
Lehmer, B. D., Alexander, D. M., Bauer, F. E., Brandt, W. N., Goulding, A. D., Jenkins, L. P., Ptak, A., Roberts, T. P. 2010, , 724, 559
Leroy, A. K., Evans, A. S., Momjian, E., et al. 2011, , 739, L25
Li, Z.-Y., & Nakamura, F. 2006, , 640, L187
Lingenfelter, R. E., Ramaty, R., & Fisk, L. A. 1971, , 8, 93
Lisenfeld, U., Völk, H. J., & Xu, C. 1996, , 306, 677
Lithwick, Y., & Goldreich, P. 2001, , 562, 279
Lonsdale, C. J., Diamond, P. J., Thrall, H., Smith, H. E., & Lonsdale, C. J. 2006, , 647, 185
Lord, S. D., Hollenbach, D. J., Haas, M. R., Rubin, R. H., Colgan, S. W. J., Erickson, E. F. 1996, , 465, 703
Lucek, S. G., & Bell, A. R. 2000, , 314, 65
Mac Low, M.-M., & Klessen, R. S. 2004, Reviews of Modern Physics, 76, 125
Magnelli, B., Elbaz, D., Chary, R. R., Dickinson, M., Le Borgne, D., Frayer, D. T., Willmer, C. N. A. 2009, , 496, 57
Mannheim, K., & Schlickeiser, R. 1994, , 286, 983
Mao, R. Q., Henkel, C., Schulz, A., Zielinsky, M., Mauersberger, R., Störzer, H., Wilson, T. L., Gensheimer, P. 2000, , 358, 433
Matzner, C. D. 2002, , 566, 302
Mauersberger, R., Henkel, C., Wielebinski, R., Wiklind, T., & Reuter, H.-P. 1996, , 305, 421
McCrady, N., & Graham, J. R. 2007, , 663, 844
McKee, C. F., & Ostriker, J. P. 1977, , 218, 148
McKee, C. F. 1989, , 345, 782
McIvor, I. 1977, , 178, 85
Melia, F., & Fatuzzo, M. 2011, , 410, L23
Melioli, C., & de Gouveia Dal Pino, E. M. 2004, , 424, 817
Melo, V. P., P[é]{}rez Garc[í]{}a, A. M., Acosta-Pulido, J. A., Mu[ñ]{}oz-Tu[ñ]{}[ó]{}n, C., & Rodr[í]{}guez Espinosa, J. M. 2002, , 574, 709
Miesch, M. S., & Bally, J. 1994, , 429, 645
Mohan, N. R., Goss, W. M., & Anantharamaiah, K. R. 2005, , 432, 1
Molinari, S., Bally, J., Noriega-Crespo, A., et al. 2011, , 735, L33
Moskalenko, I. V., Porter, T. A., & Strong, A. W. 2006, , 640, L155
Mouschovias, T. C., & Spitzer, L., Jr. 1976, , 210, 326
M[ü]{}hle, S., Seaquist, E. R., & Henkel, C. 2007, , 671, 1579
Muno, M. P., Baganoff, F. K., Bautz, M. W., et al. 2004, , 613, 326
Murphy, E. J., Porter, T. A., Moskalenko, I. V., Helou, G., & Strong, A. W. 2012, , 750, 126
Murray, N., Quataert, E., & Thompson, T. A. 2010, , 709, 191
Naylor, B. J., Bradford, C. M., Aguirre, J. E., et al. 2010, , 722, 668
Neff, S. G., Ulvestad, J. S., & Teng, S. H. 2004, , 611, 186
Noeske, K. G., Weiner, B. J., Faber, S. M., et al. 2007, , 660, L43
Nord, M. E., Henning, P. A., Rand, R. J., Lazio, T. J. W., & Kassim, N. E. 2006, , 132, 242
Norman, C. A., & Ferrara, A. 1996, , 467, 280
O’Connell, R. W., Gallagher, J. S., III, Hunter, D. A., & Colley, W. N. 1995, , 446, L1
Osterbrock, D. E. 1961, , 134, 270
Osterbrock, D. E. 1989, *Astrophysics of Gaseous Nebulae and Active Galactic Nuclei*, (Mill Valley, CA, University Science Books)
Ostriker, E. C., Stone, J. M., & Gammie, C. F. 2001, , 546, 980
Ostriker, E. C., & Shetty, R. 2011, , 731, 41
Padoan, P., Nordlund, A., & Jones, B. J. T. 1997, , 288, 145
Padoan, P., & Nordlund, [Å]{}. 2011, , 730, 40
Paglione, T. A. D., Marscher, A. P., Jackson, J. M., & Bertsch, D. L. 1996, , 460, 295
Paglione, T. A. D., & Abrahams, R. D. 2012, , 755, 106
Pan, L., & Scannapieco, E. 2010, , 721, 1765
Papadopoulos, P. P. 2010, , 720, 226
Papadopoulos, P. P., Thi, W.-F., Miniati, F., & Viti, S. 2011, , 414, 1705
Parra, R., Conway, J. E., Diamond, P. J., Thrall, H., Lonsdale, C. J., Lonsdale, C. J., Smith, H. E. 2007, , 659, 314
Pedlar, A., Muxlow, T. W. B., Garrett, M. A., Diamond, P., Wills, K. A., Wilkinson, P. N., Alef, W. 1999, , 307, 761
P[é]{}rez-Torres, M. A., Romero-Ca[ñ]{}izales, C., Alberdi, A., & Polatidis, A. 2009, , 507, L17
Persic, M., Rephaeli, Y., Braito, V., Cappi, M., Della Ceca, R., Franceschini, A., Gruber, D. E. 2004, , 419, 849
Persic, M., Rephaeli, Y., & Arieli, Y. 2008, , 486, 143
Pfrommer, C., En[ß]{}lin, T. A., & Springel, V. 2008, , 385, 1211
Pohl, M. 1994, , 287, 453
Prantzos, N., Boehm, C., Bykov, A. M., et al. 2011, Reviews of Modern Physics, 83, 1001
Rengarajan, T. N. 2005, Proc. 29th Int. Cosmic Ray Conf. (Pune), 3
Rephaeli, Y., Arieli, Y., & Persic, M. 2010, , 401, 473
Revnivtsev, M., Sazonov, S., Churazov, E., Forman, W., Vikhlinin, A., Sunyaev, R. 2009, , 458, 1142
Reynolds, S. P., & Chevalier, R. A. 1981, , 245, 912
Riechers, D. A. 2011, , 730, 108
Robishaw, T., Quataert, E., & Heiles, C. 2008, , 680, 981
Rodr[í]{}guez-Rico, C. A., Goss, W. M., Viallefond, F., et al. 2005, , 633, 198
Rodr[í]{}guez-Rico, C. A., Goss, W. M., Zhao, J.-H., G[ó]{}mez, Y., & Anantharamaiah, K. R. 2006, , 644, 914
Rovilos, E., Diamond, P. J., Lonsdale, C. J., Smith, H. E., & Lonsdale, C. J. 2005, , 359, 827
Roy, J.-R., & Kunth, D. 1995, , 294, 432
Roy, S., Rao, A. P., & Subrahmanyan, R. 2005, , 360, 1305
Roy, N., Peedikakkandy, L., & Chengalur, J. N. 2008, , 387, L18
Sakamoto, K., Mao, R.-Q., Matsushita, S., Peck, A. B., Sawada, T., Wiedner, M. C. 2011, , 735, 19
Sanders, D. B., Mazzarella, J. M., Kim, D.-C., Surace, J. A., & Soifer, B. T. 2003, , 126, 1607
Sargent, M. T., B[é]{}thermin, M., Daddi, E., & Elbaz, D. 2012, , 747, L31
Scalo, J., Vazquez-Semadeni, E., Chappell, D., & Passot, T. 1998, , 504, 835
Scalo, J., & Elmegreen, B. G. 2004, , 42, 275
Schekochihin, A. A., Cowley, S. C., Taylor, S. F., Maron, J. L., & McWilliams, J. C. 2004, , 612, 276
Schekochihin, A. A., Cowley, S. C., Kulsrud, R. M., Hammett, G. W., & Sharma, P. 2005, , 629, 139
Schekochihin, A. A., & Cowley, S. C. 2006, Physics of Plasmas, 13, 056501
Schlickeiser, R., & Miller, J. A. 1998, , 492, 352
Schlickeiser, R. 2002, *Cosmic Ray Astrophysics*, (New York: Springer)
Seaquist, E. R., & Odegard, N. 1991, , 369, 320
Shetty, R., Beaumont, C. N., Burton, M. G., Kelly, B. C., & Klessen, R. S. 2012, , 425, 720
Shraiman, B. I., & Siggia, E. D. 2000, , 405, 639
Silich, S., Tenorio-Tagle, G., & Rodr[í]{}guez-Gonz[á]{}lez, A. 2004, , 610, 226
Silich, S., Tenorio-Tagle, G., & Mu[ñ]{}oz-Tu[ñ]{}[ó]{}n, C. 2007, , 669, 952
Silich, S., Tenorio-Tagle, G., Mu[ñ]{}oz-Tu[ñ]{}[ó]{}n, C., Hueyotl-Zahuantitla, Wünsch, R., Palouš, J. 2010, , 711, 25
Smith, H. E., Lonsdale, C. J., Lonsdale, C. J., & Diamond, P. J. 1998, , 493, L17
Smith, L. J., Westmoquette, M. S., Gallagher, J. S., O’Connell, R. W., Rosario, D. J., & de Grijs, R. 2006, , 370, 513
Socrates, A., Davis, S. W., & Ramirez-Ruiz, E. 2008, , 687, 202
Sokoloff, D. D., Bykov, A. A., Shukurov, A., Berkhuijsen, E. M., Beck, R., Poezd, A. D. 1998, , 299, 189
Sorai, K., Nakai, N., Kuno, N., Nishiyama, K., & Hasegawa, T. 2000, , 52, 785
Spangler, S. R., & Gwinn, C. R. 1990, , 353, L29
Stone, J. M., Ostriker, E. C., & Gammie, C. F. 1998, , 508, L99
Strickland, D. K., & Heckman, T. M. 2007, , 658, 258
Strickland, D. K., & Heckman, T. M. 2009, , 697, 2030
Strong, A. W., & Moskalenko, I. V. 1998, , 509, 212
Strong, A. W., Moskalenko, I. V., & Ptuskin, V. S. 2007, Annual Review of Nuclear and Particle Science, 57, 285
Strong, A. W., Porter, T. A., Digel, S. W., J[ó]{}hannesson, G., Martin, P., Moskalenko, I. V., Murphy, E. J., & Orlando, E. 2010, , 722, L58
Subramanian, K., Shukurov, A., & Haugen, N. E. L. 2006, , 366, 1437
Suchkov, A., Allen, R. J., & Heckman, T. M. 1993, , 413, 542
Tacconi, L. J., Neri, R., Chapman, S. C., et al. 2006, , 640, 228
Tacconi, L. J., Genzel, R., Neri, R., et al. 2010, , 463, 781
Tagger, M., Falgarone, E., & Shukurov, A. 1995, , 299, 940
Takahashi, T., Mitsuda, K., Kelley, R., et al. 2012, , 8443, 84431Z
Tennekes, H., & Lumley, J. L. 1972, *A First Course in Turbulence*, Cambridge: MIT Press
Thompson, T. A., Quataert, E., & Murray, N. 2005, , 630, 167
Thompson, T. A., Quataert, E., Waxman, E., Murray, N., & Martin, C. L. 2006, , 645, 186
Thompson, T. A., Quataert, E., Waxman, E. 2007, , 654, 219
Thompson, T. A., Quataert, E., & Murray, N. 2009, , 397, 1410
Thornton, K., Gaudlitz, M., Janka, H.-T., & Steinmetz, M. 1998, , 500, 95
Torres, D. F. 2004, , 617, 966
Torres, D. F., Cillis, A., Lacki, B., & Rephaeli, Y. 2012, , 423, 822
Troland, T. H., & Crutcher, R. M. 2008, , 680, 457
Tsuboi, M., Inoue, M., Handa, T., Tabara, H., Kato, T., Sofue, Y., Kaifu, N. 1986, , 92, 818
Ulvestad, J. S. 2009, , 138, 1529
van Dokkum, P. G., & Conroy, C. 2010, , 468, 940
van der Laan, H. 1962, , 124, 125
Völk, H. J. 1989, , 218, 67
V[ö]{}lk, H. J., Berezhko, E. G., Ksenofontov, L. T., & Rowell, G. P. 2002, , 396, 649
Wada, K., Meurer, G., & Norman, C. A. 2002, , 577, 197
Walter, F., Riechers, D., Cox, P., Neri, R., Carilli, C., Bertoldi, F., Weiss, A., Maiolino, R. 2009, , 457, 699
Ward, J. S., Zmuidzinas, J., Harris, A. I., & Isaak, K. G. 2003, , 587, 171
Wei[ß]{}, A., Neininger, N., H[ü]{}ttemeister, S., & Klein, U. 2001, , 365, 571
Westmoquette, M. S., Gallagher, J. S., Smith, L. J., Trancho, G., Bastian, N., & Konstantopoulos, I. S. 2009, , 706, 1571
Wild, W., Harris, A. I., Eckart, A., Genzel, R., Graf, U. U., Jackson, J. M., Russell, A. P. G., Stutzki, J. 1992, , 265, 447
Williams, P. K. G., & Bower, G. C. 2010, , 710, 1462
Wills, K. A., Pedlar, A., Muxlow, T. W. B., & Wilkinson, P. N. 1997, , 291, 517
Wills, K. A., Pedlar, A., & Muxlow, T. W. B. 1998, , 298, 347
Wolfire, M. G., Tielens, A. G. G. M., & Hollenbach, D. 1990, , 358, 116
Wong, T., & Blitz, L. 2002, , 569, 157
Worrall, D. M., Marshall, F. E., Boldt, E. A., & Swank, J. H. 1982, , 255, 111
Yan, H., & Lazarian, A. 2002, Physical Review Letters, 89, 1102
Yan, H., & Lazarian, A. 2004, , 614, 757
Yoast-Hull, T. M., Everett, J. E., Gallagher, J. S., III, & Zweibel, E. G. 2013, , 768, 53
Yousef, T. A., Heinemann, T., Schekochihin, A. A., Kleeorin, N., Rogachevskii, I., Iskakov, A. B., Cowley, S. C., McWilliams, J. C. 2008, Physical Review Letters, 100, 184501
Yun, M. S., Ho, P. T. P., & Lo, K. Y. 1993, , 411, L17
Yun, M. S., Reddy, N. A., & Condon, J. J. 2001, , 554, 803
Yusef-Zadeh, F., Morris, M., & Chance, D. 1984, , 310, 557
Yusef-Zadeh, F., et al. 2009, , 702, 178
Zhao, J.-H., Desai, K., Goss, W. M., & Yusef-Zadeh, F. 1993, , 418, 235
Zweibel, E. G., & Josafatsson, K. 1983, , 270, 511
[^1]: @Beswick06 also criticized the ISM pressure $10^7\ \Kelv\ \cm^{-3}$ used by @Chevalier01, suggesting that it was much too high. But as Table \[table:Pressures\] shows, that estimate is if anything *too low* – the energy densities of the molecular gas turbulence, superwind thermal gas, magnetic fields, radiation, hydrostatic pressure, and even the thermal pressure in some H II regions are all $\ga 10^7\ \Kelv\ \cm^{-3}$. The first counterargument of @Beswick06 is that the column of ionized gas inferred from the free-free absorption cutoff of supernova remnants is too small under such pressures, unless ionized gas only fills a small fraction. But this is what I have argued in @Lacki13-LowNu – the H II regions *are* the WIM but fill only a few percent of the starburst volume. Furthermore, the H II regions may be supported by turbulent pressure, so the densities estimated by @Beswick06 might be too high. Second, they quote the small thermal pressures inferred for M82’s neutral gas. As I discussed in § \[sec:UStarburst\], the pressure is dominated by turbulence, even in high density clumps. Of course, none of this invalidates the basic conclusion that the SNRs are expanding quickly, in a low density medium.
[^2]: But this is not the case in the GCCMZ, where the amount of gamma-ray emission is so low that CR protons are evidently not penetrating molecular clouds [@Crocker11-Wild]; this is also indicated by the lack of radio emission from secondary $e^{\pm}$ in GCCMZ molecular clouds [@Jones11].
[^3]: The superwind’s “thermostat” is also a major difference with another home for extremely hot plasma, galaxy clusters, where the turbulent energy density is just a few percent of the thermal energy density [e.g., @Subramanian06]. While a starburst’s gravity is not strong enough to confine its superwind, a galaxy cluster’s gravity does dominate the dynamics of its hot plasma. Therefore, the turbulent energy density is a small fraction of the thermal energy of the virialized plasma. Thermal energy from turbulent dissipation must accumulate in the plasma for over a Hubble time to start affecting the dynamics, a duration much longer than the Gyr eddy crossing time.
[^4]: But if diffusive reacceleration operates in superwinds, it plausibly could drive turbulence and the CR energy density much closer to equipartition (see Section \[sec:DCascCalc\]).
[^5]: Strictly speaking, this ${\ensuremath{\langle \sigma v \rangle}}$ is for heavy ions colliding with H$_2$ molecules. However, at the temperatures of $100\ \Kelv$ typical of starburst molecular gas, this cross section should still roughly work [@Osterbrock61].
[^6]: For scattering off fast modes, @Schlickeiser02 finds that $D_{pp}$ is $\ln(c/v_A)$ times larger and $t_{\rm reacc}$ is $\ln(c/v_A)$ times shorter.
[^7]: This mechanism could be a way around extreme radiative losses by SNRs in cold starbursts, setting a floor to the amount of mechanical energy driven into the ISM by supernovae, but only if the CR acceleration efficiency is $\sim 10\%$ as in the Milky Way and if CR diffusion is very slow. The fact that ULIRGs lie on the far-infrared–radio correlation [@Condon91] supports the idea that the CR acceleration efficiency is the same. Thus, CRs could play an important dynamical role in cold starbursts. Note that, unlike in other scenarios where CRs are dynamically important [e.g., @Socrates08], the CRs do not directly drive a wind, and do not fill the starburst. Their dynamical role arises only because they are very inhomogeneous, concentrating their energy density into a few small bubbles.
[^8]: In these cases, rotation measure is defined as the derivative of polarization angle with wavelength squared [@Chi97; @Sokoloff98].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a simple semi-numerical model designed to explore black hole growth and galaxy evolution. This method builds on a previous model for black hole accretion that uses a semi-numerical galaxy formation model and universal Eddington ratio distribution to describe the full AGN population by independently connecting galaxy and AGN growth to the evolution of the host dark matter halos. We fit observed X-ray luminosity functions up to a redshift of $z \sim 4$, as well as investigate the evolution of the Eddington ratio distributions. We find that the Eddington ratio distribution evolves with redshift such that the slope of the low-Eddington accretion rate distribution increases with cosmic time, consistent with the behavior predicted in hydrodynamical simulations for galaxies with different gas fractions. We also find that the evolution of our average Eddington ratio is correlated with observed star formation histories, supporting a picture in which black holes and galaxies evolve together in a global sense. We further confirm the impact of luminosity limits on observed galaxy and halo properties by applying selection criteria to our fiducial model and comparing to surveys across a wide range of redshifts.'
author:
- 'Mackenzie L. Jones, Ryan C. Hickox, Simon J. Mutch, Darren J. Croton, Andrew F. Ptak, Michael A. DiPompeo'
bibliography:
- 'ap\_evol.bib'
title: 'Evolution of Black Hole and Galaxy Growth in a Semi-numerical Galaxy Formation Model'
---
Introduction {#sec:intro}
============
Despite significant progress in establishing a generalized model of the formation and evolution of galaxies with respect to their host dark matter halos, it is still uncertain how the growth of supermassive black holes (SMBH) fit into this evolutionary scheme and what impact actively accreting supermassive black holes (active galactic nuclei; AGN) have on their hosts (for reviews, see e.g., @Sil12 [@Ale12; @Fab12]).
There is observational evidence of a co-evolution between AGN and their host galaxies, as well as theoretical models that explore the impact of feedback (both stellar and AGN) on galaxy growth (e.g., @Bow06 [@Cro06; @Vol15]). @Hec04 find that the volume-averaged galaxy-black hole growth rate, for a range of black hole masses, is consistent with the observed black hole-spheroid mass relationship. Additional evidence shows that black hole mass is correlated with stellar bulge properties (e.g., @Kor13). There is also a similarity in the growth histories of black hole accretion and star formation across cosmic time, including where they peak at $z \sim 1-2$ (e.g., @Air10 [@Raf11; @Mul12]). @Che13 observed that the average black hole accretion rate is directly proportional to the star formation rate in star forming galaxies, although this may be a secondary effect due to the dependence of black hole accretion on stellar mass (@Yan17). However, other than a common supply of cold gas on kpc scales, the physical processes linking black hole growth to galaxy evolution are still not well understood (e.g., @Dim05 [@Hop06; @Ale12]).
Disentangling this co-evolution is made more difficult by the presence of observational biases (e.g., selection effects, obscuration from gas and dust, dilution from host galaxy emission) which may obscure these relationships. How an AGN is selected in a sample can be strongly impacted by these observational biases (see @Pad17 for a review; e.g., @Lau07). Some selection criteria are more complex than others; for example, color selection and emission line ratio selection rely on detailed decompositions to separate the AGN from the host galaxy properties. Others are more straightforward, such as X-ray AGN selection based on an observed or absorption corrected luminosity limit.
In principle it is possible to correct for the presence of biases in most observations by assuming an underlying black hole accretion distribution and forward modeling the observations. Using SDSS star forming galaxies, @Jon16 uncovered the underlying Eddington ratio distribution (distribution of the ratio of the instantaneous luminosity to the maximum possible accretion given by the Eddington limit; ${L}/{L}_\text{Edd}$), using a straight-forward model for black hole accretion to correct for the effects of dilution that are more prevalent in galaxies with star formation. This simple model for black hole accretion was also tested in @Jon17 at $z = 0$ to place AGN in simulated galaxies from the @Mut13 semi-numerical galaxy formation model. The results of this work showed that selection criteria introduce potential bias by selecting different host galaxy and dark matter halo properties (see also e.g., @Aza15 [@Air15]). The Eddington ratio distribution is observed directly in the X-rays to be consistent with a power law with a potential cutoff at high ${L}/{L}_\text{Edd}$ (e.g., @Air10 [@Air15]). Many theoretical models assume a simple functional form of the Eddington ratio distribution (e.g., @Con13 [@Hop09; @Hic14; @Jon16]), however, it is uncertain whether it should be dependent on mass and/or vary across cosmic time (e.g., @Cap15 [@Wei17; @Ber18]).
In this work, we use an updated version of the @Jon17 model for black hole accretion to investigate the black hole-galaxy connection as well as the role AGN selection criteria play in the observed properties of galaxy and dark matter halos. This method is computationally less expensive than in @Jon17 and includes evolution of the AGN, galaxy, and dark matter halo properties up to a redshift of $z \sim 4$. We discuss how the simulation is constructed in Section \[sec:method\], including the modeling of the evolution of the observed X-ray luminosity functions (Section \[ssec:luminosity\]). Our results are found in Section \[sec:xray\], where we also discuss potential evidence for AGN-galaxy co-evolution (Section \[ssec:evol\]) and the change in galaxy properties observed when imposing selection criteria. Our conclusion and a summary of our results are provided in Section \[sec:con\]. Throughout this paper, we assume a 1-year Wilkinson Microwave Anisotropy Probe (WMAP1; @Spe03) cold dark matter (CDM) cosmology with $\Omega_m$ = 0.25, $\Omega_\lambda$ = 0.75, and $\Omega_b$ = 0.045. All results are shown with a Hubble constant of $\text{h} = 0.7$, where $\text{h} \equiv \text{H}_0/100$ km s$^{-1}$ Mpc$^{-1}$, and all luminosities are presented in erg s$^{-1}$.
Simulation Methodology {#sec:method}
======================
Dark Matter and Galaxies {#ssec:gal}
------------------------
As in @Jon17, we use a semi-numerical galaxy formation model (@Mut13) as the foundation of our simulation. This model connects galactic growth to the formation history of the N-body Dark Matter Millennium Simulation (@Spr05) with a prescribed baryonic growth function and a physics function. These simple analytic functions dictate the availability of baryonic material to be used by stars and the efficiency by which this process occurs. The merits of using this analytic solution, as well as a description of how we repopulate the output star formation rates, are described in @Jon17.
We use 41 of the 63 available Millennium snapshots corresponding to a redshift range of $0\le z<4.12$ in steps of $\sim$ 200 – 350 Myr (the remaining 22 snapshots are at redshifts $> 4.12$). This expands on the work in @Jon17, in which we limited our investigation to $z = 0$. Within this redshift range, the galaxy formation model is able to accurately reproduce the stellar mass function and its evolution to $z \sim 4$. These evolving simulated stellar mass functions inform our black hole mass distributions and so are used as a constraint in determining the best black hole accretion distribution model at each redshift, as discussed in Section \[ssec:agn\]. For more information on the galaxy formation model, see @Mut13.
Adding in AGN activity {#ssec:agn}
----------------------
Our simulated galaxy sample consists of approximately 24 million galaxies per snapshot. We add an AGN counterpart to each galaxy using the following method. We first calculate black hole masses using the @Har04 black hole-bulge relationship using the total galaxy stellar mass (M\*) from the @Mut13 galaxy formation model as a proxy for the bulge mass (@Kor13). The merits of using this relationship to calculate our black hole masses are discussed in @Jon17 and summarized below.
The significant uncertainty and scatter found in measuring the relationship between black hole mass and total stellar mass has led to disagreements on the necessity of adding in an evolutionary component (e.g., @Dec10 [@Mer10; @Scz11; @Ben11; @Woo13; @Deg15; @Sha16]). It is possible, however, that the differences observed for redshift evolution in this relationship may be caused by sample selection bias, rather than an intrinsic evolution. The observed differences in the black hole mass-stellar mass relationship based on galaxy type adds a further complication (e.g., @Rei15 [@Bal16]). Since the @Har04 relationship bisects those found for ellipticals and AGN, we adopt this non-evolving model not only for simplicity, but as a way to handle the large intrinsic scatter in the relationship and limit the impact of potential observational biases. We investigate the impact of scatter observed in the relationship between black hole mass and stellar mass and the error associated with @Har04 by adding a scatter of $\sim0.3$ dex to our black hole masses and running a Monte Carlo simulation with an additional uncertainty corresponding to the @Har04 published errors to demonstrate the variance of our black hole mass functions (Figure \[fig:BHM\]). As a further check, we test an alternate black hole-stellar mass relationship with moderate evolution (@Mer10). We evaluate the merit of adding a black hole evolution by fitting to the X-ray luminosity function as we do for the non-evolving case (described in detail below). However, we do not find a significant enough improvement to justify this additional complexity. We thus select the simplest form of the black hole-stellar mass relationship to limit the number of free parameters and avoid adding any additional degeneracies.
\
Building on the method in @Jon17, we determine the black hole accretion by convolving a broad Eddington ratio distribution with the simulated black hole mass functions to reproduce a bolometric luminosity function. We first fit the black hole mass function calculated from the @Mut13 galaxies using a Schechter function (power law with exponential cutoff). While Eddington ratio distributions have been defined using a variety of functional forms (e.g., @Hop09Her [@Nov11; @Con13; @Hic14; @Jon17]), our Eddington ratio distributions are defined by a double power law in order to reproduce the evolution of the @Air15 X-ray luminosity functions: $$\label{equ:bpl}
\frac{dt}{d\log \text{L}_{\textrm{bol}}}= \phi^*\left[\left(\frac{\text{L}_{\textrm{bol}}}{\lambda_{\textrm{break}}\text{L}_{\textrm{Edd}}}\right)^{-\alpha_1}+\left(\frac{\text{L}_{\textrm{bol}}}{\lambda_{\textrm{break}}\text{L}_{\textrm{Edd}}}\right)^{-\alpha_2}\right],$$ where ${L}_{\text{bol}}$ is the bolometric luminosity, ${L}_{\text{bol}}/{L}_{\text{Edd}}$ is the Eddington ratio, $\phi^*$ is the amplitude, $\lambda_{\text{break}}$ marks the position of the break in the distribution, and $\alpha_1$, $\alpha_2$ are the power law slopes. In this work, our amplitude is set by the minimum Eddington ratio, $\lambda_\text{min}$, which is chosen such that the integral of our Eddington ratio distribution is one. This method is computationally inexpensive, so we can generate a wide range of Eddington ratio distributions to be used in the convolution in order to better reproduce observed luminosity functions.
\
Recovering the AGN X-ray Luminosity Function {#ssec:luminosity}
--------------------------------------------
In order to determine the intrinsic black hole activity, we can compare the output of our convolution to observed luminosity functions, such as the AGN X-ray luminosity functions (XLFs) of @Air15. We convert our bolometric luminosities into hard ($2-10$ keV) X-ray luminosities using a constant bolometric correction of ${k}_\text{bol}=44$ [@Don18], although we did consider a variety of corrections, both constant as well as tied to luminosity (e.g., @Mar04 [@Hop07; @Run12]). This differs from the prescription outlined in @Jon17, in that here we do not assume a connection between the bolometric corrections and the intrinsic Eddington ratio, as observed in the work by @Lus12. We assume a constant correction for simplicity because we find that using an Eddington ratio dependent bolometric correction requires additional free parameters that make our model unnecessarily complicated in order to fit the @Air15 XLFs. We test our simulation with an accretion-dependent bolometric correction (@Lus12), and find that the results of our analysis are consistent with a constant bolometric correction, although prone to more degeneracies in this instance. Further simplifying our process, we no longer require the addition of obscuration (outlined in @Jon17) to compare to the observed XLFs, as obscuration is corrected for in the @Air15 XLFs.
We fit the @Air15 AGN XLF parametrization at each of our redshift snapshots by running a least squares fit with the convolved Eddington ratio distribution and black hole mass function. To determine the best fits, we allow the parameters defining our broken power law Eddington ratio distribution to vary at each redshift, independent of each other. The variables defining the broken power law include the minimum Eddington ratio of the function ($\lambda_\text{min}$), the threshold between the two power laws ($\lambda_{\text{break}}$), and the slopes of the two power laws ($\alpha_1$, $\alpha_2$), as shown in Equation \[equ:bpl\]. Our model was initially tested with a varying $\alpha_2$, however, we found little scatter in the resulting XLF fits and for simplicity have adopted the constant value of $\alpha_2=2.4$ to which our variable fits converged. This has the added benefit of limiting the number of free parameters.
At each redshift we run the model multiple times with the following variations to the Eddington ratio distribution: $\alpha_1$ is allowed to vary between $0.0 \le \alpha_1 \le 1.0$ in steps of $0.01$, $\lambda_{\text{break}}$ is allowed to vary between $0.0 \le \log(\lambda_{\text{break}}) \le 4.0$ in steps of $0.1$, and $\lambda_\text{min}$ is allowed to vary between $-8.0 \le \log(\lambda_\text{min}) \le -4.5$ in steps of $0.5$. As such, there are only three parameters that may change at any time in order to reduce the number of possible degeneracies. We examined the error introduced by fitting to a published XLF by running a Monte Carlo simulation of our model fit. Rather than fitting to the @Air15 parametrization for a given redshift, we allow the binned estimates of the XLF for hard-X-rays to vary along their error bars. We determine the best-fit through $\chi^2$ minimization for each of these runs and find, e.g., the $1\sigma$ errors for the low-Eddington slope are $\delta \alpha_1\sim0.042$. The results of these fits are shown in Figure \[fig:XLF\] where we have selected a sample of redshifts covering the range studied by @Jon17 and @Air10, allowing a comparison of our convolution (solid light colors) to the @Air15 AGN XLF (solid black). For clarity, the @Air15 $z=0$ line has been included at each redshift for comparison (dot dash grey).
Once we have determined the Eddington ratio distribution that produces the best-fit convolution, we use these parameters to define the AGN properties for each galaxy in the @Mut13 model. This is very similar to what is done in @Jon17. Based on the best-fit Eddington ratio distribution at each redshift, we draw an intrinsic bolometric AGN luminosity. We then apply the same bolometric correction to our bolometric AGN luminosities as in our convolution (@Don18) to determine the AGN X-ray luminosity. With the “best fit” convolved model parameters we are able to simulate the luminosity distributions of our sample (dark colored dashed lines) and reproduce the @Air15 AGN XLFs. Scatter at low redshifts at the highest luminosities is due to the relatively low number of extreme AGN in our volume limited sample. The X-ray emission from stellar processes and hot gas in our simulation is calculated following the prescription outlined in @Jon17, in which we utilize the @Leh16 scaling relationships to determine the luminosity contribution from low mass X-ray binaries (LMXB) and high mass X-ray binaries (HMXB) from the @Mut13 galaxy parameters. With knowledge of the dark matter, galaxy, and AGN properties of our intrinsic sample across the selected redshift range, we are able to compare the properties of the full AGN population to X-ray observations.
Comparison of the Simulated AGN Population with X-ray Observations {#sec:xray}
==================================================================
Investigating the co-evolution of galaxies and AGN {#ssec:evol}
--------------------------------------------------
\
\
The intrinsic black hole activity in our simulation provides insight into black hole-galaxy co-evolution as the best-fit parameters change across our redshift range. We focus on $\alpha_1$, the slope of the low-Eddington ratio distribution that directly determines the low end of the luminosity function through our convolution. Figure \[fig:alpha\] shows how our $\alpha_1$ slope changes with redshift. We fit this relationship with the following exponential function: $$\label{equ:exp}
\alpha_1(z)=b_{0}~b_{1}^{~z}+b_{2}$$ where the best-fit parameters are determined by the IDL routine `COMFIT` to be: $b_{0}=0.49$, $b_{1}=0.41$, and $b_{2}=0.22$ (green line). This function may be used to further limit the number of free parameters defining the evolution of the low-Eddington slope. The flattening of the low-Eddington ratio distribution slope is not unexpected since the observed AGN X-ray luminosity functions are observed to flatten with increasing redshift. To fit this evolution, we convolve our black hole mass functions (which are defined by a constant stellar mass-black hole mass relationship) with an Eddington ratio distribution. In the case where the luminosity function slope is steeper at low luminosities compared to the low-mass regime of the black hole mass function, the fit is driven by the Eddington ratio distribution which will be steeper to compensate for the black hole mass function. While the low-luminosity regime of the luminosity function is not well constrained (e.g., @Air15 [@Buc15]), this pattern may point to a relationship between AGN activity and the evolving properties of the galaxies where AGN are found.
This relationship is clearer when put in the context of previous work (e.g., @Air12 [@Gab13]). @Gab13 created a hydrodynamic simulation in which they investigate the Eddington ratio distribution for gas-poor (${f}_\text{gas}=10 \%$) and gas-rich (${f}_\text{gas}=50 \%$) galaxies, corresponding to typical star-forming galaxies at redshifts of $z=0$ and $z=2$, respectively (Figure \[fig:alpha\]; orange regions). This distribution took the form of a power law, corresponding to the low-Eddington slope in our model, with $\alpha_1=0.76$ $(z=0)$ and $\alpha_1=0.24$ $(z=2)$ (following the nomenclature of Equation \[equ:bpl\]). We find agreement with our evolving Eddington ratio distribution slopes and the @Gab13 parameters based on a given galactic gas fraction. This is consistent with the @Gab13 result that the Eddington ratio distribution slope decreases with increasing gas fraction (at higher redshifts). If we fit our simulation to an alternate XLF parametrization, such as @Ued14, we find that our low-Eddington ratio distribution slopes similarly decrease with increasing redshift out to $z~2$.
We can also investigate the evolution of the average Eddington ratio, or the black hole accretion history, for our simulated sample (Figure \[fig:Edd\]; green). If black hole activity were tied to galactic activity, we expect to see a similar evolution in the specific star formation rates (sSFR; ratio of the star formation rate to the stellar mass) of our sample. It is possible to calculate the sSFR for our sample since the @Mut13 model keeps track of the star formation rates for each individual galaxy. We probe this co-evolution by plotting the average sSFR of our galaxy sample (with stellar mass $\log{M}* > 9.0 {M}_\odot$) at each redshift, compared to the @Whi12 parametrization of the star forming sequence, and a compilation of mean sSFR from @Mad14. Directly comparing the average Eddington ratios to the global specific star formation rates typically requires scaling the average Eddington ratio by a constant value, in this case we use a scale factor of 250.
While the evolution of the growth of the black holes is similar in shape to the growth of the star forming global average, our black holes are growing slower than our galaxies. This is not surprising since our model assumes that all black holes are accreting with the same underlying Eddington ratio distribution and the most massive galaxies in our simulation are not star-forming. Differences in global and individual galaxy growth are also found when comparing the black hole-bulge relationship for individual galaxies to the global growth rate (similar growth, but the amplitudes are separated by $\sim10$ percent) (e.g., @Hec04 [@Kor13]).
We also find qualitative agreement between the shape of the average Eddington ratio evolution and our simulated average sSFR with the @Whi12 line (Figure \[fig:Edd\]; purple), and observed sSFR compiled by @Mad14. Recent work suggests there may be a difference in the integrated quantities of specific star formation rate and accretion (e.g., @Air15 [@Cap18; @Yan18]). While we cannot compare directly to these works since we are calculating the average growth, there may be a similar discrepancy in the shapes of our specific star formation and accretion rate histories. However, within our error, we can only conservatively say that the average growth appears to be consistent.
Host Galaxy SFR and Mass {#ssec:color}
------------------------
In our model we simultaneously know the intrinsic properties of the black hole, galaxy, and dark matter halos. This allows us to investigate how observations of these properties are impacted by observational biases and selection effects. As in @Jon17, we examine the sSFR-M\* distribution and compare it to surveys of observed X-ray selected AGN. Our intrinsic sSFR-M\* distribution is shown in black contours in Figure \[fig:color\].
\
We again compare our simulation to the @Men16 X-ray selected AGN (orange points). In this work, however, rather than applying a main sequence correction to the @Men16 AGN and comparing it to our simulation at $z=0$, we select from our simulation the entire redshift range given by the @Men16 X-ray selected AGN $0.2 < z < 1.2$. Once we select the snapshots corresponding to the survey redshifts, we add in obscuration following the same parametrization outlined in Jones et al. (2017) in which we assign obscuration based on the @Mer14 incidence of obscuration and X-ray absorption corresponding to the NuSTAR-informed NH distribution (@Lan14 [@Lan15]). We then apply a luminosity threshold to “select” the AGN observed from our intrinsic sample. The luminosity threshold we use mimics the @Men16 limit of ${L}_\text{X}=10^{41.5}$. We find that the luminosity limited sample (green contours) covers the same dynamic range as the distribution of the Mendez X-ray selected AGN sample, more so than in @Jon17 by including the survey redshift limits. Of particular note is the consistency at high stellar masses without the need for additional AGN suppression in high mass systems, keeping our model mass independent. We further illustrate the effects of luminosity limits on the observed galaxy properties in Figure \[fig:lim\], depicted by the contour boundaries for increasing luminosity thresholds. As the luminosity limit increases, the “observed” sample selected is “biased” towards higher stellar masses and correspondingly lower average sSFR (e.g., @Aza15 [@Air15]).
\
Host Dark Matter Halos {#ssec:halo}
----------------------
As with the host galaxies, this model allows us to probe the observed properties of dark matter halos for AGN selection techniques. This is a useful tool for investigating the impact of selection effects on clustering measurements as well as studies of large scale structure. We first compared our $z =0$ simulation to the observational clustering analysis of @Ric13 and the simulation from @Cha12, as in @Jon17. While the results of this previous work were promising, this new analysis incorporates evolution and thus is better poised to directly compare to the dark matter halo distributions at the corresponding redshifts.
In this work, we directly compare the halo mass distributions of our simulated AGN rather than the HOD (the fraction of halos of a given mass that host an AGN). This takes into account the low density of halos at large masses by appropriately weighting the distribution at more common moderate halo masses. We multiply the halo mass function of our intrinsic sample at $z = 1.0$ with the HOD from @Cha12 and @Ric13 (Figure \[fig:halo\]; blue, orange lines). This weighted distribution is compared with the mass distribution of the full simulated sample (dash purple). After applying obscuration to our simulated X-ray luminosities, we also select three luminosity limited samples ($39 \le \log{L}_\text{X} \ge 45$) at $z = 1.0$ (green) to compare to the theoretical and observationally motivated mass distributions.
\
\
We find that as the luminosity limits increase, the dark matter halo mass distribution becomes increasingly “biased” toward higher halo masses. For a luminosity limit of $\log({L}_\text{X}) > 41.5$, we recover the @Cha12 weighted halo mass distribution. We are also able to recover parts of the observed @Ric13 halo mass distribution (orange) using different luminosity limits. Our inability to recover the full halo mass distribution is not unexpected, since any biases that may impact this sample selection are likely more complicated than a single luminosity limit.
We further investigate the impact of luminosity limits on our sample by calculating the average dark matter halo mass and luminosity as a function of luminosity limit and redshift. From the full simulation, five snapshots are selected that correspond to redshifts: $z\sim 0.0, 0.5, 1.0, 2.0, 3.0$. At each redshift, we select a sub-sample based on luminosity limits between $39 \le \log L_{\text{X}} \le 45$ (in steps of 1 dex) and calculate the mean halo mass and X-ray luminosity. We find, as expected, that as our luminosity limits increase the mean X-ray luminosity from our simulation also increases (diamonds; Figure \[fig:MvL\]). Our mean halo masses, however, exhibit a plateau between $42 \le \log L_{\text{X}} \le 44$. This flattening is consistent with the average halo masses and luminosities of X-ray selected AGN in observational clustering surveys (e.g., @Coi09 [@Cap10; @Sta11; @Mou12; @Kru12], Powell et al. 2019 *submitted*; select observed luminosities were renormalized to $L_{\text{X}}$\[2-10 keV\] in @Fan13cluster [@Gea19]). The observed plateau is a direct consequence of the Eddington limit: the only way to reach the highest luminosities is to have a large black hole accreting near its Eddington limit. Thus selecting AGN at high luminosities biases the sample to a narrower range of black hole masses and Eddington ratios. Likewise, at lower luminosities, there is a mix of high mass black holes accreting at low Eddington rates and low mass black holes accreting at high Eddington rates. As a result, there is a broader distribution in accretion rate and a weaker luminosity dependence for the average halo mass.
These results reaffirm the capability of our simple model to describe the properties of dark matter halos for luminosity limited AGN selection techniques. When comparing surveys and analyses of the environment where AGN reside, it is important to understand how the chosen selection methods impact the observed dark matter and host galaxy properties. This simulation provides insight into what subset of the AGN population is observed based on selection criteria.
Conclusions {#sec:con}
===========
The goal of this work is to investigate the evolution of AGN and their host galaxies using a simple prescription for AGN activity and a semi-analytic galaxy formation model from @Mut13. We have made improvements to the black hole accretion prescription from @Jon17 by fitting the @Air15 X-ray luminosity functions for redshifts $0.0 \le z < 4.12$. This involves convolving an intrinsic Eddington ratio distribution with the black hole mass functions determined directly from the galaxy formation histories, which decreases the computational cost of our model compared to that in @Jon17. From these best-fit Eddington ratio distributions we can build a simulated population of AGN corresponding to every galaxy in the @Mut13 galaxy formation model. This allows us to simultaneously investigate the properties of the galaxy, AGN, and dark matter halos.
The model described in this paper uses a forward-modeling technique in which we build a simulation of dark matter halos, galaxies, and AGN based on fiducial models of galaxy formation theory and AGN fueling. Our best-fit black hole accretion models are constrained by observations while our black hole mass functions are motivated by structure formation theory and the observed evolution of the galaxy stellar mass functions. With this prescription we are able to build a large volume of galaxy and AGN with knowledge of their properties as a whole, rather than the internal galaxy conditions. Additionally, we have selected a parametrization of the black hole accretion that is a universal shape (following the results of @Jon16) which keeps our model simple and with few variables in order to reduce the number of degeneracies. This makes the simulation more straightforward to analyze and computationally inexpensive, running in sub-seconds, a significant improvement to more complicated semi-analytic models.
Since we connect the dark matter halo mass to the galaxy stellar mass, the galaxy stellar mass to the black hole mass, and the black hole mass to the AGN activity, we can directly compare to observations and test for black hole-galaxy coevolution. A fundamental question that defines our model is *what is the underlying relationship governing AGN activity required to recreate observations across cosmic time?* We use the black hole mass functions tied to the galaxy formation simulation convolved with a double power law to describe black hole accretion in order to model the observed @Air15 XLF. This more directly tells us about the long term processes of black hole growth than observations because the volume of our simulation can be used to our advantage to smooth outliers by describing the typical black hole accretion distribution (@Hic14).
By calculating the intrinsic average AGN and galaxy activity in our evolving model, we can directly test relationships of black hole-galaxy co-evolution without the influence of observational biases. This may help dispel some of the uncertainty in the observational evidence of co-evolution due to current limitations observing the full AGN population (e.g., @Hic14 [@Pad17]). We find that the evolution of our black hole accretion distributions are consistent with the predicted evolution of galaxy gas fraction (@Gab13), which is potential evidence for a black hole-galaxy connection through a common supply of gas. In addition, we find similar evolution of the average Eddington ratio distribution and the average specific star formation rate. This may be explained by a common supply of gas, or the presence of AGN and/or stellar feedback processes that may trigger or quench star formation and AGN activity. Furthermore, since our model also contains information about the dark matter halo and galaxy properties we can apply selection criteria to our full sample to compare directly to observations with similar limitations (e.g., volume, luminosity limit, mass limit) to investigate the impact of selection biases on the observed properties used in evidence of co-evolution.
The results of our analysis are summarized as follows:
- We find that the evolution of the slope defining the Eddington ratio distribution in our model becomes flatter with increasing redshift. This shape is consistent with the picture from @Gab13 in which the black hole accretion rate varies based on the predicted behavior of galaxy gas fraction, where galaxies at higher redshifts have increased gas fraction and flatter Eddington ratio distributions.
- We find that the redshift evolution of the average Eddington ratio is broadly consistent with the evolution of the average specific star formation rate in galaxies, as well as the @Whi12 star forming sequence. This co-evolution may be due to a common supply of gas, and/or feedback processes that may quench or trigger galaxy and black hole activity.
- We confirm that selecting AGN based on a luminosity limit impacts the properties of the observed host galaxies and dark matter halos. We compare the distributions of the galaxy and halo properties of our luminosity limited simulated sample to observations of AGN selected samples (@Men16) and dark matter halo clustering studies (@Cha12 [@Ric13]). We find that different AGN selection criteria yield different galaxy and halo properties.
We have shown that a simple model of black hole accretion and galaxy evolution is able to reproduce the observed evolution of AGN and black holes across cosmic time. Given the breadth of the available AGN, galaxy, and dark matter observables built into this simulation, the natural next step would be to investigate additional properties and selection effects on the X-ray AGN population. This work has further confirmed the universality of a simple accretion model to describe black hole growth, but has also shown there can be some evolution in those model parameters (e.g., with redshift) that may be connected to host galaxy properties. Based on the assumption of a universal Eddington ratio distribution to describe the full multi-wavelength AGN population, it is possible to expand this work into a powerful multi-wavelength simulation to make predictions for the next generation of observatories.
We thank our collaborators for their insights in developing this work, as well as for constructive comments that improved the paper. We also thank our referee for comments that improved this work. This work was supported in part by the National Aeronautics and Space Administration under Grant Number NNX15AU32H issued through the NASA Education Minority University Research Education Project (MUREP) through the NASA Harriett G. Jenkins Graduate Fellowship activity. This research has made use of NASA’s Astrophysics Data System.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We consider random systems of equations $x_1+ \dots + x_k=a,$ $0 \le a \le 2$ which are interpreted as equations modulo $3.$ We show for $k \ge 15$ that the satisfiability threshold of such systems occurs where the $2-$core has density $1.$ We show a similar result for random uniquely extendible constraints over $4$ elements. Our results extend previous results of Dubois/Mandler for equations $\mod 2$ and $k=3$ and Connamacher/Molloy for uniquely extendible constraints over a domain of $4$ elements with $k=3$ arguments.
Our proof technique is based on variance calculations, using a technique introduced Dubois/Mandler. However, several additional observations (of independent interest) are necessary.
author:
- 'Andreas Goerdt, Lutz Falke'
title: 'Satisfiability thresholds beyond $k-$XORSAT'
---
Introcuction
============
Contribution
------------
Often constraints are equations of the type $f(x_1, \dots , x_k)=a$ where $a$ is an element of the domain considered and $f$ is a $k-$ary function on this domain, for example addition of $k$ elements. Given a formula, which is a conjunction of $m$ constraints over $n$ variables we want to find a solution. It is natural to assume that $f$ has the property: Given $k-1$ arguments we can always set the last argument such, that the constraint becomes true. In this case we can restrict attention to the $2-$core. It is obtained by iteratively deleting all variables which occur at most once. Thus it is the maximal subformula in which each variable occurs at least twice.
We consider the random instance $F(n, p):$ Each equation over $n$ variables is picked independently with probability $p;$ the domain size $d$ and the number of slots per equation $k$ is fixed. We consider the case $p=c/n^{k-1}$ and the number of constraints is linear in $n$ whp. (with high probability, that is probability $1\,-\,o(1), n$ large. ) The density of a formula is equal to the number of equations divided by the number of variables. The following is well known:
\[[@MOLLOY]\] \[Core\] 1. Conditional on the number of variables $n'$ and equations $m'$ of the $2-$core the $2-$core is a uniform random member of all formulas where each variable occurs at least twice.\
2. There exist $n'=n'(c)$ and $m'=m'(c)$ such that the number of variables of the $2-$core is $n'(1+o(1))$ and the number of equations $m'(1+o(1))$ whp.\
3. There exists a $T$ such that whp. for $c \le T-\varepsilon$ the $2-$core has density $\le 1-\varepsilon$ and for $c \ge T+\varepsilon$ the $2-$core has density $\ge 1+ \varepsilon.$ $T$ is determined as the solution of an analytical equation.
The expected number of solutions of the $2-$core is $d^{n-m} ,n$ the number of variables, $m$ the number of equations. When the $2-$core has density $\ge 1+ \varepsilon$ whp. no solution exists. This holds in particular when the density of $F(n, p)$ itself is $\ge 1+\varepsilon.$ The formulas considered here always have density $<1.$ In seminal work Dubois and Mandler [@DUMA] consider equations $\mod 2:$ $x_1 \, +\, \dots \, + \,x_k=a,$ $0 \le a\le 1, k=3.$ They show satisfiability whp. when the $2$-core has density $\le 1-\varepsilon.$ For larger $k\ge 15$ a full proof for this result is given in [@MITZ], Appendix C . Thus $T/n^{k-1}$ is the threshold for unsatisfiability in this case.
It is a natural conjecture that the same threshold applies to equations as discussed initially (and to some other types.) However, it seems difficult to prove the conjecture in some generality. One of the difficulties seems to be that we have $2$ parameters $k$ and $d.$ We make some progress towards this conjecture. We show it for equations $\mod 3. $ (The result is for $k>15,$ but we think it mainly technical to get it for all $k \ge 3.$)
\[mod3\] Let $F(n, p)$ be the random set of equations $\mod 3:$ $ x_1 + \dots + x_k \,=\, a , \, 0 \le a \le 2,$ $x_1+ \dots +x_k$ an ordered $k-tuple$ of variables. If $p < (T-\varepsilon)/n^{k-1}$ $F(n, p)$ is satisfiable whp. for $k>15.$
The main task is to show that a $2-$core of density $\le 1-\varepsilon$ has a solution with probability $>\varepsilon>0$ . Our proof starts as Dubois/Mandler: Let $X$ be the number of satisfying assignments of the $2-$core. Its expectation is $ \ge d^{\varepsilon n}, d=3.$ We show that $E[X^2] \le O(E[X])^2.$ This implies (by Cauchy-Schwartz (or Paley-Zygmund) inequality) that the probability to have a solution is $\ge \varepsilon >0.$ By Fact \[Core\] $F(n, p)$ has a solution with the same probability. We apply Friedgut-Bourgain’s Theorem to $F(n, p)$ to show that unsatisfiability has a sharp threshold. By this the probability becomes $1-o(1).$ In [@CRDA] Friedgut-Bourgain is applied to the $\mod 2-$case. It seems that our proof for the $\mod 3-$case is somewhat simpler (and applies to the $\mod 2-$case and other cases.)
To determine $E[X^2]$ Dubois/Mandler apply Laplace Method (one ingredient: bounding a sum through its maximum term.) The main difficulty is to bound a real function of several arguments from above. They show that their function has only one local maximum. We proceed by the same method, but substantial changes are necessary for $ k>3.$
First, we observe (cf. [@MITZ], Appendix C) that the function in question is $\le$ the [*infimum* ]{} with respect to certain other parameters. This is based on generating functions: If $f(x)=\sum c_kx^k$ then $c^k \le f(a)/a^k,\, a>0, c_i\ge 0$ (a method rarely used in the area, a notable exception is [@PUY].) Thus to bound the maximum from above we need to find suitable parameters and show that the value with respect to these parameters is less than the required upper bound . (This leads to involved, but elementary calculus. )
To make this approach work we need appropriate generating functions: $X=X_{a_1}+ \dots X_{a_{3^n}},$ where $X_{a_i}$ is the indicator random variable of the event that assignment $a_i$ makes the formula true. Then $X^2= \sum_a \sum_b \, X_aX_b.$ To get $E[X^2]$ we need to determine Prob$[X_aX_b=1].$ To this end we observe that the equation $x_1+ \dots + x_k=c$ which is true under $a$ is true exactly under those assignments $b$ such that $0k_0 + 1k_1 +2k_2 = 0 \mod 3, $ and $k_i$ is the number of slots of $x_1+ \dots +x_k$ filled with a variable $x$ with $b(x)=a(x)+i.$ Thus there are $\sum_{k_1=k_2 \mod 3} {k \choose k- k_1-k_2, k_1, k_2}$ different ways in which an equation can become true under $a, b.$ The following generating function allows us to deal with these possibilities analytically. With $\bf w_1 \,= \, \exp(2\pi \i/3)$ the primitive third root of unity and $ \bf w_2\,=w_1^2$ we define $r(x_0, x_1, x_2) = \frac{1}{3} \left[ (x_0\,+\,x_1\,+\,x_2 )^k \,+ \,
(x_0 \, + \,{ \bf w_1} x_1 + {\bf w_2} x_2)^k \,+ \, (x_0 \,+ \,{ \bf w_2} x_1 \,+ \,
{\bf w_1} x_2)^k \right]$\
then Coeff$[x_1^{k_1}x_2^{k_2}, r(1,\, x_1, \, x_2)]= {k \choose k-k_1-k_2, k_1, k_2}$ if $k_1=k_2\mod 3$ and $0$ otherwise (easy from properties $ \bf w_j.)$ In the $\mod 2-$case we use $1/2\left[(1+x)^k \,\, + \,\, (1-x)^k\right]$ instead [@MITZ], Appendix C.\
With the motivation to get an exact threshold of unsatisfiability for a type of constraint whose worst-case complexity is NP-complete, Connamacher/Molloy [@COMO] see also the very recent [@CO] introduce uniquely extendible constraints. A $k-$ary uniquely extendible constraint is a function from $D^k$ to true, false with the property: Given values from $D$ for any $k-1$ argument slots there is exactly one value for the remaining slot which makes the constraint true. (The $k> 8$ in the following result can be eliminated at the price of some additional technical effort.)
\[uni\] Let $F(n, p)$ be the random formula of uniquely extendible constraints: Each constraint is a random $k-$tuple of variables and a $k-$ary uniquely extendible constraint over $D$ and we pick with probability $p.$ For $|D|=4 $ and $p < (T- \varepsilon)/n^{k-1}$ $F(n, p)$ is satisfiable whp. for $k>8.$
The threshold $T/n^{k-1}$ is proved for $k=3$ and $|D|=4$, cf. [@CO] remark following Theorem 8. Our proof uses the technique as in the $\mod 3-$case, however the details are different. One of the contributions making is the generating polynomial\
$ p(x)=\frac{1}{d} \left[(1+x)^k+ (d-1)(1-\frac{x}{d-1})^k \right],$ as $r(x_0, x_1, x_2)$ above, not used before.
Motivation
----------
Many computational problems can be naturally formulated as conjunctions of constraints. And we are interested to find a solution of this conjunction. Algorithmic properties of these conjunctions are considered in theoretical research (with remarkable results e. g. in the realm of approximation[@HA]) and applied research, e. g. [@MEI]. An additional aspect is the investigation of conjunctions of randomly picked constraints; [@MO] is a fundamental study here. Propositional formulas in $k$-conjunctive normalform provide an example which has lead to a rich literature e. g. [@DIAZ]. One of the characteristic properties of this research is that its findings can often be related to experimental work by running algorithms on randomly generated instances.
One of the aspects of random formulas is a threshold phenomenon: If the number of constraints of a conjunction picked is less than a threshold value the conjunction is typically satisfiable, if it is more we get unsatisfiability whp. Moreover instances picked close to the threshold seem to be algorithmically hard, thus being candidate test cases for algorithms. The threshold phenomenon and the possibility to investigate it by experiments causes physics to become interested in the area e. g. [@MEZE]. On the other hand, physical approaches lead to new algorithms and classical theoretical computer science research, e. g. [@COPRI].
One of the major topics is to determine the value of the threshold in natural cases. A full solution even in the natural $k-$CNF SAT case has not been obtained, but many partial results, [@DIAZ] for $k=3.$ Note that $k-$CNF does not have the unique extendability property as possessed by the constraints considered here. And it seems to be a major open problem to get the precise threshold for constraints without unique extendibility and not similar to $2-$CNF. A mere existence result is the Friedgut-Bourgain theorem [@FRI]. Based on this theorem thresholds for formulas of constraints over domains with more than $2$ elements are considered in [@MO]. Ordering constraints are considered in [@GO], only partial results towards a threshold can be proven. In order to get definite threshold results further techniques are required. Therefore it is a useful effort to further develop the techniques with which thresholds can be proven. This is the general contribution of this paper.
A notable early exception, in that the precise threshold can be proven is the $\mod 2-$case considered above. Historically [@DUMA] is the first paper which uses variance calculation based on Laplace method in this area. Subsequently, for $k-$CNF SAT this method has lead to substantial progress in [@ACMO]. The contribution here is that $\mod 2-$proof can be refined and extended to cover other cases based on observations of independent interest. Note that random sparse linear systems over finite fields are used to construct error correcting codes, e. g. [@LMSS] or [@RU], motivating the $\mod 3-$case. A very recent study of the $\mod 2-$ case is [@ACHLIO]. More literature can be found in [@KO], but precise threshold results have not been obtained.
Contents
--------
---------------------------------------------------------------------------- --
I. Equations modulo $3$
1\. Notation and basics
2\. Outline of the proof of Theorem \[EX2EI\]
3\. Proof of Theorem \[OPT\]
3.1 Proof of Lemma \[lem1\]
3.2 Proof of Lemma \[lem2\]
3.3 Proof of Lemma \[lem3\]
3.4 Proof of Lemma \[lem4\]
4\. Proof of Theorem \[LAPLA\]
5\. Remaining proofs
5.1 Local limit consideration
5.2 The sharp threshold
II\. Uniquely extendible constraints
1\. Outline
2\. Proof of Theorem \[UNOPT\] for $d=4,\, s\ge 7 , \, \lambda \le 1-1/d$
3\. Proof of Theorem \[UNOPT\] for $d=4,\, \lambda \ge 1-1/d, s \ge 5.$
---------------------------------------------------------------------------- --
\
\
[**I. Equations modulo $3$** ]{}
Notation and basics
===================
We use the abbreviation $$\begin{aligned}
M(m, n) \,:= \, \sum_{ v_1, \dots v_n \,\ge 2}\, {m \choose v_1,\dots v_n } \mbox{ and }
N_0 \,:= \, M(km, n). \mbox{ Then } N_0\cdot 3^m \label{DEFN0}
\end{aligned}$$ is the number of all formulas with $ k $ variables per equation and $ m $ equations. We consider the uniform distribution on the set of formulas. Note that the formulas we consider are $2-$cores. (Here the same equation to occur several times. This happens with probability $o(1)$ as $m$ is linear in $n$ and can be ignored. ) Let $X$ be the number of solutions of a formula. We have $X=\sum_a\, X_a$ where $a$ stands for an assignment of the variables with $0,1, 2$ and $X_a(F)=1$ if $F$ is true under $a$ and $0$ otherwise. The expectation of $X$ is $3^{n-m}$ because given an assignment each equation is true independently with probability $1/3.$ We assume that $m=\gamma n , \gamma $ bounded above by a constant $<1.$ As $k$ is also constant, the asymptotics is only with respect to $n.$ We need to show the following theorem
\[EX2\] $\mbox{E}[X^2] \le C \cdot 3^{2(n-m)} $
We have E$[X^2] \,= \, \sum_{(a, b)}\, \mbox{E}[X_a \cdot X_b] $ where $(a, b)$ refers to all ordered pairs of assignments.
Let $\overline{W}=(W_0, W_1, W_2)$ be a partition of the set of variables into $3$ sets.We always use the notation $w_i = \sharp W_i,\, \, \bar{w}\, = \, (w_0, w_1, w_2).$ For two assignments we write $b \, = \, D(a, \overline{W}) $ iff $W_i= \{x\, | \,b(x)=a(x) +i \mod 3 \}.$ We have that $a(x_1 +\dots +x_k)=b(x_1 +\dots +x_k)$ (Here $ \, a(x_1 +\dots +x_k)$ is the value of $ x_1 +\dots +x_k$ under $a$ (analogously for $b$).) iff $\sum_{i=0, 1, 2} \,i \cdot \sharp \{ j \,| \,x_j \in W_i\} \,= \,0 \mod 3.$ This is equivalent to $ \sharp \{ j \,| \,x_j \in W_1\}\,= \, \sharp \{ j \,| \,x_j \in W_2\} \mod 3.$ Given $\bar{l}= (l_0, l_1, l_2) $ with $\sum l_i \, =\, km $ we let ${ \cal K}(\bar{l})$ be the set of all $3\times m-$matrices $(k_{i, j})_{0\le i\le 2, 1\le j \le m}$ with $k_{1, j}= k_{2, j} \mod 3 $ and each column sums to $k$, that is $\sum_i k_{i, j}\,= k$ for each $j.$ Moreover, $ \sum_j k_{i,j} \,=\, l_i $ for $ i=0, 1, 2 $ ( the $i'$th row sums to $l_i.)$
We denote $$\begin{aligned}
K(\bar{l}) :=
\sum_{(k_{i, j})\in {\cal K}(\bar{l})}\,\prod_{j=1}^m{ k \choose k_{0, j}\,, k_{1, j}\,, k_{2, j}} .\, \,
\mbox{ Then } \, \, \hat{N}(\bar{w}, \bar{l})\,:= \, K(\bar{l}) \cdot \prod_{i=0}^2 M(l_i, w_i) \label{NUFO}\end{aligned}$$ is the number of formulas $F$ true under two assignments $a, b$ with $b\,= \, D(a, \overline{W}) $ (with $w_i =\sharp W_i)$ and the variables from $W_i$ occupy exactly $l_i$ slots of $F.$ The factor $K(\bar{l})$ of $\hat{N}(\bar{w}, \bar{l})$ counts how the $l_i$ slots available for $W_i$ are distributed over the left-hand-sides of the equations. The second factor counts how to place the variables into their slots. Note that the right-hand-side of an equation cannot be chosen, it is determined by the value of the left-hand-side under $a, b.$
We abbreviate ${n \choose \bar{w}}\,= \, { n \choose w_0, w_1, w_2}. $ Given an assignment $a, $ $\bar{w}, $ and $\bar{l}, $ the number of assignment formula pairs $(b, F)$ with : There exist $\overline{W}$ with $\sharp W_i= w_i,$ such that $b \in D(a, \overline{W}),$ $F$ is true under $a$ and $b$, and the variables from $W_i$ occupy exactly $l_i$ slots of $F$ is $$\begin{aligned}
N( \bar{w}, \, \bar{l})\,
:= {n \choose \bar{w}} \cdot \hat{N}( \bar{w}, \, \bar{l}). \, \, \mbox{ This implies } \, \, \mbox{E}[X^2]\,= \, 3^n \cdot \sum_{\bar{w}, \bar{l}}\, N(\bar{w}, \,\bar{l}) \cdot \frac{1}{3^m \cdot N_0}. \label{EWQU} \end{aligned}$$
Theorem \[EX2\] follows directly from the next theorem:
\[EX2EI\] $\sum_{\bar{w}, \bar{l}} N(\bar{w}, \bar{l})/N_0 \, \le \, C\cdot 3^{(1- \gamma )n} .$
One more piece of notation: $ \omega_i= w_i/n$ usually is the fraction of variables belonging to $W_i.$ And $\lambda_i\,= l_i\,/ (km)=l_i/(k\gamma n)$ is the fraction of slots filled with a variable from $W_i.$ We use $\bar{\omega} = (\omega_0, \omega_1, \omega_2), $ and $\bar{\lambda}=(\lambda_0, \lambda_1, \lambda_2).$ Sometimes $\omega_i, \lambda_i$ stand for arbitrary reals, this should be clear form the context.
Outline of the proof of Theorem \[EX2EI\]
===========================================
First, bounds for $M(m, n)$ and $K(\bar{l}).$ We consider $q(x):= \exp(x)-x-1= \sum_{j \ge 2} \frac{x^j}{j!} $ for $x\ge 0.$ Then for $a>0$ and all $m, n$ $$\begin{aligned}
M(m, n) \,= \, \mbox{Coeff}[x^m, q(x)^n]\cdot m! \, < \, q(a)^n \cdot \frac{1}{a^m } \cdot m!\,\le \, q(a)^n \left(\frac{m}{a\cdot e}\right)^m \cdot O(\sqrt{m}) \label{M1}\end{aligned}$$ using Stirling in the form $ m! < (m/e)^m \cdot O(\sqrt{m}).$
To get rid of the $\sqrt{m}-$factor we let $Q(x):= xq'(x)/q(x)$ with $q'(x)$ the derivative of $q(x), \,q'(x)=\exp(x)-1$ for $x>0.$ Then $Q'(x)>0$ for $x>0,$ $Q(x)>x,$ and $Q(x)\longrightarrow 2$ for $x \longrightarrow 0.$ Thus, for $y>2$ the inverse function $Q^{-1}(y)>0$ is defined and differentiable. Lemma \[MLOC\] is proved in Section \[REPRO\].
\[MLOC\] Let $Cn \ge m \ge (2+\varepsilon)n, \, \, C, \varepsilon>0$ constants. Then $$\begin{aligned}
M(m, n) \,= \, \Theta(1)\cdot \left( \frac{ m}{ae}\right)^m\cdot q(a)^n \mbox{ with } a \mbox{ defined by } Q(a)=\frac{m}{n} \nonumber\end{aligned}$$
Throughout we use $s=s(k, \gamma)$ uniquely defined by $Q(s) = k \gamma= k \gamma n/n=km/n.$ Note that for $k \ge 3$ we can assume that $k \gamma >2$ and $s$ always exists. We have $Q(s) \ge s.$ We often write $Q$ instead of $Q(s).$ Recall $N_0=M(km, n)$ and we get a tight bound on the number of formulas (cf. (\[DEFN0\]).)
\[N0\] $ N_0 \,= \, \Theta(1) \left(k \gamma n/(s e) \right)^{k \gamma n}\cdot q(s)^n. $
We treat the sum $ K(\bar{l})$ similarly to $M(m, n).$ Instead of $q(x)$ we use the function, $$\begin{aligned}
r(\bar{x}):= \sum_{k_1 = k_2 \mod 3} { k \choose k_0, k_1, k_2} x_0^{k_0}x_1^{k_1} x_2^{k_2}, \, \, \bar{x}=(x_0, x_1, x_2).\, \mbox{ Then} \, \,\nonumber \\
{K(\bar{l})} \, =\,
\sum_{(k_{i, j})\in {\cal K}(\bar{l})}\,\prod_{j=1}^m{ k \choose k_{0, j}\,, k_{1, j}\,, k_{2, j}} \, \, = \, \mbox{Coeff}[\bar{x}^{\bar{l}},\, r(\bar{x})^m\, ] \,
< \,\frac{ r(\bar{c})^m}{\bar{c}^{\bar{l}}} \label{K1}\end{aligned}$$ with the notation $ \bar{x}^{\bar{l}} = \prod_i x_i^{l_i}$ and $\bar{c}=(c_0, c_1, c_2)>0,$ meaning $c_i >0$ for all $i.$
For calculations it is useful to have a different representation of $r(\bar{x}).$ Let $\i$ be the imaginary unit, and ${\bf w_1}\,:= -1/2+ (\sqrt{3}/2) \i$ is the primitive third root of unity, ${ \bf w_2} := -1/2- (\sqrt{3}/2 )\i\,= \, {\bf w_1}^2.$ We have $$\begin{aligned}
r(\bar{x}) =
\frac{1}{3} \left[ (x_0\,+\,x_1\,+\,x_2 )^k \,+ \,
(x_0 \, + \,{ \bf w_1} x_1 + {\bf w_2} x_2)^k \,+ \, (x_0 \,+ \,{ \bf w_2} x_1 \,+ \,
{\bf w_1} x_2)^k \right] \label{RCOM}\end{aligned}$$ The preceding equation is well known and easy to prove from basic properties of roots of unity. Note that in derivatives $\frac{d}{dx_i} r(\bar{x})$ the roots of unity are treated as constants.
For $ x_i, y_i >0$ we define (convention $\alpha^\alpha=1$ for $\alpha=\omega_i$ or $\alpha=\lambda_i$ and $\alpha =0)$ $$\begin{aligned}
\Psi( \bar{\omega}\, ,\, \bar{\lambda}\, ,\, \bar{x} \,,\, \bar{y} \, )\,= \,
\prod_{i=0,1,2}\left( \frac{q(x_i )}{\omega_i q(s)} \right)^{\omega_i} \cdot
\left[ \prod_{i=0, 1, 2} \left(\frac{\lambda_i s }{x_i y_i}\right)^{\lambda_i } \right]^{k \gamma}
r(y_0, y_1, y_2)^\gamma \nonumber \end{aligned}$$ With $\omega_i=\lambda_i=1/3, a_i=s(k, \gamma)=s, $ and $c_i=1,$ we have $\Psi(\overline{\omega}, \overline{\lambda}, \bar{a}, \bar{c})=
3\cdot (1/3)^{k\gamma}\cdot ((1/3)3^k)^\gamma= 3^{1-\gamma} $ (use (\[RCOM\]).)
\[LEMBA\] $ N(\bar{w}, \bar{l})/N_0 \, \, < \, \, \Psi( \bar{\omega}\, ,\, \bar{\lambda}\, ,\, \bar{a} \,,\, \bar{c} \, )^n\cdot O(n)^{3/2}$ for any $\bar{a}, \bar{c}, a_i, c_i>0.$
${n \choose \bar{w}} \le \prod_i (1/\omega_i)^{\omega_i n}$ for all $\bar{w}, $ ([@MITZEN], page 228 ) $\prod_{i=0, 1, 2} M(l_i, w_i)/N_0 \, \le \, \\
\prod_i \left( (l_i/(a_ie))^{l_i} q(a_i)^{w_i} O(\sqrt{l_i})\right)\cdot (es/(k\gamma n))^{k\gamma n}\cdot 1/q(s)^n \cdot O(1)$ with (\[M1\] ) and Corollary \[N0\]. Observe that $l_i= \lambda_i k \gamma n, \sum_i \lambda_i=1, \sum \omega_i=1.$ Concerning $K(\bar{l})$ apply (\[K1\]).
For reals $a, b$ we let ${\cal U}_\varepsilon(a, b)=\{(c, d)| \, \, |c-a|, |d-b|<\varepsilon \}$ be the open square neighborhood of $(a, b).$ The notation $\bar{\lambda}, \bar{\omega}\in {\cal U}_\varepsilon(a, b)$ is used to mean $(\lambda_1, \lambda_2), (\omega_1, \omega_2)\in {\cal U}_\varepsilon(a, b).$ Theorem \[OPT\] is proved in Section \[PROOPT\].
\[OPT\] For any $\bar{\lambda}>0 $ there exist $\bar{a}, \bar{c}\,>\,0$ such that:\
(1) $\Psi( \bar{\omega}\, ,\, \bar{\lambda}\, ,\, \bar{a} \,,\, \bar{c} \, )\, \le \, 3^{1-\gamma}. $\
(2) For any $\varepsilon>0, $ if $\bar{\lambda} \notin {\cal U}_\varepsilon(1/3, 1/3)$ then $\Psi( \bar{\omega}\, , \bar{\lambda}\, , \bar{a} \,, \bar{c} )\, \le \, 3^{1-\gamma} - \delta $ for a $\delta>0.$\
\[SUMOPT\] Let $U= {\cal U}_\varepsilon(1/3, 1/3)$ then $\sum_{\bar{\lambda} \notin U, \lambda_i>0 , \bar{\omega} } N(\bar{w}, \bar{l})/N_0 \, < \, C \cdot 3^{(1-\gamma)n}. $
The sum has only $O(n^4)$ terms. With Lemma \[LEMBA\] and Theorem \[OPT\] (2) we see that each term is bounded above by $ (3^{1-\gamma}- \delta)^n O(n)^{3/2}.$
To treat $(\lambda_1, \lambda_2) $ close to $(1/3, 1/3)$ we need a lemma analogous to Lemma \[MLOC\] for $K(\bar{l}).$ Let the function $R(x_1, x_2) \,= \, (R_1(x_1, x_2),\, R_2( x_1, x_2))$ be defined by $R_i(x_1, x_2) = \\= x_i r_{x_i}(1, x_1, x_2)/r(1, x_1, x_2)$ for $i=1, 2,$ $r_{x_i}(1, x_1, x_2)$ is the partial derivative of $r(1, x_1, x_2)$ wrt. $x_i.$ The Jacobi Determinant of $R(x_1, x_2) $ is $>0$ at $ x_1=x_2=1$ (proof Subsection \[LOLICO\].) Thus there is a neighborhood of $(1, 1)$ in which $R( x_1, x_2)$ is invertible and the inverse function is differentiable. We have that $R(1, 1)= (k/3, k/3).$ Thus for a suitable $\varepsilon $ and $(\lambda_1, \lambda_2) \in {\cal U}_\varepsilon(1/3, 1/3)$ we can define $(c_1, c_2)$ by $R(c_1, c_2)= (k \lambda_1, k\lambda_2).$ Moreover, $c_i=c_i(\lambda_1, \lambda_2)$ is differentiable. Lemma \[KLOC\] is proved in Subsection \[LOLICO\].
\[KLOC\] There is an $\varepsilon>0$ such that for $(\lambda_1, \lambda_2) \in {\cal U}_\varepsilon(1/3, 1/3)$ $$\begin{aligned}
K(\bar{l}) \,= \, O\left( \frac{1}{n}\right) \cdot \frac{r(1, c_1, c_2)}{c_1^{l_1}c_2^{l_2}} \mbox{ with } R(c_1, c_2)=(k\lambda_1, k\lambda_2) \mbox{ defining } c_1, c_2. \nonumber \end{aligned}$$
\[PSIINU\] There is $\varepsilon >0$ such that for $(\omega_1, \omega_2), ( \lambda_1, \lambda_2 ) \in {\cal U}_\varepsilon(1/3, 1/3)$ $$\begin{aligned}
\frac{N(\bar{w}, \bar{l})}{N_0} \, \, \le \,
\,O \left(\frac{1}{n^2}\right) \Psi( \bar{\omega}\, ,\, \bar{\lambda}\, ,\, \bar{a} \,,\, \bar{c} \, )^n \nonumber \end{aligned}$$ where $ Q(a_i)= l_i / w_i = \lambda_ik \gamma /\omega_i$ and $c_0=1$ and $ R(c_1, c_2)= (\lambda_1k, \lambda_2k). $
[*Comment.*]{} Observe that $\lambda_ik \gamma /\omega_i\,\approx \, k\gamma, \, a_i\approx s, c_i \approx 1.$
Our restriction on $\bar{\omega}$ implies that ${n \choose \bar{w}}\le O(1/n)\prod_i (1/\omega_i)^{\omega_i n}$ (Stirling), giving us one $O(1/n).$ We get $\prod_i M(l_i, w_i)/N_0 \le \\
\prod_i \left( (l_i/(a_ie))^{l_i} q(a_i)^{w_i} \right)\cdot (es/(k\gamma n))^{k\gamma n}\cdot 1/q(s)^n \cdot O(1)$ applying Corollary \[N0\] and Lemma \[MLOC\] for the $M(l_i, w_i).$ Concerning $K(\bar{l})$ apply Lemma \[KLOC\] which gives us a second factor $O(1/n).$ Otherwise the proof is as the proof of Lemma \[LEMBA\].
Lemma \[LOCMAX\] is proved in Section \[PROLAPLA\].
\[LOCMAX\] The function $ \Psi( \bar{\omega}\, ,\, \bar{\lambda}\, ,\, \bar{a} \,,\, \bar{c} \, )$ with $a_i, c_i$ given by $ Q(a_i) = \lambda_ik\gamma /\omega_i$ and $c_0=1$ and $ R(c_1, c_2)= (\lambda_1k, \lambda_2k) $ has a local maximum with value $3^{1-\gamma}$ for $\lambda_i= \omega_i=1/3.$ In this case we get $a_i=s$ and $c_i=1.$
\[SUMOMEGA\] Let $U= {\cal U}_\varepsilon(1/3, 1/3), \varepsilon$ small enough. Then $\sum_{\bar{\omega} \notin U, \bar{\lambda} \in U, \lambda_i>0 } N(\bar{w}, \bar{l})/N_0 < \, C \cdot 3^{(1-\gamma)n}. $
Let $\varepsilon>0$ be such that $\Psi(\bar{\omega}, \bar{\lambda}, \bar{a}, \bar{c}) \le 3^{1-\gamma} $ for $\bar{a}, \bar{c} $ as specified in Lemma \[LOCMAX\] and $\bar{\omega}, \bar{\lambda}\in U.$ Let $\bar{\omega } \notin U, \bar{\lambda} \in U.$ We show $\Psi(\bar{\omega}, \bar{\lambda}, \bar{a}, \bar{c}) \le 3^{1-\gamma} - \delta$ for some $\bar{a}, \bar{c}.$ This implies the claim as in the proof of Corollary \[SUMOPT\].
Let $\varepsilon'<\varepsilon/3$ and $U'= {\cal U}_{\varepsilon'}(1/3, 1/3).$ For $ \bar{\lambda} \notin U'$ the claim follows with Theorem \[OPT\] (2). For $ \bar{\lambda} \in U'$ we show that $\Psi(\bar{\omega}, \bar{\lambda}, \bar{a}, \bar{c}) \le 3^{1- \gamma}- \delta$ for $a_i=s$ and $ c_0=1, R(c_1, c_2)=(k \lambda_1, k \lambda_2).$ (Recall $\bar{\lambda} \in U.$) For $\Psi:= \Psi(\bar{\lambda}, \bar{\lambda}, \bar{a}, \bar{c})$ with $\bar{a}, \bar{c}$ as required by Lemma \[LOCMAX\] we have $\Psi \le 3^{1-\gamma}.$ Note, $Q(a_i)=\lambda_ik\gamma/\lambda_i= k \gamma$ which implies $a_i=s$ and $c_0=1, R(c_1, c_2)= (k \lambda_1, k \lambda_2).$ Therefore all $a_i-$terms cancel and $\Psi= \prod (1/\lambda_i)^{\lambda_i}\cdot \left(\prod (\lambda_i/c_i)^{\lambda_ik\gamma}\right)p(\bar{c})^\gamma \le 3^{1-\gamma}. $
As $\bar{\omega} \notin U$ whereas $\bar{\lambda} \in U'$ and $\varepsilon' \le \varepsilon/3$ we have that $\prod (1/\omega_i)^{ \omega_i } \, \le \prod (1/\lambda_i)^{\lambda_i}- \delta'$ for a $\delta' >0$ (proof omitted.) Then $\Psi(\bar{\omega}, \bar{\lambda}, \bar{a}, \bar{c}) \le \Psi - \delta'\left(\prod
(\lambda_i/c_i)^{\lambda_i k \gamma }\right)p(\bar{c})^\gamma .$ If $\Psi \le 3/2$, we are done. Otherwise we have that $\left(\prod_i(\lambda_i/c_i)^{\lambda_ik\gamma}\right)p(\bar{c})^\gamma $ is bounded below by $1/2$ (as $\prod (1/\lambda_i)^{\lambda_i} \le 3)$ and the claim follows, with with $\delta = (1/2)\delta'.$
Theorem \[LAPLA\] is proved in Section \[PROLAPLA\] by Laplace method.
\[LAPLA\] Let $U={\cal U}_\varepsilon(1/3, 1/3).$ There is an $\varepsilon>0$ such that\
$\sum_{\bar{\lambda} , \bar{\omega} \in U} N(\bar{w}, \bar{l})/N_0 \,< \, C \cdot 3^{(1-\gamma)n}.$\
[*Proof of Theorem \[EX2EI\].*]{} Pick $\varepsilon $ such that Theorem \[LAPLA\] applies. Use Corollary \[SUMOPT\], Corollary \[SUMOMEGA\], and Theorem \[LAPLA\] and the sum of all terms $N(\bar{w}, \bar{l})/N_0$ with $l_i>0$ is $\le C \cdot 3^{(1- \gamma)n}.$ Terms with an $l_i=0$ do not add substantially to the sum (proof omittted.)
Proof of Theorem \[OPT\] {#PROOPT}
========================
We use the notation $\bar{x}= (x_0,x_1, x_2), \bar{y}=(y_0, y_1, y_2)$ and define $$\begin{aligned}
\mbox{OPT}_1(\bar{x} , s)\,= \, \frac{ q(sx_0)}{q(s)} \,+ \, \frac{q(sx_1)}{q(s)} + \frac{ q(sx_2)}{q(s)} \nonumber \\
\mbox{OPT}_2(\bar{x}, \bar{y} , s)\,= \, \left( \frac{1}{x_0y_0+x_1y_1+ x_2y_2} \right)^{Q} , \, x_0y_0+x_1y_1+ x_2 y_2 >0\nonumber \\
\mbox{OPT}_3(\bar{y}, s)\,= \, \left( y_0+y_1+y_2 \right )^Q \,+ \,
2\cdot \left( y_0^2 \,+\,y_1^2\,+\, y_2^2\,- y_0 y_1\,- \, y_0 y_2\, - \, y_1 y_2 \right)^{1/2\cdot Q} \nonumber \\
\mbox{OPT}(\bar{x}, \bar{y}, s) \,=\,
\mbox{OPT}_1(\bar{x} , s) \cdot
\mbox{OPT}_2 ( \bar{x}, \bar{y} , s)\cdot
\mbox{OPT}_3( \bar{y}, s) . \nonumber \end{aligned}$$ Observe that OPT$(1, 1, 1, 1, 1,1 , s)= 3(1/3)^Q 3^Q=3=$ OPT$_1(1, 1, 1, s),$ OPT$(1,0,0, 1,0,0, s)=1\cdot (1/1)^Q \cdot 3=3=$OPT$_3(1, 0, 0,s).$ The following lemma shows the idea of OPT.
\[LEMOPT\] Given $\bar{\lambda}>0$ and let $\lambda$ be the maximum of the $\lambda_i.$ Let $ \, a_i, c_i \,>0$ be such that $P_i:= a_ic_i= \lambda_i/\lambda.$ Then $$\begin{aligned}
\Psi:= \Psi(\bar{\omega}, \bar{\lambda}, \bar{a}\cdot s, \bar{c}) \le \frac{1}{3^\gamma}
\mbox{OPT}(\bar{a}, \bar{c}, s). \nonumber \end{aligned}$$
The factors of $\Psi$ one by one: The first factor: The AGM-inequality gives\
$ \prod_{i=0, 1, 2} \left(\frac{q(a_i s)}{\omega_i q(s)} \right)^{\omega_i} \,\le \, \mbox{OPT}_1(\bar{a}, s). $ (Applies for $\omega_i=0$, too.)
The second factor: We have $P_0+P_1+P_2= a_0c_0+ a_1c_1+ a_2c_2 = 1/\lambda$ and $\lambda_i/a_ic_i= \lambda$ for $i=0, 1, 2.$ Recall $Q=k\gamma, $ and the second factor of $\Psi \,=$ $$\begin{aligned}
\prod_{i=0, 1, 2} \left( \frac{\lambda_is}{a_isc_i}\right)^{\lambda_ik\gamma} \,= \, \lambda^{k \gamma} =
\left(\frac{1}{a_0c_0+a_1c_1+a_2c_2}\right)^Q\,= \mbox{OPT}_2(\bar{a}, \bar{c}, s). \nonumber \end{aligned}$$
The third factor: We let $C_1=\sum_i c_i$ and $C_2= \sum_i c_i^2 -c_0c_1-c_0c_2-c_1c_2.$ Then $r(\bar{c}) =|r(\bar{c})| \le (1/3)(C_1^k\,+ \,2 C_2^{k/2})$ by the triangle inequality and as $|c_0 +{\bf w_1}c_1 + {\bf w_2}c_2|= [(c_0-1/2\cdot (c_1+c_2))^2+ (\sqrt{3}/2( c_1-c_2))^2]^{1/2}=C_2^{1/2}.$ Then $|r(\bar{c})|^\gamma \le 1/3^\gamma (C_1^k+2C_2^{k/2})^\gamma \le 1/3^\gamma (C_1^{k\gamma}+2^\gamma C_2^{\gamma k/2})\le1/3^\gamma \mbox{OPT}_3(\bar{c}, s)$ as $Q=k\gamma,$ and as $x^\gamma$ is concave (by $\gamma <1)$ we have $(y+z)^\gamma \le y^\gamma + z^\gamma.$
The following picture shows OPT$(1, a, a, 1, c, c, s), \, \, 0 \le a, c \le 1.$ The $\le 3-$area is dark. We have a path from $a=c=0$ to $a=c=1$ through this area. Therefore, for all $P$ with $0 \le P \le 1$ we have $0 \le a, c\le 1$ with $P=ac$ such that OPT$(1, a, a, 1, c, c, s) \le 3.$ In the notation of Lemma \[LEMOPT\] this corresponds to $\lambda_0 \ge \lambda_1=\lambda_2$ (and visualizes Theorem \[OPT\] for this case.) The following four lemmas are the technical core of our proof.
{width="49.00000%"} {width="49.00000%"}
\[lem1\] Let $s\ge 8, A(x)\,=\,A(x, s) := \,(7/10)Q \cdot x.$\
(a) OPT$(y)\,\,:=$ OPT $\left( 1, A(y), \, A(y), \, 1, y, \, y, \, s \right)$ is strictly decreasing for $ 0 \le y \le 1/(2Q).$ The start value is OPT$(0)=3.$\
(b) Given $0\le y \le 1/(2Q),$ OPT$(z):= $ OPT$(1, A(y+z), \, A(y-z), \,1, y+z, \, y-z, \, s)$ is decreasing in $0\le z \le y.$
\[lem2\] Let $ s\, \ge 7 \, \,$, and $ \frac{7}{20} \le A \le 1-\frac{1}{Q}.$ Then\
OPT$(z):= $ OPT$\left(1, A,\, A, \, 1, 1/(2Q)+z, \, 1/(2Q)-z \, , \, s \right) \le 3 \,- \, \delta$ for $0 \le z \le 1/(2Q).$
\[lem3\] Let $\, s\, \ge \, 7,$ and $ 1/(2Q) \le C \le 1/2.$ Then\
OPT$(z):= $ OPT$\left(1, 1-1/Q,\, 1-1/Q, \,1, C+z, \, C-z\,, s \right)\, \le 3 \,- \, \delta$ for $0 \le z \le C.$
\[lem4\] Let $s\ge 15$ and $ A(x)\,=\, A(x, s) \, := \, 1+ 7/(10Q)\cdot x \,- \, 7/(10Q).$\
(a) OPT$(y):= $OPT$\left(1, A(y), A(y), 1, y, y, s \right) $ is strictly increasing in $4/10 \le y \le 1.$ The final value is OPT$(1)=3.$\
(b) Given $4/10 \le y \le 1,$ OPT$(z):=$OPT$\left(1, A(y+z), A(y-z),1, y+z \, , y-z \, , \, s \right) $ is decreasing in $0 \le z \le \min\{y, 1-y\}.$
[*Proof of Theorem \[OPT\] from the preceding lemmas.*]{} We prove Theorem \[OPT\] for $\lambda_0 \ge \lambda_1 \ge \lambda_2>0 $ first. We denote $P_i:=\lambda_i/ \lambda_0, \, \mbox{ then } 1\ge P_1\ge P_2>0.$
[*Case 1: $P_1\, +\, P_2 \le \frac{7}{20 Q}$.*]{} With $A(x)$ from Lemma \[lem1\] we have $A(x)\cdot x \,= \, (7/10)Q \cdot x^2.$ Thus there exist $y_1 \ge y_2$ with $P_i= A(y_i)\cdot y_i.$ We represent $y_i$ such that Lemma \[lem1\] is applicable. $$\begin{aligned}
y\,:= \, \frac{y_1 + y_2}{2} ,\,\, z \,:= \, \frac{y_1 - y_2}{2}.
\mbox{ Then } y_1 = y+z , y_2 = y-z , 0 \le z \le y. \nonumber \end{aligned}$$ We show $y\le \frac{1}{2Q}$ and Lemma \[lem1\] applies to $y, z.$ $$\begin{aligned}
\frac{7}{20Q}\, \ge \, P_1+P_2 \,= \, \frac{7}{10}Q (y_1^2+ y_2^2) \Longrightarrow y_1^2+ y_2^2 \,\le \, \frac{1}{2Q^2}. \nonumber \\
(y_1+ y_2)^2 \le 2 y_1^2+2 y_2^2 \le \frac{1}{Q^2} \mbox{ and } y= \frac{y_1+y_2}{2}\le \frac{1}{2Q}. \nonumber \end{aligned}$$ With $a_0=c_0=1, a_1=A(y+z), a_2=A(y-z), c_1=y+z, c_2=y-z$ we have $a_ic_i=P_i.$ By Lemma \[LEMOPT\] $\Psi:= \Psi(\bar{\omega},\bar{\lambda}, \bar{a}\cdot s, \bar{c})\le 1/3^\gamma \mbox{OPT}, \,\,
\mbox{OPT}:= \mbox{OPT}(\bar{a}, \bar{c},s).$ If $P_1 \ge \varepsilon $ for an $\varepsilon>0$ we have OPT$< 3-\delta'$ by Lemma \[lem1\] and Theorem \[OPT\] holds.
For smaller $P_1$ we have OPT $ \le 3,$ approaching $3$. Only (1) of Theorem \[OPT\] holds. To get (2) for small $P_1$ we argue as follows: For $P_1$ approaching $0$ we see that $c_1$ and $c_2$ approach $0 .$ We consider the treatment of the factor $r(\bar{c})$ in the proof Lemma \[LEMOPT\]. Both $C_1$ and $ C_2$ from this proof approach $1$ in this case. Therefore we have a $\delta'>0$ such that $(C_1^k+2C_2^{k/2})^\gamma \le C_1^{k\gamma}+2^\gamma C_2^{k\gamma/2}- \delta'.$ As $a_0=c_0=1$ the first two factors of OPT do not approach $0.$ And we have $\Psi(\bar{\omega},\bar{\lambda}, \bar{a}\cdot s, \bar{c})\le (1/3^\gamma)(\mbox{OPT}- \delta'') \le 3^{1-\gamma}-\delta$ and Theorem \[OPT\] (2) holds.
[*Case 2: $\frac{7}{20Q} \le P_1\, +\, P_2 \le \left(1\,- \, \frac{1}{Q}\right) \frac{1}{Q}.$*]{} To use Lemma \[lem2\] we define $A$ by $A\cdot \frac{1}{Q}=P_1+P_2.$ and $A$ is as required by Lemma \[lem2\]. We need to find an appropriate $z.$ As $ P_1 \ge P_2 $ there is a $ y\ge \frac{1}{2}$ such that $ P_1=A\frac{1}{Q}y $ and $ P_2=A\frac{1}{Q}(1-y).$ With $ y=\frac{1}{2}+ z' $ and $ 1-y = \frac{1}{2}-z' , z' \le \frac{1}{2},$ and $ P_1= A\left(\frac{1}{2Q}+ \frac{z'}{Q}\right) \, , \, P_2\,= \, A\left(\frac{1}{2Q}-\frac{z'}{Q} \right) $ Lemma \[lem2\] applies with $z=z'/Q.$ Again we set $a_0=c_0=1$ and $a_1=a_2=A, c_1= \frac{1}{2Q}+ z, c_2= \frac{1}{2Q}- z.$ By Lemma \[LEMOPT\] $\Psi(\bar{\omega},\bar{\lambda}, \bar{a}\cdot s, \bar{c})\le 3^{1-\gamma} -\delta.$
[*Case 3: $ \left( 1 \,- \, \frac{1}{Q}\right) \frac{1}{Q} \le P_1+P_2 \le
1\,- \, \frac{1}{Q} . $*]{} Let $C$ be given by $\left( 1 \,- \, \frac{1}{Q}\right)\cdot C \,= \, \frac{P_1 + P_2}{2}.$ Then $C$ is as required by Lemma \[lem3\]. We have a $0 \le z' \le \frac{1}{2}$ such that $$\begin{aligned}
P_1 \,= \, \left( 1 \,- \, \frac{1}{Q}\right)\cdot C \cdot 2\left(\frac{1}{2}\, +\, z'\right)\, \, =
\left( 1 \,- \, \frac{1}{Q}\right)\cdot (C \, +\, 2Cz'), \nonumber \\
P_2 \,= \, \left( 1 \,- \, \frac{1}{Q}\right)\cdot (C - 2Cz').\nonumber \end{aligned}$$ With $z=2Cz'\le C $ Lemma \[lem3\] applies. We set $a_0=c_0=1$ and $a_1=a_2=1-1/Q$ and $c_1=C+z, c_2=C-z$ and finish the argument as in Case 2.
[*Case 4: $ P_1 + P_2 \ge 1- \frac{1}{Q}.$* ]{} With $A(x)$ as from Lemma \[lem4\] we have $A(x)\cdot x\,= \, \left( 1- \frac{7}{10Q} \right) x \,+ \, \frac{7}{10Q}x^2$ and $A(x)x $ increases from $0$ to $1$ for $0\le x \le 1.$ Let $y_i$ be such that $P_i\,=\,A(y_i)\cdot y_i.$ Then $y_2 \le y_1 \le 1$ and we can represent $y_i$ such that Lemma \[lem4\] is applicable. $$\begin{aligned}
y := \frac{y_1+ y_2}{2} , z:= \frac{y_1- y_2}{2},\mbox{ and } y_1=y+z, y_2=y-z, z \le y, 1-y. \nonumber \end{aligned}$$ We show that $1\ge y \ge 4/10$ and Lemma \[lem4\] applies to $y, z.$ We have $ P_1+P_2=A(y_1)y_1 + A(y_2)y_2 = y_1+y_2 + 7/(10Q)(y_1^2+y_2^2-y_1 - y_2) \le y_1+y_2.$ Therefore $y = (y_1+y_2)/2 \ge 1/2(1-1/Q)\ge 4/10 $ as $Q \ge s \ge 15.$ Setting $a_0=c_0=1, a_1=A(y+z), a_2=A(y-z), c_1=y+z, c_2=y-z$ implies the claim.
Now, assume the $\lambda_i$ are ordered in a different way. We apply the permutation leading from $ \lambda_0 \ge \lambda_1 \ge \lambda_2$ to the ordering considered to the $P_i , a_i , c_i$ above. The first two factors of $\Psi$ do not change, only $r(\bar{c})$ may change. But, Lemma \[LEMOPT\] still applies. The three factors, OPT$_1$, OPT$_2$, OPT$_3$ of OPT$(\bar{a}, \bar{c}, s)$ do not change. This refers to $C_1, $ and $C_2, $ too, and the argument above for $P_1$ small applies, too.
In the proofs to come in the following four subsections we use the notation $$\begin{aligned}
L(a, s)= \frac{q(as)}{q(s)}= \frac{\exp(as)-as-1}{\exp(s)-s-1}, \, K(a, s)= \frac{q'(as)}{q'(s)}\,= \,\frac{\exp(as)-1}{\exp(s)-1}, \nonumber \\
M(a, s)= \frac{\exp(as)}{\exp(s)}. \mbox{Then } aK(a, s) \le L(a, s) \le K(a, s) \le M(a, s) \,,\, 0 \le a \le 1 . \, \label{BAKL} \end{aligned}$$ [*Proof of (\[BAKL\].)* ]{} $p(x):=q'(x), \,K:= K(a, s), L:= L(a, s).$ For $a=0$ or $a=1$ we have $aK=L.$ For $a>0,$ $aK \le L \Longleftrightarrow ap(as)/q(as) \le p(s)/q(s) \Longleftrightarrow asp(as)/q(as) \le sp(s)/q(s).$ The preceding inequality holds trivially for $a=1.$ We show that $as p(as)/q(as)$ is strictly increasing in $a>0.$ We observe that $q(x)/(xp(x)) = 1/x- 1/p(x).$ The derivative is of the last expression is $<0$ iff $x^2+2< \exp(x)+1/\exp(x).$ For $x=0$ we have equality and several differentiations show the inequality.
For $a=0, L\le K$ is true. For $a>0$ $L \le K \Longleftrightarrow 1- sa/p(sa) \le 1-s/p(s).$ The last inequality follows from $a \ge p(sa)/p(s)$ for $0 \le a \le 1.$ This follows from convexity. $K(a, s) \le M(a, s)$ is very easy to show.
$$\begin{aligned}
\mbox{We also have } aK(a, s) \le \frac{7}{10} L(a, s), \,\mbox{ for } 0 \le a \le \frac{1}{2}, s\ge 4 \, \, \, \mbox{ (proof omitted.) } \label{BAKL7/10} \\
\mbox{ We recall } Q(x)= \frac{x q'(x)}{q(x)}=\frac{x (\exp(x)-1)}{\exp(x)-x-1}, Q=Q(s)=k\gamma , \, Q(s)> s. \nonumber \end{aligned}$$
Proof of Lemma \[lem1\]
------------------------
[**Lemma \[lem1\] (repeated)**]{} Let $s\ge 8, A(x)\,=\,A(x, s) := \,(7/10)Q \cdot x.$\
(a) OPT$(y)\,\,:=$ OPT $\left(1, A(y), \, A(y),1, \, y, \, y, s \right)$ is strictly decreasing for $ 0 < y \le 1/(2Q).$ The start value is OPT$(0)=3.$\
(b) Given $0\le y \le 1/(2Q),$ OPT$(z):= $ OPT$(1, A(y+z), \, A(y-z), \,1, \, y+z, \, y-z, \, s)$ is decreasing in $0\le z \le y.$
[*Proof of (a).*]{} We have $$\begin{aligned}
\mbox{OPT}(y) =
\left( 1+2L(A(y), s) \right)
\left(\frac{1}{1+2A(y)\cdot y}\right)^Q \left(\left(1\,+ \, 2y\right)^Q \,+ \,2\left(1\,- \, y \right)^Q \right). \nonumber \\
\mbox{ We write OPT}_1(y)\,= \, 1+2L(A(y), s).
\mbox{ Clearly OPT}(0)\,= \, 3 .\nonumber \\
\mbox{ We have } A' := \frac{d}{dy} \,A(y)\,= \,\frac{7}{10} Q. \, \, \, \, \mbox{ And }
\frac{d}{dy} \ln \mbox{ OPT }(y) \,>=< \, 0 \Longleftrightarrow \nonumber \\
\frac{A' \cdot 2 \cdot K(A(y), s)\,} {\mbox{ OPT}_1(y)}\,- \,
\frac{2A(y)\,+ \, 2A'\cdot y}{1+2A(y) \cdot y}\,+ \,\frac{2(1+2y)^{Q-1}- 2(1-y)^{Q-1}}{\left(1\,+ \, 2y\right)^Q \,+
\,2\left(1\,- \, y \right)^Q } \nonumber \\ \,>=<\,0 \label{EQKEY}\end{aligned}$$ The relationship (\[EQKEY\]) is obtained by taking the derivative and dividing by $Q.$ To get the first summand we look into the definition of $Q$ (the formula after \[BAKL7/10\].) $$\begin{aligned}
\frac{d}{dy} \ln \mbox{ OPT }_1(y)= \frac{ \frac{A' s\cdot 2 (\exp( A(y)s )-1)}{q(s)}} {\mbox{ OPT}_1(y)}\, ,\,
\frac{1}{Q(s)} \frac{A's\cdot 2 (\exp(A(y)s)-1)}{q(s)}= A'\cdot 2 K(A(y), s) \nonumber \end{aligned}$$
Observe that the first and third term of (\[EQKEY\]) is $\ge 0$ for $0 \le y \le 1$ whereas the second term is $\le 0.$ Moreover, $ A'\cdot y = A(y). $ We have that $\frac{d}{dy} \ln \mbox{ OPT }(y)\,<\,0$ if the following two inequalities both hold: $$\begin{aligned}
\frac{A' \cdot 2 \cdot K(A(y), s)}{\mbox{ OPT}_1(y)}\,< \, \frac{\, \frac{7}{10}A' \cdot y}{1+2A(y)\cdot y} \label{klsymugl1} \\
\frac{2(1+2y)^{Q-1}- 2(1-y)^{Q-1}}{\left(1\,+ \, 2y\right)^Q \,+ \,2\left(1\,- \, y \right)^Q }\,
< \frac{\frac{33}{10}A'\cdot y}{1+2A(y) \cdot y} \, \, \label{klsymugl2}\end{aligned}$$ Note that for $y=0$ both sides of the first inequality are equal to $0$ and of the second inequality, too. The derivative of OPT$(y)$ is $=0$ for $y=0.$
[*Comment:*]{} It is important to split up the left-hand-side of inequality (\[EQKEY\]), otherwise the calculations get very complicated. Equally important is the step leading to (\[EQKEY\]). Analogous steps will occur several times.
[*Proof of (\[klsymugl1\]) for $ 0 < y \le 1/(2Q) \, \, ,\,\, s \ge 7 $ .*]{} We abbreviate $K:=K(A(y), s)\, \, , L:=L(A(y),s).$ Note OPT$_1(y)\, = \, 1+2L.$ As $ A'>0$ we show $$\begin{aligned}
\frac{2 \cdot K}{1+2L} \,< \, \frac{\frac{7}{10}y}{1+2A(y)\cdot y}
\Longleftrightarrow 2K+ 4K\cdot A(y)\cdot y-2\frac{7}{10} L\cdot y\,< \, \frac{7}{10}y. \nonumber \\
\mbox{ By (\ref{BAKL7/10}) we know } K \cdot A(y) \le \frac{7}{10} L \mbox{ for } s \ge 4 \mbox{ as} A(y) \le \frac{1}{2}
\mbox{ ( by } y \le \frac{1}{2Q}. ) \end{aligned}$$ Thus (\[klsymugl1\]) follows from $ 2K+ 2K \cdot A(y) \cdot y < \frac{7}{10}y.$ As $ 2K \cdot A(y) \cdot y \, \le \, 2K$ and $2K$ is convex and $2K=0$ for $y=0$ we show that $2K<(7/20) y$ for $y=1/(2Q).$ For $y=1/(2Q)$ we have $ A(y)=7/20$ and $ 2K= 2(\exp((7/20)s)-1)/(\exp(s)-1).$ As $1/(2Q)=(\exp(s)-s-1)/(2s(\exp(s)-1))$ we have for $y=1/(2Q)$ $$\begin{aligned}
2K\,< \,\frac{7}{20}y \Longleftrightarrow 2 \left(\exp \left(\frac{7}{20}s\right)-1 \right)\,<\,\frac{7}{40} \frac{\exp(s)-s-1}{s} \nonumber \end{aligned}$$ This last inequality holds for $s \ge 7 $ (but not for $s\le 4.$ )\
[*Proof of (\[klsymugl2\]) for $y \le 1/Q $ and $s \ge 2.$* ]{} Inequality (\[klsymugl2\]) is equivalent to $$\begin{aligned}
2(1+2y)^{Q-1}- 2(1-y)^{Q-1} \, < \, \nonumber \\
< A(y)\left[\frac{33}{10}\left[\left(1\,+ \, 2y\right)^Q \,+ \,2\left(1\,- \, y\right)^Q\right] \,- \, 2y\cdot \left[ 2(1+2y)^{Q-1}- 2(1-y)^{Q-1} \right]\right] \,\, \label{klsymugl210} \\
\mbox{ The right-hand-side of (\ref{klsymugl210}) is } \,\ge \, \nonumber \\
A(y)\left[\frac{33}{10}\left[\left(1\,+ \, 2y\right)^Q \,+ \,2\left(1\,- \, y\right)^Q\right] \,- \, \frac{33}{10}y\cdot \left[ 2(1+2y)^{Q-1}- 2(1-y)^{Q-1} \right]\right] \nonumber \\
= \, \frac{33}{10} A(y) \left[\left(1\,+ \, 2y\right)^{Q-1}(1+2y-2y) \,+ \,2\left(1\,- \, y \right)^{Q-1}(1-y+y)\right]\nonumber \\
= \, \frac{33}{10} A(y) \left[\left(1\,+ \, 2y\right)^{Q-1} \,+ \,2\left(1\,- \, y \right)^{Q-1}\right].\nonumber \\
\mbox{ And (\ref{klsymugl210}) follows from }
\frac{ 2(1+2y)^{Q-1}- 2(1-y)^{Q-1}}{(1+2y)^{Q-1} + 2(1-y)^{Q-1}} \, < \, \frac{33}{10} A(y) \label{klsymugl21} \end{aligned}$$ For $y=0$ both sides of (\[klsymugl21\]) are equal to $0.$ We show that $33/10\cdot A'\,> \, $ the derivative with respect to $y$ of the left-hand-side of (\[klsymugl21\].) By elementary calculation $$\begin{aligned}
\frac{d}{dy} \frac{ 2(1+2y)^{Q-1}- 2(1-y)^{Q-1}}{(1+2y)^{Q-1} + 2(1-y)^{Q-1}}\,
=\,\frac{18 \cdot (Q-1)(1+y-2y^2)^{Q-2}}{\left[(1+2y)^{Q-1} + 2(1-y)^{Q-1}\right]^2}. \nonumber \\
\mbox{ We need to show }
\frac{33}{10}\frac{7}{10}Q\left[ (1+2y)^{Q-1} + 2(1-y)^{Q-1}\right]^2\,> \, 18 (Q-1)(1+y-2y^2)^{Q-2}. \nonumber \\
\mbox{ Enlarging the right-hand-side, } 1\le 1+y-2y^2 \mbox{ for } y \le 1/Q \le 1/s \le 1/2 \mbox{(by } Q(s)\ge s) \nonumber \\
\mbox{ we show } 33\cdot 7\left[ (1+2y)^{Q-1} + 2(1-y)^{Q-1}\right]^2\,> \,1800 (1+y-2y^2)^{Q-1} \,\nonumber \\
= \, 1800\left((1+2y)(1-y)\right)^{Q-1} \nonumber \\
\Longleftrightarrow
231\left[(1+2y)^{2(Q-1)}\, + \, 4\left(( 1+2y)(1-y) \right)^{Q-1} \, + \, 4(1-y)^{2(Q-1)} \right]\,> \nonumber \\
> \, 1800\left((1+2y)(1-y)\right)^{Q-1}
\Longleftrightarrow \mbox{ (Division by } \left((1+2y)(1-y)\right)^{Q-1} \mbox{)}\nonumber \\
\Longleftrightarrow \, \, \left(\frac{1+ 2y}{1-y}\right)^{Q-1}\, + \, 4\,+ \, \,4 \left(\frac{1-y}{1+2y}\right)^{Q-1} \,> \, 1800/231. \nonumber \\
\mbox{ Rescaling the fraction to } x \mbox{ the preceding inequality follows from } \nonumber \\
x+4\frac{1}{x} > 1800/231-4\,= 3.79 \dots\, \mbox{ true for } \, x > 0 .\nonumber \end{aligned}$$
[*Proof of (b).*]{} We assume $0 \le y \le 1/(2Q)$ and $ 0 < z \le y.$ $$\begin{aligned}
A(y+z)= \frac{7}{10}Q\cdot (y+z)\, \,\,\,, \, A(y+z)\cdot (y+z) \,= \, \frac{7}{10}Q \cdot (y+z)^2 \nonumber \\
A(y+z)\cdot (y+z)\,+ \, A(y-z)\cdot (y-z)\,= \, \frac{7}{10}Q\cdot 2(y^2+z^2) \nonumber \\
\mbox{OPT}(z)\,=
\, \left(1 \,+ \, L(A(y+z), s)\,+ \, L(A(y-z), s) \right) \cdot \nonumber \\
\cdot \left(\frac{1}{1 + \frac{7}{10}Q\cdot 2 (y^2+z^2) }\right)^Q\cdot
\left( (1+2y )^Q \,+ \, 2\cdot \left((1-y)^2 \,+ \, 3z^2 \right)^{Q/2}\right). \nonumber \end{aligned}$$
$$\begin{aligned}
\frac{d}{dz} \ln \mbox{OPT}(z) \,>=< \, 0 \Longleftrightarrow \nonumber \\
\frac{ \frac{7}{10}Q \cdot K(A(y+z), s)\,- \,\frac{7}{10}Q \cdot K(A(y-z), s) }
{1\,+ \, L(A(y+z), s)\,+ \, L(A(y-z), s) \,} \, \, - \, \, \frac{\frac{7}{10}Q 4 z}{1 + \frac{7}{10}Q\cdot 2 (y^2+z^2) }
\,+ \, \nonumber \\
\frac{6z \cdot ((1-y)^2+3z^2)^{Q/2-1}}{(1+2y )^Q \,+ \, 2\cdot ((1-y)^2+3z^2)^{Q/2} } \, \, >=<\,\,0. \nonumber \end{aligned}$$
The first term of the sum is obtained as the first term of (\[EQKEY\].) The first and third term of the left-hand-side of the preceding inequality are $\ge 0$ for $0\le z \le y$ whereas the second term is $\le 0.$
Analogously to (\[klsymugl1\]) and (\[klsymugl2\]) $\frac{d}{dz} \ln \mbox{OPT}(z)\,<\,0$ is implied by $$\begin{aligned}
\frac{ \frac{7}{10} Q \left[ K(A(y+z), s)\,- \, K(A(y-z), s) \right]}
{1\,+ \, L(A(y+z), s)\,+ \, L(A(y-z), s) }\, \, <\, \, \frac{\frac{9}{10}\frac{28}{10}Q z}{1 + \frac{14}{10}Q (y^2+z^2) }
\label{klasymugl1} \\
\frac{6z \cdot ((1-y)^2+3z^2)^{Q/2-1}}{(1+2y )^Q \,+ \, 2\cdot ((1-y)^2+3z^2)^{Q/2} }
\,< \,
\frac{\frac{1}{10}\frac{28}{10}Q z}{1 + \frac{14}{10}Q (y^2+z^2) } \label{klasymugl2}\end{aligned}$$
[*Proof of (\[klasymugl1\]) for $y \le 1/(2Q)$ and $ s \ge 3.5.$*]{} The denominator of the right-hand-side fraction is maximal for $y=z=1/(2Q).$ In this case it is $1+ 7/(10Q)< 1+1/Q.$ We lower the denominator of the left-hand-side simply to $1.$ The claim follows from $$\begin{aligned}
K(A(y+z), s)\,- \, K(A(y-z), s)\,< \, \frac{\frac{18}{5}z}{1+\frac{1}{Q}} \nonumber \end{aligned}$$ The left-hand-side of the preceding inequality is convex in $z$ for all $y<1/(2Q)$ (based on the convexity of $\exp(x)-\exp(-x).$) For $z=0$ both sides are $=0.$ Therefore it is sufficient to show that the inequality holds for $z=y$ where $y \le 1/(2Q).$ Setting $z=y$ yields $K(A(y-z), s)=0$ and we show $$\begin{aligned}
K(A(2y), s)\,\,< \, \frac{\frac{18}{5}y}{1+\frac{1}{Q}} \nonumber \end{aligned}$$ Again by convexity of the left-hand-side it is sufficient to show the inequality for $y=1/(2Q).$ In this case we need to show $$\begin{aligned}
K ( A(1/Q),s )\,= \, \frac{\exp\left(\frac{7}{10}s\right)-1}{\exp(s)-1}\,<\,\frac{18}{10}\frac{1}{Q+1} \nonumber \\
\mbox{ By (\ref{BAKL}) we know }\frac{\exp\left(\frac{7}{10}s\right)-1}{\exp(s)-1}\,\le \, \exp\left( -\frac{3}{10}s\right). \nonumber \\
\mbox{ And } \exp\left( -\frac{3}{10}s\right) \,< \, \frac{18}{10}\frac{1}{Q+1} \mbox{ holds (proof omitted) for } s\ge 3.5. \nonumber \end{aligned}$$
[*Proof of (\[klasymugl2\]) for $ s\ge 8.$* ]{} We show $$\begin{aligned}
\frac{(1+2y )^Q \,+ \, 2\cdot ((1-y)^2+3z^2)^{Q/2} }{6 z \cdot ((1-y)^2+3z^2)^{Q/2-1}}
\,> \,
\frac{1 + \frac{14}{10}Q (y^2+z^2) }{\frac{1}{10}\frac{28}{10}Q z }. \nonumber \\
\mbox{ Canceling } z \mbox{ in the denominator , setting } z=y \mbox{ on the right-hand-side, this follows from }\nonumber \\
\frac{(1+2y )^Q} {6 \cdot ((1-y)^2+3z^2)^{Q/2-1}}\,+ \, \frac{1}{3}((1-y)^2+3z^2)\,> \, \frac{1+\frac{14}{5}Qy^2}{\frac{28}{100}Q} \,= \,
\frac{100}{28Q} + 10y^2 \nonumber \\
\mbox{As } (1-y)^2+3z^2\,\le \, 1-2y+4y^2\,< 1 \mbox{ by }y\le 1/(2Q), Q\ge s \ge 8
\nonumber \\ \mbox{ this follows from }
\frac{1}{6}(1+2y)^Q \,+ \, \frac{1}{3}(1-y)^2\,>\, \frac{100}{28Q} + 10y^2 \nonumber \end{aligned}$$ The last inequality holds for $Q \ge 8, y \ge 0$ and then the claim holds as $Q \ge s.$
Proof of Lemma \[lem2\]
-----------------------
[**Lemma \[lem2\] (repeated)**]{} Let $ s\, \ge 7 \, \,$ and $ \frac{7}{20} \le A \le 1-\frac{1}{Q}.$ Then\
OPT$(z):= $ OPT$\left(1, A,\, A, \, 1, 1/(2Q)+z, \, 1/(2Q)-z\, , s\right) \le 3 \,- \, \delta$ for $0 \le z \le 1/(2Q).$
OPT$(z)\,= \,$ $$\begin{aligned}
\left( 1\,+ \ 2\cdot L(A, s) \right)
\left(\frac{1}{1+\frac{A}{Q}}\right)^Q
\cdot \left[ \left( 1+\frac{1}{Q} \right)^Q \,+ \, 2\left(\left(1-\frac{1}{2Q}\right)^2+ 3z^2\right)^{1/2\cdot Q} \right] \nonumber \end{aligned}$$ is increasing in $z.$ We show the claim for $z=1/(2Q).$ Let from now on OPT$(A)\,= \, $ OPT$(1, A, A, 1, 1/Q, 0, s)\,= \, $ $$\begin{aligned}
\left( 1\,+ \ 2\cdot L(A, s) \right)
\left(\frac{1}{1+\frac{A}{Q}}\right)^Q
\cdot \left[ \left( 1+\frac{1}{Q} \right)^Q \,+ \, 2\left( 1-\frac{1}{Q} + \frac{1}{Q^2}\right)^{1/2\cdot Q} \right] \nonumber \end{aligned}$$
First, we show that OPT$(A)$ has exactly one extremum in $0 \le A\le 1$ which is a minimum. $$\begin{aligned}
\frac{d}{dA} \ln \mbox{OPT}(A) \, > =< \,0 \Longleftrightarrow
\frac{2K(A, s)}{1+2L(A, s)} \,- \, \frac{\frac{1}{Q}}{1+ \frac{A}{Q}}\,>=< \,0 .\, \nonumber \end{aligned}$$ Concerning the first term of the preceding sum we refer to the explanation following (\[EQKEY\].) For $A=0$ the first term is $=0$ and the derivative is $<0.$ For $A=1$ the first term is $=2/3,$ whereas the second term is $1/(Q+1)\,<\,2/3$ for $Q>s>2,$ and the derivative is $>0.$ We show that the derivative is $=0$ for exactly one $ 0 < A < 1 $ which must be a minimum.
The second fraction of the derivative is decreasing in $A.$ We check that the first fraction is increasing. Abbreviating $L=L(A, s), \, L' \, = \, \frac{d}{dA}L(A, s) $ and analogously for $K,$ we get $$\begin{aligned}
\frac{d}{dA} \frac{2K(A, s)}{1+2L(A, s)} \,> \,0 \Longleftrightarrow
2K' (1+2L) \,> \, 2K2L' \Longleftrightarrow \nonumber \\
\mbox{ (Multiplication with }( \exp(s)-s-1)(\exp(s)-1), \mbox{division by } 2 \mbox{ and } s \mbox{.)} \nonumber \\
(\exp(s)-s-1) \exp(sA) \,+ \,\exp(sA) 2( \exp(sA)-sA-1 ) \,> \, 2(\exp(sA)-1)^2 \nonumber \\ \Longleftrightarrow
(\exp(s)-s-2sA)\exp(sA) \,>\,-\exp(sA)\,+ \, 2 \nonumber \\
\Longleftrightarrow \exp(s)-s-2sA \,> \,-1 + 2/\exp(sA) \nonumber \\
\mbox{ which is true for } s >2, 0< A <1 \mbox{ by convexity of } 2/\exp(sA). \nonumber \end{aligned}$$
We need to show the claim for the boundary values $A=7/20 \, $ and $ A=1-1/Q.$ First, $A=7/20:$ $$\begin{aligned}
1 \,+ \, 2 L(A, s) \, \le 1 + 2 M(A, s) \, = \, 1\,+ \, 2\exp(-13/20 \cdot s) \, \, \, \mbox{ (by (\ref{BAKL}).)}\nonumber \\
\mbox{ With the derivative of the logarithm and the Mean Value Theorem we can show that } \nonumber \\
\left(\frac{1+\frac{1}{Q}}{1+\frac{A}{Q}} \right)^Q \mbox{ is increasing in } Q \mbox{ towards its limit } \exp(13/20) .
\, \frac{ 2\left( 1-\frac{1}{Q} \,+ \, \frac{1}{Q^2} \right)^{1/2\cdot Q}}{\left(1+ \frac{A}{Q}\right)^Q} \nonumber \\
\mbox{ is decreasing in }Q =Q(s)\ge 2 \mbox{ (proof by standard calculus methods) and therefore }\nonumber \\
\mbox{ also in } s \mbox{ towards its limit } 2\exp(-17/20).\mbox{ For } Q=7 \mbox{ we get a value } \le 0.9 \nonumber \\
\mbox{ Therefore, for all } \ge s \, \ge 7 \mbox{( as } Q(s)\ge s) \nonumber \\ \mbox{OPT}(A)\,
< \, (1+ \, 2\exp(-13/20 \cdot 7))( \exp(13/20)+0.9 ) \,= \, 2.87\dots \, \,. \nonumber \end{aligned}$$
Now, $A=1-1/Q:$ $$\begin{aligned}
1\,+ \, 2L(A, s)\,\le \, 1 \,+ \, 2 M(A, s) \,= \,1\,+\, 2 \exp\left(-\frac{s}{Q}\right) \nonumber \\
\,=\, 1+2\exp\left( -\frac{\exp(s)-s-1}{\exp(s)-1}\right)
\mbox{decreasing in } s \mbox{ to } 1+2\exp(-1). \nonumber \\
\mbox{ For } s=7 \mbox{ we get } \, 1+2L(A, s) \le 1.7404 \dots \nonumber \\
\left( \frac{1+\frac{1}{Q}}{1+\frac{A}{Q}} \right)^Q\,= \, \left( \frac{1+\frac{1}{Q}}{1\,+\,\frac{1}{Q}\,- \, \frac{1}{Q^2}}\right)^Q
\mbox{ is decreasing in } Q =Q(s) \nonumber \\ \mbox{ (elementary proof omitted) and therefore in}\,\, s \mbox{ to } 1.\nonumber \\
\mbox{ For } Q=7 \mbox{ we get } 1.1344 \dots .\mbox{ As } Q(s)\ge s \, \mbox{this bound applies to } s=7 \mbox{, too.} \nonumber \\
\frac{ 2\left ( 1-\frac{1}{Q}+\frac{1}{Q^2} \right )^{1/2\cdot Q}}{\left(1+\frac{1}{Q}- \frac{1}{Q^2}\right)^Q}
\mbox{is again decreasing (proof omitted) in } \nonumber \\ Q
\mbox{ and } s \mbox{ to } 2\exp(-3/2).
\mbox{ For } Q=7 \mbox{ we get } 0.564 \dots \nonumber \\
\mbox{Altogether for } Q(s)\ge s \ge 7 \nonumber \\ \mbox{OPT}(A) \le 1.741\cdot(1.135+ 0.565)\,= 2.9597 . \nonumber \end{aligned}$$
Proof of Lemma \[lem3\]
------------------------
[**Lemma \[lem3\] (repeated)**]{} Let $\, s\, \ge \, 7$ and $ 1/(2Q) \le C \le 1/2.$ Then\
OPT$(z):= $ OPT$\left(1, 1-1/Q,\, 1-1/Q, \,1, C+z, \, C-z\,, s\right)\, \le 3 \,- \, \delta$ for $0 \le z \le C.$
We abbreviate $A=1-1/Q.$ First, analogously to the proof of Lemma \[lem2\] we can restrict attention to $z=C.$ OPT$(z)\,= \, $ $$\begin{aligned}
=\,\left(1+2L(A, s)\right) \left( \frac{1}{1+2AC}\right)^Q \left[(1+2C)^Q\,+ \, 2((1-C)^2+3z^2)^{1/2\cdot Q}\right] \nonumber \\
\mbox{ Let from now on OPT}(C)\,= \, \mbox{OPT}(1, A,A, 1, 2C, 0, s) \,= \,\nonumber \\
=\,\left(1+2L(A, s)\right) \left( \frac{1}{1+2AC}\right)^Q \cdot \left[(1+2C)^Q\,+ \, 2(1+ 4C^2-2C)^{1/2\cdot Q}\right] \nonumber \end{aligned}$$
OPT$(C)$ has exactly one extremum, which is a minimum for $0 \le C \le 1.$ $$\begin{aligned}
\frac{d}{dc} \ln \mbox{OPT}(C) \,>=<\, 0 \Longleftrightarrow \nonumber \\
- \frac{2A}{1+2AC} \,+ \, \frac{2(1+2C)^{Q-1} \,+ \, (8C-2)(1+ 4C^2-2C)^{1/2 \cdot Q-1}}{(1+2C)^Q \,+ \, 2(1+ 4C^2-2C)^{1/2\cdot Q}}\, >=< \,0
\Longleftrightarrow \nonumber \\
2A \left((1+2C)^Q\,+ \, 2(1+ 4C^2-2C)^{1/2Q}\right) \,- \nonumber \\
-\,2A C \left(2(1+2C)^{Q-1}\,+ \, (8C-2)(1+ 4C^2-2C)^{1/2 \cdot Q-1}\right)\, = \nonumber \\
\,= \, 2A\left[(1+2C)^{Q-1}\,+ \, (2-2C)(1+ 4C^2-2C)^{1/2 \cdot Q-1}\right]\, <\,\, = \, \, > \nonumber \\
\,<\,\,=\,\, >\,2(1+2C)^{Q-1}\,+ \, (8C-2)(1+ 4C^2-2C)^{1/2 \cdot Q-1} \Longleftrightarrow \nonumber \\
2A\,<\, \, =\, \, > \, \frac{2(1+2C)^{Q-1}\,+ \, (8C-2)(1+ 4C^2-2C)^{1/2 \cdot Q-1}}{(1+2C)^{Q-1}\,+(2-2C)(1+ 4C^2-2C)^{1/2 \cdot Q-1}} \, \Longleftrightarrow \, \nonumber \\
A\,<\, \, =\, \, > \, \frac{(1+2C)^{Q-1}\,+ \, (4C-1)(1+ 4C^2-2C)^{1/2 \cdot Q-1}}{(1+2C)^{Q-1}\,+(2-2C)(1+ 4C^2-2C)^{1/2 \cdot Q-1}} \nonumber \end{aligned}$$ For $C=0$ the right-hand-side fraction is equal to $0<A$ and OPT$(C)$ is decreasing. For $C=1$ the right-hand-side fraction is greater than $1>A$ and OPT$(C)$ is increasing.
Next we show that the preceding fraction is increasing in $0<C<1,$ and equality is attained for only one $C$ which must be a minimum. $$\begin{aligned}
\mbox{ Rewriting } 4C-1 \,= \, (2-2C)+6C-3 \mbox{ the fraction is rewritten as } \nonumber \\
1 \,+ \, \frac{(6C-3)(1+ 4C^2-2C)^{1/2 \cdot Q-1}}{(1+2C)^{Q-1}\,+(2-2C)(1+ 4C^2-2C)^{1/2 \cdot Q-1}} \nonumber \\
\mbox{ Rescaling }1/2\cdot Q-1 \mbox{ to } Q \mbox{ ( then } Q-1 \mbox{ scales to } 2Q+1 ) \mbox{ and } 2C \mbox{ to }C
\mbox{ we get } \nonumber \\
1 \,+ \, \frac{(3C-3)(1+ C^2-C)^Q}{(1+C)^{2Q+1}\,+(2-C)(1+ C^2-C)^Q} \, \, \nonumber \\
\mbox{ Dividing through } 3(C-1)(1+ C^2-C)^Q \mbox{ the preceding fraction is certainly increasing if } \nonumber \\
\frac{(1+C)^{2Q+1}}{3(C-1)(1+ C^2-C)^Q} \mbox{ and } \frac{2-C}{3(C-1)} \mbox{ are both decreasing for } 0 <C <2, C \neq 1. \nonumber \end{aligned}$$ The second fraction is easily seen to be decreasing. We show that the inverse of the first fraction is increasing. The numerator of its derivative is $$\begin{aligned}
\left[ (1+C^2-C)^Q \,+ \,( C-1)(2C-1)Q(1+C^2-C)^{Q-1}\right] \cdot (1+C)^{2Q+1}\, \,- \, \nonumber \\
- \, (C-1)(1+C^2-C)^Q\cdot (2Q+1)(1+C)^{2Q} \,\,
\,= \, (1+C)^{2Q}(1+C^2-C)^{Q-1}\cdot \nonumber \\
\left[(1+C)(1+C^2-C) \,+ \, (1+C)(C-1)(2C-1)Q \,- \, (2Q+1)(C-1)(1+C^2-C)\right] \nonumber \\
\mbox{ The expression in square brackets can be rewritten as} \nonumber \\
(1+C)(C-1)(2C-1)Q\,- \, 2Q(C-1)(1+C^2-C)\,+\, (-C+1+C+1)(1+C^2-C) \nonumber \\
=\,\, Q(1-C)^2\,+ \, 2(1+C^2-C)\,>\, 0 \nonumber \end{aligned}$$
Now it is sufficient to show the claim for the boundary values, $C=1/(2Q)$ and $C=1/2.$ The first case is contained in Lemma \[lem2\]. Let $C=1/2.$ We proceed as in the proof of Lemma \[lem2\], case $A=1-1/Q.$ $$\begin{aligned}
1+2L(A, s) \le 1.7404 \mbox{ for } s\ge 7 \nonumber \\
\left( \frac{1+2C}{1\,+\,2CA} \right)^Q\,= \, \left( \frac{2}{2\,- \, \frac{1}{Q}}\right)^Q
\mbox{ is decreasing in } Q =Q(s) \nonumber \\ \mbox{ (elementary proof omitted) and therefore in}\,\, s \mbox{ to } \exp(-1/2).
\nonumber \\
\mbox{ For } Q=7 \mbox{ we get } 1.67993\dots .\mbox{ As } Q(s)\ge s \, \mbox{this bound applies to } s=7 \mbox{, too.} \nonumber \\
\frac{ 2(1+ 4C^2-2C)^{1/2\cdot Q}}{(1+2AC)^Q } \,= \,\frac{2}{(2-\frac{1}{Q})^Q} \mbox{ decreasing to } 0 \nonumber \\
\mbox{ For } Q=7 \mbox{ we get } 0.02624\dots \nonumber \\
\mbox{ Altogether OPT}(C) \le 1.75\cdot(1.68+0.027)=2.98 \mbox{ for } s\ge 7.\nonumber \end{aligned}$$
Proof of Lemma \[lem4\]
-----------------------
[ **Lemma \[lem4\] (repeated)**]{} Let $s\ge 15$ and $ A(x)\,=\, A(x, s) \, := \, 1+ 7/(10Q)\cdot x \,- \, 7/(10Q).$\
(a) OPT$(y):= $OPT$\left(1, A(y), A(y), 1, y, y, s \right) $ is strictly increasing in $4/10 \le y < 1.$ The final value is OPT$(1)=3.$\
(b) Given $4/10 \le y \le 1,$ OPT$(z):=$OPT$\left(1, A(y+z), A(y-z),1, y+z, y-z, s \right) $ is decreasing in $0 \le z \le \min\{y, 1-y\}.$\
[*Proof of (a).*]{} We have OPT$(y) \,= \,$ $$\begin{aligned}
\left( 1+2L( A(y), s) \right)
\left(\frac{1}{1+2A(y)\cdot y}\right)^Q \left(\left(1\,+ \, 2y\right)^Q \,+ \,2\left(1\,- \, y \right)^Q \right) \nonumber \\
\mbox{ We write OPT}_1(y)= 1+2L(A(y), s). \mbox{ Clearly OPT}(1)\,= \, 3 \nonumber \\
\mbox{ We have } A' := \frac{d}{dy} \,A(y)\,= \,\frac{7}{10}\frac{1}{Q}. \nonumber \\
\frac{d}{dy} \ln \mbox{ OPT }(y) \,>=< \, 0 \Longleftrightarrow \mbox{(See comment to (\ref{EQKEY}.))} \nonumber \\
\frac{A' \cdot 2 \cdot K(A(y), s) }{\mbox{ OPT}_1(y)}\,- \,
\frac{2A(y)\,+ \, 2A'\cdot y}{1+2A(y) \cdot y}\,+ \,\frac{2(1+2y)^{Q-1}- 2(1-y)^{Q-1}}{\left(1\,+ \, 2y\right)^Q \,+ \,2\left(1\,- \, y \right)^Q }\,>=<\,0 \nonumber \end{aligned}$$ Observe that the first and third term of the preceding sum are $\ge 0$ for $0 \le y \le 1$ whereas the second term is $\le 0.$
We have that $\frac{d}{dy} \ln \mbox{ OPT }(y)>0$ if the following two inequalities both hold: $$\begin{aligned}
\frac{A' \cdot 2 \cdot K(A(y), s) }{\mbox{ OPT}_1(y)}\,> \, \frac{\, 2A'\cdot y}{1+2A(y)\cdot y} \label{grosymugl1} \\
\frac{2(1+2y)^{Q-1}- 2(1-y)^{Q-1}}{\left(1\,+ \, 2y\right)^Q \,+ \,2\left(1\,- \, y \right)^Q }\,>
\frac{2A(y)}{1+2A(y) \cdot y} \, \, \label{grosymugl2}\end{aligned}$$ Note that for $y=1$ both sides of the first inequality are equal to $7/(10Q) \cdot 2/3$ and of the second inequality $2/3.$ Therefore the derivative of OPT$(y)$ is $=0$ for $y=1.$\
[*Proof of (\[grosymugl1\]) for $1>y \ge \ 0\, \, ,\, \, \, s \ge 4$ .*]{} Let $ K= K(A(y), s)$ and $ L=L(A(y), s) .$ $$\begin{aligned}
\mbox{ As } A'>0 \mbox{ we need to show } \frac{K}{1+2L} \,> \, \frac{y}{1+2A(y)\cdot y}.\nonumber \\
\Longleftrightarrow K\,+ \, 2K\cdot A(y)y \,- \, 2L\cdot y\, > \, y \nonumber \\
\mbox{ As } K \ge L \, \, \mbox { by (\ref{BAKL}) this follows from } \nonumber \\
K\left( 1\, +\, 2A(y)\cdot y\, - 2y \right) \, \,
=\, K \left( 1+ 2\frac{7}{10Q}y^2 - 2\frac{7}{10Q}y)\right) \, > \, y \label{grosymugl11}\end{aligned}$$ For $y=1$ both sides of (\[grosymugl11\]) are $=1.$ For $y=0$ (\[grosymugl11\]) holds as $K>0$ in this case.
$K$ considered as a function in $y$ is convex, increasing and $> 0.$ The second term on the left-hand-side of (\[grosymugl11\]), $ 1- 2\frac{7}{10Q}y^2 + 2\frac{7}{10Q} y, $ is convex, $>0,$ and increasing for $y>1/2.$ Therefore the left-hand-side of (\[grosymugl11\]) is convex for $1/2<y<1.$ We next show that the derivative of the left-hand-side at $y=1$ is $<1.$ This implies that (\[grosymugl11\]) holds for $1/2 \le y <1.$
$$\begin{aligned}
\frac{d}{dy} K \left( 1 \,+\,2\frac{7}{10Q}y^2\,- \,2\frac{7}{10Q}y)\right)\,= \,
\frac{s\cdot \frac{7}{10Q}\cdot \exp(sA(y))}{\exp(s)-1} \nonumber \\
\cdot
\left(1\, \,\,+ 2\frac{7}{10Q}y^2\,- \,2\frac{7}{10Q}y\right)
\, + \, \frac{\exp(sA(y))-1}{\exp(s)-1}\cdot\left( 4\frac{7}{10Q}y\,- \,2\frac{7}{10Q}\right). \nonumber \\
\mbox{ Plugging in }\,\, y=1 \, \, \mbox{ yields }
\frac{7}{10Q} \left(\frac{s \exp(s)}{\exp(s)-1} + 2\right) \label{grosymugl12} \end{aligned}$$
For $s=4$ (\[grosymugl12\]) is $0.9837 \dots< 1. $ As (\[grosymugl12\]) is in decreasing in $s $ (proof omitted) (\[grosymugl11\]) holds for all $s \ge 4$ and $1/2 \le y <1.$
$ 1- 2\frac{7}{10Q}y^2 + 2\frac{7}{10Q} y$ is decreasing for $y<1/2.$ Therefore, for $0 \le y \le 1/2,$ we can bound the left-hand-side of (\[grosymugl11\]) from below by $$\begin{aligned}
K \left[ \left( 1 + 2\frac{7}{10Q}y^2 - 2\frac{7}{10Q}y \right)_{y=1/2} \right] \nonumber\end{aligned}$$ This function (the argument $y$ occurs only in $K$) is convex in $y$ . For $y=1/2$ it is $>y$ by the previous argument. For $y=1$ it is $<y.$ Therefore it is $>y$ for $0 \le y \le 1/2.$ The claim is shown.\
[*Proof of (\[grosymugl2\]) for $y \ge 4/10 $ and $ s \ge 3.5 .$* ]{} Inequality (\[grosymugl2\]) is equivalent to $$\begin{aligned}
2(1+2y)^{Q-1}- 2(1-y)^{Q-1} \, > \, \nonumber \\ > 2 A(y)\left[ \left( 1\,+ \, 2y\right)^Q \,+ \,2\left(1\,- \, y \right)^Q \,- \, y\cdot \left[ 2(1+2y)^{Q-1}- 2(1-y)^{Q-1} \right]\right] \,= \, \nonumber \\
= \, 2A(y)\left[ (1+2y)^{Q-1}(1+2y- 2y) \,+\, 2(1-y)^{Q-1} (1-y +y)\right) \nonumber \\ \,= \, 2A(y)\left[ (1+2y)^{Q-1} + 2(1-y)^{Q-1}\right]
\Longleftrightarrow \frac{ (1+2y)^{Q-1}- (1-y)^{Q-1}}{(1+2y)^{Q-1} + 2(1-y)^{Q-1}} \,> \, A(y) \label{grosymugl21} \end{aligned}$$ For $y=1$ both sides of (\[grosymugl21\]) are equal to $1.$ For $y<1$ (\[grosymugl21\]) can be rewritten as $$\begin{aligned}
\left(\frac{1+2y}{1-y}\right)^{Q-1}\,> \, \frac{2A(y)+1}{1-A(y)}.
\mbox{ With } y=\frac{4}{10}
\mbox{ this becomes } 3^{Q-1}\, > \, \frac{30}{7}Q \,- \, 2 . \nonumber \end{aligned}$$ The preceding inequality holds for $Q>s \ge 3.5.$ and we have the claim for $y=4/10.$
To show the claim for $4/10<y<1$ we show that the left-hand-side of (\[grosymugl21\]) is concave in $y.$ The derivative of the left-hand-side is $$\begin{aligned}
9(Q-1) \frac{(1+y-2y^2)^{Q-2}}{\left[ (1+2y)^{Q-1} \,+ \, 2(1-y)^{Q-1} \right]^2 } \nonumber \end{aligned}$$ This is a decreasing function in $y\ge 4/10$ because the numerator is decreasing in this case whereas the denominator is increasing and $>0.$\
[*Proof of (b).*]{} Some preparatory calculations: $$\begin{aligned}
A(y+z)\,= \, A(y)\,+ \, \frac{7}{10Q} z \, \,, \,\, \, \, A(y-z)\,= \, A(y)\,- \, \frac{7}{10Q} z \, \, \,\nonumber \\
A(y+z) \cdot (y+z) \,= \, A(y)y +A(y)z +\frac{7}{10Q}zy + \frac{7}{10Q}z^2 \nonumber \\
A(y-z) \cdot (y-z) \,= \, A(y)y -A(y)z - \frac{7}{10Q}zy + \frac{7}{10Q}z^2 \nonumber \\
A(y+z) \cdot (y+z) \,+\, A(y-z) \cdot (y-z) \, \,= \, 2A(y) \cdot y +\frac{14}{10Q}z^2 \nonumber \end{aligned}$$ We denote $$\begin{aligned}
\mbox{OPT}_1(z)\,= \, 1 \,+ \,L(A(y+z), s) \,+ \, L(A(y-z), s) \nonumber \\
\mbox{Then OPT}(z)\,= \, \mbox{OPT}_1(z)\cdot \hfill \nonumber \\
\cdot \left(\frac{1}{1 + 2A(y) \cdot y +\frac{14}{10Q}z^2 }\right)^Q\cdot
\left( (1+y )^Q \,+ \, 2\cdot \left((1-y)^2 \,+ \, 3z^2 \right)^{Q/2}\right). \nonumber \end{aligned}$$
We proceed to show that $\frac{d}{dz} \ln$ OPT$(z)\,< \,0$ for $z>0.$ Some derivatives first. $$\begin{aligned}
\frac{d}{dz} \,A(y+z)\,= \,\frac{7}{10}\frac{1}{Q}, \, \mbox{\quad \quad }
\frac{d}{dz} \,A(y-z)\,= \,- \,\frac{7}{10}\frac{1}{Q}, \, \nonumber \\
\frac{d}{dz} \left(1 + 2A(y) \cdot y+\frac{14}{10Q}z^2 \right)\, \,= \, \frac{28}{10Q}z \,\nonumber \\
\frac{d}{dz} \left( ( 1 + y )^Q \,+ \, 2\cdot ((1-y)^2+3z^2)^{Q/2} \right) \,= \, 6z \cdot Q \cdot ((1-y)^2+3z^2)^{Q/2-1}. \nonumber \\
\frac{d}{dz} \ln \mbox{OPT}(z) \,>=< \, 0 \Longleftrightarrow \mbox{ (Recall comment to (\ref{EQKEY}).)} \nonumber \\
\frac{ \frac{7}{10Q} K(A(y+z), s)\,-\frac{7}{10Q} K(A(y-z), s) }
{\mbox{OPT}_1(z) } \, \, - \, \, \frac{\frac{28}{10Q}z}{1+2yA(y)+\frac{14}{10Q}z^2}
\,+ \, \nonumber \\
\frac{6z \cdot ((1-y)^2+3z^2)^{Q/2-1}}{(1+y )^Q \,+ \, 2\cdot ((1-y)^2+3z^2)^{Q/2} } \, \, >=<\,\,0. \nonumber \end{aligned}$$ Observe that the first and third term of the preceding inequality are $\ge 0$ for $0 \le z \le \min \{y, 1-y \},$ whereas the second term is $\le 0.$
We have that $\frac{d}{dz} \ln \mbox{ OPT }(z)<0$ if the following two inequalities both hold: $$\begin{aligned}
\frac{ \frac{7}{10Q} \left( K(A(y+z), s)\,-\, K(A(y-z), s) \right) }
{ \mbox{OPT}_1(z) } \, \, <\, \,
\frac{\frac{11}{10Q}z}{1+2yA(y)+
\frac{14}{10Q}z^2} \label{groasymugl1} \\
\frac{6z \cdot ((1-y)^2+3z^2)^{Q/2-1}}{(1+y )^Q \,+ \, 2\cdot ((1-y)^2+3z^2)^{Q/2} }
\,< \,
\frac{\frac{17}{10Q}z}{1+2yA(y)+\frac{14}{10Q}z^2} \label{groasymugl2}\end{aligned}$$ Note that for $z=0$ both sides of the preceding inequalities are equal to $0$ and the derivative of $\ln \mbox{OPT}(z)$ is $=0.$ Moreover, we have $1+2yA(y)+\frac{14}{10Q}z^2\,\le \,1+2y $ and the inequalities follow when they are shown with the denominator $1+2y$ in the right-hand-side fraction. To get this, observe that $$\begin{aligned}
2yA(y)+ \frac{14}{10Q}z^2\,= \, 2y+ \frac{14}{10Q}\left(y^2-y+z^2\right) \le 2y , \nonumber \\
\mbox{as } z\le \min\{y, 1-y\} \mbox{ we have } z^2 \le y(1-y) \mbox{ or } y(y-1)+z^2\le 0. \nonumber \end{aligned}$$
[*Proof of (\[groasymugl1\]) for $0< z < \min \{y, 1-y \} ,0 \le y \le 1, s \ge 5 $ .*]{} We enlarge the left-hand-side of (\[groasymugl1\]) first: $$\begin{aligned}
K(A(y+z), s)\,-\, K(A(y-z), s) \,= \,
\,\frac{1}{\exp(s)-1}\left( \exp( A(y+z)\cdot s)\,- \, \exp( A(y-z)\cdot s)\right) \nonumber \\
= \, \, \frac{ \exp(A(y)s) }{ \exp(s)-1 } \left[\exp\left(\frac{7}{10Q}sz \right) \, - \,
\exp\left(-\, \frac{7}{10Q}sz\right) \right] \nonumber \end{aligned}$$
$$\begin{aligned}
\mbox{OPT}_1(z)\,= \, 1\,+ \,L(A(y+z), s) \,+ \, L(A(y-z), s) \, \, \,= \, 1 + \frac{1}{\exp(s)-s-1}\cdot \nonumber \\
\cdot [\exp(A(y+z)s) \,- \, A(y+z)s\,-1 \,+ \,
\exp(A(y-z)s) \,- \, A(y-z)s\,-1 ] \nonumber \\
\ge \, \mbox{( As } A(y+z), A(y-z)\le 1 \mbox{.)} \nonumber \\
1 \,\, + \,\, \frac{1}{\exp(s)-1} [\exp(A(y+z)s) \,+ \, \,
\exp(A(y-z)s) \,- \, 2s\,-2 ] \, \, = \, \frac{1}{\exp(s)-1} \cdot \nonumber \\
\cdot \left[\exp(s)\,-2s\,-3\, + \, \exp(A(y)s)\left( \exp\left(\frac{7}{10Q}sz \right) \,+ \,
\exp\left(-\frac{7}{10Q}sz \right) \right) \right] \nonumber \\
\ge \, \mbox{ (As } A(y)s \le s \mbox{ and } s\ge 2 \mbox{ so that } \exp(s)-2s-3>0\mbox{.)}\nonumber\\
\frac{\exp(A(y)s)}{\exp(s)-1} \left[\frac{\exp(s)\,-2s\,-3}{\exp(s)}\, + \, \exp\left(\frac{7}{10Q}sz\right) \,+ \,
\exp\left(-\frac{7}{10Q}sz \right) \right] \nonumber \\
\ge \, \frac{\exp(A(y)s)}{\exp(s)-1} \left[ 0.9 \, + \, \exp\left(\frac{7}{10Q}sz\right) \,+ \,
\exp\left(-\frac{7}{10Q}sz \right) \right], \nonumber \end{aligned}$$
as $(\exp(s)-2s-3)/\exp(s) \ge 0.9$ for $s \ge 5.$ The denominator of the right-hand-side of (\[groasymugl1\]) is enlarged by $1+2y\le 3.$ We set $$\begin{aligned}
u= \exp\left(\frac{7}{10Q}sz\right)>1 \mbox{ and show (simple algebra from (\ref{groasymugl1})) }
\frac{u-\frac{1}{u}}{0.9+ u+ \frac{1}{u}} \, < \, \frac{11}{3\cdot 7}z \nonumber \\
\mbox{ We have } z= (\ln u)\frac{10}{7}\frac{Q}{s}\, >\, (\ln u)\frac{10}{7} \mbox{ ( by } Q>s.) \nonumber \\
\mbox{ Therefore it is enough to show } \frac{u -\frac{1}{u}}{0.9+ u+ \frac{1}{u}}\,<\, (\ln u)\frac{10}{7}\frac{11}{21} \nonumber \\
\mbox{ Elementary means show that this is true for } u>1. \nonumber \end{aligned}$$
[*Proof of (\[groasymugl2\] ) for $ s\ge 15, 1 \ge y \ge 2/10 \, , \, 0< z \le \min \{y, 1-y \} .$*]{} Inequality (\[groasymugl2\]) follows from $$\begin{aligned}
\frac{6z \cdot ((1-y)^2+3z^2)^{Q/2-1}}{(1+y)^Q}
\,< \,
\frac{\frac{17}{10 Q}z}{1+2y} \nonumber \\
\Longleftrightarrow
60Q(1+2y)((1-y)^2+3z^2)^{Q/2-1} \, < \, 17(1+y)^Q \label{groasymugl21} \end{aligned}$$
For $ y \le 1/2 $ we have $ z \le y $ and (\[groasymugl21\]) follows from $$\begin{aligned}
60Q(1+2y)(1-2y+4y^2)^{Q/2-1} \, < \,17 (1+y)^Q \nonumber \end{aligned}$$ The preceding inequality holds for $ Q \ge s \ge 15 $ and $ 1/2\ge y \ge 2/10$ (proof omitted.)
For $y \ge 1/2$ we have $z \le 1-y$ and (\[groasymugl21\]) follows from $$\begin{aligned}
60Q(1+2y)(4(1-y)^2)^{Q/2-1} \, < \, 17 (1+y)^Q \nonumber \end{aligned}$$ This inequality holds for $Q\ge s \ge 10$ and $y\ge 1/2$ (details omitted.)
Proof of Lemma \[LOCMAX\] and Theorem \[LAPLA\] {#PROLAPLA}
===============================================
We consider $\Psi( \bar{\omega}\, ,\, \bar{\lambda} )\, =\, \Psi( \bar{\omega}\, ,\, \bar{\lambda}\, ,\, \bar{a} \,,\, \bar{c} \, )$ as function of $w_i, \lambda_i,$ $i=1, 2$ in a neighborhood of $(\omega_1, \omega_2)=(\lambda_1, \lambda_2)=(1/3, 1/3).$ The parameters $a_i, c_i$ are given by $ Q(a_i) = \lambda_i k\gamma /\omega_i,$ and $ c_0=1, \, \,\, \, R(c_1, c_2)= (\lambda_1k, \lambda_2k).$ Subsection \[LOLICO\] shows that this is well defined and $a_i, c_i$ is differentiable in $\lambda_i, \omega_i.$ For $\lambda_i=1/3, \omega_i=1/3$ we have $a_i=s, c_i=1 \, (Q(s)=k\gamma$ defining $s.$) We show that the partial derivatives of $\ln \Psi ( \bar{\omega}\, ,\, \bar{\lambda} )$ are $0$ for $\omega_i= \lambda_i=1/3$ and the Hessian matrix is negative definite. This implies Lemma \[LOCMAX\].
For $i=1, 2$ the first derivatives are, with $a_i', c_i' $ denoting the right derivatives of $a_i, c_i$ resp. and recalling that $Q(x)= \frac{xq'(x)}{q(x)}, q(x)=\exp(x)-x-1, R(x_1,x_2)= $\
$\left( \frac{x_1r_{x_1}(1, x_1, x_2)}{r(1, x_1, x_2)}\,, \, \frac{x_2r_{x_2}(1, x_1, x_2)}{r(1, x_1, x_2)} \right) $ $$\begin{aligned}
\frac{d \ln \Psi(\bar\omega, \bar\lambda)}{d \omega_i} &=&
- \ln q(a_0) + \omega_0 \frac{a_0' q'(a_0)}{q(a_0)} + \ln \omega_0 + 1 + \nonumber \\
&& \quad + \ln q(a_i) + \omega_i \frac{a_i' q'(a_i)}{q(a_i)} - \ln \omega_i -1 - \nonumber \\
&& \quad - k \gamma \lambda_0 \frac{a_0'}{a_0} - k \gamma \lambda_i \frac{a_i'}{a_i} \nonumber \\
&=& \ln \omega_0 - \ln \omega_i + \ln q(a_i) - \ln q(a_0) \quad (\textrm{ using } Q(a_i) = k\gamma\lambda_i/\omega_i) \label{eqn:diffWi}. \end{aligned}$$
$$\begin{aligned}
\frac{d \ln \Psi(\bar\omega, \bar\lambda )}{d \lambda_i} &=&
\omega_0 \frac{a_0' q'(a_0)}{q(a_0)} + \omega_i \frac{a_i' q'(a_i)}{q(a_i)} + \nonumber \\
&& \quad k \gamma \Bigg( - \ln \lambda_0 - 1 + \ln a_0 - \lambda_0 \frac{a_0'}{a_0} + \nonumber \\
&& \quad \quad + \ln \lambda_i + 1 - \ln a_i - \lambda_i \frac{a_i'}{a_i} - \nonumber \\
&& \quad \quad - \ln c_i - \lambda_1 \frac{c_1'}{c_1} - \lambda_2 \frac{c_2'}{c_2}
\Bigg) + \nonumber \\
&& \quad \gamma \frac{c_1' r_{c_1}(1, c_1,c_2) + c_2' r_{c_2}(1, c_1, c_2)}{r(1, c_1, c_2)} \nonumber \\
&=& k \gamma ( \ln \lambda_i - \ln \lambda_0 + \ln a_0 - \ln a_i - \ln c_i ) \label{eqn:diffLi} \\
& & \quad ( \mbox{ using } \, \, \, R(c_1,c_2) = (k\lambda_1, k\lambda_2), Q(a_i) = k\gamma\lambda_i/\omega_i) ) \nonumber \end{aligned}$$
For $\bar\lambda = \bar\omega = (1/3, 1/3)$ the terms in (\[eqn:diffWi\]) and (\[eqn:diffLi\]) yield $0$.
The second derivatives (with $i,j \in \{1,2\}, i \not= j)$ are (observe that some of the subsequent terms are equal as the derivative does not depend on the ordering of the variables) $$\begin{aligned}
\frac{d^2 \ln \Psi(\bar\omega, \bar\lambda)}{d \lambda_i,\lambda_i} &=&
k \gamma \left( \frac{1}{\lambda_i} + \frac{1}{\lambda_0} +
\frac{a_0'}{a_0} - \frac{a_i'}{a_i} - \frac{c_i'}{c_i} \right)
\label{eqn:diffLiLi} \\
\frac{d^2}{d \lambda_i,\lambda_j} &=&
k \gamma \left( \frac{1}{\lambda_0} + \frac{a_0'}{a_0} - \frac{c_i'}{c_i} \right)
\label{eqn:diffLiLj} \\
\nonumber \\
\frac{d^2\ln \Psi(\bar\omega, \bar\lambda)}{d \omega_i,\omega_i} &=&
- \frac{1}{\omega_0} - \frac{1}{\omega_i} + \frac{a_i' q'(a_i)}{q(a_i)}
- \frac{a_0' q'(a_0)}{q(a_0)}
\label{eqn:diffWiWi} \\
\frac{d^2\ln \Psi(\bar\omega, \bar\lambda)}{d \omega_i,\omega_j} &=&
- \frac{1}{\omega_0} - \frac{a_0' q'(a_0)}{q(a_0)}
\label{eqn:diffWiWj}\end{aligned}$$
$$\begin{aligned}
\frac{d^2\ln \Psi(\bar\omega, \bar\lambda)}{d \omega_i, \lambda_i} &=&
\frac{a_i' q'(a_i)}{q(a_i)} - \frac{a_0' q'(a_0)}{q(a_0)} \label{eqn:diffWiLi} \\
\frac{d^2}{d \omega_i, \lambda_j} &=&
- \frac{a_0' q'(a_0)}{q(a_0)} \label{eqn:diffWiLj} \\
\nonumber \\
\frac{d^2\ln \Psi(\bar\omega, \bar\lambda)}{d \lambda_i, \omega_i} &=&
k \gamma \left( \frac{a_0'}{a_0} - \frac{a_i'}{a_i} \right) \label{eqn:diffLiWi} \\
\frac{d^2 \ln \Psi(\bar\omega, \bar\lambda)}{d \lambda_i, \omega_j} &=&
k \gamma \frac{a_0'}{a_0} \label{eqn:diffLiWj}\end{aligned}$$
In (\[eqn:diffLiLi\]) - (\[eqn:diffLiWj\]) we need several $a_i'$ and $c_i'.$ We get these from the defining equations $Q(a_i)$ and $R(c_1,c_2).$
#### Derivative of $a_0.$
By $Q(a_i) = k \gamma \lambda_i/\omega_i$ we have $$\frac{a_0 q'(a_0)}{q(a_0)} = \frac{k \gamma \lambda_0}{\omega_0} \quad \Leftrightarrow \quad
\frac{a_0}{k \gamma \lambda_0} = \frac{q(a_0)}{\omega_0 q'(a_0)} \nonumber$$ Taking the derivative of both sides wrt. $\omega_i$ yields $$\begin{aligned}
\frac{a_0'}{k \gamma \lambda_0} =
\frac{a_0' q'(a_0) \omega_0 q'(a_0) -
q(a_0) \left( -q'(a_0) + \omega_0 a_0' q''(a_0) \right)}{\omega_0^2 q'(a_0)^2} \nonumber \\
=\, \frac{a_0'}{\omega_0}\, +\, \frac{q(a_0)}{\omega_0^2 q'(a_0)} \,- \,
\frac{ a_0' q''(a_0)q(a_0)}{\omega_0 q'(a_0)^2 } \nonumber \\
\Longleftrightarrow \frac{a_0' q'(a_0)}{q(a_0)} =
\frac{1}{\omega_0 \left( \frac{\omega_0}{k \gamma \lambda_0} +
\frac{q''(a_0) q(a_0)}{q'(a_0)^2} - 1 \right)} \nonumber\end{aligned}$$ The last step is obtained by collecting all terms with $a_0'$ on the left, multiplying with $q'(a_0)/q(a_0)$ and dividing through the term in brackets. We define $$C(x) := \left( \frac{q(x)}{x q'(x)} + \frac{q''(x) q(x)}{q'(x)^2} - 1 \right). \nonumber$$ Using $Q(a_i) = k \gamma \lambda_i / \omega_i$ the preceding equation becomes $$\begin{aligned}
\frac{{a_0'}_{,\omega_1} q'(a_0)}{q(a_0)} = \frac{{a_0'}_{,\omega_2} q'(a_0)}{q(a_0)} =
\frac{1}{\omega_0 \left( \frac{q(a_0)}{a_0 q'(a_0)} +
\frac{q''(a_0) q(a_0)}{q'(a_0)^2} - 1 \right)} =
\frac{1}{\omega_0 C(a_0)}. \nonumber \end{aligned}$$ We use equation $Q(a_i) = k \gamma \lambda_i / \omega_i$ again to get $$\begin{aligned}
k \gamma \frac{{a_0'}_{\omega_1}}{a_0} = k \gamma \frac{{a_0'}_{\omega_2}}{a_0} =
\frac{1}{\lambda_0 \left( \frac{q(a_0)}{a_0 q'(a_0)} +
\frac{q''(a_0) q(a_0)}{q'(a_0)^2} - 1 \right)} =
\frac{1}{\lambda_0 C(a_0)} \nonumber \end{aligned}$$
#### Derivative of $a_1.$
As for $a_0$ we get $$\begin{aligned}
k \gamma \frac{{a_1'}_{\omega_1}}{a_1} =
- \frac{1}{\lambda_1 \left( \frac{q(a_1)}{a_1 q'(a_1)} +
\frac{q''(a_1) q(a_1)}{q'(a_1)^2} - 1 \right)} =
- \frac{1}{\lambda_1 C(a_1)} \nonumber \\
\frac{{a_1'}_{\omega_1} q'(a_1)}{q(a_1)} =
- \frac{1}{\omega_1 \left( \frac{q(a_1)}{a_1 q'(a_1)} +
\frac{q''(a_1) q(a_1)}{q'(a_1)^2} - 1 \right)} =
- \frac{1}{\omega_1 C(a_1)}. \nonumber \end{aligned}$$ The remaining $a_i-$derivatives can be calculated in a similar way. For $\omega_i = \lambda_i = \frac{1}{3}$ (then $ a_i = s, c_i=1)$ we get $$k\gamma\frac{a_i'}{a_i} \quad \textrm{and} \quad
\frac{a_i' q'(a_i)}{q(a_i)} \quad \textrm{is} \quad \frac{3}{C(s)} \mbox{ for } i=0 \mbox{ and } -\frac{3}{C(s)} \mbox{ for } i=1, 2 \label{MAT1}.$$
#### Derivatives of $c_i$
By $R(c_1,c_2) = (k\lambda_1, k\lambda_2)$ we have $$\frac{c_1 r_{c_1}(1, c_1,c_2)}{r(1, c_1,c_2)} = k \lambda_1 \quad \Longleftrightarrow \quad
\frac{c_1}{k} = \frac{\lambda_1 r(1, c_1,c_2)}{r_{c_1}(1, c_1, c_2)} \nonumber$$ Taking the derivative wrt. $\lambda_1$ leads to (omitting the argument $1$) $$\begin{aligned}
{c_1'}_{\lambda_1} \left( \frac{1}{k} - \lambda_1 +
\lambda_1 \frac{r(c_1,c_2) r_{c_1,c_1}(c_1,c_2)}{r_{c_1}(c_1,c_2)^2} \right) = \nonumber \\
\quad = \lambda_1 {c_2'}_{\lambda_1} \left( \frac{r_{c_2}(c_1,c_2)}{r_{c_1}(c_1,c_2)} -
\frac{r(c_1,c_2) r_{c_1,c_2}(c_1,c_2)}{r_{c_1}(c_1,c_2)^2} \right)
+ \frac{r(c_1,c_2)}{r_{c_1}(c_1, c_2)} \label{eqn:diffC1L1}\end{aligned}$$ Also by $R(c_1,c_2) = (k\lambda_1, k\lambda_2)$ we have $$\frac{c_2 r_{c_2}(c_1,c_2)}{r(c_1,c_2)} = k \lambda_2 \quad \Longleftrightarrow \quad
\frac{c_2}{k \lambda_2} = \frac{r(c_1,c_2)}{r_{c_1}} \nonumber$$ Taking the derivative wrt. $\lambda_1$ again leads to $$\begin{aligned}
{c_2'}_{\lambda_1} \left( \frac{1}{k \lambda_2} - 1 +
\frac{r(c_1,c_2) r_{c_2,c_2}(c_1,c_2)}{r_{c_2}(c_1,c_2)^2} \right) = \nonumber \\
\quad = \lambda_1 {c_1'}_{\lambda_1} \left( \frac{r_{c_1}(c_1,c_2)}{r_{c_2}(c_1,c_2)} -
\frac{r(c_1,c_2) r_{c_2,c_1}(c_1,c_2)}{r_{c_2}(c_1,c_2)^2} \right)
\label{eqn:diffC2L1}\end{aligned}$$ Again we consider the point $\lambda_1 = \lambda_2 = \frac{1}{3}$ then $ c_1 = c_2 = 1$ and equations (\[eqn:diffC1L1\]) and (\[eqn:diffC2L1\]) yield $$2 {c_1'}_{\lambda_1} = {c_2'}_{\lambda_1} + 9 \quad \textrm{and} \quad
2 {c_2'}_{\lambda_1} = {c_1'}_{\lambda_1}. \nonumber$$ Therefore we have $\frac{{c_1'}_{\lambda_1}}{c_1} = 6$ and $\frac{{c_2'}_{\lambda_1}}{c_2} = 3$. Analogously for the derivatives wrt. $\lambda_2$ we get $\frac{{c_1'}_{\lambda_2}}{c_1} = 3$ and $\frac{{c_2'}_{\lambda_2}}{c_2} = 6$.
Putting the derivatives together we get from (\[eqn:diffLiLi\]) - (\[eqn:diffLiWj\]) the following Hessian-Matrix of $\ln \Psi(\bar\omega,\bar\lambda)$ at the point $\omega_i=\lambda_i=1/3$ , abbreviating $D=3/C(s), $ $$H = \left(
\begin{array}{cccc}
-2(\frac{1}{3}+D) & -(\frac{1}{3}+D) & 2D & D \\
-(\frac{1}{3}+D) & -2(\frac{1}{3}+D) & D & 2D \\
2D & D & -2(\frac{8}{3}k\gamma+D) & -(\frac{8}{3}k\gamma+D) \\
D & 2D & -(\frac{8}{3}k\gamma+D) & -2(\frac{8}{3}k\gamma+D)
\end{array}
\right) \nonumber$$ $H $ is negative definite iff $-H$ is positive definite.
A matrix $A = A^T = (a_{ij}) \in \mathbb{R}^{n \times n}$ is positive definite iff the determinants of ist $n$ main-sub-matrices $S_i$ are positive. $$S_1 = a_{11}, \quad
S_2 = \left( \begin{array}{cc} a_{11} & a_{12} \\ a_{21} & a_{22}\end{array} \right),
..., \quad
S_k = \left( \begin{array}{ccc} a_{11} & \hdots & a_{1k} \\
\vdots && \vdots \\
a_{k1} & \hdots & a_{kk}
\end{array} \right),
..., \quad
S_n = A$$ \[lem:jacobi\]
By Lemma \[lem:jacobi\] $-H$ is positive definite, as $D>0$ as $C(x)>0$ for $x>0, $ and $$\begin{aligned}
\det S_1 &=& 2\left( \frac{1}{3} + D \right) > 0 \\
\det S_2 &=& 3\left( \frac{1}{3} + 2D + 3D^2 \right) > 0 \\
\det S_3 &=& \frac{16}{9} k \gamma + \frac{2}{3} D +
\frac{32}{3} k \gamma D + 2D^2 + 16 k \gamma D^2 > 0 \\
\det S_4 = \det (-H)&=& \frac{64}{9} k^2 \gamma^2 +
D^2 + 64 k^2 \gamma^2 D^2 + 16 k \gamma D^2 + \\
&& \quad + \frac{128}{3} k^2 \gamma^2 D + \frac{16}{3} k \gamma D > 0.\end{aligned}$$
[**Theorem \[LAPLA\] (repeated)** ]{} Let $U={\cal U}_\varepsilon(1/3, 1/3).$ There is an $\varepsilon>0$ such that\
$\sum_{\bar{\lambda} , \bar{\omega} \in U} N(\bar{w}, \bar{l})/N_0 \,< \, C \cdot 3^{(1-\gamma)n}.$\
For $ \bar{\lambda} , \bar{\omega} \in U $ and $a_i$ given by $Q(a_i)= \lambda_ik\gamma/\omega_i$ and $c_0=1$ and $R(c_1, c_2)= ( \lambda_1k,\, \lambda_2k)$ we have $ \frac{N(\bar{w}, \bar{l})}{N_0} \, \, \le \,
\,O \left(\frac{1}{n^2}\right) \Psi( \bar{\omega}\, ,\, \bar{\lambda}\, ,\, \bar{a} \, ,\, \bar{c} \, )^n $ by Corollary \[PSIINU\]. Let $\bar{x}=(x_1, \dots, x_4)$ and $h(\bar{x})\,= \, \ln \Psi( \bar{\omega}\, ,\, \bar{\lambda}\, ,\, \bar{a} \, ,\, \bar{c} \, )$ with $\omega_1=x_1, \omega_2=x_2, \lambda_1=x_3, \lambda_2=x_4 $ and $a_i, c_i$ as before for $\bar{\omega}, \bar{\lambda} \in U.$ Let $\overline{1/3}= (1/3, 1/3, 1/3, 1/3)$ then $h(\overline{1/3})=\ln 3^{1-\gamma},\, h_{x_i}(\overline{1/3})\,= \, 0 \, $ and $-\mbox{Hess}(h)(\overline{1/3})$ $, \mbox{Hess}(h)$ the Hessian matrix of $h, $ is positive definite ( proved above, note $\mbox{Hess}(h)(\overline{1/3})=H.)$ We abbreviate $h_{i, j}\,= \, h_{x_i, x_j}( \overline{1/3})$ and by Taylor’s Theorem we have for $\sum_i \,x_i^2 \,\rightarrow \,\,0$ $$\begin{aligned}
h(\overline{1/3}+ \bar{x}) \,= \, h(\overline{1/3}) \,- \, \frac{1}{2} \, \, \sum_i\, \sum _j -h_{i,j} x_i x_j\, + \, o(\sum_i x_i^2) \nonumber \\
\le \,h(\overline{1/3}) \,- \, \frac{1}{2} \, \, \left( \sum_i \,- (h_{i,i} + \delta) x_i^2\, + \, \sum_i\,
\sum _{j \neq i} -h_{i,j} x_i x_j \right)\label{TAYLOR1}\end{aligned}$$ with $\delta $ arbitrarily small for $\sum_i \,x_i^2 $ small enough. We pick $\delta$ such that $-(\mbox{Hess}(h)(\overline{1/3})+ \delta I )$ is still positive definite.
We consider (\[TAYLOR1\]) with $ x_1=w_1/n-1/3,\,\, \, x_2=w_2/n-1/3$ and $x_3=l_1/(k\gamma n)-1/3\, \,\, x_4= l_2/(k \gamma n)-1/3.$ Then $$\begin{aligned}
\sum_{\bar{\omega}, \bar{\lambda} \in U} \, \Psi ( \bar{\omega}\, ,\, \bar{\lambda}\, ,\, \bar{a} \, ,\, \bar{c} \, )^n\,
= \,\sum_{\bar{\omega}, \bar{\lambda} \in U} \exp ( h(x_1, x_2, x_3, x_4) n) \nonumber \\
\, \le \, 3^{(1-\gamma)n} \cdot \sum_{\bar{\omega}, \bar{\lambda} \in U} \exp
\left[\,- \, \frac{1}{2}\left( \, \sum_i \, - (h_{i,i} + \delta) x_i^2\, + \, \sum_i\, \sum _{j \neq i}-h_{i,j} x_i x_j \right)n \right] \label{TAYLOR2}\end{aligned}$$ Note that $\omega_i=w_i/n,\, \lambda_i=l_i/(k\gamma n), \, w_i, l_i $ integer. We distribute the factor $n$ into the $x_i$ multiplying each $x_i$ with $\sqrt{n}:$ $x_1 \sqrt{n}\,= \, w_1/\sqrt{n}-\sqrt{n}/3$ and $x_3\sqrt{n}\,= \, l_1/(k \gamma\sqrt{n})-\sqrt{n}/3.$ As $w_i, l_i$ are integers, the sum in (\[TAYLOR2\]) multiplied with $1/(\sqrt{n}^4(k\gamma)^2)$ is a Riemannian sum of the integral $\int \int \int \int \exp[-(1/2) ( - (h_{i,i} + \delta) x_i^2\, + \, \sum_i\, \sum _{j \neq i}-h_{i,j} x_i x_j)]dx_1 dx_2 dx_3 dx_4$ with bounds $-\infty, \, \, \infty$ for each $x_i.$ Following [@dBRU], page 71, the integral evaluates to $(2\pi)^2/\sqrt{D}$ where $D>0$ is the determinant of $(-h_{i, j})+ \delta I.$ Thus for $n$ large the sum in (\[TAYLOR2\])is $(2\pi)^2/\sqrt{D}(1+o(1)) \sqrt{n}^4(k\gamma)^2\,= \, O(n^2).$ The claim follows.
Remaining proofs {#REPRO}
================
Local limit consideration {#LOLICO}
-------------------------
[**Lemma \[MLOC\] (repeated)**]{} Let $Cn \ge m \ge (2+\varepsilon)n, \, \, C, \varepsilon>0$ constants. Then $$\begin{aligned}
M(m, n) \,= \, \Theta(1)\cdot \left( \frac{ m}{a e}\right)^m\cdot q(a)^n \mbox{ with } a \mbox{ defined by } Q(a)=\frac{m}{n} \nonumber\end{aligned}$$
As $Q(x)$ is increasing the assumptions for $m/n$ imply that $a$ is bounded away from $0$ and $\infty.$ Let $ X \,= \, X(x) $ be a random variable with Prob$[ X=j ]\,= \, (x^j/j!)/ q(x), \,$ for $ j \ge 2, $ and let $X_1, \dots ,X_n $ be independent copies of $X.$ Then $$\begin{aligned}
\sum_{ l_i \ge 2} {m \choose l_1, \dots , l_n} \,= \, \mbox{Prob}[ X_1 + \dots + X_n\,= \, m ] \cdot \frac{q(x)^n}{x^m}\cdot m!. \nonumber \end{aligned}$$ We have E$[X]= xq'(x)/q(x)=Q(x).$ We pick $x=a$ then E$[X]=m/n, $ E$[X_1+ \dots + X_n]\,= \, m.$ The bounds on $a$ imply that $C > \mbox{VAR}[X] > {\varepsilon}>0$ (constants ${\varepsilon}, C $ not the same as above.) Therefore the Local Limit Theorem for lattice type random variables , cf. [@DUR] , Theorem 5. 2, page 112, implies that Prob$[ X_1 + \dots + X_n\,= \, m ]= \Theta\left(\frac{1}{\sqrt{m }}\right) .$ Applying Stirling’s formula in the form $ m! = \Theta (\sqrt{m}) \left(\frac{m}{e}\right)^m $ yields the claim.
We come to Lemma \[KLOC\]. First we show that $R(c_1, c_2)\, = \, (R_1(c_1, c_2), R_2(c_1, c_2))\,=\,(k \lambda_1, k\lambda_2)$ with $R_i(x_1, x_2)= \frac{x_i r_{x_i}(1, x_1, x_2)}{r(1, x_2, x_2)}$ defines $c_i=c_i(\lambda_1, \lambda_2)$ and that $c_i$ is differentiable with respect to $\lambda_i$ for $(\lambda_1, \lambda_2)\in {\cal U}_\varepsilon(1/3, 1/3).$ By the theory of implicit function of several variables we need to show that the Jacobian Determinant of $R(x_1, x_2)$ is $\neq 0$ for $x_1=x_2=1.$ The Jacobian Matrix of $R(x_1, x_2)$ is , omitting the arguments $x_i,$ recalling that $r=r(1, x_1, x_2)$ is our polynomial, $$J = \frac{1}{r^2}\left(
\begin{array}{cc}
(r_{x_1}+x_1r_{x_1, x_1})r-x_1 r_{x_1}^2 \quad & \quad x_1r_{x_1, x_2}r-x_1 r_{x_1}r_{x_2} \\
& \\
x_2r_{x_1, x_2}r-x_2 r_{x_1}r_{x_2} \quad & \quad (r_{x_2}+x_2r_{x_2, x_2})r-x_2 r_{x_2}^2
\end{array}
\right). \nonumber$$
For $x_1=x_2=1$ we get the following values: $r=r(1, 1, 1)= 3^{k-1}, \, r_{x_1}= r_{x_2}=k3^{k-2}, \, r_{x_1, x_1}=r_{x_2, x_2}= r_{x_1, x_2} = k(k-1)3^{k-3}.$ >From this we get that the determinant of $J$ for $x_1=x_2=1$ is $.... \neq 0.$\
[**Lemma \[KLOC\] (repeated)**]{} There is an $\varepsilon>0$ such that for $(\lambda_1, \lambda_2) \in {\cal U}_\varepsilon(1/3, 1/3)$ $$\begin{aligned}
K(\bar{l}) \,= \, O\left( \frac{1}{n}\right) \cdot \frac{r(1, c_1, c_2)}{c_1^{l_1}c_2^{l_2}} \mbox{ with } R(c_1, c_2)=(k\lambda_1, k\lambda_2) \mbox{ defining } c_1, c_2. \nonumber \end{aligned}$$
The previous consideration shows that $(c_1, c_2)$ is close to $(1, 1)$ and well-defined. Let $(X, Y)=(X(x_1, x_2), Y(x_1, x_2))$ be the random vector with $$\mbox{Prob}[(X, Y)\,=\,(k_1, k_2)]\,= \, \frac{ { k \choose k-k_1-k_2, \,k_1, \,k_2} x_1^{k_1}x_2^{k_2}}{r(1, x_1, x_2)} \mbox{ if } k_1=k_2 \mod 3$$ and $0$ otherwise. Then $E(X, Y)\,= \,( R_1(x_1, x_2),\, R_2(x_1, x_2)).$ We consider $m$ independent copies $(X_i\, , Y_i)$ of $(X,\, Y)$ with $(x_1,\,x_2)= (c_1,\, c_2).$ Then $E\left[\sum_i\, (X_i, Y_i)\right]= (k \lambda_1 m, \,k\lambda_2 m)=(l_1, l_2).$ Let $DCo$ be the determinant of the covariance matrix of $(X, Y).$ We show below that for $(c_1,c_2)$ close to $(1, 1)$ we have that $DCo >0$ for constants. The Local Limit Theorem for lattice random vectors [@INDER], Theorem 22.1, Corollary 22.2 with $k=2$ shows that Prob$\left[\sum_i\, (X_i, Y_i) \,= \, (k\lambda_1 m, \,k\lambda_2 m)\right]\,= \, \Theta(1/m).$ This implies the claim.
The covariance matrix of $(X, Y)$ is defined as $$Co \,= \, \left(
\begin{array}{cc}
EX^2-(EX)^2\quad & \quad E[XY]-\, E[X]E[Y] \\
& \\
E[XY]-\, E[X]E[Y] \quad & \quad EY^2-(EY)^2
\end{array}
\right). \nonumber$$ For $(X, Y)\,=\,(X(x_1, x_2), Y(x_1, x_2))$ we get $$\begin{aligned}
EX^2\,= \, \frac{x_1 (x_1 r_{x_1, x_1}(1, x_1, x_2)\,+ \, r_{x_1}(1, x_1, x_2)}{r(1, x_1,x_2))}, \, \\
EY^2\,= \, \frac{x_2 (x_2r_{x_2, x_2}(1, x_1, x_2)\,+ \, r_{x_2}(1, x_1, x_2)}{r(1, x_1,x_2))}, \, \nonumber \\
E[XY]\,= \, \frac{x_1x_2r_{x_1x_2}(1, x_1, x_2)}{r(1, x_1, x_2)}. \end{aligned}$$ This leads to a matrix similar to the Jacobian Matrix above: For $x_1=x_2=1$ its determinant is positive.
The sharp threshold
-------------------
To prove the sharp threshold we apply a general theorem. Let $A \subseteq \{0,\, 1\}^N$ and let $a_m$ be the number of elements of $A$ with exactly $m$ $1'$s. We let $\mu_p(A) \,= \, \sum_{m=0}^N \, a_m \cdot p^m \cdot (1-p)^{N-m}$ be the probability of $A,$ note $a_m \le {N \choose m}.$ If $A$ is a non-trivial, monotone set we have that $\mu_p(A)$ is a strictly increasing, continuous, differentiable function in $0 \le p \le 1.$ In this case for $0 \le \tau \le 1$ we have that $p_\tau$ is well defined by $ \mu_{p_\tau}(A)=\tau.$ Not let $A=(A_n)_{n\ge 1}$ and let be $A_n$ be monotone. We say that $A$ has a coarse threshold iff there exist constants $0 < \rho < \tau < 1$ such that $(p_{\tau}- p_{\rho})/p_{\rho} \, \ge \varepsilon$ for a constant $\varepsilon$ (and infinitely many $n.$) We can assume that $p_{\tau}=O(p_\rho)$ otherwise the threshold is clearly coarse. Moreover, we assume that $p_{1-o(1)}\,= \,o(1).$
There exist functions $\delta=\delta(C, \tau)>0$ and $K=K(C, \tau)$ such that the following holds: Let $A=A_n$ with $A \subseteq \{0,\, 1\}^N$ be a monotone set with $\tau \le \mu_p(A) \le 1-\tau$ for constant $1/2> \tau > 0$ and assume that $p \cdot \frac{d\mu_p(A)}{dp} < C.$ Then at least one of the following two possibilities holds:\
1.$$\begin{aligned}
\mbox{Prob}_p[a \in A\, \, ; \, \, \exists b \in A \, \, , |b| \le K \, \, , b \subseteq a] \,> \, \delta\end{aligned}$$ 2. There exists $b\, \in \{0, 1\}^N
, \, b \notin \, A\, , |b| \le K$ such that the conditional probability $$\begin{aligned}
\mbox{Prob}_p[a \in A \,| \, b \subseteq a ]\,> \, \mbox{Prob}_p[A]\,+ \, \delta .\end{aligned}$$
\[Thresh\] $A=(A_n)$ has a sharp threshold if $ p_{1-o(1)}=O(p_\tau)$ for all $\tau>0,$ and for each $1/2>\tau >0, \, \delta >0, \varepsilon >0 , K, $ $p_\tau < p < p_{1-\tau}$ and all sufficiently large $n$ the following two statements hold:\
1. $$\begin{aligned}
\mbox{Prob}_p[a \in A\, \, ; \, \, \exists b \in A \, \, , |b| \le K \, \, , b \subseteq a]\, \,<\, \delta. \end{aligned}$$ 2. If $b\, \in \{0, 1\}^N
, \, b \notin \, A \, , |b| \le K$ with the conditional probability $\mbox{Prob}_p[a \in A \,| \, b \subseteq a ]\,> \, \mbox{Prob}_p[A]\,+ \, \delta $ then $\mbox{Prob}_{p(1+\varepsilon)}[A] \,> \, 1-\tau$
Assume, that $A$ has a coarse threshold. Let $1>\alpha> \beta>0$ be such that $(p_{\alpha}- p_{\beta})/p_{\beta} \, \ge \varepsilon.$ We abbreviate $q=(p_\alpha+ p_\beta)/2.$ By strict monotonicity of $\mu_p(A)$ we have $\mu_q(A)\,= \, \gamma $ for a $\alpha> \gamma > \beta.$ We have that $\frac{\gamma-\beta}{q-p_\beta} \,= \, \frac{d\mu_p(A)}{dp}|p=p^*$ for a $p _\beta< p^*<q$ (by the Mean Value Theorem.) We have that $(q-p_\beta)/p^* \, \ge \varepsilon'$ as $p^*=O(p_\beta).$ Therefore $\frac{\gamma -\beta}{q-p_\beta} \cdot p^*\,= \, \left(\frac{d\mu_p(A)}{dp}|p=p^*\right) \cdot p^* \le C$ for a constant $C.$ The preceding theorem applies to $p^*.$ Our assumption implies that the first item of the theorem does not hold.
Therefore the second item of the preceding theorem must hold for $p=p^*.$ We have that $p^*+\frac{p_\alpha-p_\beta}{ 2}\, < p_\alpha.$ Therefore $p^*\left( 1+ \frac{p_\alpha-p_\beta}{p^* \cdot 2}\right) < p_\alpha.$ Moreover $\frac{p_\alpha-p_\beta}{p^* \cdot 2}\,> \, \varepsilon''$ as $p^* =O(p_\beta).$ Our second assumption shows that the preceding statement cannot hold. Therefore the second item of the preceding theorem does not hold, too. Therefore $A$ cannot have a coarse threshold.
Let $F(n, p)$ be the random formula of equations $y_1+ \dots + y_k= a \mod 3,$ $0 \le a \le 2$ over $n$ variables where each equation is picked with probability $c/n^{k-1}$ independently.
\[THRESH\] Unsatisfiability of $ F(n, p) $ has a sharp threshold.
We apply Corollary \[Thresh\]. Let $p=c/n^{k-1}.$ Observe that $F(n, p)$ is unsatisfiable whp. for $c>1$ by expectation calculation. Concerning the first item of the corollary we show that $F(n, p)$ does not contain a subformula over a bounded number of variables such that each variable occurs at least twice. The expected number of such subformulas over $1\le l\le B,$ $B$ constant variables is bounded above by $ {n \choose l }\cdot\left(c/n^{k-1}\right)^{2l/k}\le O(1) \cdot n^{(2/k-1) l}.$ As $k\ge 3$ and $l\ge 1$ the geometric series shows that the expectation of the number of such subformulas with $\le B$ variables is $o(1).$ As each unsatisfiable formula contains a subformula where each variable occurs at least twice we have no unsatisfiable subformula of bounded size whp. The first item of the corollary holds.
Concerning the second item, let $B$ be a fixed satisfiable formula and let $p < 1/n^{k-1}.$ We assume that $\mbox{Prob}[\mbox{UNSAT} (B \cup F(n, p)) ] > \mbox{Prob}[\mbox{UNSAT}(F(n, p)) ]+ \delta. $ UNSAT$(F)$ is the event that $F$ is unsatisfiable. With high probability $F(n, p)$ contains only equations with $1$ or none variables from $B$ (as $p< 1/n^{k-1}$ and the number of variables of $B$ is constant. )
Consider a fixed satisfiable formula $F$ over the variables not in $B$ We pick each equation with exactly one variable in $B$ with probability $p=c/n^{k-1}$ independently. We assume that the resulting random formula is unsatisfiable with probability $\delta>0.$ We show that this implies that the random instance obtained from $F$ by adding [*each* ]{} equation with probability $ \varepsilon / n^{k-1},$ independently, $\varepsilon>0$ constant. is unsatisfiable with high probability. This directly implies that the second item of Corollary \[Thresh\] holds.
Consider a fixed variable $x$ of $F.$ We throw in the equations containing $x$ with $ \varepsilon / n^{k-1},$ We show below that the resulting random formula is unsatisfiable with probability $\delta'>0,$ $\delta'$ constant. Throwing [*each*]{} equation with probability $\varepsilon/n^{k-1},$ the expected number of variables $x$ such that the equations containing $x$ lead to unsatisfiability of $F$ is $\delta' n.$ For $x \neq x'$ the equations with $x$ or $ x'$ are nearly independent. Tschebycheff’s inequality shows that we even have a linear number of variables $x$ whose equations yield unsatisfiability whp.
We show the statement above concerning the fixed variable $x.$ When throwing in the equations with one variable in $B$ with $p=c/n^{k-1}$ we get with probability $\delta$ a set $U$ such that $F\cup B \cup U$ is unsatisfiable. With probability slightly lower, but still constant $>0$ we can assume that $U$ is of bounded size. Now consider a satisfying assignment $a$ of $B.$ We replace the variable from $B$ in each equation by its value under $a$ and get a set of equations with $k-1$ variables each. When we add these equations to $F$ the resulting formula is unsatisfiable.
Now consider our variable $x$ from $F$ and throw in each equation containing $x$ with probability $\varepsilon/n^{k-1}.$ With constant probability $>0$ we get the a set $U'$ obtained from a set $U$ as above by replacing the variable from $B$ by $x.$ With the same probability we get $U_0$ instead of $U'$ where $U_0$ is obtained as follows: Let $E$ be an equation of $U$ such that the variable from $B$ has the value $j$ in the satisfying assignment $a$ from $B.$ The variable from $B$ is replaced with $x$ in $E$ and we subtract $j$ from the right hand side. The resulting formula is unsatisfiable for all assignments which have $x=0.$ $U_1$ is defined by adding $1-j$ to the right hand-side. The resulting formula is unsatisfiable for $x=1.$ $U_2$ is defined by adding $2-j$ and the resulting formula is unsatisfiable for $x=2.$ With constant probability $>0$ we get one such set $U_j.$
To get unsatisfiability for all $3$ values of $x$ we observe that with probability roughly $\delta^3$ we get three sets $U, V, W$ with one variable in $B$ which are disjoint and each of them causes unsatisfiability. This implies that with constant probability $>0$ we get three sets $U_0, V_1, W_2$ of equations with $x.$ The resulting formula is unsatisfiable for any value of $x.$
[**II. Uniquely extendible constraints** ]{}
Outline
=======
A uniquely extendible constraint $C$ on a given domain $D$ is a function from $D^k$ to true, false with the following restriction: For any argument list with a gap at an arbitrary position, like $(d_1, \dots d_{i-1}, -, d_{i+1}, \dots , d_k)$ there is a unique $d \in D$ such that\
$C(d_1,\dots d , \dots , d_k)$ evaluates to true. Note that $C(d_1 ,\dots , d, \dots, d_k) = $ true implies that $C(d_1, \dots ,d', \dots ,d_k)=$ false for $d \neq d'.$ The random constraint is a uniform random member from the set of all uniquely extendible constraints over $D.$ Let $\Gamma$ be the set of all such constraints. Typical examples of such constraints are linear equations with $k$ variables, modulo $|D|.$ A threshold result analogous to Lemma \[THRESH\] can be proved by similar arguments based on symmetry properties of uniquely extendible constraints.
Given a set of $n$ variables a clause is an ordered $k$-tuple of variables equipped with a uniquely extendible constraint. The number of all formulas with $m$ clauses is $M(km, n) \cdot |\Gamma|^m, $ we denote $N_0=M(km, n)$ (notation cf. (\[DEFN0\]).) A random formula is a uniform random element of the set of all formulas. The random variable $X$ gives the number of solutions of a formula and E$[X] = (1/d)^{(1-\gamma)n}, m = \gamma n.$ This follows from symmetry considerations. For two assignments $a, b$ we study E$[X_aX_b]$ where $X_a$ is $=1$ iff the formula is true under $a.$ It turns out that E$[X_a X_b]$ depends only on the number of variables which have different values under $a, b.$ Let DIFF$(a, b)=$ the set of variables with different values under $a$ and $b.$
Given a $k$-tuple $a$ of values from $D$ and another $k$-tuple $b$ differing from $a$ in exactly $i,\,\ 0 \le \, i\, \le k,$ slots, we let $p_i$ be the probability that the random constraint is true under $b$ conditional on the event that it is true under $a.$ The following very simple generating polynomial for the ${k \choose i} \cdot p_i$ is the observation making our proof possible.
\(a) (From [@COMO]) $p_0=1, \, \, p_{i+1}\,= \, \frac{1}{d-1} \left(1 - p_i \right).$\
(b) $$\begin{aligned}
\mbox{ Let } p(z) \,\,= \,\,\frac{1}{d}\left( (1+z)^k \,+ \, (d-1) \left( 1-\frac{z}{d-1} \right)^k \right) \,
\mbox{ then }
p(z) \,\,= \,\, \sum_i {k \choose i} p_i \cdot z^i \nonumber \end{aligned}$$
\(b) We need to show that $p_i\,= \,\frac{1}{d}\left( 1+ (-1)^i \left( \frac{1}{d-1}\right)^{i-1}\right).$ This holds for $i=0, i=1.$ For $\,i >1 \,$ we get by induction: $$\begin{aligned}
p_i\,= \frac{1}{d-1}(1-p_{i-1}) \, \,\,= \,\, \,\frac{1}{d-1}\left( 1\,- \, \frac{1}{d}\left( 1+(-1)^{i-1}\left(\frac{1}{d-1}\right)^{i-2}\right)\right) \,=\, \nonumber \\
\,= \, \frac{1}{d-1} \, \,\, - \,\, \, \frac{1}{d(d-1)}- \frac{1}{d}(-1)^{i-1}\left(\frac{1}{d-1}\right)^{i-1}\,\,= \, \,
\frac{1}{d}\left( 1+ (-1)^i\left(\frac{1}{d-1}\right)^{i-1}\right). \nonumber \end{aligned}$$
$$\begin{aligned}
\mbox{ We let } C_j = \frac{|\Gamma|}{d} \cdot {k \choose j} \cdot p_j \mbox{ for } 0 \le j \le k \, \, , \, \, \,
K(l)\,= \, \sum_{j_1+ \dots + j_m=l}\,C_{j_1}\cdots C_{j_m} . \nonumber \\
\mbox{ Then } \hat{N}(w, l) \,= \, M(l, w)M(km-l, n-w) K(l) \nonumber \end{aligned}$$
is the number of formulas $F$ true under two assignments $a, b$ with $|\mbox{DIFF}(a, b)|=w$ and the variables with different values occupy exactly $l$ slots of $F.$ The factors ${k \choose j}$ of $C_{j}$ count how to distribute the $l$ slots. The factor $ M(l, w)M(km-l, n-w)$ counts how to place the variables into these slots. The factors $\frac{|\Gamma|}{d} \cdot p_{j}$ count the number of constraints such that the formula becomes true under $a, b.$ Given an assignment $a$ the number of assignment formula pairs $(b, F)$ with $|\mbox{DIFF}(a, b)|=w,$ $F$ is true under $a, b,$ and the variables from $\mbox{DIFF}(a, b)$ occupy exactly $l$ slots is $$\begin{aligned}
N(w, l) = {n \choose w}(d-1)^w\cdot \hat{N}(w, l).\, \, \mbox{ And E}[X^2]= d^n\sum_{w, l}\, N(w, l)\cdot \frac{1}{N_0\cdot |\Gamma|^m}
\nonumber \end{aligned}$$ The next theorem is analogous to Theorem \[EX2EI\].
\[UNMAIN\] $\sum_{w, l} \, N(w, l)/(N_0 |\Gamma|^m) \, \, \le C d^{(1-2\gamma)n}\, \, , \, \, $ $k\ge 8, \, \, m=(1-\gamma)n.$
We let $\lambda=l/km$ and $\omega=w/n$ with $w, l$ always having the meaning above. The proof of Theorem \[UNMAIN\] follows the pattern of Theorem \[EX2EI\]. We omit all steps referring to the summation, they are quite analogous. The details to bound the summands are however different. We have $$\begin{aligned}
K(l) = \mbox{Coeff} [z^l, \, p(z)^m ]
\cdot \left( \frac{|\Gamma|}{d}\right)^m \le \left( \frac{p(c)|\Gamma|}{d}\right)^m\cdot \frac{1}{c^l} \mbox{ for } c>0. \nonumber \end{aligned}$$
We define $\Psi(\omega, \, \lambda, \, x, \, y, \,z\,):= \,$ $$\begin{aligned}
\left(\frac{(d-1)q(x)}{q(s)\omega}\right)^\omega
\left( \frac{q(y)}{q(s)( 1-\omega )} \right)^{1- \omega} \cdot \left( \frac{\lambda s}{xz} \right)^{\lambda k \gamma}
\left( \frac{ (1- \lambda ) s }{y}\right)^{ ( 1-\lambda ) k \gamma}\cdot \left (\frac{p(z)}{d} \right)^\gamma. \nonumber \end{aligned}$$ We have $\Psi(1-1/d,\, 1-1/d, \, s, \, s,\, d-1 )\,= \, d^{1-2\gamma}\quad, $ $s$ is given by $Q(s)=k\gamma,$ cf. discussion around Lemma \[MLOC\]. As Lemma \[LEMBA\] we have the next Lemma; the subsequent Theorem is as Theorem \[OPT\].
$ N(w, l)/(N_0 |\Gamma|^m) \, \le \, \Psi(\omega, \lambda, a, b, c)\cdot O(n) \mbox{ for } a, b, c >0.$
Observe that for $Q(s)=k\gamma \ge 8$ we have $s\ge 7.$
\[UNOPT\] Let $ d=4 $ and $ s \ge 7.$ For any $\lambda>0$ there exist $a, b, c>0$ such that:\
(1) $\Psi(\omega, \lambda, a, b, c) \, \le \, d^{1-2\gamma}.$\
(2) For any $\varepsilon >0$, $\lambda $ not $\varepsilon-$close to $1-1/d,$ $\Psi(\omega, \lambda, a, b, c)\, \le \, d^{1-2\gamma}- \delta.$
Two reals $a, b$ are $\varepsilon-$close iff $|a-b |< \varepsilon.$ To treat $\lambda$ close to $(d-1)/d$ we consider the function $P(z)= zp'(z)/p(z)$ (cf. discussion after Corollary \[SUMOPT\].) We have $P(d-1)= k(1-1/d)$ and the derivative $P'(d-1)>0.$ Thus we can define $c=c(\lambda)$ for $\lambda$ $\varepsilon-$close to $1-1/d$ by $P(c)=k \lambda.$ And $c(\lambda)$ is differentiable. As Lemma \[KLOC\], Corollary \[PSIINU\], and Lemma \[LOCMAX\] we get the next $3$ items. To prove Lemma \[UNMAX\] the Hessian matrix of $\Psi(\omega, \lambda, a, b ,c)$ is considered (calculation analogously to [@MITZ].)
There is an $\varepsilon>0$ such that for $\omega, \lambda\, \, $ $\, \, \varepsilon-$close to $1-1/d$ we have $K(l) = O(1/\sqrt{n}) \cdot \left( p(c)|\Gamma|/d\right)^m\cdot 1/c^l$ with $P(c)=k\lambda.$
There is an $\varepsilon \, > \,0$ such that for $\omega, \lambda $ being $\varepsilon-$close to $1-1/d$ $N(w, l)/(N_0 |\Gamma|^m) \le O(1/n) \cdot \Psi(\omega, \, \lambda,\, a,\, b, \, c)$ with $Q(a)=l/w, Q(b)=(km-l)/(n-w), P(c)= \lambda k.$
\[UNMAX\] The function $\Psi(\omega, \lambda, a, b ,c)$ with $a, b,c$ given by $Q(a)=l/w, Q(b)=(km-l)/(n-w), P(c)= \lambda k$ has a local maximum with value $d^{1-2\gamma} $ for $\lambda= \omega=1-1/d.$ In this case we have $a=b=s$ and $c=d-1.$
$$\begin{aligned}
\mbox{ We define }\, \, \mbox{OPT}_1(x, y, s)\,= \, (d-1) \cdot \frac{ q(sx)}{q(s)} \,+ \, \frac{q(sy)}{q(s)} , \quad \quad \nonumber \\
\mbox{OPT}_2(x, y, z, s)\,\,= \,\,
\, \left( \frac{1}{y+xz} \right)^{Q} , \, y+xz>0\nonumber \\
\mbox{OPT}_3( z, s)\,= \, (1+z)^Q \,+ \, (d-1)\cdot \left| 1\,- \frac{z}{d-1}\right|^Q \, \, , \, Q=Q(s) \nonumber \\
\mbox{OPT}(x, y, z, s) \,=\, \mbox{OPT}_1(x,y, s) \cdot \mbox{OPT}_2(x, y, z, s) \cdot
\mbox{OPT}_3(z, s) . \nonumber \end{aligned}$$
As Lemma \[LEMOPT\] we have the next Lemma. We prove Theorem \[UNOPT\] based on this lemma. We cannot proceed analogously to the proof of Theorem \[OPT\] because the polynomial $p(z)$ is not as symmetric as $r(x_0, x_1, x_2).$ The two cases $\lambda$ small (in Section \[SMALL\]) and $\lambda$ large (in Section \[LARGE\]) are treated separately.
\[lagrkl\] $$\begin{aligned}
\mbox{ Let } \, a, b, c\,>0 \, \, \,\mbox{be such that } \frac{\lambda}{1-\lambda}\,= \,\frac{a c} {b} . \, \,
\mbox{ Then } \, \,
\Psi(\omega, \lambda, as, bs, c ) \, \, \le \, \, \frac{1}{d^{2\gamma}} \mbox{OPT}(a, b, c, s). \,\, \nonumber \end{aligned}$$
Proof of Theorem \[UNOPT\] for $d=4,\, s\ge 7 , \, \lambda \le 1-1/d$ {#SMALL}
======================================================================
We restrict attention to $d=4$ fix $b=1$ and consider $c, a $ with $0 \le c \le 3$ and $ 0 \le a \le 1.$ With these values OPT$(a, b, c, s)$ leads to the following notation used in this Section. $$\begin{aligned}
\mbox{OPT}_1(a, s)\,= \, 3 \cdot \frac{q(sa)}{q(s)}\,+ \,1 \, \,, \, \, \mbox{OPT}_2(a, c, s)\,= \, \left( \frac{1}{1+ac} \right)^{Q} \nonumber \\
\mbox{OPT}_3( c, s)\,= \, (1+c)^Q \,+ \, 3\cdot \left( 1\,- \frac{c}{3}\right)^Q \, \,\nonumber \\
\mbox{OPT}(a, c, s) \,=\, \mbox{OPT}_1(a, s) \cdot \mbox{OPT}_2(a, c, s) \cdot
\mbox{OPT}_3( c, s) . \nonumber\end{aligned}$$ The values of OPT$(a, c, s)$ at the corners of the rectangle for $0\le c\le 3,\, \, 0 \le a \le 1 $ are: $$\begin{aligned}
\mbox{OPT}(0, 0, s)\,= \, 4\, \, \,\,\,,\, \mbox{OPT}(0, \, 3 , \, s)\,=\, 4^Q \nonumber \\
\mbox{OPT}(1, 0, s)\,= \, 4^2 \, \, \,\,,\, \, \mbox{OPT}(1, \, 3, \, s)\,=\, 4 \label{ecken}\end{aligned}$$
{width="49.00000%"} {width="49.00000%"}
We prove four lemmas. Observe that $A(c, s)$ in Lemma \[flagekl\] is a flat linear function in $c\ge 0$ from $A(0, s)=1-\frac{7}{10Q}$ to $A(3, s)=1.$
\[flagekl\] $$\begin{aligned}
\mbox{ Let } s \ge 7 \mbox{ and let }\,A(c)\,= \, A(c, s) \,\,=\,\, \frac{7}{Q\cdot 10\cdot 3}\cdot c \,\,+1 \, \, - \frac{7}{10 \cdot Q}. \nonumber \\
\mbox{Then } \,
\mbox{OPT}(A(c), \, c, \, s\,) \,\mbox{ is strictly increasing in } \, \, 1 \le c <3. \nonumber \end{aligned}$$
$A(c, s)$ in the subsequent Lemma is a steep linear function starting at $A(0, s)=0.$
\[stgekl\] $$\begin{aligned}
\mbox{ Let } s\ge 6 \mbox{ and let }\, A (c)\,\,=\,\, A(c, s) \,\,=\,\, \frac{Q}{2}\cdot c\, \,. \nonumber \\
\mbox{ Then }\, \mbox{OPT}(A(c), \, c, \, s) \, \mbox{ is strictly decreasing for } 0\, <c\, \le \, \frac{1}{Q} \nonumber \end{aligned}$$
\[einmi\] (a) For each constant $0 \le a \le 1$ OPT$(a, c, s)$ as a function in $c$ with $ 0 \le c \le 3$ has a unique local minimum.\
(b) For each constant $0 \le c \le 3$ OPT$(a, c, s)$ as a function in $a$ with $ 0 \le a \le 1$ has a unique local minimum.
\[pukl\] Let $ s\, \ge 6\,$ then OPT$(a, \, \, c, \, \, s\,) \,< \, 4-\delta $ for $ (a, \, c)\,= \, $ $$\begin{aligned}
=\,\left(\frac{1}{2}, \, \, \frac{1}{Q} \, \right)\, ,\, \left(\frac{1}{2}, \, \frac{2}{Q}\,\right),\,
\,\left(\frac{2}{3}, \, \frac{2}{Q} \,\right) , \, \left(\frac{2}{3}, \, \frac{3}{Q} \,\right), \,
\left( 1\, \, - \, \, \frac{7}{15Q}\, , \, \frac{3}{Q} \right), \, \,
\, \left(1\, - \, \frac{7}{15Q}\, \, , \, 1 \,\right) \nonumber \end{aligned}$$
[*Proof of Theorem \[UNOPT\] for $\lambda\le 1-1/d.$ (cf. proof of Theorem \[OPT\] after Lemma \[lem4\].)*]{} We have $ \lambda \le 1-1/d \Longleftrightarrow \lambda /(1-\lambda)\le d-1.$ Using Lemma \[lagrkl\] we need to show that for each $P\le d-1$ we have a decomposition $P=ac$ such that OPT$(a, c, s)\le 4$ of $4- \delta.$ Lemma \[flagekl\] treats $ 1-7/(15Q) \le P \le d-1.$ Lemma \[pukl\] together with Lemma \[einmi\] treat $ 1-7/(15Q)\ge P \ge 1/(2Q).$ Finally Lemma \[stgekl\] treats $1/(2Q) \ge P > 0.$ Observe that OPT$(0, 0, s)=4$ and we need to look into the proof of Lemma \[lagrkl\] to get the $-\delta$ required for small $P.$
Proof of Lemma \[flagekl\]
--------------------------
[**Lemma \[flagekl\] (repeated)**]{} [*$$\begin{aligned}
\mbox{ Let } s \ge 7 \mbox{ and let }\,A(c)\,= \, A(c, s) \,\,=\,\, \frac{7}{Q\cdot 10\cdot 3}\cdot c \,\,+1 \, \, - \frac{7}{10 \cdot Q}. \nonumber \\
\mbox{Then } \,
\mbox{OPT}(A(c), \, c, \, s\,) \,\mbox{ is strictly increasing in } \, \, 1 \le c <3. \nonumber \end{aligned}$$* ]{}
$$\begin{aligned}
\mbox{ Some notation:} \, \mbox{PPLUS3}(x, y)\,= \, (1\,+ \, x)^y \,+ \, 3\left(1-\frac{x}{3}\right)^y, \,\, x \le 3 \nonumber \\
\mbox{PMINUS}(x, y) \,= \,(1\,+ \, x)^y \,-\, \left(1-\frac{x}{3}\right)^y, \,\, x \le 3 \label{notPPPM} \\
\frac{d}{dc}\ln \mbox{PPLUS3}(c, Q)\,= \, Q \cdot \frac{\mbox{PMINUS}(c, Q-1)}{\mbox{PPLUS3}(c, Q)}, \, \nonumber \\
A\, \,=\,\, A(c), \, \, A' \, =\, \frac{d}{dc}\,A\,= \frac{7}{10\cdot 3 \cdot Q} \, \,\, , \, \, \,
\frac{d}{dc}\ln \mbox{OPT}(A, c, s)\, = \, \nonumber \\
\,=\,
\frac{3\frac{\exp(As)-1)}{\exp(s)-s-1} \cdot s A'}{\mbox{OPT}_1(A, s)}\,- \,
Q\cdot \frac{A'c\,+ \,A}{1+Ac}\,+ \, Q \cdot \frac{\mbox{PMINUS}(c, Q-1)}{\mbox{PPLUS3}(c, Q)} \, >=<0 \nonumber \\
\Longleftrightarrow \, \frac{3\frac{\exp(As)-1}{\exp(s)-1}\cdot A'}{\mbox{OPT}_1(A, s)}\, - \, \frac{A+A'c}{1+Ac}\,+ \,
\frac{\mbox{PMINUS}(c, Q-1)}{\mbox{PPLUS3}(c, Q)} \, >=<\,0 \,\,\,\,\,\;\;\, \, \, \label{flagekl1} \\
\mbox{(Division with } Q\,= \, \frac{s(\exp(s)-1)}{\exp(s)-s-1} \mbox{.)} \nonumber \\
\mbox{ For } c\,= \,3 \mbox{ the derivative is } =0, \,\, A(3, s)=1\, \, \mbox{( OPT}(1, 3, s) =4.) \nonumber \end{aligned}$$
We split the right-hand-side of (\[flagekl1\]) into two additive terms. Inequalities (\[flagekl2\]) and (\[flagekl3\]) imply that the $\frac{d}{dc} $ OPT$(A, c, s)>0.$ For $c=3$ both left-hand-sides are $=0.$ $$\begin{aligned}
\frac{3\frac{\exp(As)-1)}{\exp(s)-1} \cdot A'}{\mbox{OPT}_1(A, s)}\,- \,
\frac{A'c}{1+Ac}\, > \, 0 \label{flagekl2} \\
- \, \frac{A}{1+Ac}\,+ \, \frac{\mbox{PMINUS}(c, Q-1)}{\mbox{PPLUS3}(c, Q)} \,>\,0 \label{flagekl3}\end{aligned}$$
[*Proof of (\[flagekl2\]) for $0 \le c <3 \, $ and $ s \ge 7 .$* ]{} $$\begin{aligned}
K\,:= \, \frac{\exp(sA)-1}{\exp(s)-1}, \,\, \, L:= \frac{\exp(sA)-sA -1}{\exp(s)-s-1} \nonumber \\
\mbox{ We need to show } \frac{3KA'}{3L+1}\, \, > \, \frac{A'c}{1+Ac} \, \,
\Longleftrightarrow \, \, \, \frac{3K}{3L+1}\, \, > \, \frac{c}{1+Ac} \nonumber \\
\Longleftrightarrow \, \, 3\left( K+ KAc -Lc \right) \,> \,c \, \, \, \,
\mbox{ (For } c= 3 \mbox{ both sides are } \,=3 .) \label{flagekl21} \end{aligned}$$ By (\[BAKL\]) we have $ L \, \le \, K$ and (\[flagekl21\]) is implied by $$\begin{aligned}
3K\left( 1+ Ac -c \right) \,> \,c \, \, \quad \mbox{ (For } c=3 \mbox{ both sides are } =3.\mbox{)} \label{flagekl22}\end{aligned}$$ $K \ge 0 $ is increasing and convex, $1+ Ac -c $ is $>0,$ and increasing for $c>3/2,$ and convex. Therefore the left-hand-side of (\[flagekl22\]) is convex for $c\ge 3/2.$ Therefore, for $ c\ge 3/2,$ it follows from $$\begin{aligned}
\left(\frac{d}{dc} 3K\left( 1+ Ac -c \right) \right)_{|c=3}\, < \,\left( \frac{d}{dc}c \right)_{|c=3}\,= \,1. \label{flagekl23}\\
\frac{d}{dc} 3K\left( 1+ Ac -c \right)\,\, \, \, \, \,
\,= \, \frac{3\exp(sA)\cdot s \cdot \frac{7}{30Q}}{\exp(s)-1}\cdot(1+Ac-c) \,+ \, \nonumber \\
\frac{3 ( \exp (sA)-1)}{\exp(s)-1}\cdot \left( \frac{7}{30Q}c + \frac{7}{30Q}c +1- \frac{7}{10Q} -1\right). \nonumber \\
\mbox{ Therefore } \, \frac{d}{dc} 3K\left( 1+ Ac -c \right)_{|c=3}\,= \, \nonumber \\
\frac{7\exp(s)(\exp(s)-s-1)}{10(\exp(s)-1)^2}\,\,+ \, \, \frac{21(\exp(s)-s-1)}{10s(\exp(s)-1)}\, \, \, \, \label{flagekl24}\end{aligned}$$ For $s=7$ we get that (\[flagekl24\]) is $<0.995.$ Moreover it is decreasing in $s$ (proof omitted) and (\[flagekl23\]) holds for all $s \ge 7$ and $c \ge 3/2.$ For $c\le 3/2 $ we argue as in the proof of Lemma \[lem4\](a) cf. the argument following (\[grosymugl12\]).\
[*Proof of (\[flagekl3\]) for $1\, \le c \,< \,3 \,$ and $ s \ge 5 $ .*]{} $$\begin{aligned}
\mbox{ We need to show } \frac{A}{1+Ac}\,\,< \, \frac{\mbox{PMINUS}(c, Q-1)}{\mbox{PPLUS3}(c, Q)} \, \nonumber \\
\Longleftrightarrow A \, \cdot \, \mbox{PPLUS3}(c, Q)\,< \, (1+Ac) \mbox{PMINUS}(c, Q-1) \nonumber \\
\Longleftrightarrow A\cdot\left(\mbox{PPLUS3}(c, Q) \,- \, c \cdot \mbox{PMINUS}(c, Q-1)\right)
\,= \, A\cdot \mbox{PPLUS3}(c, Q-1)\,\nonumber \\
< \, \mbox{PMINUS}(c, Q-1) \nonumber \\
\Longleftrightarrow A \,< \, \frac{\mbox{PMINUS}(c, Q-1) }{\mbox{PPLUS3}(c, Q-1)}\,=\, \frac{(1+c)^{Q-1}- (1-\frac{c}{3})^{Q-1}}{(1+c)^{Q-1}+ 3 (1-\frac{c}{3})^{Q-1}}\;\;\;\;\label{flagekl31} \\
\mbox{(For } c=3 \mbox{ both sides of (\ref{flagekl31}) are }= 1.) \nonumber \end{aligned}$$
For $ c=1 $ inequality ( \[flagekl31\]) becomes $$\begin{aligned}
1-\frac{7}{15Q} \,< \,
\frac{2^{Q-1}- \left(\frac{2}{3} \right)^{Q-1}}{2^{Q-1}+ 3 \left( \frac{2}{3} \right)^{Q-1}} \,= \, 1- \frac{4\left( \frac{1}{3} \right)^{Q-1}}{1+3\left(\frac{1}{3}\right)^{Q-1}} \nonumber \\
\mbox{As }\, 4\left( \frac{1}{3} \right)^{Q-1}\,< \, \frac{7}{15Q} \mbox{ for } \, \, Q \ge 5 ,\, \,
\mbox{ (\ref{flagekl31}) holds for }
c=1 \mbox{ and } s \ge 5 \, \mbox{ as } Q \ge s. \nonumber \end{aligned}$$
To show that (\[flagekl31\]) holds for all $3\,> \,c \ge\, \, 1$ we show that the right-hand-side is concave for $c>1.$ $$\begin{aligned}
\mbox{ Numerator of } \frac{d}{dc} \frac{\mbox{PMINUS}(c, Q-1) }{\mbox{PPLUS3}(c, Q-1)} \,= \, \nonumber \\
(Q-1)\cdot \left((1+c)^{Q-2} \,+ \, \frac{1}{3} \left( 1- \frac{c}{3}\right)^{Q-2}\right)\cdot
\left((1+c)^{Q-1} \,+ \, 3 \left( 1- \frac{c}{3}\right)^{Q-1}\right) \,- \, \nonumber \\
\,- \, (Q-1)\cdot \left((1+c)^{Q-1} \,-\, \left( 1- \frac{c}{3}\right)^{Q-1}\right)\cdot
\left((1+c)^{Q-2} \,- \, \left( 1- \frac{c}{3}\right)^{Q-2}\right) \ \,\nonumber \\
\,= \, (Q-1) \cdot (1+c)^{Q-2}\cdot \left( 1 - \frac{c}{3} \right)^{Q-2}\cdot \left(\left( \frac{1}{3}+1\right)(1+c)
\,+ \, \left(1- \frac{c}{3}\right)\cdot (1+3) \right) \, \, \nonumber \\
\,= \, (Q-1)\cdot(1+c)^{Q-2}\cdot \left( 1 - \frac{c}{3} \right)^{Q-2} \cdot \left( \frac{1}{3}+3+2\right) \;\;\;\;\;\label{flagekl32} \\
\mbox{ We have that }(1+c)\cdot \left( 1 - \frac{c}{3}\right)
\mbox{ is decreasing for } c > 1,
\mbox{ and PPLUS3}(c, Q-1) \mbox{ is increasing } . \nonumber \\
\mbox{Therefore the right-hand-side of
(\ref{flagekl31}) is concave. } \nonumber \end{aligned}$$
Proof of Lemma \[stgekl\]
--------------------------
\
\
[**Lemma \[stgekl\] (repeated)**]{} $$\begin{aligned}
\mbox{ Let } s\ge 6 \mbox{ and let }\, A (c)\,\,=\,\, A(c, s) \,\,=\,\, \frac{Q}{2}\cdot c\, \,. \nonumber \\
\mbox{ Then }\, \mbox{OPT}(A(c), \, c, \, s) \, \mbox{ is strictly decreasing for } 0\, <c\, \le \, \frac{1}{Q} \nonumber \end{aligned}$$
Analogously to (\[flagekl2\]) and (\[flagekl3\]) this follows from (\[stegekl1\]) and (\[stegekl2\].) (Notation cf. ( \[notPPPM\].) $$\begin{aligned}
\mbox{ with } A=A(c),\, \, A' \,= \,\, Q/2 \, \,\frac{3\frac{\exp(As)-1)}{\exp(s)-1}}{\mbox{OPT}_1(A,\, s)}A'\,- \, \frac{A'c}{1+Ac}\,\,< \, \,0, \,
\label{stegekl1} \\
- \, \frac{A}{1+Ac}\,+ \, \frac{\mbox{PMINUS}(c, Q-1)}{\mbox{PPLUS3}(c, Q)} \,<\,0 \label{stegekl2}.\end{aligned}$$
[*Proof of (\[stegekl1\]) for $ s \ge 5.55 \, \,$ and $ 0 \,<\, c \, \le \,1/Q$* ]{} $$\begin{aligned}
K\,:= \, \frac{\exp(sA)-1}{\exp(s)-1}, \,\, \, L:= \frac{\exp(sA)-sA -1}{\exp(s)-s-1} \nonumber \\
\mbox{ We need to show } \frac{3KA'}{3L+1}\, \, < \, \frac{A'c}{1+Ac} \, \,
\Longleftrightarrow \, \, \frac{3K}{3L+1}\, \, < \, \frac{c}{1+Ac} \nonumber \\
\Longleftrightarrow \, \, 3\left( K+ KAc -Lc \right) \,< \,c
\mbox{ For } c=0 \mbox{ both sides are } \, =0 .\label{stegekl11} \end{aligned}$$ As $AK\le L$ by (\[BAKL\]) we get that (\[stegekl11\]) is implied by $3 \cdot K < c.$ For $c=0$ both sides of $3 \cdot K < c$ are $0.$ The left-hand-side is convex. It is sufficient to show $3 \cdot K < c.$ for $c=1/Q.$ Plugging in the definition of $1/Q$ for $c$ and $ A(1/Q \, , s)\,=\, 1/2 $ into $K$ we need to show $$\begin{aligned}
\frac{3\exp(s/2)-1}{\exp(s)-1}< \frac{\exp(s)-s-1}{ s (\exp(s)-1)} \,
\Longleftrightarrow \, 3 s\exp(s/2) \, < \,\exp(s)-1 \nonumber \end{aligned}$$ For $s\ge 6$ the preceding inequality holds by simple consideration.\
[*Proof of (\[stegekl2\]) for $s\ge 2 \, \, $ and $c \le 1/Q$*]{} Analogously to the proof of (\[flagekl31\]) we need to show $$\begin{aligned}
A\,= \, \frac{Q}{2}c \,> \, \frac{\mbox{PMINUS}(c, Q-1) }{\mbox{PPLUS3}(c, Q-1)}\,=\,
\frac{(1+c)^{Q-1}- (1-\frac{c}{3})^{Q-1}}{(1+c)^{Q-1}+ 3 (1-\frac{c}{3})^{Q-1}}\;\;\;\;\label{stgekl21} \\
\mbox{ For } c=0 \mbox{ both sides of the preceding inequality are } = 0 \nonumber \end{aligned}$$ We show, that $A'\,>$ the derivative wrt. $c$ of the right-hand-side of (\[stgekl21\]). Using (\[flagekl32\]) we need to show
$$\begin{aligned}
\frac{Q}{2} \cdot \left( (1+c)^{Q-1} + 3\left(1-\frac{c}{3}\right)^{Q-1}\right)^2\, \, > \, \,
(Q-1)\cdot (1+c)^{Q-2}\cdot \left( 1- \frac{c}{3} \right)^{Q-2}\cdot \frac{16}{3}. \nonumber \\
\mbox{Note }\, Q\cdot (1+c)^{Q-1}\cdot \left( 1- \frac{c}{3} \right)^{Q-1}\cdot \frac{16}{3}
\ge (Q-1)\cdot (1+c)^{Q-2}\cdot \left( 1- \frac{c}{3} \right)^{Q-2}\cdot \frac{16}{3} \nonumber \\
\mbox{ as } \, \, \, (1+c) \cdot \left(1-\frac{c}{3}\right) \ge 1 \mbox{ for } 0 \le c \le 1/Q < 2 .\nonumber \\
\mbox{Enlarging the right-hand-side it is sufficient to show } \nonumber \\
3\left( (1+c)^{Q-1} + 3\left(1-\frac{c}{3}\right)^{Q-1}\right)^2\,>\,
(1+c)^{Q-1}\cdot \left( 1- \frac{c}{3} \right)^{Q-1}\cdot 32 . \nonumber \\
\Longleftrightarrow 3\left(1+ 3\left(\frac{1-\frac{c}{3}}{1+c}\right)^{Q-1}\right)^2\,> \, 32\cdot \left(\frac{1-\frac{c}{3}}{1+c}\right)^{Q-1} \nonumber \end{aligned}$$
Setting $x= \left(\frac{1- \frac{c}{3}}{1+c}\right)^{Q-1}$ it is easy to see that the preceding inequality holds for $x \ge 0,$ and therefore clearly for $ c \le 1/Q < 3 .$
Proof of Lemma \[pukl\] and Lemma \[einmi\]
--------------------------------------------
Lemma \[einmi\] follows by elementary consideration, see the analogous situation in the proof of Lemma \[lem2\](a) and Lemma \[lem3\] (a).
[**Lemma \[pukl\] (repeated)**]{} Let $ s\, \ge 6\,$ then OPT$(a, \, \, c, \, \, s\,) \,< \, 4-\delta $ for $ (a, \, c)\,= \, $ $$\begin{aligned}
=\,\left(\frac{1}{2}, \, \, \frac{1}{Q} \, \right)\, ,\, \left(\frac{1}{2}, \, \frac{2}{Q}\,\right),\,
\,\left(\frac{2}{3}, \, \frac{2}{Q} \,\right) , \, \left(\frac{2}{3}, \, \frac{3}{Q} \,\right), \,
\left( 1\, \, - \, \, \frac{7}{15Q}\, , \, \frac{3}{Q} \right), \, \,
\, \left(1\, - \, \frac{7}{15Q}\, \, , \, 1 \,\right) \nonumber \end{aligned}$$
The claim for $a = \frac{1}{2} , \, \, c\,= \, \frac{1}{Q}$ is included in Lemma \[stgekl\].
$$\begin{aligned}
\mbox{FIRSUM}(a, c, s)\,= \, (1+c)^Q\mbox{OPT}_2(a, c, s)\,\,= \, \, \left(\frac{1+c}{1+ac}\right)^Q \nonumber \\
\mbox{SECSUM}(a, c, s)\,= \, 3\left(1- \frac{c}{3}\right)^Q \mbox{OPT}_2(a, c, s)\,= \,
3\left(\frac{1-\frac{c}{3}}{1+ac}\right)^Q
\mbox{ then }\nonumber \\
\mbox{OPT}(a, c, s) \,= \, \mbox{OPT}_1(a, s)\cdot \left[ \mbox{FIRSUM}(a, c, s)\,+ \, \,\mbox{SECSUM}(a, c, s)\right]. \nonumber \\
\mbox{For } x , y\ge 0 \, \mbox{ we have } \mbox{FIRSUM}\left(x, \frac{y}{Q}, s \right) \, =\,\,
\left(\frac{1+ \frac{y}{Q}}{1+x\cdot \frac{y}{Q}}\right)^Q \,= \, \, \, \nonumber \\
=\, \, \left( 1 \, + \, \frac{\frac{y}{Q}(1-x)}{1+ x\frac{y}{Q}}\right)^Q \, \,
\le \, \exp\left( \frac{y(1-x)}{1+x\frac{y}{Q}}\right) \le \,\exp(y(1-x)) \label{bFIR} \\
\mbox{ We have that OPT}_1(a, s) \mbox{is decreasing in } s \mbox{ for constant } a < 1. \label{OPT1dec}\end{aligned}$$
$$\begin{aligned}
\mbox{ Let } \, a=\frac{1}{2}, \, \, c\,= \, \frac{2}{Q}. \nonumber \\
\mbox{ We have by (\ref{bFIR}) }
\mbox{FIRSUM}(a, c, s) \, < \,\exp(1) \nonumber \\
\mbox{SECSUM} (a, c, s) \mbox{ is decreasing in } s\ge 0. \nonumber \\
\mbox{(As can be shown by elementary means.)} \, \, \nonumber \\
\mbox{OPT}_1(a, s) \left( \mbox{SECSUM}(a, c, s)+ \exp(1) \right) \,< \,3.913 \,\mbox{ for } s=5 \nonumber \\
\mbox{ and decreasing in } s \mbox{ with (\ref{OPT1dec})} \nonumber \end{aligned}$$
$$\begin{aligned}
\mbox{ Let } \, a=\frac{2}{3}, \, \, c\,= \, \frac{2}{Q}. \nonumber \\
\mbox{ We have } \, \mbox{FIRSUM}(a, c, s) \, < \,\exp(2/3) \nonumber \\
\mbox{SECSUM} (a, c, s) \mbox{ is decreasing in } s\ge 0. \, \, \nonumber \\
\mbox{OPT}_1(a, s) \left( \mbox{SECSUM}(a, c, s)+ \exp(2/3) \right) \,< \,3.962 \,\mbox{ for } s=4 \nonumber \\
\mbox{ and decreasing in } s \mbox{ with (\ref{OPT1dec})} \nonumber \end{aligned}$$
$$\begin{aligned}
\mbox{ Let } \, a=\frac{2}{3}, \, \, c\,= \, \frac{3}{Q}. \mbox{ We have }
\mbox{FIRSUM}(a, c, s) \, < \,\exp(1) \nonumber \\
\mbox{SECSUM} (a, c, s) \mbox{ is decreasing in } s\ge 2. \, \, \nonumber \\
\mbox{OPT}_1(a, s) \left( \mbox{SECSUM}(a, c, s)+ \exp(1) \right) \,< \,3.985 \,\mbox{ for } s=6 \nonumber \\
\mbox{ and decreasing in } s \mbox{ with (\ref{OPT1dec})} \nonumber \end{aligned}$$
$$\begin{aligned}
\mbox{ Let } a=1-7/(15Q), \, c=3/Q .\nonumber \\
\mbox{OPT}_1(a, s) \mbox{ is increasing in } s \mbox{ to } 3\exp(-7/15)+1. \nonumber \\
\mbox{FIRSUM }(a, c, s), \, \mbox{SECSUM}(a, c, s) \mbox{are both decreasing in } s. \nonumber \\
(3\exp(-7/15)+1)\left( \mbox{SECSUM}(a, c, s)+ \mbox{FIRSUM}(a, c, s) \right)
\,< \,3.9 \mbox{ for } s=4 \nonumber \end{aligned}$$
The case $a=1\, - \, \frac{7}{15Q}$ and $c=1\,$ is included in Lemma \[flagekl\].
Proof of Theorem \[UNOPT\] for $d=4,\, \lambda \ge 1-1/d, s \ge 5.$ {#LARGE}
====================================================================
We fix $a=1.$ Observe that $B(1/c)$ in the subsequent lemma goes from $1$ to $1-1/(2Q)$ for $c\ge 3.$
\[lagr\] Let $ B(x) \, = \, B(x, s) \, = \, 1+3/(2Q)x-1/(2Q). $ Then OPT$(1, B(1/c), c, s) $ is strictly decreasing in $c \ge 3.$
[*Proof of Theorem \[UNOPT\] for $\lambda\ge 1-1/d.$*]{} We have $\lambda/(1-\lambda) \ge d-1.$ For each $P \ge d-1$ we have $c$ such that $P \, =\, \frac{c}{B(1/c)}.$ As OPT$(1, 1, 3, s)=4$ the Theorem follows.
[*Proof of Lemma \[lagr\].*]{} We rewrite OPT$(1, B(1/c), c, s)$ first. We multiply OPT$_2$ with $c^Q$ and OPT$_3$ with $1/c^Q$ and get (using $c\ge 3$ to get rid of the absolute value) OPT$(1, B(1/c, s), c, s) \,= \,$ $$\begin{aligned}
=\, \left( 3+ \frac{q(B(1/c, s) s)}{q(s)}\right)\left(\frac{1}{\frac{B(1/c, s)}{c}+1}\right)^Q \left(
\left(\frac{1}{c}+1\right)^Q +
3\left( \frac{1}{3}- \frac{1}{c}\right)^Q \right) .\nonumber \\
\mbox{ We substitute } c \mbox{ for } 1/c \mbox{ in the preceding equation. The claim follows from }
\nonumber \\
\left( 3+ \frac{q(B(c))s)}{q(s)}\right)\left(\frac{1}{B(c)c+1}\right)^Q \left(
\left(c+1\right)^Q +
3\left(\frac{1}{3}-c\right)^Q \right) \nonumber \\
\mbox{increases in } 0\, <\, c \,<\, 1/3. \label{lagro}\end{aligned}$$
We use the following notation in the sequel: $$\begin{aligned}
{\mathrm{OPT}}_1(b,s) = 3 + \frac{q(sb)}{q(s)}\, , \, {\mathrm{OPT}}_2(b,c,s) \, = \,\left( \frac{1}{bc+1} \right)^Q , \nonumber \\
{\mathrm{OPT}}_3(c,s) = (c+1)^Q+\,3\left(\frac{1}{3}-c\right)^Q , \, c \le \frac{1}{3} \nonumber \\
{\mathrm{OPT}}(b, c, s)= {\mathrm{OPT}}_1(b, s){\mathrm{OPT}}_2(b, c, s){\mathrm{OPT}}_3(c, s). \nonumber \\
\mbox{ For } b=1, c= \frac{1}{3} \mbox{ we have } {\mathrm{OPT}}(b, c, s)=4. \mbox{ We abbreviate } \nonumber \\
\mbox{PM}(x, y)= (x+1)^y- 3\left(\frac{1}{3}-x\right)^y , \, x \le \frac{1}{3} \nonumber \\
\mbox{PP}(x, y)= (x+1)^y+ 3\left(\frac{1}{3}-x\right)^y, \, x \le \frac{1}{3} \nonumber \\
B\,= B(c, s), \quad B' = {\frac{\partial}{\partial c}}B = \frac{3}{2Q}, \quad q'(x) = \exp(x)-1 \nonumber \\
q(x)=\exp(x)-x-1, \, \, \, \, \, {\frac{\partial}{\partial c}}\ln({\mathrm{OPT}}(B,c,s) > = <0 \nonumber \\
\Longleftrightarrow \frac{\frac{B'sq'(sB)}{q(s)}}{3+\frac{q(sB)}{q(s)}}\,-Q\frac{B'c+B}{1+Bc}
+ Q\frac{\mbox{PM}(c, Q-1)}{\mbox{PP}(c, Q)} > 0 \nonumber \\
\Longleftrightarrow \frac{\frac{B'q'(sB)}{q'(s)}}{3+\frac{q(sB)}{q(s)}} - \frac{B'c+B}{1+Bc}
+ \frac{\mbox{PM}(c, Q-1)}{\mbox{PP}(c, Q)} >=< 0. \mbox{( Division by } Q.)
\label{gro_ableit}\end{aligned}$$
{width="49.00000%"} {width="49.00000%"}
For $c=\frac{1}{3}$ we have ${\mathrm{OPT}}(B, 1/3, s)=4$, and the derivative is $0$. We split (\[gro\_ableit\]) into two additive terms. The following two inequalities directly imply (\[lagro\].)
$$\begin{aligned}
\frac{\frac{B'q'(sB)}{q'(s)}}{3+\frac{q(sB)}{q(s)}} - \frac{B'c}{1+Bc} & > & 0 \label{gro_teil1}\\
\frac{\mbox{PM}(c, Q-1)}{\mbox{PP}(c, Q)}- \frac{B}{1+Bc} & > & 0 \label{gro_teil2}\end{aligned}$$
[*Proof of (\[gro\_teil1\]) for $\, s>2 \, .$*]{} Let $K = \frac{q'(sB)}{q'(s)}$ and $L= \frac{q(sB)}{q(s)}.$ By (\[BAKL\]) we have $L\le K,$ and as $B' > 0$ it is sufficient to show $$\frac{K}{3+K} > \frac{c}{1+Bc} \Leftrightarrow K(1+Bc-c) > 3c. \nonumber$$
For $c=\frac{1}{3}$ both sides of the preceding inequality are $=1$. It is easy to observe that $K(1+Bc-c)$ is convex in $c$ for $c>1/(3\cdot 2).$ (Cf. proof of Lemma \[lem4\]) and $3c$ is a linear function. If at $c=\frac{1}{3}$ the derivative of $3c$ is greater than the derivate of $K(1+Bc-c)$ the second intersection of both sides (if any) lies at some point $c>\frac{1}{3}$ and the claim holds for $1/(3\cdot 2<c<\frac{1}{3}$. For $c<1/(3\cdot 2) $ we argue as in the proofs of the Lemmas mentioned above. Therefore it is sufficient to show that at $c=\frac{1}{3}$ $${\frac{\partial}{\partial c}}K(1+Bc-c) < {\frac{\partial}{\partial c}}3c. \nonumber$$ We have $$K' = \frac{B's\exp(sB)}{\exp(s)-1}.$$ and at $c=1/3$ $$\begin{aligned}
K'(\underbrace{1+Bc-c}_{=1}) + \underbrace{K}_{=1}(\underbrace{B'c+B-1}_{=1/2Q}) & < & 3 \nonumber \\
\Leftrightarrow \frac{3(\exp(s)-s-1)\exp(s)}{2(\exp(s)-1)^2}+\frac{\exp(s)-s-1}{2s(\exp(s-1)} & < & 3 \label{last}\end{aligned}$$ We omit the proof that inequality (\[last\]) holds for $s\ge 2.$\
[*Proof of (\[gro\_teil2\]) for $s\ge 5$.*]{} As in (\[flagekl31\]) inequality (\[gro\_teil2\]) is equivalent to $$B < \frac{\mbox{PM}(c,Q-1)}{\mbox{PP}(c,Q-1} \label{gro_teil2_umform}$$
The left hand side is a linear function in c and the right hand side a strictly increasing, concave function in $c.$ For $c=\frac{1}{3}$ both sides of (\[gro\_teil2\_umform\]) are $1$. So we must show that (\[gro\_teil2\_umform\]) holds for $c=0$. Setting $c=0$ leads to $$1-\frac{1}{2Q} < \frac{1-3\left( \frac{1}{3} \right)^{Q-1}}{1+3\left( \frac{1}{3} \right)^{Q-1}}
= \frac{1- \left(\frac{1}{3}\right)^{Q-2}}{1+ \left(\frac{1}{3}\right)^{Q-2}}. \nonumber$$ For $Q=5$ we get $\frac{9}{10} < \frac{13}{14}$. We omit the argument that the last inequality holds for all $Q\ge 5$ and therefore as $Q>s$ for all $s \ge5.$
[0000]{} J. Diaz et al. On the satisfiability threshold of formulas with three literals per clause. Theoretical Computer Science 410 (2009) 2920 - 2934. M Molloy. Cores in random hypergraphs and boolean formulas. Random Stuctures and Algorithms 27, 2005, 124 - 135. J. Hastad. Some optimal inapproximability results. J. ACM 48, 2001, 798 – 859. R. Durrett. Probability Theory: Theory and Examples. Wadsworth and Brooks 1991. M Dietzfelbinger et al. Tight thresholds for Cuckoo Hashing via XORSAT. CoRR, 2009, abs/0912.0287. See also Proceedings ICALP 2010, LNCS 6198, 213 - 225. H. Connamacher, M. Molloy. The exact satisfiability threshold for a potentially in tractable random constraint satisfaction problem. Proceedings 45th FoCS 2004, 590 - 599. Michael Molloy. Models for Random Constraint Satisfaction Problems. SIAM J. Comput. 32(4), 2003, 935-949. O. Dubois, J. Mandler. The $3-$XORSAT satisfiability threshold. Proceedings 43rd FoCS 2003, 769. N. Creignou, H. Daud[é]{}. The SAT-UNSAT transition for random constraint satisfaction problems. Discrete Mathematics 309 (8), 2085 - 2099. V. F. Kolchin. Random graphs and systems of linear equations in finite fields. Random Structures and Algorithms 5, 1995, 425 - 436. A. Braunstein, M. Mezard, R. Zecchina. Survey propagation: an algorithm for satisfiability. arXiv:cs/0212002. Amin Coja-Oghlan, Angelica Y. Pachon-Pinzon. The Decimation Process in Random k-SAT. In Proceedings ICALP (1) 2011, 305-316. Ehud Friedgut. Hunting for sharp thresholds. Random Struct. Algorithms 26(1-2), 2005, 37-51. Andreas Goerdt. On Random Betweenness Constraints. Combinatorics, Probability and Computing 19(5-6), 2010, 775-790 D. Achlioptas, C. Moore. Random k-SAT: Two Moments Suffice to Cross a Sharp Threshold. SIAM J. Comput. 36(3), 2006, 740-762 V. Puyhaubert. Generating functions and the satisfiability threshold. Discrete Mathematics and Theoretical Computer Science 6, 2004, 425, - 436. H. Connamacher. Exact thresholds for DPLL on random XOR-SAT and NP-complete extensions of XOR-SAT. Theoretical Computer Science 2011. A. Meisels, S. E. Shimony, G. Solotorevsky. Bayes Networks for estimating the number of solutions to a CSP. Proceedings AAAI 1997, 179 - 184. M. Luby, M. Mitzenmacher, A. Shokrollahi, D. A. Spielman. Efficient erasure coeds. IEEE Trans. Inform. Theory 47(2), 2001, 569 - 584. T. J. Richardson, R. Urbanke. Modern Coding Theory. Cambridge University Press, 2008. Dimitris Achlioptas, Morteza Ibrahimi, Yashodhan Kanoria, Matt Kraning, Mike Molloy and Andrea Montanari. The Set of Solutions of Random XORSAT Formulae. In Proceedings SoDA 2012. N. G. de Bruijn. Asymptotic Methods in Analysis. North Holland 1958, N. Bhattacharya, R. Ranga Rao. Normal Approximation and Asymptotic Expansions. Robert E. Krieger Publishing Company, 1986. M. Mitzenmacher, Eli Upfal. Probability and Computing: Randomized Algorithms and Probabilistic Analysis. Cambridge University Press 2005.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We address quantum invisibility in the context of electronics in nanoscale quantum structures. We make use of the freedom of design that quantum corrals provide and show that quantum mechanical objects can be hidden inside the corral, with respect to inelastic electron scattering spectroscopy in combination with scanning tunneling microscopy, and we propose a design strategy. A simple illustration of the invisibility is given in terms of an elliptic quantum corral containing a molecule, with a local vibrational mode, at one of the foci. Our work has implications to quantum information technology and presents new tools for nonlocal quantum detection and distinguishing between different molecules.'
author:
- 'J. Fransson'
- 'H. C. Manoharan'
- 'A. V. Balatsky'
title: Detection and Cloaking of Molecular Objects in Coherent Nanostructures Using Inelastic Electron Tunneling Spectroscopy
---
As we approach the quantum limit for a wide range of experiments and technologies, it is rather natural to ask whether it is possible to hide information from the measurements. A recent theoretical study suggested that a cloak of invisibility is in principle possible in optical applications, using the freedom of design offered by metamaterials in order to redirect electromagnetic fields [@pendry2006]. It has since been shown that a copper cylinder may be hidden inside a metamaterial cloak constructed according to the theoretical prescription [@schurig2006].
Physical and chemical signal detection is based on the interactions between the probe and the sample. Those interactions are probed within a certain frequency range, which itself depends on the specific property we are probing, e.g spin or charge response. Thus, by eliminating or weakening the effect of the relevant interactions within the operational frequency range, the objects can be made invisible from the detector. In this sense, one has been able to show that metamaterials can be used as a cloaking device. One example is a copper cylinder, that was made *invisible* in the microwave frequency band range. Here, we discuss detection and invisibility within the THz band. The discussion is based on the ability to intentionally engineer the properties of quantum structures with tailored properties in the THz band range. This high frequency regime is accessible for nano structures and it opens the window for detection of species down to single molecules and atoms, based on their chemical (vibrational) signatures.
In this Letter, we address invisibility in the context of electronics in nanoscale quantum structures, for which we use quantum corrals as a prototype coherent device. We take advantage of the freedom of design that quantum corrals provide and show that quantum mechanical objects located inside the corral can be made invisible with respect to certain measurement. We first need to ask how specific species of quantum matter can be detected. This can be accomplished through chemical identity of atoms and molecules for which molecular vibrations function as a *fingerprint* of species identity. In this case, the relevant frequency scale is THz rather than MHz. Remarkably, scanning tunneling microscopy (STM) can provide access to these frequency scales through the inelastic electron tunneling spectroscopy (IETS) process, and can simultaneously give access to the challenging length scales necessary to probe and manipulate individual quanta of matter and excitations. We propose a method for detecting and manipulating quantum invisibility based on THz cloaking of molecular identity in coherent nanostructures.
Using IETS combined with STM we thus have an experimental tool that enables identification of molecules through detection of certain *fingerprint frequencies* which correspond to vibrational modes of the molecules. Recently, CO molecules were used on Cu(111) surfaces for molecular assembly and quantum corrals [@moon2008] and well-known molecular excitations exist for this materials system. Quantum corrals with elliptic geometries have recently been employed for discussion of mirage effects [@manoharan2000; @aligia2001; @hallberg2002; @gadzuk2003; @correa2005; @rossi2006; @moonNP2008], where a magnetic adsorbate atom or molecule is located in one focus of the ellipse while the (STM) measurement is performed nonlocally in the second, empty, focus.
To be specific, we consider an elliptic quantum corral containing a single molecule with a local vibrational mode, located at one of the foci; we argue that the vibrational mode should be measurable at the second, unoccupied, focus through IETS combined with STM. Theoretically, it has been shown that vibrational modes in a molecular structure adsorbed on a metallic surface would be measurable by means of IETS in a narrow energy range around the vibrational mode [@balatsky2006; @franssonIETSfriedel2007]. Moreover, the inelastic scattering generates Friedel oscillations in the local surface density of states (DOS), thus enabling spatial imaging of the inelastic signatures. In the present context, the standing waves generated inside the corral by the inelastic scattering at the occupied focus are reflected and reassembled at the unoccupied focus, thus creating a finite ghost response. Within the same framework and exploiting engineering abilities of the quantum corral for another purpose, we here suggest that the molecule can be made undetectable.
As an experimentally realizable system [@moon2008], we consider elliptical quantum corrals comprised of $N$ CO molecules on a Cu surface. The CO molecules generate a vibrational mode $\omega_0\simeq4$ meV caused by translational frustration when placed on a metallic, e.g. Cu, surface. The presence of the quantum corral generates a modified LDOS in a neighborhood around itself. Using the scattering theory proposed by Fiete and Heller [@rodberg1967; @heller2003], we calculate the modified LDOS built up by the quantum corral. Our calculations indicate that the IETS contribution from molecules on the walls is small with respect to the signal of the focal molecule for these geometries. Thus, the contribution from the vibrations in the wall does not significantly contribute to the effect we are studying, but merely provides a small modification to the overall background. Therefore, we neglect the inelastic contribution from the wall and, for simplicity, use the elastic scattering theory in the construction of the LDOS generated by the molecules in the wall of the quantum corral. We employ a $T$-matrix approach [@rodberg1967; @heller2003] to generate the electronic structure imposed on the surface by the elliptical quantum corral. The corral is generated by placing molecules equidistant along the ellipse $(x/a)^2+(y/b)^2=R^2$. The two foci in the ellipse are located at $\bfr_\pm=(\pm c,0)$ where $c=\sqrt{a^2-b^2}$. Having established the electronic structure of the quantum corral, we proceed by including the inelastic effects from a molecular impurity located at the *molecular focus* $\bfr_+$ (the second focus at $\bfr_-$ is henceforth referred to as the *empty focus*).
All vibrational modes have the fingerprint frequency $\omega_0$ and are assumed to be the same as they originate from the same type of molecules. We use the Hamiltonian for the local vibrational modes, coupled to electrons via Holstein coupling [@holstein1959] with interactions assumed to occur only at the single impurity site, so that $$\begin{aligned}
\Hamil=\sum_{\bfk\sigma}\dote{\bfk}\cdagger{\bfk}\cc{\bfk}
+\omega_0\bdagger{}\bc{}
+\lambda\sum_{\bfk\bfk'\sigma}\csdagger{\bfk\sigma}\cs{\bfk'\sigma}(\bdagger{}+\bc{}).
\label{eq-Ham}\end{aligned}$$ Here, a surface electron is created (annihilated) by $\cdagger{\bfk}\ (\cc{\bfk})$ at the energy $\dote{\bfk}$. The strength of the electron-vibron interaction is given by the parameter $\lambda$, whereas $\omega_0$ is the mode of the bare vibron which is created (annihilated) by $\bdagger{}\ (\bc{})$.
![a) Typical electronic structure of the quantum corral at the ellipse foci $\pm c$. The underlying ellipse is given by $(x/a)^2+(y/b)^2=R^2$ with $R=57.21$ Å, and $a/b=1.5$, comprised of 40 molecules. b) Local DOS at the ellipse foci at the Fermi level as function of the ellipse radius $R$. The electronic structures were calculated using quadratic energy dispersion $E_\text{surf}(k)-E_F=E_0+\hbar^2k^2/2m^*$, with $E_0\simeq-0.45$ eV for Cu(111), and $m^*=0.38m$, where $m$ is the free electron mass.[]{data-label="fig-qc1"}](N0_comb){width="8.5cm"}
The features we are considering are expected to be seen in the second derivative of the tunneling current with respect to the bias voltage $V$ in real space, i.e. $\partial^2I(\bfr,V)/\partial V^2$. This quantity is directly proportional to the frequency derivative of the local DOS. In second order perturbation theory (sufficient for weak electron-vibron coupling), this amounts to taking the frequency derivative of the correction to the density of states, $\delta N(\bfr,\omega)$, due to the influence of the impurity scattering. The real space electron Green function (GF) is given by$$\begin{aligned}
G(\bfr,\bfr';\omega)=&G_0(\bfr,\bfr';\omega)
\nonumber\\&
+G_0(\bfr,\bfr_+;\omega)\Sigma(\omega)G_0(\bfr_+,\bfr';\omega),
\label{eq-GF}\end{aligned}$$ where $G_0(\bfr,\bfr')$ is the GF for electronic structure of the quantum corral without the molecular impurity [@franssonIETSQC2008], whereas $\bfr_+$ is the position of the impurity inside the quantum corral. The self-energy $\Sigma$ is given by $$\begin{aligned}
\Sigma(i\omega)=&
\lambda^2\sum_\bfk
\biggl[\frac{n_B(\omega_0)+f(\dote{\bfk})}{i\omega+\omega_0-\dote{\bfk}}
\nonumber\\&\hspace{2cm}
+\frac{n_B(\omega_0)+1-f(\dote{\bfk})}{i\omega-\omega_0-\dote{\bfk}}\biggr],
\label{eq-S}\end{aligned}$$ where $n_B(x)$ and $f(x)$ are the Bose and Fermi functions, respectively.
For later reference, we denote the unperturbed LDOS of the quantum corral by $N_0(\bfr,\omega)=-\im{G_0(\bfr,\bfr;\omega)}/\pi$ and the correction due to the electron-vibron coupling by $$\begin{aligned}
\delta N(\bfr,\omega)=&
-\frac{1}{\pi}\im{\{G_0(\bfr,\bfr_+;\omega)\Sigma(\omega)G_0(\bfr_+,\bfr;\omega)\}}.\end{aligned}$$ The dressed LDOS $N(\bfr,\omega)$ is then given by $N(\bfr,\omega)=N_0(\bfr,\omega)+\delta N(\bfr,\omega)$. Generally, the differential conductance $\partial I(\bfr,V)/\partial V$ between the STM and the structure is directly proportional to the LDOS $N(\bfr,\omega)$, and the second bias voltage derivative $\partial ^2I(\bfr,V)/\partial V^2\propto\partial N(\bfr,\omega)/\partial\omega$. The latter derivative, $\partial ^2I(\bfr,V)/\partial V^2$, is used for inelastic tunneling spectroscopy (IETS) using STM. It is therefore sufficient to study the LDOS and its frequency derivative in order to account for the expected qualitative results in an STM measurement.
Typically, the electronic structure of the elliptic quantum corral is an oscillating function of the energy. Fig. \[fig-qc1\] (a) illustrates a typical DOS at the two foci. The geometry in this particular case gives rise to a peak in the DOS at the Fermi level $\dote{F}$. Whether the local DOS at the foci peaks at the Fermi level is a geometrical matter of the structure. This is illustrated in Fig. \[fig-qc1\] (b), which displays the local Fermi level DOS at the foci for varying ellipse radius $R$, and it is clear that the DOS at the Fermi level is an oscillating function of the radius as well. The width of the density peaks are related to the fact that the quantum corral is comprised of discrete atoms or molecules and is not an entirely closed structure, although the wall nevertheless create quantum confinement. The wave functions, thus built up within the confinement potential, have a non-negligible leakage to the surrounding environment and this generates a broadening of the confined states. The level broadening is accounted for in the employed $T$-matrix approach [@rodberg1967; @heller2003]. The general conclusion drawn is that the quantum corral can be engineered, or designed, to meet certain requirements set by the desired functionality.
Quantum invisibility in our context would be the ability to hide information from detection, which in a broader sense implies information storage away from a perturbing environment. Hiding information away from reading technology would in principle mean an extremely small interaction with the surroundings and would, hence, allow for extremely long decoherence times. We might also think of other applications of this ability to make physical objects invisible from the measurements. For instance, we can create devices which can individually distinguish between different types of molecules, or other types of molecular or mesoscopic objects, by determining whether the quantum object carries a predefined fingerprint frequency.
In particular, the functionality we propose in this Letter is the ability to detect and/or hide molecules adsorbed at one focus inside the elliptic quantum corral, which have their fingerprint frequency within a certain frequency range. It is well known that perturbing the electronic structure in one focus generates a mirror response at the second focus, which presumably is empty [@manoharan2000; @gadzuk2003]. However, the remote projection of an object at one focus is efficient only when the local DOS at the foci peaks, or is significant, at the particular energy of interest. The reason is straightforward: A large density allows for a large response while the response becomes suppressed for a small density. This rule thus enables a dual functionality of the device we propose, namely both detection within a certain frequency range and alternately quantum invisibility of molecular objects.
![(Color online) a) Size and energy dependence of the response $|\partial_\omega\delta N(\bfr_-,\omega)|^2$ at the empty focus $\bfr_-=(-c,0)$. b) Size dependence of the restricted deviation $\sigma_{\omega_0}$ at the empty focus. Here we use $R_0=57.21$ Å, $\omega_0=4$ meV, at $T=4$ K. The restricted deviation has been calculated by integrating over the interval $(\omega_0-\delta,\omega_0+\delta)$, with $\delta=4$ meV.[]{data-label="fig-d2I"}](d2I_sigma){width="8.5cm"}
In order to quantify our assertions and to make contact with realistic experimental situations, consider the elliptic quantum corral corresponding to the local DOS at the foci given in Fig. \[fig-qc1\] (a). From the above discussion we thus expect the response to the inelastic scattering to be an oscillating function of the ellipse radius. Moreover, following the argument in Refs. [@balatsky2006; @franssonIETSfriedel2007] we expect sharp features in the response $\partial ^2I/\partial V^2\propto\partial\delta N(\bfr,\omega)/\partial\omega$ at the inelastic resonances $\omega=\pm\omega_0$ for low temperatures. These expected features of the response to the inelastic scattering are verified in Fig. \[fig-d2I\] (a), where we plot the response $|\partial\delta N(\bfr,\omega)/\partial\omega|^2$ at the empty focus $\bfr_-=(-c,0)$ as a function of the ellipse radius and energy. It is readily seen that the response is an oscillating function of the radius and that it peaks around the fingerprint frequency $\pm\omega_0$.
We characterize the quality of the response from the inelastic scattering by defining the integrated deviation from the unperturbed response by $$\begin{aligned}
\sigma(\bfr,R;\omega,\omega_0)=\sqrt{\int
\biggl(\frac{\partial}{\partial\omega}\delta N(\bfr,\omega)\biggr)^2d\omega}.
\label{eq-dev}\end{aligned}$$ A large value of this function clearly describes that the response to the inelastic scattering significantly deviates from the unperturbed response. Moreover, since the response $\partial\delta N(\bfr,\omega)/\partial\omega$ is expected to be large only in a narrow interval around the inelastic resonances $\pm\omega_0$ and small anywhere outside, this function provides a good measure of the deviation caused by the electron-vibron coupling. In the experimental situation, however, one would measure the response in a finite interval around the inelastic resonance, e.g. in the interval $(\omega_0-\delta,\omega_0+\delta)$, $\delta>0$. It therefore makes sense to restrict the integration in Eq. (\[eq-dev\]) to this interval, and we call this function $\sigma_\delta(\bfr,\bfR;\omega,\omega_0)$. In Fig. \[fig-d2I\] (b) we plot this restricted deviation at the empty focus $\bfr_-$ as function of the ellipse radius $R$, corresponding to the setup in Fig. \[fig-d2I\] (a). This plot clearly illustrates the possibility to engineer the structure in order to maximize measurability of the electron-vibron response.
The ability to purposely engineer the quantum structure has wider implications than the above discussion about optimization of the response from the inelastic mode. For a given quantum structure, the underlying electronic structure is an oscillating function of the energy. As a such, one can use the presence or absence of underlying electronic density for the purpose of displaying or hiding the inelastic response from the impurity at the molecular focus.
![(Color online) Dependence of the response $|\partial_\omega\delta N(\bfr,\omega)|^2$ on the vibron mode $\omega_0$ at a) the empty focus $\bfr=(-c,0)$, and b) the molecular focus $\bfr=(c,0)$, in a corral with a density peak in the foci at the Fermi level. Here we use $R_0=57.21$ Å, at $T=4$ K. The two branches correspond to the two inelastic resonances at $\omega=\pm\omega_0$.[]{data-label="fig-Q1"}](Q2_R5721_cpm_w0){width="8.5cm"}
Assume that the corral is optimized for a peaked electron density at the Fermi level in the foci. As seen in Fig. \[fig-d2I\], there is a detectable response at the empty focus whenever the vibrational mode of the impurity at the molecular focus lies within the width of the density peak. If, on the other hand, the vibrational mode fall outside the width of the density peak the response at the empty focus becomes less pronounced, and eventually invisible. A typical example of this discussion is illustrated in Fig. \[fig-Q1\], which shows the inelastic response at (a) the empty, and (b) molecular focus for varying vibrational modes of the impurity at the molecular focus. The two branches correspond to the inelastic resonances at $\pm\omega_0$. The plot clearly illustrates that the response in the empty focus decays as the vibrational mode starts to fall outside the width of the density peak, and eventually for sufficiently large vibrational modes the impurity becomes invisible to the probe at the empty focus. The returning signal for even larger modes arise because of the presence of another density peak at larger energies. All these features occur in the empty focus although the inelastic response at the molecular focus is sufficiently strong for detection. By the same token, we can design the corral to have a low electron density at the Fermi level in the foci. In this situation, impurities with low vibrational modes become invisible to the probe at the empty focus (see Fig. \[fig-Q2\]). The asymmetry in the response with respect to resonance frequency is due to the fact that the underlying electron density is not symmetric around the Fermi level for the given geometry.
![(Color online) Same as in Fig. \[fig-Q1\] for a corral with low electron density in the foci at the Fermi level. Here we use $R_0=60.21$ Å. []{data-label="fig-Q2"}](Q2_R6021_cpm_w0){width="8.5cm"}
In summary, we have considered the engineering of quantum structures for precisely designing functionality. By way of example, using elliptic quantum corrals with radii between 50 and 65 Å, comprising about 40 atoms or molecules, we show that the geometry of elliptical quantum corrals can be optimized for large or small inelastic response at one focus, generated by the vibrational mode or *fingerprint frequency* of an impurity at the other focus. Such optimization opens possibilities for detection devices that can be used to distinguish between different species of molecules. Moreover, functional design of the quantum structures enables cloaking, or quantum invisibility, of quantum objects.
The authors thank I. Grigorenko for useful discussions. This work has been supported by US DOE, LDRD and BES, and was carried out under the auspices of the NNSA of the US DOE at LANL under Contract No. DE-AC52-06NA25396 (J.F. and A.V.B), by the US DOE at SLAC under Contract No. DE-AC02-76SF00515 (H.C.M), and by the NSF (H.C.M). J.F. thanks Swedish Research Council (VR) for support.
[20]{} J. B. Pendry, D. Schurig, and D. R. Smith, Science [**312**]{}, 1780 (2006). D. Schurig, J. J. Mock, B. J. Justice, S. A. Cummer, J. B. Pendry, A. F. Starr, and D. R. Smith, Science [**314**]{}, 977 (2006). C. R. Moon, L. S. Mattos, B. K. Foster, G. Zeltzer, W. Ko, and H. C. Manoharan, Science [**319**]{}, 782 (2008).
H. C. Manoharan, C. P. Lutz, and D. M. Eigler, Nature, [**403**]{}, 512 (2000). J. W. Gadzuk and M. Plihal, Phys. Rev. B [**68**]{}, 235413 (2003). C. R. Moon, C. P. Lutz, and H. C. Manoharan, Nature Physics, [**4**]{} 454 (2008).
A. A. Aligia, Phys. Rev. B, [**64**]{}, 121102(R) (2001). K. Hallberg, A. A. Correa, and C. A. Balseiro, Phys. Rev. Lett. [**88**]{}, 066802 (2002). A. A. Correa, F. A. Reboredo, and C. A. Balseiro, Phys. Rev. B, [**71**]{}, 035418 (2005). E. Rossi and D. K. Morr, Phys. Rev. Lett. [**97**]{}, 236602 (2006).
A.V. Balatsky, I. Vekhter, and J. -X. Zhu. Rev. Mod. Phys. [**87**]{}, 373 (2006). J. Fransson and A. V. Balatsky, Phys. Rev. B, [**75**]{}, 195337 (2007).
L. S. Rodberg and R. M. Thaler, *Introduction to the Quantum Theory of Scattering*, (Academic, New York, 1967). G. A. Fiete and E. J. Heller, Rev. Mod. Phys. [**75**]{}, 933 (2003). T. Holstein, Ann. Phys. (N.Y.) [**8**]{}, 325 (1959).
J. Fransson, H. C. Manoharan, and A. V. Balatsky, unpublished (2008).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We address the question which additional information on the source shape and dynamics can be extracted from three-particle Bose-Einstein correlations. For chaotic sources the true three-particle correlation term is shown to be sensitive to the momentum dependence of the saddle point of the source and to its asymmetries around that point. For partially coherent sources the three-pion correlator allows to measure the degree of coherence without contamination from resonance decays. We derive the most general Gaussian parametrization of the two- and three-particle correlator for this case and discuss the space-time interpretation of the corresponding parameters.'
address: |
Institut für Theoretische Physik, Universität Regensburg,\
D-93040 Regensburg, Germany
author:
- 'U. Heinz and Q.H. Zhang[@home]'
title: 'What can we learn from three-pion interferometry?'
---
PACS numbers: 25.75+r, 13.85 Hd, 24.10-i.
Introduction {#sec1}
============
Two-particle Bose-Einstein interferometry (also known as Hanbury Brown-Twiss intensity interferometry) as a method for obtaining information on the space-time geometry and dynamics of relativistic heavy ion collisions has recently received intensive theoretical and experimental attention. Detailed theoretical investigations (for a recent review see Ref. [@He96]) have shown that high-quality two-particle correlation data can reveal not only the geometric extension of the particle-emitting source but also its dynamical state at particle freeze-out. This information is encoded in the second central space-time moments of the “emission function” $S(x,K)$, i.e. of the Wigner phase-space density of the source. For chaotic sources, certain linear combinations of these moments can be extracted from the two-particle correlation function $C_2(\bbox{q},\bbox{K})$ by fitting it to a Gaussian in the relative momentum $q$ of the pair [@BDH94; @CSH95; @HTWW96]. These second space-time moments give the size of the regions of homogeneity [@MS88; @CSH95] which effectively contribute to the emission of particle pairs with a given pair momentum $K$; collective dynamics of the source results in a characteristic $K$-dependence of these homogeneity regions [@P84; @WSH96; @HTWW96].
More detailed information on the space-time structure of the source may be hidden [@WH96] in possible non-Gaussian features of the correlation function $C_2(\bbox{q},\bbox{K})$ even if they are hard to extract; due to the symmetry under $q\to -q$, however, only even space-time moments of the source are accessible via two-particle correlations. In this paper we will extend previous studies of multi-particle correlations [@Zajc1; @Pratt3; @Zhang1; @Zhang3; @APW93] and show that three-pion correlations provide in principle additional information on the space-time characteristics of the source which cannot be obtained from two-particle interferometry. We show in particular in Sec. \[sec2\] that for completely chaotic sources the true three-pion correlations are determined by the phase of the two-particle exchange amplitude [@APW93; @HV96] which drops out from the two-particle cross section. This phase is shown to be sensitive to the rate at which the saddle point $\bar x(K)$ of the source, from which most pairs with momentum $K$ are emitted, moves as $K$ changes, and to the asymmetries of the emission function around this saddle point via its [*third*]{} central space-time moments. Unfortunately, this phase turns out to be generically small, and its sensitivity to these asymmetries is very weak, making them extremely hard to measure.
In the absence of such a non-trivial phase, three-particle correlations can still be used to test the chaoticity of the emitting source. To this end we derive the expressions for two- and three-particle correlations for chaotic and partially coherent sources and establish their respective relationships. Our treatment differs from previous studies of multi-particle Bose-Einstein correlations in that we consistently express the correlation functions through the source Wigner density, even for partially coherent sources. This enables us to relate the shape of the correlators as functions of the various relative momenta to certain space-time features of the source. To the best of our knowledge the corresponding relations for partially coherent sources (Eqs. (\[x20\]) - (\[x22\])) are new.
Chaotic sources {#sec2}
===============
For a chaotic source, the two-pion correlation function $C_{2}(p_i,p_j)$ can be expressed as [@He96; @APW93] $$\begin{aligned}
C_{2}(p_i,p_j) &=& {P_2(p_i,p_j) \over P_1(p_i)\, P_1(p_j)}
\nonumber\\
&=& 1 + {\left \vert \int d^4x \, S(x,K_{ij})\,
e^{i q_{ij}\cdot x} \right \vert^2
\over
\int d^4x\, S(x,p_i) \ \int d^4y\, S(y,p_j)}
= 1 + \frac{|\rho_{ij}|^2}{\rho_{ii} \rho_{jj}} \, .
\label{1}
\end{aligned}$$ Here $P_2(p_i,p_j)$ is the two-pion inclusive cross section, and $P_1(p_i)$ is the single-particle inclusive spectrum. $S(x,p)$ is the single-particle Wigner density of the source, i.e. the quantum mechanical analogue of its phase-space distribution. The average and relative 4-momenta $K_{ij}=(p_i+p_j)/2$ and $q_{ij}=p_i-p_j$ satisfy the constraint $q_{ij}\cdot K_{ij} = 0$ which results from the on-shell nature of the observed momenta $p_i$. The two-particle exchange amplitude $\rho_{ij}$ is defined as [@APW93; @CH94] $$\begin{aligned}
\rho_{ij} &=& \rho(q_{ij},K_{ij}) = \sqrt{E_i\, E_j}
\langle \hat a^\dagger(p_i) \hat a(p_j) \rangle
\nonumber\\
&=& \int d^4x \, S(x,K_{ij}) \,
e^{i q_{ij}\cdot x} \equiv f_{ij}\, e^{i\phi_{ij}} \, .
\label{2}
\end{aligned}$$ From (\[2\]) it follows that $\rho_{ij}=\rho^*_{ji}$ and thus $f_{ij}=f_{ji}$ and $\phi_{ij}=-\phi_{ji}$. Correspondingly, $\phi_{ii}=0$, $\rho_{ii} = f_{ii}$, and $f_{ij}$ must be an even function of $q_{ij}$ while $\phi_{ij}$ is odd in $q_{ij}$.
The single-pion spectrum can be written as $$P_1(p_{i}) = \int d^4x \, S(x,p_i) = f_{ii}
\label{3}$$ while the true two-pion correlation function is defined by $$R_2(i,j) \equiv R_2(p_i,p_j) = C_2(p_i,p_j) - 1
= \frac{f_{ij}^{2}}{f_{ii} \, f_{jj}} \, .
\label{4}$$ Similarly, the true three-pion correlation function is given by [@APW93; @W89; @BBMST90; @SB92; @PRW92; @Cramer; @Eggers] $$\begin{aligned}
R_3(p_1,p_2,p_3) &=& C_3(p_1,p_2,p_3) - R_2(1,2) - R_2(2,3) - R_2(3,1) - 1
\nonumber\\
&=& 2 \frac{{\rm Re\, }(\rho_{12}\,\rho_{23}\,\rho_{31})}
{f_{11}\,f_{22}\,f_{33}}
\label{6} \\
&=& 2 \frac{f_{12}\,f_{23}\,f_{31}}
{f_{11}\,f_{22}\,f_{33}} \,
\cos(\phi_{12}+\phi_{23}+\phi_{31}) \, .
\nonumber
\end{aligned}$$ Since the real parts $f_{ij}$ of the exchange amplitudes $\rho_{ij}$ can be extracted from the two-pion correlator, for chaotic sources the only additional information contained in the 3-pion correlation function resides in the phase [@APW93] $$\label{6a}
\Phi \equiv \phi_{12}+\phi_{23}+\phi_{31} \, ;$$ it is a linear combination of the phases of the three exchange amplitudes $\rho_{12}$, $\rho_{23}$, and $\rho_{31}$ which enter the true 3-pion correlator $R_3$. This phase is odd under interchange of any two particles. It can be isolated by normalizing $R_3$ with respect to the true 2-pion correlator $R_2$: $$r_3(p_1,p_2,p_3)
= \frac{R_3(p_1,p_2,p_3)}{\sqrt{R_2(1,2) R_2(2,3) R_2(3,1)}}
= 2 \cos \Phi \, .
\label{7}$$
In order to understand which space-time features of the source affect the phase $\Phi$ (and thus the normalized true 3-pion correlation function $r_3$) we expand the exchange amplitude $\rho_{ij}$ for small values of $q_{ij} = p_i - p_j$ [@BDH94; @CSH95]. We define the average of an arbitrary space-time function $f(x)$ with the source distribution $S(x,K_{ij})$ as $$\langle f(x) \rangle_{ij} = \frac{\int d^4x \, f(x) \, S(x,K_{ij})}
{\int d^4x \, S(x,K_{ij})} \, .
\label{8}$$ This average is a function of the pair momentum $K_{ij}$. Using (\[2\]) we thus get $$\rho_{ij} = P_1(K_{ij}) \left[
1 + i \langle q_{ij}{\cdot}x \rangle_{ij}
- \frac{1}{2} \left\langle (q_{ij}{\cdot}x)^2 \right\rangle_{ij}
- \frac{i}{6} \left\langle (q_{ij}{\cdot}x)^3 \right\rangle_{ij}
+ O\left(q_{ij}^4\right)
\right] \, .
\label{9}$$ Separating real and imaginary parts we find, after a little algebra, $$f_{ij} = P_1(K_{ij}) \left[
1 - \frac{1}{2} \left\langle (q_{ij}{\cdot}{\tilde x}_{ij})^2
\right\rangle_{ij}
+ O\left(q_{ij}^4\right)
\right]
\label{10}$$ and $$\phi_{ij} =
q_{ij}{\cdot}\langle x \rangle_{ij}
- \frac{1}{6} \left\langle (q_{ij}{\cdot}{\tilde x}_{ij})^3
\right\rangle_{ij}
+ O\left(q_{ij}^5\right) \, ,
\label{11}$$ where $$\tilde{x}_{ij} = x -\langle x \rangle_{ij} = x - \bar x (K_{ij})
\label{12}$$ is the distance to the “saddle point” of the source, i.e. to the point of maximum emission for pions with momentum $K_{ij}$. According to Eqs. (\[10\]) and (\[4\]), the two-pion correlator is sensitive to the second central (i.e. saddle-point corrected) space-time moments of the emission function $S(x,K_{ij})$ [@BDH94; @CSH95], with higher order corrections from all even central space-time moments. The phase $\Phi$, on the other hand, contains information on the odd space-time moments. Expanding $S(x,K_{ij})$ around the average momentum $K$ of the pion triplet, $$\begin{aligned}
K &=& {p_1 + p_2 + p_3 \over 3} = {K_{12} + K_{23} + K_{31} \over 3} \, ,
\label{13}\\
K_{ij} &=& K + {1\over 6} \left( q_{ik} + q_{jk} \right)\, ,
\qquad i\ne j \ne k\, ,
\label{14}
\end{aligned}$$ and using $q_{12}+q_{23}+q_{31}=0$, we find from Eqs. (\[6a\]) and (\[11\]) $$\begin{aligned}
\Phi &=& \frac{1}{2} \, q_{12}^\mu \, q_{23}^\nu
\left[ {\partial \langle x_\mu \rangle \over \partial K^\nu}
- {\partial \langle x_\nu \rangle \over \partial K^\mu}
\right]
\nonumber\\
&-& {1\over 24} \left[ q_{12}^\mu q_{12}^\nu q_{23}^\lambda
+ q_{23}^\mu q_{23}^\nu q_{12}^\lambda \right]
\left[ {\partial^2 \langle x_\mu \rangle \over
\partial K^\nu \partial K^\lambda}
+ {\partial^2 \langle x_\nu \rangle \over
\partial K^\lambda \partial K^\mu}
+ {\partial^2 \langle x_\lambda \rangle \over
\partial K^\mu \partial K^\nu} \right]
\label{15}\\
&-& {1\over 2} q_{12}^\mu q_{23}^\nu (q_{12} + q_{23})^\lambda
\, \langle \tilde x_\mu \tilde x_\nu \tilde x_\lambda \rangle
+ O(q^4) \, .
\nonumber
\end{aligned}$$ Here the average without subscripts $$\langle f(x) \rangle = \frac{\int d^4x \, f(x) \, S(x,K)}
{\int d^4x \, S(x,K)}
\label{16}$$ denotes the space-time average with the emission function evaluated at the mean momentum $K$ of the pion triplet, and $$\tilde{x} = x - \langle x \rangle = x - \bar x(K)\, .
\label{17}$$ Eq. (\[15\]) is the main new result of this Section. One easily checks that has it the correct symmetries under particle exchange. It should be noted that, due to the on-shell constraint $q_{ij}\cdot
K_{ij}=0$, only three of the four components $q_{ij}^\mu$ are independent. The resulting relation $$\label{17a}
(q^0)_{ij} = \bbox{q_{ij}}\cdot \bbox{\beta}_{ij},
\quad \text{with}\quad
\bbox{\beta}_{ij} = \bbox{K}_{ij}/(K^0)_{ij}\, ,$$ can be used to eliminate the redundant $q$-components in Eq. (\[15\]), thereby mixing spatial and temporal components of the corresponding coefficients. This is a well-known problem also for the two-pion correlator (see, e.g., [@He96]) which prohibits a clean model-independent separation of the spatial and temporal widths of the source.
Eq. (\[15\]) features two types of contributions to the phase $\Phi$: The formally leading contribution enters at second order in the relative momenta $q_{ij}$ and is proportional to the rate $\partial \bar x_\mu(K) / \partial K^\nu$ with which the saddle point of the emission function changes as a function of the pion momentum $K$. This term will in general be non-zero even for emission functions with a purely Gaussian $x$-dependence. It gives rise to a leading $q^4$-dependence of the normalized true three-particle correlator $r_3
= 2 \cos \Phi$. At order $q^3$ the phase $\Phi$ receives additional contributions from the second $K$-derivatives of the saddle point as well as from the third central space-time moments $\langle \tilde
x_\mu \tilde x_\nu \tilde x_\lambda \rangle$ of the source. The latter are the leading contributions from a possible asymmetry of the emission function $S(x,K)$ around its saddle point $\bar x(K)$; they vanish for purely Gaussian emission functions. We see that they enter the normalized three-particle correlator $r_3$ at order $q^5$ in a mixture with the $K$-dependence of the saddle point. This renders their isolation essentially impossible.
In contrast to the widths of the emission function, which affect the two-pion correlator at [*second*]{} order in the relative momentum, the additional structural information which can (in principle) be extracted from the (normalized) three-pion correlator is seen to enter at most at [*fourth*]{} order in $q$. Their measurement is thus very sensitive to an accurate removal of all leading $q^2$-dependences by proper normalization to the two-particle correlators. To achieve this looks like an extremely difficult experimental task. We are therefore somewhat pessimistic about the short-term prospects of extracting additional structural information about the source from three-pion correlations.
If the phase $\Phi$ and the information it contains about the source are inaccessible, what else can three-pion correlations be used for experimentally? The answer is that one can test the assumption that the source is chaotic. This has been pointed out previously in Refs. [@BBMST90; @PRW92] where specific simple parametrizations for the two- and three-particle correlators (as well as for higher order correlations) were assumed and the relationship between the various parameters was studied. We will here derive more general expressions which, in principle, permit such a test without making any simplifying assumptions about the shape of the source.
Before proceeding to the discussion of Bose-Einstein correlations from partially coherent sources, we would like to close this Section with a few short remarks on the effects from resonance decays. It is well known [@FW77; @GKW79] that partial coherence in the source leads to incomplete correlations in the two-particle sector, in the sense that $R_2(q,K)$ at vanishing relative momentum $q=0$ does not approach the ideal value $R_2(0,K)=2$ for chaotic sources. In actual experiments there are, however, other possible reasons for apparently incomplete two-particle correlations. Most importantly, pions from the decay of long-lived resonances contribute to the correlator only at very small values of $q$ and thus (due to limited 2-track resolution) may escape detection in the correlation signal while fully contributing to the single-particle spectrum, thereby reducing the apparent correlation strength even for a completely chaotic source [@Z86; @SOPW92; @CLZ96; @WH96a]. In a Gaussian parametrization of the exchange amplitude this can be implemented by writing instead of Eq. (\[10\]) for $q_{ij}\ne 0$ $$\label{18}
f_{ij} = \lambda^{1/2}(K_{ij})\, P_1(K_{ij})\,
\exp\left[ - {1\over 2} q_{ij}^\mu q_{ij}^\nu R_{\mu\nu}(K_{ij})
\right] \, ,$$ where, up to second order in $q$, $R^{\mu\nu}(K_{ij}) = \langle
\tilde x^\mu_{ij} \tilde x^\nu_{ij} \rangle_{ij}$, with the source average on the r.h.s. being taken only over the “core” of pions from direct emission and from the decays of short-lived resonances [@He96; @WH96a; @C96]. The two-particle correlator then becomes $$\label{19}
R_2(i,j) = \lambda(K_{ij})\,
{P_1^2(K_{ij}) \over P_1(p_i) \, P_1(p_j)} \,
\exp\left[ - q_{ij}^\mu q_{ij}^\nu R_{\mu\nu}(K_{ij}) \right] \, ,$$ and for vanishing relative momenta $q$ the three-particle correlation function assumes the value $$\label{20}
C_3(p_1{=}p_2{=}p_3{=}K) = 1 + 3\, \lambda(K) + 2\, \lambda^{3/2}(K) \, .$$ Note, however, that the expression (\[7\]) for the normalized true three-pion correlation function is not affected by resonance decay contributions and remains unchanged. This will no longer be true for partially coherent sources.
Partially coherent sources {#sec3}
==========================
Expressions for the $n$-particle inclusive spectra from partially coherent sources have been previously derived, with differing methods, in Refs. [@APW93; @W89; @BBMST90; @SB92; @PRW92; @Cramer]. In the covariant current formalism of Refs. [@GKW79; @CH94] one decomposes the classical source current which creates the free pions in the final state into a coherent and a chaotic term: $$J(x) = J_{\rm coh}(x) + J_{\rm cha}(x) \, .
\label{x5}$$ Following the treatment of Ref. [@CH94] this leads to the following definition of the single-particle Wigner density (“emission function”) of the source: $$\begin{aligned}
S(x,K) &=& \int {d^4y\over 2 (2\pi)^3}\, e^{-iK\cdot y}
\langle J^*(x+{\textstyle{y\over 2}})
J(x-{\textstyle{y\over 2}}) \rangle
\nonumber\\
&=& S_{\rm coh}(x,K) + S_{\rm cha}(x,K) \, ,
\label{x6}
\end{aligned}$$ with
\[x7\] $$\begin{aligned}
\label{x7a}
S_{\rm coh}(x,K) &=&
\int {d^4y\over 2 (2\pi)^3}\, e^{-iK\cdot y}\,
J^*_{\rm coh}(x+{\textstyle{y\over 2}})
J_{\rm coh}(x-{\textstyle{y\over 2}}) \, ,
\\
\label{x7b}
S_{\rm cha}(x,K) &=&
\int {d^4y\over 2 (2\pi)^3}\, e^{-iK\cdot y}\,
\langle J^*_{\rm cha}(x+{\textstyle{y\over 2}})
J_{\rm cha}(x-{\textstyle{y\over 2}}) \rangle \, .
\end{aligned}$$
The average on the r.h.s. of the definition (\[x7b\]) for the chaotic part of the emission function is defined as in Ref. [@CH94], and we used $$\label{x8}
\langle J^*_{\rm cha}(x)\, J_{\rm coh}(y) \rangle = 0 \, .$$ The Wigner density of the full source is thus the sum of a coherent and a chaotic contribution; no mixed terms occur because the chaotic and coherent source currents do not interfere. This allows to carry over the intuitive and very successful Wigner function language for fully chaotic sources to the case of partially or completely coherent sources.
We now write $$\begin{aligned}
\rho_{ij} &=& \int d^4x\, S(x,K_{ij}) \, e^{iq_{ij}\cdot x}
\nonumber\\
&=& \rho_{ij}^{\rm cha} + \rho_{ij}^{\rm coh}
\equiv F_{ij}\, e^{i\Phi_{ij}} + f_{ij}\, e^{i\phi_{ij}} \, ,
\label{x11}
\end{aligned}$$ where $K_{ij} = (p_i+p_j)/2,\ q_{ij} = p_i-p_j$, and
\[x12\] $$\begin{aligned}
F_{ij}\, e^{i\Phi_{ij}}
&=& \int d^4x\, S_{\rm cha}(x,K_{ij})\, e^{iq_{ij}\cdot x} \, ,
\label{x12b}\\
f_{ij}\, e^{i\phi_{ij}}
&=& \int d^4x\, S_{\rm coh}(x,K_{ij}) \, e^{iq_{ij}\cdot x} \, .
\label{x12c}
\end{aligned}$$
As shown in Ref. [@APW93] this yields the two-pion correlation function in the form $$C_2(p_i,p_j) = 1 + R_2(i,j)
= 1+ \frac{F_{ij}^2 + 2 f_{ij} F_{ij}
\cos(\Phi_{ij}-\phi_{ij})}
{(f_{ii}+F_{ii})(f_{jj}+F_{jj})} \, ,
\label{x14}$$ while the three-particle correlation is given by $$\begin{aligned}
C_3(p_1,p_2,p_3) &=& \frac{P_3(p_1,p_2,p_3)}
{P_1(p_1)\, P_1(p_2)\, P_1(p_3)}
= 1 + R_2(1,2) + R_2(2,3) + R_2(3,1)
\nonumber\\
&+& \frac{2}{P_1(p_1)\, P_1(p_2)\, P_1(p_3)}
\Bigl( F_{12} F_{23} F_{31} \cos(\Phi_{12}+\Phi_{23}+\Phi_{31})
\nonumber\\
&+& f_{12} F_{23} F_{31} \cos(\phi_{12}+\Phi_{23}+\Phi_{31})
\nonumber\\
&+& F_{12} f_{23} F_{31} \cos(\Phi_{12}+\phi_{23}+\Phi_{31})
\nonumber\\
&+& F_{12} F_{23} f_{31} \cos(\Phi_{12}+\Phi_{23}+\phi_{31})
\Bigr) \, .
\label{x15}
\end{aligned}$$ Similar expressions were derived in Ref. [@SB92]. The two- and three-particle correlations are seen to vanish for completely coherent sources ($F_{ij}\to 0\ \forall i,j$). In the opposite limit ($f_{ij}\to 0\ \forall i,j$) one recovers the results from Sec. \[sec2\] for completely chaotic sources.
The representations (\[x11\]) and (\[x12\]) permit us to write down for $F_{ij},\, f_{ij}$ and $\Phi_{ij},\, \phi_{ij}$ similar small-$q$ expansions as in Eqs. (\[10\]) and (\[11\]); the corresponding averages are defined with respect to the chaotic and coherent parts, respectively, of the Wigner function (\[x6\]). In the true two-pion correlation function $R_2(i,j)$ of Eq. (\[x14\]), the first term thus contains information on the second central space-time moments of $S_{\rm cha}(x,K_{ij})$ while the second term mixes the second moments of $S_{\rm cha}(x,K_{ij})$ and $S_{\rm
coh}(x,K_{ij})$ in a rather nontrival way. Since the number of measurable parameters in $R_2(i,j)$ is the same as before, this implies a relative loss of information: the second space-time moments of $S_{\rm cha}$ and $S_{\rm coh}$ can neither be separated nor do they simply combine to the second central moments of the total source $S = S_{\rm cha} + S_{\rm coh}$.
This complication goes hand in hand with a similar one in the three-pion correlator: Defining the true three-pion correlator as before, $$\begin{aligned}
R_3(1,2,3) &=& C_3(p_1,p_2,p_3) - 1 - R_2(1,2) - R_2(2,3) - R_2(3,1)
\nonumber\\
&=&\frac{2}{(f_{11}+F_{11})(f_{22}+F_{22})(f_{33}+F_{33})}
\nonumber\\
&&\times \Bigl(
F_{12} F_{23} F_{31} \cos(\Phi_{12}+\Phi_{23}+\Phi_{31})
+ f_{12} F_{23} F_{31} \cos(\phi_{12}+\Phi_{23}+\Phi_{31})
\nonumber\\
&& \ \
+ F_{12} f_{23} F_{31} \cos(\Phi_{12}+\phi_{23}+\Phi_{31})
+ F_{12} F_{23} f_{31} \cos(\Phi_{12}+\Phi_{23}+\phi_{31})
\Bigr) ,
\label{x16}
\end{aligned}$$ one sees that, in contrast to Eq. (\[7\]) for chaotic sources, the phase factors can no longer be isolated by normalizing $R_3$ with a proper combination of two-particle correlators $R_2$. This means that, in a samll-$q$ expansion, $R_3(1,2,3)$ contains leading terms of second order in $q$ which are independent of those occurring in the two-particle correlator. On the one hand, those terms supplement the incomplete information from $R_2$ on the second space-time moments of the source; on the other hand, they render the measurements of source asymmetries impossible.
The full reconstruction of all the (in principle) measurable information obviously requires a measurement of $R_2(i,j)$ and $R_3(1,2,3)$ as a function of all nine components of $\bbox{p}_1,\bbox{p}_2,\bbox{p}_3$ . In view of the technical complexity (both experimental and theoretical) of such a program this is not likely to happen soon. It must, however, be mentioned that simple one- or two-parameter Gaussian parametrizations as suggested in Refs. [@BBMST90; @PRW92; @HV96] are not sufficient for this purpose because they very strongly prejudice the form of the source.
To pursue this last point a little further, let us define the (momentum-dependent) chaotic fraction of the single particle spectrum $$\epsilon(p_i)=\frac{F_{ii}}{f_{ii}+F_{ii}}
= \frac{\int d^4x\, S_{\rm cha}(x,p_i)}{\int d^4x\, S(x,p_i)}
\label{x17}$$ The coherent fraction is accordingly $f_{ii}/(f_{ii}+F_{ii})=1-
\epsilon(p_i)$. For vanishing relative momentum $q_{ij}=0 \ (i,j=1,2,3)$, we then have $$\begin{aligned}
R_2(p,p) &=& \epsilon(p)\bigl(2-\epsilon(p)\bigr)\, ,
\nonumber\\
R_3(p,p,p) &=& 2\, \epsilon^2(p) \, \bigl(3-2\epsilon(p)\bigr) \,.
\label{x18}
\end{aligned}$$ For completely chaotic sources, $\epsilon(p) = 1$, we recover the results of Sec. \[sec2\]. For partially coherent sources, the normalized three pion correlator $r_3$ at vanishing $q$ is given by $$r_3(p,p,p) = \frac{R_3(p,p,p)}{\bigl(R_2(p,p)\bigr)^{3/2}}
= 2 \sqrt{\epsilon(p)} \frac{(3-2\epsilon(p))}
{(2-\epsilon(p))^{3/2}}
\label{x19}$$ which, in general, deviates from the chaotic limit $r_3(p,p,p)=2$.
It would thus seem to be a simple matter to check the limits of $R_2$ and $R_3$ for vanishing relative momenta and construct the ratio (\[x19\]) in order to see whether or not the source contains a coherent component. In practice, however, the $q=0$ limit can not be measured directly, but requires an extrapolation of data at finite $q$ to zero relative momenta. It is well known that such an extroplation can be very sensitive to the assumed functional behavior of the correlator at small $q$. As we will now show our results provide a basis for a reasonable parametrization of $R_2$ and $R_3$ for small $q$.
To this end we start from Eqs. (\[x14\]) and (\[x16\]) together with the small $q$ expansions (\[10\]), (\[11\]). Noting that $R_2$ must vanish for $q\to \infty$, a parametrization which is correct up to second order in $q$ is given by $$\begin{aligned}
R_2(i,j) &\approx& \epsilon^2(K_{ij})\,
\exp\left[ - q_{ij}^\mu q_{ij}^\nu R_{\mu\nu}(K_{ij}) \right]
\nonumber\\
&+& 2 \epsilon(K_{ij})(1-\epsilon(K_{ij}))
\exp\left[ -{\textstyle{1\over 2}} q_{ij}^{\mu} q_{ij}^{\nu}
\left(R_{\mu\nu}(K_{ij}) + r_{\mu\nu}(K_{ij})\right) \right]
\cos\left(q_{ij}{\cdot}s(K_{ij})\right)\, .
\label{x20}
\end{aligned}$$ It follows from Eqs. (\[10\]), (\[11\]) that here
\[x21\] $$\begin{aligned}
R^{\mu\nu}(K_{ij})
&=& \langle \tilde{x}_{ij}^\mu \tilde{x}_{ij}^\nu \rangle_{ij}^{\rm cha}
\, ,
\label{x21a}\\
r^{\mu\nu}(K_{ij})
&=& \langle \tilde{x}_{ij}^\mu \tilde{x}_{ij}^\nu \rangle_{ij}^{\rm coh}
\, ,
\label{x21b}\\
s^{\mu}(K_{ij})
&=& \langle x^\mu \rangle^{\rm cha}_{ij}
- \langle x^\mu \rangle^{\rm coh}_{ij}
\, .
\label{x21c}
\end{aligned}$$
Eq. (\[x20\]) neglects an additional factor $P^2(K_{ij})/P(p_i)
P(p_j)$ which is unity for exponentical single particle spectra [@CSH95]. Eq. (\[x20\]) differs from the parametrization suggested in Ref. [@PRW92] by the factor $\cos\left(q_{ij}\cdot
s(K)\right)\, \exp\left[ - {1\over 2} q_{ij}^{\mu} q_{ij}^\nu
r_{\mu\nu}(K_{ij}) \right]$; the parametrization of Ref. [@PRW92] is thus not general enough. (It essentially assumes that the coherent part of the source is pointlike (in space [*and*]{} time!) and localized at the saddle point of the chaotic part of the source.) Note that from Eq. (\[x20\]) one must still eliminate the redundant $q$-component via the on-shell constraint (\[17a\]).
The three-pion correlator can similarly parametrized as $$\begin{aligned}
R_3(p_1,p_2,p_3) &=& 2 \epsilon^2(K)
\exp \left[ - \left( q_{12}^\mu q_{12}^\nu + q_{23}^\mu q_{23}^\nu
+ {\textstyle{1\over 2}} (q_{12}^\mu q_{23}^\nu + q_{12}^\mu q_{23}^\nu)
\right) R_{\mu\nu}(K) \right]
\nonumber\\
&\times& \Bigl[\ \ \epsilon(K)
\nonumber\\
&& \ + (1-\epsilon(K))\, \cos(q_{12}{\cdot}s(K))\,
\exp\left( {\textstyle{1\over 2}} q_{12}^\mu q_{12}^\nu
(R_{\mu\nu}(K)-r_{\mu\nu}(K))\right)
\nonumber\\
&& \ + (1-\epsilon(K))\, \cos(q_{23}{\cdot}s(K))\,
\exp\left( {\textstyle{1\over 2}} q_{23}^\mu q_{23}^\nu
(R_{\mu\nu}(K)-r_{\mu\nu}(K))\right)
\nonumber\\
&& \ + (1-\epsilon(K))\, \cos((q_{12}+q_{31}){\cdot}s(K))
\nonumber\\
&&\qquad \times \exp\left( {\textstyle{1\over 2}}(q_{12}+q_{23})^\mu
(q_{12}+q_{23})^\nu (R_{\mu\nu}(K)-r_{\mu\nu}(K))
\right) \Bigr] \, .
\label{x22}
\end{aligned}$$ This again generalizes the parametrizations given in Refs. [@BBMST90; @PRW92]; according to Eqs. (\[10\]), (\[11\]), it is correct up to the second order in $q$ if one approximates $P_1^2(K_{ij}/P_1(p_i) /P_1(p_j) \approx 1$ as well as $\epsilon(K_{ij}) \approx \epsilon(K)$. The parametrizations of Ref. [@BBMST90; @PRW92] are recovered in the limit of a pointlike coherent source, $r_{\mu\nu}(K)=0$, and assuming $s(K)=0$. (The first of these two assumptions is explicity stated in Ref[@BBMST90].) One can easily convince oneself that at $q_{12}=0$, for example, the term $\cos(q_{23}{\cdot}s(K))\, \exp\left[\frac{1}{2}q_{23}^\mu
q_{23}^\nu \left( R_{\mu\nu}(K)-r_{\mu\nu}(K) \right) \right]$ enters $R_3(q_{23})$ with a different weight than $R_2(q_{23})$. Thus $R_{3}$ provides additional information which allows to separate $R_{\mu\nu}(K)$ from $r_{\mu\nu}(K)$ and thereby the widths of the chaotic and coherent parts of the source.
In practice, one must also take into account resonance decays. Since it follows from the discussion at the end of in Sec. \[sec2\] that the longlived resonances do not affect the intercept (\[x19\]) of the normalized true three-pion correlator, and it was shown in Refs. [@He96; @WH96a; @C96] that expression (\[x21a\]) remains essentially valid if the chaotic part of the emission function is restricted to the “core” of direct pions and short-lived resonance decays, we expect Eqs. (\[x20\]) - (\[x22\]) to be practically useful even when resonance decays are included.
Conclusions {#sec4}
===========
We have studied the question to what extent three-pion Bose-Einstein correlations can provide independent information about the space-time structure of the emitting source which cannot be extracted from two-pion correlations. For chaotic sources we found that the three-pion correlator depends on the phase of the two-particle exchange amplitude which drops out from the two-particle cross section. This phase can be isolated by proper normalization of the true three-pion correlator with respect to the two-pion correlator. It was shown to be sensitive to the momentum dependence of the point of highest emissivity in the source and to the asymmetries of the emission function around that point. However, this sensitivity is weak (it enters only at 4th order in the relative momenta $q_{ij}$), and the corresponding source properties are hard to measure.
We then proceeded to study sources which are not completely chaotic but contain a coherent component. We showed that in this case the emission function can be written as a sum of two Wigner densities describing the chaotic and coherent components, respectively, and expressed the two- and three-pion correlation functions via these chaotic and coherent Wigner densities. We showed that a comparison of two- and three-pion correlators allows for a determination of the degree of coherence in the source, without contaminations from resonance decays. To this end one must study the respective correlation functions at vanishing relative momenta of all particles. To facilitate the extraction of this limit from experimental data we derived in Eqs. (\[x20\]) and (\[x22\]) the most general parametrizations for the two-and three-pion correlation functions at small relative momenta. These new parametrizations are based on our expressions of the correlation functions in terms of the Wigner density of the source; they are exact up to second order in the relative momenta, i.e. for emission functions $S(x,K)$ with a Gaussian $x$-dependence. After eliminating the redundant $q$-components, they are seen to depend on 16 parameters which are all functions of the average momentum $\bbox{K}$ of the pion pair resp. triplet. To determine all these parameter functions, a complete study of the two-and three-particle spectra as functions of all 6 + 9 = 15 momentum components is necessary. (The 16th parameter, $\epsilon(K)$, describes the degree of coherence and enters the normalization of the correlation functions at vanishing relative momenta.) This is certainly not an easy task, and it might be worthwhile to study whether, for certain simple but not too unrealistic models for the emission function, it is not possible to obtain simpler parametrizations (for example by exploiting certain symmetries of the source).
Our results show that in the case of partially coherent sources the three-pion correlator contains independent information on the second space-time moments of the source which cannot be extracted from the two-pion correlator. This information is needed to separate the space-time characteristics (lengths of homogeneity or effective widths) of the chaotic and coherent parts of the emission function. To extract it in practice will not be easy, but the theoretical framework by which this should be done has been presented here.
The authors would like to express their gratitude to T. Csörgő, H. Heiselberg, A.P. Vischer, and U.A. Wiedemann for stimulating remarks. This work arose from discussions at the Workshop on Particle Interferometry in High Energy Heavy Ion Reactions (HBT96) at the ECT\* in Trento, Sept. 16 - 27, 1996. We would like to thank the ECT\* for their hospitality and for providing such a fruitful and stimulating atmosphere. Q.H.Z. gratefully acknowledges support by the Alexander von Humboldt Foundation through a Research Fellowship. The work of U.H. was supported in part by BMBF, DFG, and GSI.
[100]{} Humboldt Research Fellow; on leave from China Center of Advanced Science and Technology (CCAST). U. Heinz; in: [*Correlations and Clustering Phenomena in Subatomic Physics*]{}, ed. by M.N. Harakeh, O. Scholten, and J.H. Koch, NATO ASI Series B, (Plenum, New York, 1996), in press (Los Alamos eprint archive nucl-th/9609029) G. Bertsch, P. Danielewicz, and M. Herrmann, Phys. Rev. C[**49**]{}, 442 (1994). S. Chapman, P. Scotto and U. Heinz, Phys. Rev. Lett. [**74**]{}, 4400 (1995); and Heavy Ion Phys. [**1**]{}, 1 (1995). U. Heinz, B. Tomášik, U.A. Wiedemann, and Wu Y.-F., Phys. Lett. B[**382**]{}, 181 (1996). A.N. Makhlin and Yu.M. Sinyukov, Z. Phys. C[**39**]{}, 69 (1988); S.V. Akkelin and Yu.M. Sinyukov, Phys. Lett. B[**356**]{}, 525 (1995). S. Pratt, Phys. Rev. Lett. [**53**]{}, 1219 (1984). U.A. Wiedemann, P. Scotto and U. Heinz, Phys. Rev. C[**53**]{}, 918 (1996). U.A. Wiedemann and U. Heinz, Los Alamos eprint archive nucl-th/9610043, submitted to Phys. Rev. Lett. W.A. Zajc, Phys. Rev. D[**35**]{}, 3396 (1987). S. Pratt, Phys. Lett. B[**301**]{}, 159 (1993). W.Q. Chao, C.S. Gao and Q.H. Zhang; J. Phys. G[**21**]{}, 847 (1995). Q.H. Zhang, W.Q. Chao, C.S. Gao, Phys. Rev. C[**52**]{}, 2064 (1995). I.V. Andreev, M. Plümer, and R.M. Weiner, Int. J. Mod. Phys. A[**8**]{}, 4577 (1993). H. Heiselberg and A.P. Vischer, nucl-th/9609023. S. Chapman and U. Heinz, Phys. Lett. B[**340**]{}, 250 (1994). R.M. Weiner, Phys. Lett. B[**232**]{}, 278 (1989); B[**242**]{}, 547 (1990). M. Biyajima, A. Bartl, T. Mizoguchi, N. Suzuki, and O. Terazawa, Prog. Theor. Phys. [**84**]{}, 931 (1990). N. Suzuki and M. Biyajima, Prog. Theor. Phys. [**88**]{}, 609 (1992). M. Plümer, L.V. Razumov and R.M. Weiner, Phys. Lett. B[**286**]{}, 335 (1992). J.G. Cramer, Phys. Rev. C[**43**]{}, 2798 (1991); J.G. Cramer and K. Kadija, Phys. Rev. C[**53**]{}, 908 (1996). H.C. Eggers, P. Lipa, P. Carruthers and B. Buschbeck, Phys. Lett. B[**301**]{}, 298 (1993). G.N. Fowler and R.M. Weiner, Phys. Lett. B[**70**]{}, 201 (1977); and Phys. Rev. D[**11**]{}, 3118 (1978). M. Gyulassy, S.K. Kauffmann, and L. Wilson, Phys. Rev. C[**20**]{} (1979) 2267. W.A. Zajc, in: Hadronic Matter in Collision, edited by P. Carruthers and D. Strottman (World Scientific, Singapore, 1986), p. 43. B.R. Schlei, U. Ornik, M. Plümer, and R.M. Weiner, Phys. Lett. B[**293**]{}, 275 (1992); J. Bolz, U. Ornik, M. Plümer, B.R. Schlei, and R.M. Weiner, Phys. Lett. B[**300**]{}, 404 (1993); and Phys. Rev. D[**47**]{}, 3860 (1993). T. Csörgő, B. Lörstad, and J. Zimányi, Z. Phys. C[**71**]{}, 491 (1996). U.A. Wiedemann and U. Heinz, nucl-th/9611031, submitted to Phys. Rev. C. T. Csörgő, talk given at the HBT96 Workshop, ECT\* Trento, 16 - 27 Nov. 1996.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we present a generalization of Berreman’s model for the elastic contribution to the surface free-energy density of a nematic liquid crystal in presence of a sawtooth substrate which favours homeotropic anchoring, as a function of the wavenumber of the surface structure $q$, the tilt angle $\alpha$ and the surface anchoring strength $w$. In addition to the previously reported non-analytic contribution proportional to $-q\ln q$, due to the nucleation of disclination lines at the wedge bottoms and apexes of the substrate, the next-to-leading contribution is proportional to $q$ for a given substrate roughness, in agreement with Berreman’s predictions. We characterise this term, finding that it has two contributions: the deviations of the nematic director field with respect to a reference field corresponding to the isolated disclination lines, and their associated core free energies. Comparison with the results obtained from the Landau-de Gennes model shows that our model is quite accurate in the limit $wL>1$, when strong anchoring conditions are effectively achieved.'
author:
- 'O. A. Rojas-Gómez'
- 'J. M. Romero-Enrique'
title: 'Generalized Berreman’s model of the elastic surface free energy of a nematic liquid crystal on a sawtoothed substrate'
---
Introduction
============
The behaviour of nematic liquid crystals in the presence of microstructured substrates has been subject of intensive research in the recent times [@lee; @kim; @ferjani]. This problem has practical applications such as the design of zenithally bistable devices [@Brown_2000; @Parry-Jones1; @Parry-Jones2; @davidson; @evans], or the trapping of colloidal particles in specified sites [@nuno; @ohzono]. It is well known that the nematic director field, in presence of the structured substrate, may be distorted, leading to an elastic contribution to the free energy. Since the seminal work by Berreman [@berreman; @gennes], this problem has been extensively studied and generalized in the literature [@barbero1; @barbero2; @barbero3; @Brown_2000; @Kitson_2002; @Fukuda; @patricio5; @harnau; @kondrat; @harnau2; @harnau3; @barbero4; @Yi_2009; @poniewierski; @romero]. Wetting and filling transitions by nematic on these grooved surfaces have also been studied [@bramble; @patricio; @patricio2; @patricio4]. When the substrate presents cusps, topological defect nucleate on them [@barbero1; @barbero2; @barbero3; @poniewierski; @romero], and Berreman’s expression of the elastic contribution to the free-energy density, which is proportional to the wavenumber of the substrate structure $q$, breaks down because of the emergence of a non-analytical contribution proportional to $-q\ln q$ associated to the nucleated defects [@romero]. This result constrasts with the phenomenology observed in smooth substrates, as sinuosidal substrates, where in absence of topological defects a suitable generalization of Berreman’s model works [@barbero4; @patricio4].
In this paper we will complete the characterization of the elastic contribution to the surface free-energy density for sawtoothed substrates [@romero]. Beyond the $-q\ln q$ term previously mentioned, we find that the next-to-leading contribution follows Berreman’s scaling with the wavenumber $q$. The origin of this term is twofold: the deviations of the nematic director field with respect to the distortions imposed by the presence of the nucleated disclination lines, and the defect core contributions. We estimate both contributions, finding a fairly good agreement with the reported values in Ref. [@romero].
The paper is organized as follows. The problem is set up in Section II, where we identify the different contributions to the elastic contribution to the surface free-energy density. Sections III and IV are devoted to the estimation of these contributions, and the obtained results will be discussed in Section IV. We will end up with the conclusions in Section V.
The model
=========
We consider a nematic liquid in contact with a sawtooth substrate characterized by the angle $\alpha$ and the length side $L$ (see Fig. \[fig1\]). The substrate favours homeotropic anchoring of the molecules. We assume traslational symmetry along $z$ axis and a periodic distribution of wedges and cusps along the $x$ axis. Under these conditions, the nematic director field ${\mathbf{n}}(\mathbf{r})$ shows only in-plane distortions [@romero], so it can be parametrized by using the angle $\theta$ between the local director and the $y$ axis, yielding ${\mathbf{n}}(\mathbf{r})=(-\sin\theta(\mathbf{r}),\cos\theta(\mathbf{r}),0)$. Far from the substrate, the bulk nematic phase orients homogeneously along either the $x$ axis (perpendicular texture $N^\perp$) or the $y$ axis (parallel texture $N^\parallel$). The nematic order of the system can be locally represented by a traceless symmetric order parameter second-rank tensor ${\mathbf{Q}}$, with Cartesian components ${Q_{ij}}= \frac 3 2 S[n_in_j - \frac 1 3
\delta_{ij}]+ \frac 1 2 B [l_il_j - m_i m_j]$, where $S$ is the nematic order parameter, which measures the orientational ordering along the nematic director, and $B$ the biaxiality parameter, which measures the ordering of the molecules on the orientations perpendicular to $\mathbf{n}$, characterized by the eigenvectors $\mathbf{l}$ and $\mathbf{m}$.
As in previous works [@romero; @patricio; @patricio2], the system will be described within the Landau-de Gennes (LdG) framework. The order parameter tensor profile is obtained by minimizing the surface free-energy density functional $$\begin{aligned}
&& {\cal F}=
\frac{1}{\lambda}\int_{-\lambda/2}^{\lambda/2} {\mathrm{d}}x
\int_{|x|\tan\alpha}^\infty {\mathrm{d}}y
\Bigg[\frac{3-2\tau}{4\tau-3}\left(\frac{2\operatorname{Tr}\tilde{\mathbf{Q}}^2}{3}-1
\right) \nonumber\\&& - \frac{2}{4\tau-3}\left(\frac{4}{3} \operatorname{Tr}\tilde{\mathbf{Q}}^3 -1\right)+ \frac{\tau}{4\tau-3}\left(\frac{4}{9}
[\operatorname{Tr}\tilde{\mathbf{Q}}^2]^2 -1\right) \nonumber\\
&&+ \frac {1}
{3+2\kappa}[ \partial_k \tilde Q_{ij} \partial_k \tilde Q_{ij} +
\kappa \partial_j
\tilde Q_{ij}
\partial_k \tilde Q_{ik}] \Bigg]\nonumber \\ &&
-\frac{1}{\lambda}\int_{-\lambda/2}^{\lambda/2}{\mathrm{d}}x \frac 2 3 w \operatorname{Tr}[\tilde{\mathbf{Q}}\cdot\tilde{\mathbf{Q}}_{\mathrm{surf}}]
\Big |_{y=|x|\tan\alpha}
\label{free_energy}\end{aligned}$$ where $\lambda\equiv 2L\cos\alpha$ is the substrate periodicity wavelength, $\tilde{\mathbf{Q}}={\mathbf{Q}}/S_b$, where $S_b=S_b(T)$ is the bulk value of the nematic order parameter at the temperature $T$, and $\tau=S_b(T)/S_b(T_{NI})$ is the ratio between the nematic order parameter at the temperature $T$ and the nematic-isotropic transition temperature $T_{NI}$. The positions are measured in units of $\sqrt{2}\xi$, where $\xi$ is the nematic correlation length along the local nematic director axis. Finally $\kappa$ is the ratio between the relevant elastic parameters ($\kappa>-3/2$) and $w$ the (reduced) anchoring strength. Homeotropic alignment of the nematic is favoured by setting $\tilde{\mathbf{Q}}_{\mathrm{surf}}=(3\boldsymbol{\nu}\otimes
\boldsymbol{\nu}-1)/2$, being $\boldsymbol{\nu}=((x/|x|)\sin\alpha,-\cos\alpha,0)$ the outwards normal vector to the substrate. The global minimum of the functional Eq. (\[free\_energy\]) yields the mean-field equilibrium surface free-energy density, $f$, in appropiated units. The contribution due to the elastic deformations induced by the substrate structure, ${f_{\mathrm{elastic}}}$, can be obtained as ${f_{\mathrm{elastic}}}=f-r\sigma_{NW}(w)$, where the surface roughness is $r=1/\cos\alpha$ and $\sigma_{NW}(w)$ is the nematic-flat substrate interfacial tension. In a similar way as it was obtained at nematic isotropic coexistence in Ref. [@patricio4], the interfacial tension for the LdG model has the expression: $$\begin{aligned}
&&\sigma{_{\mathrm{NW}}}=
\frac{\sqrt{2}\sqrt{(\tau-1)(\tilde{S}(0)+1)^2+\tilde{S}(0)^2}}
{6\tau^2\sqrt{4\tau-3}}
\nonumber\\ &&\times
\left(2\tau^2\tilde{S}(0)^2-\tau(2\tau+1)\tilde{S}(0)-4\tau^2+7\tau-3
\right)\nonumber\\
&&-\frac{\sqrt{2}\sqrt{4\tau-3}(-4\tau^2+6\tau-3)}{6\tau^2\sqrt{4\tau-3}}
\nonumber\\
&&-\sqrt{2}\frac{1-3\tau+2\tau^2}{2\tau^{5/2}\sqrt{4\tau-3}}\nonumber\\
&&\times\ln\left(\frac{\tau\tilde{S}(0)+
\tau-1+\sqrt{\tau}\sqrt{(\tau-1)(\tilde{S}(0)+1)^2+\tilde{S}(0)^2}
}{2\tau-1+\sqrt{\tau}
\sqrt{4\tau-3}}\right)\nonumber\\
&&-w\tilde{S}(0)
\label{sigmanw}\end{aligned}$$ where $\tilde{S}(0)=S(0)/S_b$ and $S(0)$ is the nematic order parameter at the substrate. The value of $\tilde{S}(0)$ is obtained as the largest solution of the equation: $$(3-2\tau)\tilde{S}^2(0)-2\tilde{S}^3(0)+\tau\tilde{S}^4(0)=1-\tau+
\frac{4\tau-3}{2}w^2
\label{s0}$$ At coexistence $\tau=1$, and Eq. (\[sigmanw\]) reduces to: $$\sigma{_{\mathrm{NW}}}=\frac
{\sqrt{2}(2\tilde{S}(0)+1)(\tilde{S}(0)-1)^2}{6}-w\tilde{S}(0)
\label{sigmanwcoex}$$ where $\tilde{S}(0)=(1+\sqrt{1+2\sqrt{2}w})/2$.
A systematic study of this system via full minimization of the LdG functional was done in Ref. [@romero]. For this purporse, the functional Eq. (\[free\_energy\]) was numerically minimized by using a conjugate-gradient method. The numerical discretization of the continuum problem is performed with a finite element method [@zienkiewicz] combined with adaptive meshing in order to resolve the different length scales that may emerge in the problem [@patricio3]. It was found that the $N^\perp$ texture has lower free energy if $\alpha<\pi/4$ owing to lesser distortion. Conversely, the $N^\parallel$ texture has lower free energy for $\alpha>\pi/4$, in agreement with earlier predictions [@barbero1; @barbero2; @barbero3; @poniewierski]. For large $wL$, strong anchoring conditions are effectively achieved, leading to the nucleation of disclination lines characterized by non-half-integer winding numbers along the ridges and wedges of the substrate [@barbero1; @barbero2; @barbero3; @poniewierski; @romero]. As a consequence, the elastic contribution to the surface free-energy density has the following scaling [@romero] $${f_{\mathrm{elastic}}}\approx
- \frac{\mathcal{K}(\alpha)}{2\pi}q\ln \frac{q\cos\alpha}{\pi} +
\frac{q}{2\pi}B(\alpha,w)
\label{scaling}$$ where $q=2\pi/\lambda=\pi/L\cos\alpha$ is the substrate periodicity wavenumber, $\mathcal{K}(\alpha)$ is defined as: $$\mathcal{K}(\alpha)=\begin{cases}
\frac{K\pi \alpha^2}{\left(\frac{\pi}{2}
\right)^2-\alpha^2} & N^\perp\ \textrm{texture}\\
\\
K\pi\frac{\frac{\pi}{2}-\alpha}{\frac{\pi}{2}
+\alpha} & N^\parallel\ \textrm{texture}\\
\end{cases}
\label{defkalpha}$$ being $K=(9/2)(2+\kappa)/(3+2\kappa)$ the reduced bulk elastic constant associated to bend and splay distortions. From the numerical results, the function $B$ is found to depend on the substrate rougness (i.e. $\alpha$) and nematic texture, as well as the anchoring $w$, but asymptotically *not* on $L$ for large $wL$. However, as $L$ increases, the complete minimization becomes very time-consuming. On the other hand, we do not get information about the origin of $B(\alpha,w)$.
In this paper we will introduce an alternative way to obtain ${f_{\mathrm{elastic}}}$ from the functional Eq. (\[free\_energy\]). We divide the minimization domain into three regions (see Fig. \[fig1\]): most of the domain will correspond to the region $I$, formed by the points which are far enough from the substrate. The neighbourhood of the substrate will be split into two regions: the region $II$, formed by the union of the circular sections of radii $\xi<R_c\ll L$ centered at each wedge and apex; and region $III$, which are the points which are at a distance smaller than $\eta \sim \xi$ to the substrate, but at distances larger than $R_c$ from any substrate ridge. Our hypothesis is that, for large $wL$, the minimization of the surface free-energy functional restricted to each region (subject to appropriated boundary conditions), gives a good approximation to the complete minimization of ${\mathcal{F}}$. On the other hand, we anticipate that this analysis will give us some insight in the different contributions to $B(\alpha,w)$.
We start with the minimization of region $III$. As it was argued in Ref. [@romero], large $wL$ leads to strong anchoring conditions. So, in order to minimize the surface free-energy density, we impose to the angle field $\theta(\mathbf{r})$ to be constant along its boundary, so the nematic director field is homogeneous and equal to the normal to the substrate. Consequently, the minimization of the free-energy functional in this region will lead to a homogeneous director field normal to the substrate, although the nematic order parameter $S$ at each point will depend on its distance to the substrate. We impose the following fixed boundary conditions for $S$: $S=S_b$ at the boundary between regions $I$ and $III$, and the equilibrium nematic order parameter profile for the flat wall case at the boundary between regions $II$ and $III$. Assuming that $\eta$ is large enough, this situation is completely equivalent to the flat case, so the minimum value of the surface free-energy density in this region, $f_{III}$, will be $\sigma_{NW}(1-2R_c/L)/\cos\alpha$, with corrections of order of $\exp(-\eta/\xi)$. Next Sections will be devoted to the evaluation of the minimum values of the surface free-energy functionals at the remaining regions, $f_I$ and $f_{II}$.
Evaluation of $f_I$
===================
The variations of $S$ are restricted to the neighborhood of the substrate of a width typically of order of $\xi$, and inside the defect cores. So, in region $I$, $S$ takes the bulk value $S_b$ everywhere [@romero]. Thus, the surface free-energy functional to minimize in region $I$ reduces to a Frank-Oseen functional: $$\mathcal{F}_I\approx
\frac{K}{2\lambda}\int_{I} {\mathrm{d}}x {\mathrm{d}}y
|\boldsymbol \nabla \theta|^2
\label{mfofreeenergy}$$ where the integration is restricted to region $I$, $K$ is the reduced elastic constant, and $\theta$ is the orientation field. The minimization of Eq. (\[mfofreeenergy\]) yields to the Laplace equation for $\theta$, $\nabla^2\theta(\mathbf{r})=0$. In the far field, we impose Dirichlet boundary conditions $\lim_{y\to \infty} \theta(\mathbf{r})=
\alpha_\infty$, where $\alpha_\infty=0$ for the $N^\perp$ texture and $\alpha_\infty=\pi/2$ for the $N^\parallel$ texture. Along the contours $x=
\pm \lambda/2$, we should impose periodic boundary conditions. However, we impose instead Dirichlet boundary conditions $\theta=\alpha_\infty$ along these contours, as we know from the full LdG model minimization that these are the conditions satisfied by the mean-field solution [@romero]. Finally, we assume strong anchoring conditions along the boundary between regions $I$ and $III$: $\theta(x,y=x^2\tan\alpha/|x|,z)=\alpha_\infty+(x/|x|)
(\alpha-\alpha_\infty)$. As discussed above, this condition will be accurate if $wL\gg 1$. Their contribution to the free-energy density, $f_I$, comes from a contour integration of the mean-field solution via [@romero]: $$\begin{aligned}
f_I &=&
\frac{K(\alpha-\alpha_\infty)}{\lambda} \int_{C_1} \boldsymbol{\nu} \cdot
\boldsymbol{\nabla} \theta {\mathrm{d}}\mathbf{s}
\label{contour_integral}\end{aligned}$$ where $C_1$ is the contour parallel to the boundary between regions $I$ and $III$ between a wedge and apex (see Fig. \[fig1\]).
As argued in Refs. [@barbero1; @barbero2; @barbero3; @romero], the presence of cusps in the substrate induces the formation of disclination lines, which are the responsible of the non-Berreman scaling of the elastic contribution to the surface free-energy density. In this Section, we will complete that analysis, evaluating the next-to-leading contribution.
Singular contribution
---------------------
The solution $\theta(\mathbf{r})$ to the Laplace equation subject to the boundary conditions mentioned above can be split into two terms: a singular contribution $\theta_s(\mathbf{r})$ due to the periodic array of disclination lines nucleated at the ridges of the substrate, and a non-singular contribution $\theta_{ns}(\mathbf{r})$. A representation of the singular contribution for each texture is given by [@romero]: $$\begin{aligned}
\theta_s^\perp&=&\frac{-\alpha}{\frac{\pi}{2}-\alpha}\Bigg(-\arctan
\left[\tan\frac{qx}{2}\coth\frac{qy}{2}\right]\nonumber\\
&+&\arctan
\left[\tan\frac{qx}{2}\right]\Bigg)
\label{thetasingperp}\\
&+&\frac{\alpha}{\frac{\pi}{2}+\alpha}\Bigg(-\arctan
\left[\tan\frac{qx}{2}
\tanh\frac{q(y-L\sin\alpha)}{2}\right]
\nonumber\\
&+&\arctan\left[\tan\frac{qx}{2}\right]\Bigg)
\nonumber\\
\theta_s^\parallel&=& \frac{\pi}{2}+
\Bigg(-\arctan
\left[\tan\frac{qx}{2}\coth\frac{qy}{2}\right]\nonumber\\
&+&\arctan
\left[\tan\frac{qx}{2}\right]\Bigg)
\label{thetasingpar}\\
&-&\frac{\frac{\pi}{2}-\alpha}{\frac{\pi}{2}+\alpha}\Bigg(-\arctan
\left[\tan\frac{qx}{2}
\tanh\frac{q(y-L\sin\alpha)}{2}\right]
\nonumber\\
&+&\arctan\left[\tan\frac{qx}{2}\right]\Bigg)
\nonumber\end{aligned}$$ Their contribution to the free-energy density, $f_I^s$, comes from a contour integration of these solutions: $$\begin{aligned}
f_I^s &=&
\frac{K(\alpha-\alpha_\infty)}{\lambda} \int_{C_1} \boldsymbol{\nu} \cdot
\boldsymbol{\nabla} \theta_{s} {\mathrm{d}}\mathbf{s}
\label{contour_integral2}\end{aligned}$$ In Ref. [@romero] it was estimated the large-$L$ behaviour of Eq. (\[contour\_integral2\]), leading to the non-Berreman term ${\mathcal{K}}
(\alpha)q\ln(L/R_c)$ term. However, after some algebra it is possible to obtain explicitely $f_I^s$ from the solutions Eqs. (\[thetasingperp\]) and (\[thetasingpar\]) via Eq.(\[contour\_integral2\]) as: $$f_I^s=\frac{\mathcal{K}(\alpha)q}{2\pi}
\left(-\ln qR_c +\ln\left[2\cosh\left(\frac{\pi}
{2}\tan\alpha\right)\right]-\alpha\tan\alpha \right)
\label{fIsing}$$ where ${\mathcal{K}}(\alpha)$ depends on the texture and substrate geometry as Eq. (\[defkalpha\]) and we neglected terms of order $q^3 (R_c)^2$ and $q\eta$. As the dependence on the nematic texture comes from $\mathcal{K}(\alpha)$, $f_I^s$ will be minimum for the $N^\perp$ (resp. $N^\parallel$) texture for $\alpha<\pi/4$ (resp. $\alpha>\pi/4$).
Two remarks are pertinent at this point. First, we note that, although $f_I^s$ may depend on the explicit representation of the singular solution, the leading non-Berreman contribution is independent on this representation. The reason for this fact is that this leading contribution arises from the behaviour close to the wedges and apexes of $\theta_s$, which must asymptotically approach to the corresponding to an isolated disclination line [@romero]. Secondly, the next-to-leading contribution gives a first contribution to $B(\alpha,w)$, which we will denote as $B_{I,s}(\alpha)$, given by the expression: $$B_{I,s}(\alpha)={\mathcal{K}}(\alpha)\left(\ln\left[\frac{2}{\pi}
\cosh\left(\frac{\pi}
{2}\tan\alpha\right)\cos\alpha\right]-\alpha\tan\alpha\right)
\label{defbIs}$$
Non-singular contribution
-------------------------
The non-singular part of the mean-field solution, $\theta_{ns}$, is solution of the Laplace equation $\nabla^2 \theta_{ns}=0$, subject to the boundary conditions $\theta_{ns}=0$ in the far field, i.e. $y\to \infty$ and along the boundaries $x=\pm \lambda/2$. On the other hand, $\theta_{ns}=
\alpha-\theta_s$ along $C_1$. We do not have an explicit expression for $\theta_{ns}$ (however, there is an implicit expression via a Schwarz-Christoffel transformation [@barbero1], see below), so we have to resort to numerical methods. We have used two different techniques: a finite element method, analogous to the method outlined in the Section II to solve the LdG model, but minimizing instead the functional Eq. (\[mfofreeenergy\]) subject to the boundary conditions for $\theta_{ns}$ mentioned above; and as an alternative, the boundary element method [@brebbia; @katsikadelis]. In this technique the solution $\theta_{ns}$ inside the region $I$ can be written as: $$\begin{aligned}
&&\theta_{ns}(\mathbf{r})=\oint_{\partial I} {\mathrm{d}}\mathbf{s} \Bigg(
[\boldsymbol{\nu}(\mathbf{s})\cdot \boldsymbol{\nabla}_{\mathbf{s}}
\theta_{ns}(\mathbf{s})]
G(\mathbf{s},\mathbf{r})\nonumber\\
&&-\theta_{ns}(\mathbf{s})
\left[\boldsymbol{\nu}(\mathbf{s})\cdot \boldsymbol{\nabla}_{\mathbf{s}}
G(\mathbf{s},\mathbf{r})\right]\Bigg)
\label{bem1}\end{aligned}$$ where the contour integral over the boundary $\partial I$ of region $I$ is counter-clockwise, $\boldsymbol{\nu}(\mathbf{s})$ is the outwards normal to the boundary at $\mathbf{s}$ and $G(\mathbf{s},\mathbf{r})$ is the fundamental solution of the Laplace equation $G(\mathbf{s},\mathbf{r})=-\ln\left(|\mathbf{s}-\mathbf{r}|\right)/2\pi$. As we impose Dirichlet boundary conditions, the second term in the right-hand side of Eq. (\[bem1\]) is known. On the other hand, the normal derivative of $\theta_{ns}$ at the boundary is obtained by solving the integral equation [@brebbia; @katsikadelis]: $$\begin{aligned}
&&\oint_{\partial I} {\mathrm{d}}\mathbf{s}
[\boldsymbol{\nu}(\mathbf{s})\cdot \boldsymbol{\nabla}_{\mathbf{s}}
\theta_{ns}(\mathbf{s})]
G(\mathbf{s},\mathbf{r})=\frac{\theta_{ns}(\mathbf{r})}{2}\nonumber\\
&&+\oint_{\partial I} {\mathrm{d}}\mathbf{s} \theta_{ns}(\mathbf{s})
\left[\boldsymbol{\nu}(\mathbf{s})\cdot \boldsymbol{\nabla}_{\mathbf{s}}
G(\mathbf{s},\mathbf{r})\right]
\label{bem2}\end{aligned}$$ where now $\mathbf{r}\in \partial I$. In order to solve Eq. (\[bem2\]), we discretize the boundary in a set of straight segments (the boundary elements). We use the constant boundary element approach [@katsikadelis], so we assume that both $\theta_{ns}$ and its normal derivative are constants along each boundary element. Introducing this approximation to Eq. (\[bem2\]), we obtain a set of linear algebraic equations for the normal derivatives of $\theta_{ns}$. Once we solve this set of equations, and introducing the same approximation in Eq. (\[bem1\]) we obtain the non-singular orientational field $\theta_{ns}$ inside region $I$.
Once evaluated $\theta_{ns}$, its contribution to the surface free-energy density can be obtained from: $$\begin{aligned}
&& f_I^{ns} =
\frac{K(\alpha-\alpha_\infty)}{\lambda} \int_{C_1} \boldsymbol{\nu} \cdot
\boldsymbol{\nabla} \theta_{ns} {\mathrm{d}}\mathbf{s}
\label{nsfreeenergy1}\\
&& =\frac{K}{2\lambda}
\int_I {\mathrm{d}}\mathbf{r} |\boldsymbol{\nabla} \theta_{ns}|^2
+\frac{K}{\lambda}
\int_{C_1} (\alpha-\theta_s(\mathbf{s}))\boldsymbol{\nu} \cdot
\boldsymbol{\nabla} \theta_{s} {\mathrm{d}}\mathbf{s}
\label{nsfreeenergy2}\end{aligned}$$ where the first result is more appropriated for the boundary element technique, while the second is more appropriate for the finite element method (note that the last term in the second result can be evaluated numerically with high accuracy by standard methods as we know analytically $\theta_s(\mathbf{r})$).
The numerical minimization is performed in the cell shown in Fig. \[fig2\]. In order to minimize finite-size effects, the cell height $H$ is taken to be at least four times the value of $L$ (note that $\theta_s$ decays exponentially to $\alpha_\infty$ for $y\gg 2L\cos\alpha/\pi$). We checked that the value of $\lambda f_I^{ns}$ is independent of $\lambda$, as expected, so the non-singular contribution to $B(\alpha,w)$, $B_{I,ns}(\alpha)$, is related to $f_I^{ns}$ via $B_{I,ns}(\alpha)=\lambda f_I^{ns}$. The numerical results are shown for both the $N^\perp$ and $N^\parallel$ textures in Fig. \[fig3\]. The agreement between the results obtained from the finite element method and boundary element method is excellent, as for the evaluation of $\theta_{ns}
(\mathbf{r})$, which takes non-negligible values only above the contour $C_1$ and vanishes close to the wedges and apexes (see Fig. \[fig4\]). Our results show that the non-singular contribution corresponding to the $N^\perp$ (resp. $N^\parallel$) texture is smaller than the contribution associated to the $N^\parallel$ (resp. $N^\perp$) texture for $\alpha<\pi/4$ (resp. $\alpha>\pi/4$).
Exact evaluation of $f_I$
-------------------------
In the previous paragraphs we have outlined how to obtain $\theta(\mathbf{r})=\theta_{s}(\mathbf{r})+\theta_{ns}(\mathbf{r})$, and from that, to obtain $f_I=f_{I}^{s}+f_{I}^{ns}$, as well as $B_I(\alpha)=
B_{I,s}(\alpha)+B_{I,ns}(\alpha)$. However, it is possible to obtain $f_I$ directly without knowing the explicit form of $\theta(\mathbf{r})$ by using a Schwarz-Christoffel transformation [@barbero1; @davidson2]: $$\begin{aligned}
&&z=\int {\mathrm{d}}\bar{\zeta} \frac{C}{(\tilde{\zeta}+1)^{1/2-\alpha/\pi}
\tilde{\zeta}^{2\alpha/\pi}(\tilde{\zeta}-1)^{1/2-\alpha/\pi}}
\nonumber\\
&&
=C'\zeta^{1-
\frac{2\alpha}{\pi}} {_2F_1}\left(\frac{1}{2}-\frac{\alpha}{\pi},\frac{1}{2}
-\frac{\alpha}{\pi},\frac{3}{2}-\frac{\alpha}{\pi},\zeta^2\right)+C''
\label{sct}\end{aligned}$$ where $z=x+iy$, ${_2F_1}(a,b,c,z)$ is the Gauss hypergeometric function and $C$, $C'$ and $C''$ complex constants. This conformal transformation maps the minimization cell for $H\to \infty$ into the upper half $\zeta$-plane (see Fig. \[fig2\]), transforming the origin into the origin and the edges $z=\pm L\cos \alpha+iL\sin\alpha$ into $\zeta=\pm 1$, respectively. These conditions fix the values of $C'$ and $C''$, so the Schwarz-Christoffel transformation finally reads: $$\begin{aligned}
&&z=\frac{L e^{i\alpha}}{\Gamma\left(\frac{3}{2}-\frac{\alpha}{\pi}\right)
\Gamma\left(\frac{1}{2}+\frac{\alpha}{\pi}\right)}\zeta^{1-
\frac{2\alpha}{\pi}}\nonumber\\&&\times
{_2F_1}\left(\frac{1}{2}-\frac{\alpha}{\pi},\frac{1}{2}
-\frac{\alpha}{\pi},\frac{3}{2}-\frac{\alpha}{\pi},\zeta^2\right)
\label{sct2}\end{aligned}$$ where $\Gamma(x)$ is the gamma function. Eq. (\[sct2\]) can be formally inverted, so $\zeta=\zeta(z/L;\alpha)=x'(x/L,y/L;\alpha)+iy'(x/L,y/L;\alpha)$. We will consider the limit $\eta/L\to 0$, and $R_c/L$ small but finite. In this approach, the boundary of zone $II$ becomes under the Schwarz-Christoffel transformation the real axis in the $\zeta$-plane, rounded around $\zeta=0$ and $\pm 1$. The expansion of Eq. (\[sct2\]) around these values show that the circles of radii $R_c$ around the origin and the edges $z=\pm L\cos \alpha + iL\sin \alpha$ map into circles (up to corrections of order of $(R_c/L)^2$) of radii $\epsilon_1$ for $\zeta=\pm 1$ and $\epsilon_2$ for $\zeta=0$ given by: $$\begin{aligned}
&&\epsilon_1=\frac{1}{2}\left(\frac{1+\frac{2\alpha}{\pi}}
{1-\frac{2\alpha}{\pi}}\right)^{\frac{1}{\frac{1}{2}+\frac{\alpha}{\pi}}}
\left(\frac{R_c}{L}\Gamma\left[\frac{3}{2}-\frac{\alpha}{\pi}
\right]\Gamma\left[\frac{1}{2}+\frac{\alpha}{\pi}\right]\right)^{\frac{1}
{\frac{1}{2}+\frac{\alpha}{\pi}}}\nonumber
\\
&&\epsilon_2=\left(\frac{R_c}{L}\Gamma\left[\frac{3}{2}-\frac{\alpha}{\pi}
\right]\Gamma\left[\frac{1}{2}+\frac{\alpha}{\pi}\right]\right)^{\frac{1}{1-
\frac{2\alpha}{\pi}}}\label{epsilon12}\end{aligned}$$ As the Schwarz-Christoffel transformation is conformal, and $\theta(\mathbf{r})$ is harmonic, we are going to find the solution to the Laplace equation in the half $\zeta$-plane, $\tilde\theta$, subject to the boundary conditions $\tilde\theta=\alpha_\infty$ for $|x'|>
1+\epsilon_1$, and $\tilde\theta=\alpha_\infty+(x'/|x'|)(\alpha-\alpha_\infty)$ for $\epsilon_2<|x'|<1-\epsilon_1$. The solution $\tilde\theta(x',y')$ in the image of the region $I$ on the $\zeta$-plane, $I'$, is given by [@barbero1]: $$\begin{aligned}
&&\tilde\theta(x',y')=\alpha_\infty+\frac{\alpha-\alpha_\infty}{\pi}\arctan
\frac{y'}{x'-1}\nonumber\\
&&-\frac{2(\alpha-\alpha_\infty)}{\pi}\arctan\frac{y'}{x'}+
\frac{\alpha-\alpha_\infty}{\pi}\arctan
\frac{y'}{x'+1}
\label{tildetheta}\end{aligned}$$ From the solution $\tilde\theta$, we can obtain $\theta(x,y)=\tilde\theta[x'(x/L,y/L;\alpha),y'(x/L,y/L;\alpha)]$. We note that we do not have an explicit expression for $x'$ and $y'$ as functions of $x$ and $y$, so we cannot give an analytic expression for $\theta(x,y)$. However, we can evaluate exactly $f_I$ since: $$\begin{aligned}
&&f_I=\frac{K}{2\lambda}\int_{I} {\mathrm{d}}x {\mathrm{d}}y
|\boldsymbol \nabla \theta|^2=\frac{K}{2\lambda}\int_{I'} {\mathrm{d}}x' {\mathrm{d}}y'
|\boldsymbol \nabla' \tilde \theta|^2
\nonumber\\
&&=\frac{K}{2\lambda}
\int_{\cal B} \tilde \theta (\boldsymbol \nu' \cdot \boldsymbol
\nabla' \tilde \theta) {\mathrm{d}}\mathbf{s}'
\label{mfofreeenergy2}\end{aligned}$$ where $\cal B$ is the image in the $\zeta$-plane of the boundary of the zone $I$. After an straightforward calculation, we get the expressions for $f_I$ and $B_I(\alpha)$: $$\begin{aligned}
&&f_I=\frac{q\mathcal{K}(\alpha)}{2\pi}\Bigg[-\ln \frac{qR_c\cos\alpha}{\pi}
\nonumber\\&&-
\ln\left(\Gamma\left[\frac{3}{2}-\frac{\alpha}{\pi}
\right]\Gamma\left[\frac{1}{2}+\frac{\alpha}{\pi}\right]\right)
\nonumber\\&&-
\left(\frac{1}{2}-\frac{\alpha}{\pi}\right)\ln\left(\frac{\frac{\pi}{2}+\alpha}
{\frac{\pi}{2}-\alpha}\right)\Bigg]
\label{theoreticalfI}\\
&&B_I(\alpha)=-\mathcal{K}(\alpha)\Bigg[\left(\frac{1}{2}-\frac{\alpha}{\pi}\right)\ln\left(\frac{\frac{\pi}{2}+\alpha}
{\frac{\pi}{2}-\alpha}\right)\nonumber\\
&&+
\ln\left(\Gamma\left[\frac{3}{2}-\frac{\alpha}{\pi}
\right]\Gamma\left[\frac{1}{2}+\frac{\alpha}{\pi}\right]\right)\Bigg]
\label{theoreticalfI2}\end{aligned}$$ Fig. \[fig5\] shows an excellent agreement between the theoretical prediction for $B_I(\alpha)$ from Eq. (\[theoreticalfI2\]) and the results obtained in the previous subsections, except close to $\alpha=\pi/2$. The latter may be due to numerical uncertainties in the evaluation of $B_{I,ns}$, since either it diverges ($N^\perp$ texture) or vanishes ($N^\parallel$ texture) in that limit. As happened for $B_{I,s}$, the dependence on the texture comes from ${\cal K}(\alpha)$, leading to the same conclusions about the relative stability of the nematic textures with the angle $\alpha$, in agreement with the results obtained in Ref. [@barbero1].
Evaluation of $f_{II}$
======================
In Ref. [@romero] it was observed that there is a dependence on $w$ of the next-to-leading contribution to the surface free-energy density. However, in the previous Section we have shown that the contribution from region $I$ only depends on $\alpha$. So, we anticipate that this dependence comes from region $II$, where inhomogoneities of both the nematic order parameter $S$ and orientational $\theta$ fields are observed.
The free energy of region $II$ can be evaluated as the sum of the contributions of the regions around each cusp (either wedge or apex). In each of these regions and if $R_c\ll L$, we anticipate that the orientational field far from the cusp behaves asymptotically as that of the isolated disclination line which has been nucleated at the substrate wedge or apex: $\theta \sim I
\phi$, where $I$ is the topological charge of the disclination line and $(r,\phi)$ are the polar coordinates taking as origin the cusp. The values of $I$ are fixed by the (strong) anchoring conditions on the substrate and depend on the nematic texture: $I_1=-\alpha/(\pi/2-\alpha)$ (resp. $I_1=+1$) at the wedge bottom and $I_2=\alpha/(\pi/2+\alpha)$ (resp. $-(\pi/2-\alpha)/(\pi/2+\alpha)$) at the apexes for the $N^\perp$ (resp. $N^\parallel$) [@barbero1; @romero]. However, $S$ decreases from the bulk value far from the origin and vanishes as $r\to 0$ in order to remove the free-energy singularity associated to the defect core. Although we cannot solve analytically this problem, we can estimate using an ansatz the free energy associated to this combined distortion of $\theta$ and $S$, as shown in the Appendix. However, we resort here to a full minimization of the LdG model Eq. (\[free\_energy\]) restricted to region $II$. At $r=R_c$, we impose Dirichlet boundary conditions to ${\mathbf{Q}}$, where $S$ takes the value corresponding to the order-parameter profile corresponding to a planar wall at the distance between the boundary point and the closest substrate, the biaxiality parameter $B=0$ and $\theta=\alpha_\infty+I(\phi-\pi/2)$. Alternatively we used free boundary conditions $\boldsymbol \nu \cdot \boldsymbol \nabla {\mathbf{Q}}=0$, leading to similar results. Fig. \[fig6\] shows typical textures obtained after minimization. At distances $r$ larger than a few correlation lengths, $S$ decays to the bulk value except in the neighbourhood of the substrates, where takes approximately the value corresponding to the flat substrate profile. On the other hand, the orientational field deforms continuously in order to satisfy anchoring conditions on each side of the wedge or the apex. So, we anticipate that the contribution to the surface free-energy density $f_{II}$ will scale with $R_c$ as: $$\lambda f_{II}=4\sigma_{nw} R_c + {\cal K}(\alpha) \ln R_c + B_{II}
\label{analisis1}$$ The first contribution arises from the inhomogeneities of $S$ close to the substrates, and the second one from the asymptotic behaviour of $\theta$ for $r>1$: by using the Frank-Oseen functional Eq. (\[mfofreeenergy\]), and taking into account that $\boldsymbol \nabla \theta=I\mathbf{u}_\phi/r$ (where $\mathbf{u}_\phi$ is the azimuthal unit vector), we find that the free-energy contribution is proportional to $(K/2) I^2\Delta
\phi \ln R_c$, where $\Delta\phi$ is the opening angle of the wedge or apex. From this result we get the second contribution in Eq. (\[analisis1\]) by noting that ${\cal K}(\alpha)=K I_1^2(\pi-2\alpha)/2+ KI_2^2(\pi+2\alpha)/2$. The remaining contribution will give the core free energy per unit length associated to the disclination lines.
Expression (\[analisis1\]) can be used to extract $B_{II}$ from the numerical minimization. So, fixing the value of $\alpha$, and for each $w$ (between 0 and 1.5) we considered a range of values of $R_c$ between $10$ and $90$. The minimization was performed by the mesh-adaptive finite-element method used previously [@patricio3], taking $\tau=1$ and $\kappa=2$ (this choice is motivated to compare with results reported in the literature [@patricio; @romero]). After substrating $4\sigma_{nw}R_c$ to $\lambda f_{II}$, the numerical results clearly show a logarithmic dependence on $R_c$, with a slope approximately equal to ${\cal K}(\alpha)$ (see left panel in Fig. \[fig7\]). Next step was to substract the subdominant contribution ${\cal K}(\alpha)\ln R_{\infty}$. Now the numerical results are nearly independent of $R_c$ (see right panel in Fig. \[fig7\]). The value of $B_{II}$ is estimated as the mean value of these results, with an errorbar given by the dispersion of the numerical data around the average.
Discussion
==========
The results obtained in the previous Sections can be combined as follows: $$f=f_I+f_{II}+f_{III}=\frac{\sigma_{NW}}{\cos\alpha}+f_{elastic}
\label{finalf0}$$ where $$\begin{aligned}
&&f_{elastic}=\frac{q}{2\pi} {\cal K}(\alpha)\Bigg[-\ln
\frac{q\cos\alpha}{\pi}\nonumber\\
&&-\left(\frac{1}{2}-\frac{\alpha}{\pi}\right)\ln\left(\frac{\frac{\pi}{2}
+\alpha}{\frac{\pi}{2}-\alpha}\right)\nonumber\\
&&-
\ln\left(\Gamma\left[\frac{3}{2}-\frac{\alpha}{\pi}
\right]\Gamma\left[\frac{1}{2}+\frac{\alpha}{\pi}\right]\right)\Bigg]
\nonumber\\
&&+\frac{q}{2\pi} B_{II}(\alpha,w)
\label{finalf}\end{aligned}$$ up to corrections of order of $q^2$. Consequently: $$\begin{aligned}
&&B(\alpha,w)=B_I(\alpha)+B_{II}(\alpha,w)=\nonumber\\
&&
-{\cal K}(\alpha)\Bigg[
\left(\frac{1}{2}-\frac{\alpha}{\pi}\right)\ln\left(\frac{\frac{\pi}{2}
+\alpha}{\frac{\pi}{2}-\alpha}\right)\nonumber\\
&&+
\ln\left(\Gamma\left[\frac{3}{2}-\frac{\alpha}{\pi}
\right]\Gamma\left[\frac{1}{2}+\frac{\alpha}{\pi}\right]\right)\Bigg]
\nonumber\\
&&+ B_{II}(\alpha,w)
\label{finalb}\end{aligned}$$ This is the main result of our paper. Now we can check this prediction by comparing these results with those obtained from the full minimization of the LdG model of a nematic in contact with a sawtoothed substrate [@romero]. Fig. \[fig8\] shows the comparison between the results reported in the Ref. [@romero] ($\tau=1$, $\kappa=2$) and the calculated ones in this paper for $\alpha=\pi/6$ ($N^\perp$ texture) and $\alpha=\pi/3$ ($N^\parallel$ texture). For large $w$, the agreement is good, although our results slightly overestimate those from the full minimization. On the other hand, for small $w$ the curves obtained from the full minimisation converge towards our prediction as $L$ increases. So, our approximation is accurate even for moderate values of $L$, despite the assumptions involved in our approach. Consequently, the scheme considered in this paper is an alternative to the full minimization technique, which is quite expensive from a computational point of view, when $wL>1$.
Finally, it is interesting to note that the elastic contribution to the surface free-energy density, Eq. (\[finalf\]), can be expressed as follows: $$\lambda f_{elastic}=2\times \left(2\pi K \frac{I_1}{2} \frac{I_2}{2}
\ln\frac{\gamma(\alpha)}{L}\right)+ B_{II}(w,\alpha)
\label{felastic2}$$ where we note that ${\cal K}(\alpha)=-4 \pi K (I_1/2) (I_2/2)$, with $I_1$ and $I_2$ being the topological charges associated to the disclination lines at the wedges and apexes, respectively. The characteristic length $\gamma(\alpha)$, which absorbs the contribution to $B$ from region $I$, is defined as: $$\gamma(\alpha)=\Gamma\left[\frac{3}{2}-\frac{\alpha}{\pi}
\right]\Gamma\left[\frac{1}{2}+\frac{\alpha}{\pi}\right]
\left(\frac{\frac{\pi}{2}
+\alpha}{\frac{\pi}{2}-\alpha}\right)^{\frac{1}{2}-\frac{\alpha}{\pi}}
\label{defgamma}$$ This lengthscale decays continuously from $\pi/2$ for $\alpha\to 0$ to $1$ for $\alpha\to \pi/2$. We can understand the first contribution in Eq. (\[felastic2\]) as the interaction between the disclination line at $x=0$ with half its topological charge and the apexes at $x=\pm L\cos\alpha$, again with half their topological charges [@chaikin]. As the absolute values of the topological charges are always smaller or equal than $1$, the nucleation of disclinations at the cusps of the surface is favourable with respect to bulk disclinations (which only may have half-integer values). With this intepretation, $\gamma(\alpha)$ can be used to define an effective core radius $\gamma(\alpha)$, and the sum of the core energies is given by $B_{II}(\alpha,w)$.
Conclusions
===========
In this paper we have analyzed elastic contribution to the surface free-energy density for a nematic in contact to sawtoothed substrates in the strong anchoring regime, i.e. $wL>1$. We have extended the analysis done in Ref. [@romero]. So, in addition to the leading contribution proportional to $-q\ln q$, with $q$ being the substrate periodicity wavenumber $q$, we have characterized the next-to-leading term. This term has two contributions: one associated to the deviation of the orientational field with respect to the contribution of the array of disclination lines nucleated at the cusps of the substrate, and the core free-energy associated to them. We anticipate that our analysis can be generalized for other substrate shapes when the nematic texture presents topological defects induced by the structure, as, for example, in crenellated substrates [@nuno2]. Furthermore, our results may be used to predict accurately the location of first-order wetting transitions in nematic liquid crystals in contact to general substrates [@patricio; @patricio2; @nuno2]. However, we must note that our analysis is restricted to bulk nematics liquid crystals with in-plane deformations in presence of grooved substrates. The effect of the substrate structure in the nematic texture in partially filled configurations (i.e. with an isotropic fluid in bulk), the influence of twist nematic deformations [@patricio5] or the effect of the structure in the $z$ direction deserve further study, but this is beyond the scope of this work.
The authors wish to thank Dr. P. Patrício and Dr. N. M. Silvestre for very stimulating conversations and technical advice on the finite-element method, and Prof. M. M. Telo da Gama, Prof. L. F. Rull and Prof. A. O. Parry for enlightening discussions. We acknowledge the support from MICINN (Spain) through Grant FIS2009-09326, and Junta de Andalucía (Spain) through Grant No. P09-FQM-4938, both co-funded by EU FEDER.
*Ansatz* for the evaluation of $f_{II}$
=======================================
In order to estimate the free energy associated to the region $II$, we can make use of an ansatz. As argued in the paper, the nematic order parameter $S$ must vanish at the cusps of the substrate to avoid the divergence of the free energy, but it must converge to the bulk value far from the wedges and apexes. So, in the region around each cusp, we can suppose that $S$ depends only on the radial distance $r$ from the wedge or apex. In particular, we use the following ansatz [@tasinkevych]: $S(r)=S_b(1-e^{-\frac{r}{\beta}})$, where $S_b$ is the bulk nematic order parameter and $\beta$ is a lengthscale to be determined later. We will assume that the biaxiality parameter $B=0$ everywhere. Regarding the orientational field $\theta$, we assume that has the expression corresponding to a disclination line placed at the origin: $\theta=\alpha_\infty+I(\phi-\pi/2)$, where $I$ is the winding number associated to the disclination line, and $\alpha_\infty$ is the far-field value of the orientational field ($0$ for the $N^\perp$ texture, and $\pi/2$ for the $N^\parallel$ texture). Substituting this *ansatz* in the LdG functional Eq. (\[free\_energy\]), this expression reduces to a function of $\beta$: $$\begin{aligned}
&&\lambda f\approx\frac{13}{144}(\pi\mp2\alpha)\beta^{2}-2w(R_c-\beta)
+\frac{1}{7}\Bigg\{9I^{2}(\pi\mp2\alpha)\Big[\gamma
\nonumber\\&&+\ln\Big(\frac{R_c}{2\beta}\Big)\Big]+
\frac{11}{16}(\pi\mp2\alpha)
+\frac{1}{8}\Big(\frac{\sin[(I-1)(\pi\mp2\alpha)]}{(I-1)}\nonumber\\&&
\mp 18I\sin(2\alpha)
-3\big[\sin(-2I(\pi\mp\alpha)+(I-1)\pi)\nonumber\\
&&-\sin(-2I(\pm\alpha)+(I-1)\pi)\big]\Big)\Bigg\}
\label{free_energy_ansatz}\end{aligned}$$ where we have considered the case $\tau=1$ and $\kappa=2$ (for other values, we can get straightforwardly analogous expressions). In this expression, the upper sign corresponds to the wedge situation (so $I=I_1$), and the lower one to the apex (so $I=I_2$), and $\gamma$ is the Euler constant. Minimizing Eq. (\[free\_energy\_ansatz\]) with respect to $\beta$, we obtain the optimal value for this lengthscale: $$\beta=-\frac{72}{13}\frac{w}{(\pi\mp2\alpha)}+\sqrt{\left(\frac{72}{13}
\frac{w}{(\pi\mp2\alpha)}\right)^{2}+\frac{648}{91}I^{2}}
\label{ansatz17}$$ Substituting Eq. (\[ansatz17\]) into Eq. (\[free\_energy\_ansatz\]), we get the free energy estimate associated to each region around a cusp, and provides an upper limit to the real value of $\lambda f$.
In order to have an estimate of $B_{II}$, we substract to the obtained values of $\lambda f$ through this ansatz the surface and elastic contributions $-2w R_c$ and $(K/2) I^2 (\pi\mp 2\alpha)\ln R_c=(9/7) I^2
(\pi\mp 2\alpha)\ln R_c$, respectively. After this, the estimate of $B_{II}$ is obtained as the sum of the results obtained for the wedge and apex. Fig. \[fig9\] shows the comparison of this estimates with the results obtained from the full minimization, for $\alpha=\pi/6$ ($N^\perp$ texture) and $\alpha=\pi/3$ ($N^\parallel$) texture. They have the same order of magnitude, although our *ansatz* overestimate the value of $B_{II}$, and qualitatively the dependence of $B_{II}$ with $w$ is recovered.
[33]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
B.-W. Lee and N. A. Clark, Science [**291**]{}, 2576 (2001). J.-H. Kim, M. Yoneya and H. Yokoyama, Nature [**420**]{}, 19 (2002). S. Ferjani, Y. Choi, J. Pendery, R. G. Petschek and C. Rosenblatt, Phys. Rev. Lett. [**104**]{}, 257801 (2010). , , , , ****, ().
, , , ****, ().
, , , ****, ().
A. J. Davidson, C. V. Brown, N. J. Mottram, S. Ladak and C. R. Evans, Phys. Rev. E [**81**]{}, 051712 (2010).
C. R. Evans, A. J. Davidson, C. V. Brown and N. J. Mottram, J. Phys. D: Appl. Phys. [**43**]{}, 495105 (2010).
N. M. Silvestre, P. Patrício and M. M. Telo da Gama, Phys. Rev. E [**69**]{}, 061402 (2004).
T. Ohzono and J.-i. Fukuda, Nat. Commun. [**3**]{}, 701 (2012).
D. W. Berreman, Phys. Rew. Lett. [**28**]{}, 1683 (1972)
P.G. de Gennes and J. Prost, *The Physics of Liquid Crystals*, 2nd ed. (Oxford University Press, Oxford, 1995)
G. Barbero, Lett. Nuovo Cimento Soc. Ital. Fis. [**29**]{}, 553 (1980)
G. Barbero, Lett. Nuovo Cimento Soc. Ital. Fis. [**32**]{}, 60 (1981)
G. Barbero, Lett. Nuovo Cimento Soc. Ital. Fis. [**34**]{}, 173 (1982).
, ****, ().
, , , ****, ().
P. Patrício, M. M. Telo da Gama, and S. Dietrich, Phys. Rev. Lett. 88, 245502 (2002)
L. Harnau, S. Kondrat and A. Poniewierski, Phys. Rev. E [**72**]{}, 011701 (2005)
S. Kondrat, A. Poniewierski and L. Harnau, Liq. Cryst. [**32**]{}, 95 (2005).
L. Harnau and S. Dietrich, Europhys. Lett. [**73**]{}, 28 (2006).
L. Harnau, S. Kondrat and A. Poniewierski, Phys. Rev. E [**76**]{}, 051701 (2007).
G. Barbero, A.S. Gliozzi, M. Scalerandi and L.R. Evangelista, Phys. Rew. E [**77**]{}, 051703 (2008)
, , , , , , ****, ().
A. Poniewierski, Eur. Phys. J. E [**31**]{}, 169 (2010).
J. M. Romero-Enrique, C.-T. Pham and P. Patrício, Phys. Rev. E [**82**]{}, 011707 (2010).
J. P. Bramble, S. D. Evans, J. R. Henderson, C. Anquetil, D. J. Cleaver and N. J. Smith, Liq. Cryst. [**34**]{}, 1059 (2007).
P. Patrício, C.-T. Pham and J. M. Romero-Enrique, Eur. Phys. J. E [**26**]{}, 97 (2008).
P. Patrício, J. M. Romero-Enrique, N. M. Silvestre, N. R. Bernardino and M. M. Telo da Gama, Mol. Phys. [**109**]{}, 1067 (2011).
P. Patrício, N. M. Silvestre, C.-T. Pham and J. M. Romero-Enrique, Phys. Rev. E [**84**]{}, 021701 (2011).
O.C. Zienkiewicz and R.L. Taylor, *The Finite Element Method*, 5th ed. (Butterworth-Heineman, Oxford, 2000).
P. Patrício, M. Tasinkevych and M.M. Telo da Gama, Eur. Phys. J. E [**7**]{}, 117 (2002).
C. A. Brebbia and J. Domínguez, *Boundary elements: an introductory course* 2nd ed. (Computational Mechanics Publications, Southampton, 1992).
J.T. Katsikadelis, *Boundary Elements: Theory and Applications* (Elsevier, Amsterdam, 2002)
M. Tasinkevych, N.M. Silvestre, P. Patricio and M.M. Telo da Gama, Eur. Phys. J. E [**9**]{}, 341 (2002)
P.M. Chaikin y T.C. Lubensky, *Principles of Condensed Matter Physics*, (Cambridge University Press, Cambridge, 1997)
A. J. Davidson and N. J. Mottram, Eur. J. Appl. Math. [**23**]{}, 99 (2011).
N. M. Silvestre, Z. Eskandari, P. Patrício, J. M. Romero-Enrique and M. M. Telo da Gama, Phys. Rev. E (submitted) (2012).
![Schematic picture of the geometry of the system, characterized by the side length $L$ and the angle $\alpha$. The different regions $I$, $II$ and $III$ are outlined. See text for explanation.[]{data-label="fig1"}](fig1.eps){width=".95\columnwidth"}
![Left panel: Minimization cell for the evaluation of $f_{I,ns}$. Right panel: Mapping of the minimization cell under the Schwarz-Christoffel transformation in the limit $H\to \infty$.[]{data-label="fig2"}](fig2.eps){width="50.00000%"}
![Plot of $\lambda f_{I,ns}/K$ as a function of $\alpha$ for the $N^\perp$ and $N^\parallel$ textures, by using the finite-element method (F.E.M.) and the boundary element method (B. E. M.). []{data-label="fig3"}](fig3.eps){width=".95\columnwidth"}
![(Color online) Plot of $\theta_{ns}$ as a function of $x$ and $y$ obtained from the finite-element method for $\alpha=\pi/6$ and $L=0.25$. For the sake of clarity it is only represented in half a minimization cell $0<x<L\cos\alpha$. []{data-label="fig4"}](fig4.eps){width=".95\columnwidth"}
![Plot of $B_{I}/{\cal K}(\alpha)$ as a function of $\alpha$. The continuous broad line corresponds to the exact expression, and the symbols correspond to the estimates from the numerical data for $f_{I,ns}$ (from the boundary element method): circles for the $N^\perp$ texture and diamonds for the $N^\parallel$ texture. Inset: plot of $-B_{I}/K$ as a function of $\alpha$. The meaning of the symbols is the same as in the main panel. []{data-label="fig5"}](fig5.eps){width=".95\columnwidth"}
![(Color online) Typical textures obtained from the full minimization of the LdG function in region $II$ for $\tau=1$, $\kappa=2$, $w=1.0$, $L=90$ and $\alpha=\pi/6$ (upper panels) and $\alpha=\pi/3$ (lower panels). The color map corresponds to the reduced nematic order parameter field $\tilde S=S/S_b$, and the segments correspond to the nematic director field $\mathbf{n}$.\[fig6\]](fig6.eps){width=".95\columnwidth"}
![Left panel: representation of $\lambda f_{II}-4\sigma_{NW} R_c$ as a function of $R_c$ for $w=1$ and: $\alpha=\pi/6$ (squares) and $\alpha=\pi/3$ (circles). The straight lines correspond to the logarithmic regressions of the numerical data. Right panel: representation of $B_{II}(\alpha,w)$ as a function of $R_c$. The meaning of the symbols is the same as in the left panel.[]{data-label="fig7"}](fig7.eps){width="0.95\columnwidth"}
![(Color online) Comparison between the values of $B(\alpha,w)$ obtained from the full minimization of the LdG model [@romero] and the results obtained in the present work. Symbols correspond to $B(\alpha,w)$ obtained from the full minimization of the LdG model for different values of $L$ and $\alpha=\pi/6$ in the $N^\perp$ texture (black symbols), and $\alpha=\pi/3$ in the $N^\parallel$ texture (orange or light gray symbols). The wide blue (black) and wide red (dark gray) lines with errorbars correspond to the predictions from this work for $\alpha=\pi/6$ and $\pi/3$, respectively.[]{data-label="fig8"}](fig8.eps){width="0.95\columnwidth"}
![ Comparison between the values of $B_{II}(\alpha,w)$ obtained from the full minimization of the LdG model and the *ansatz* (see text). Symbols correspond to $B_{II}(\alpha,w)$ obtained from the full minimization of the LdG model for $\alpha=\pi/6$ in the $N^\perp$ texture (circles), and $\alpha=\pi/3$ in the $N^\parallel$ texture (up triangles). The wide continuous and dashed lines correspond to the estimates from the *ansatz* for $\alpha=\pi/6$ and $\alpha=\pi/3$, respectively.[]{data-label="fig9"}](fig9.eps){width="0.95\columnwidth"}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'How can we separate structural information from noise in large graphs? To address this fundamental question, we propose a graph summarization approach based on Szemerédi’s Regularity Lemma, a well-known result in graph theory, which roughly states that every graph can be approximated by the union of a small number of random-like bipartite graphs called “regular pairs”. Hence, the Regularity Lemma provides us with a principled way to describe the essential structure of large graphs using a small amount of data. Our paper has several contributions: (i) We present our summarization algorithm which is able to reveal the main structural patterns in large graphs. (ii) We discuss how to use our summarization framework to efficiently retrieve from a database the top-$k$ graphs that are most similar to a query graph. (iii) Finally, we evaluate the noise robustness of our approach in terms of the reconstruction error and the usefulness of the summaries in addressing the graph search task.'
address:
- 'DAIS, Ca’ Foscari University, Via Torino 155, Venice, Italy'
- 'ECLT, Ca’ Foscari University, S. Marco 2940, Venice, Italy'
author:
- Marco Fiorucci
- Francesco Pelosin
- Marcello Pelillo
bibliography:
- 'references.bib'
title: |
Separating Structure from Noise in Large Graphs\
Using the Regularity Lemma
---
Regularity Lemma ,Graph summarization ,Structural patterns ,Noise ,Randomness ,Graph similarity search 00-01,99-00
Introduction
============
Recent years are characterized by an unprecedented quantity of available network data which are produced at an astonishing rate by an heterogeneous variety of interconnected sensors and devices. This high-throughput generation calls for the development of new effective methods to store, retrieve, understand and process massive network data. To tackle this challenge, we introduce a framework for summarizing large graphs based on *Szemerédi’s Regularity Lemma*[@Szemeredi75Graphs], which roughly states that any sufficiently large graph can almost entirely be partitioned into a bounded number of random-like bipartite graphs, called *regular pairs*. The partition resulting from the Regularity Lemma gives rise to a summary, called *reduced graph*, which inherits many of the essential structural properties of the original graph [@Komlos2002; @Komlos1996].
In this paper, we posit that the Regularity Lemma can be used to summarize large graphs revealing its main structural patterns, while filtering out noise, which is common in any real-world networks. In its original form, the lemma is an existential predicate, but during the last decades various constructive algorithms have been proposed [@Alon94; @CzygrinowRödl; @Frieze1999APartition]. However, despite being polynomial in the size of the underlying graph, all these algorithms have a hidden tower-type dependence on an accuracy parameter. To overcome this limitation, in the last years we have proposed some simple heuristics that, most of the times, allowed us to construct a regular partition [@Fiorucci17; @PelilloRevealing17; @Sperotto07].
The main contribution of this paper is the introduction of a new heuristic algorithm which is characterized by an improvement of the summary quality both in terms of reconstruction error and of noise filtering. In particular, we first build the reduced graph of a graph $G$, and then we “blow-up” the reduced graph to obtain a graph $G'$, called *reconstructed graph*, which is close to $G$ in terms of the $l_p$-reconstruction error. We study the noise robustness of our approach in terms of the reconstruction error by performing an extensive series of experiments on both synthetic and real-world data. As far as the synthetic data are concerned, we generate graphs with a cluster structure, where the clusters are perturbed with different levels of noise. As far as the real-world data are concerned, we add spurious edges in accord with different noise probabilities. The aim of this series of experiments is to assess if the framework is able to separate structure from noise. In the ideal case, the distance between $G$ and $G'$ should be only due to the filtered noise.
Moreover, in the second part of the paper, we use our summarization algorithm to address the *graph search* problem defined under a similarity measure. The aim of graph search is to retrieve from a database the top-$k$ graphs that are most similar to a query graph. Since noise is common in any real-world dataset, the biggest challenge in graph search is developing efficient algorithms suited for dealing with large graphs containing noise in terms of missing and adding spurious edges. In our approach, all the graphs contained in a database are compressed off-line, while the query graph is compressed on-line. Thus, graph search can be performed on the summaries, and this allows us to speed up the search process and to reduce storage space. Finally, we evaluate the usefulness of our summaries in addressing the graph search problem by performing an extensive series of experiments. In particular, we study the quality of the answers in terms of the found top-$k$ similar graphs, and the scalability both in the size of the database and in the size of the query graphs.
#### **Related Works**
The first contribution of our paper is the introduction of a principled framework for summarizing large graphs with the aim of preserving their main structural patterns. Previous related works presented methods which mainly built summaries by grouping the vertices into subsets, such that the vertices within the same subset share some topological properties. The works in [@SCHAEFFER200727; @Newman] introduced methods for partitioning the vertices into non-overlapping clusters, so that vertices within the same cluster are more connected than vertices belonging to different clusters. A graph summary can be constructed by considering each cluster as a *supernode*, and by connecting each pair of supernodes with a *superedge* of weight equals to the sum of the cross-cluster edges. However, since graph summarization and clustering have different goals, this approach is suited only if the input graph has a strong community structure. In [@LeFreve], the summary is generated by greedily grouping vertices, such that the normalized reconstruction error between the adjacency matrix of the input graph and the adjacency matrix of the *reconstructed graph* is minimized. Since in their work they exploited heuristic algorithms, they can not give any guarantees on the quality of the summary. The work in [@Riondato2017] proposed a method of building a summary with quality guaranty by minimizing the *$l_p$-reconstruction error* between the adjacency matrix of the input graph and the adjacency matrix of the reconstructed graph. Since both approaches aim to minimize a distance measure between the input and the reconstructed graph, they are not the best choice for summarizing noisy graphs. By contrast, our goal is to develop a graph summarization algorithm which is robust against noise. For a more detailed picture on how the field has evolved previously, we refer the interested reader to the survey of Liu et al. [@Liu2017GraphSurvey].
The second contribution of our paper consists in addressing the graph search problem using the proposed summarization framework. Locating the occurrences of a query graph in a large database is a problem which has been approached in two main different ways, based on subgraph isomorphism and approximate graph matching respectively. Ullman [@Ullmann76] put one of the first milestones in subgraph isomorphism. He proposed an algorithm which decreases the computational complexity of the matching process by reducing the search space with backtracking. Recently, Carletti et al. [@Carletti17] introduced an algorithm for graph and subgraph isomorphism which scales better than Ullmann’s one. In particular, Carletti et al’s algorithm, which may be considered as the state-of-the-art in exact subgraph matching, can process graphs of size up to ten thousand nodes. However, since subgraph isomorphism is a NP-complete problem, the algorithms based on exact matching are prohibitively expensive for querying against a database which contains large graphs. Moreover, due to the noise contained in any real-world networks, it is common to mismatch two graphs which have the same structure but different levels of noise. Indeed, these contributions are focused on exact matching and, even if they proposed efficient solutions, they are not noise robust. By contrast, our goal is to develop an efficient graph search algorithm which is robust against noise. Hence, approaches based on approximate graph matching are more suitable for addressing the graph search problem. Indeed, in this category lies the most effective graph similarity search algorithms. Most of the time, the searching phase is conducted under the *graph edit distance* ($GED$) constraint [@Liang17; @Zheng15; @Zheng13]. The graph edit distance $GED(g_1,g_2)$ is defined as the minimum number of edit operations (adding, deletion and substitution) that modify $g_1$ step-by-step to $g_2$ (or vice versa). In [@Zheng15] and in [@Zhang10], the authors underline the robustness of $GED$ against noise due to its error-tolerant capability. Unfortunately, the $GED$ computation is NP-hard, and most existing solutions adopt a *filtering-verification* technique. In particular, first, a pruning strategy is used to filter out false positive matches, and then the remaining candidates are verified by computing $GED$. In this context, the work of Liang and Zhao [@Liang17] represents the state-of-the-art. They provided a partition-based $GED$ lower bound to improve the filter capability, and a multi-layered indexing approach to filter out false positives in an efficient way. Their algorithm can deal with databases with a high number of graphs, but cannot handle large graphs due to the complexity of $GED$ computation. Instead, our algorithm is designed to scale both in the size of the databases and in the size of the graphs.
#### **Roadmap**
The paper is organized as follows. In section \[preliminaries\], we provide the basic concepts and notations used in the next sections as well as the formal definition of graph summary. In section \[RL\], we present a short theoretical description of Szemerédi’s Regularity Lemma and we describe how to find a regular partition. Section \[sumAlg\] is devoted to the description of our summarization framework, while in section \[graphSearch\], we introduce the formal definition of graph search problem, and we discuss how to use our framework to speed up the search process and to reduce the storage space. In section \[experiments\], we report an extensive experimental evaluation of the noise robustness of our approach, in terms of the reconstruction error, and of the usefulness of the summaries in addressing the graph search task. Finally, we drawn our conclusions in section \[conclusions\].
Preliminaries
=============
Let $G = (V,E)$ be an undirected graph without self-loops. The *edge density* of a pair of two disjoint vertex sets $C_i, C_j \subseteq V$ is defined as: $$d(C_i,C_j) = \frac{e(C_i,C_j)}{\left\vert{C_i}\right\vert \left\vert{C_j}\right\vert }$$ where $e(C_i, C_j)$ denotes the number of edges of $G$ with an endpoint in $C_i$ and an endpoint in $C_j$.
Given a positive constant $\varepsilon > 0$, we say that the pair $(C_i,C_j)$ of disjoint vertex sets $ C_i,C_j \subseteq V $ is [*$\varepsilon$-regular*]{} if for every $ X \subseteq C_i $ and $ Y \subseteq C_j $ satisfying $\left\vert{X}\right\vert > \varepsilon \left\vert{C_i}\right\vert \mbox{and} \left\vert{Y}\right\vert > \varepsilon \left\vert{C_j}\right\vert$ we have $$\left\vert{d(X,Y) - d(C_i,C_j)}\right\vert < \varepsilon~.$$ This means that the edges in an $\varepsilon$-regular pair are distributed fairly uniformly, where the deviation from the uniform distribution is controlled by the tolerance parameter $\varepsilon$.
![\[fig:approximation\]Example of the reduced graph (summary) construction. Left: the original graph. Right: the *reduced graph* which contains eight $\varepsilon$-regular classes pairs. The density of each pair is expressed by the thickness of the edge that connects the classes of that pair. If a pair is $\varepsilon$-irregular the corresponding classes are not connected by an edge.](imgs/approximation.png){width="80.00000%"}
A partition $\mathcal{P} = \{C_0, C_1, \dots , C_k\}$, with $C_0$ being the exceptional set[^1] is called *$\varepsilon$-regular* if:
1. it is equitable: $|C_1| = |C_2| = \dots = |C_k|$;
2. $|C_0| < \varepsilon|V|$;
3. all but at most $\varepsilon k^2$ of the pairs $(C_i, C_j)$ are $\varepsilon$-regular, ($1 \leq i < j \leq k$).
In this paper, we propose a summarization algorithm which, given an undirected graph $G=(V,E)$, iteratively builds a summary, called *reduced graph* (Figure \[fig:approximation\]), defined as follows.
Given an $\varepsilon$-regular partition $P = \{ C_{1}, C_{2},$\
$\dots, C_{k} \}$ of a graph $G = (V, E)$ and $0 \leq d' \leq 1$, the reduced graph of $G$ is the undirected weighted graph $R = (V_{R}, E_{R}, w)$, where $V_{R} = P$, $E_{R} \subseteq V_{R}^{2}$ and $w: E_{R} \rightarrow \mathbb{R}$ is defined as follows:
$$w((C_{i}, C_{j})) =
\begin{cases}
d(C_{i}, C_{j}) & \text{if } (C_{i}, C_{j}) \text{ is $\varepsilon$-regular and } d(C_{i}, C_{j}) \geq d', \\
0 & \text{otherwise.}
\end{cases}$$
We are now ready to state the Regularity Lemma which provides us a principled way to develop a summarization algorithm with the aim of separating structure from noise in a large graph.
Szemerédi’s Regularity Lemma {#RL}
============================
In essence, Szemerédi’s Regularity Lemma states that given an $\varepsilon > 0$, every sufficiently dense graph $G$ can be approximated by the union of a bounded number of quasi-random bipartite graphs, where the deviation from randomness is controlled by the tolerance parameter $\varepsilon$. In other words, we can partition the vertex set $V$ into a bounded number of classes $C_0, C_1, . . . , C_k$, such that almost every pair $(C_i, C_j)$ behaves similarly to a random bipartite graph, ($1 \leq i < j \leq k$).
For every positive real $\varepsilon$ and for every positive integer $m$, there are positive integers $N = N(\varepsilon,m)$ and $M = M(\varepsilon,m)$ with the following property: for every graph $G=(V,E)$, with $\left\vert{V}\right\vert \geq N$, there is an $\varepsilon$-regular partition of $G$ into $k + 1$ classes such that $m \leq k \leq M$.
The strength of the Regularity Lemma is corroborated by the so-called Key Lemma, which is an important theoretical result introduced by Komlos et al. [@Komlos1996]. It basically states that the reduced graph does inherit many of the essential structural properties of the original graph. Before presenting its original formulation, another kind of graph needs to be defined, namely the *fold graph*. Given an integer $t$ and a graph $R$ (which may be seen as a reduced graph), let $R(t)$ denote the graph obtained by “blowing up” each vertex $j$ of $V(R)$ to a set $A_j$ of $t$ independent vertices, and joining $u \in A_x$ to $v \in A_y$ if and only if $(x,y)$ is an edge in $R$. Thus, $R(t)$ is a graph in which every edge of $R$ is replaced by a copy of the complete bipartite graph $K_{tt}$. The following lemma shows a link between the reduced graph $R$ and $R(t)$.
\[KeyLemma\] Given $d > \varepsilon > 0$, a graph $R$, and a positive integer $m$, let us construct a graph $G$ by performing the following steps:
1. replace every vertex of $R$ by $m$ vertices;
2. replace the edges of $R$ with $\epsilon$-regular pairs of density at least $d$.
Let $H$ be a subgraph of $R(t)$ with $h$ vertices and maximum degree $\Delta >0$, and let $\delta = d-\varepsilon$ and $\varepsilon_0 = \delta^\Delta/(2+\Delta)$. If $\varepsilon \leq \varepsilon_0$ and $t-1 \leq \varepsilon_0m$, then $H$ is embeddable into $G$ (i.e. $G$ contains a subgraph isomorphic to $H$). In fact, we have: $$\left\vert\left\vert{H\,\to\,G}\right\vert\right\vert > (\varepsilon_0m)^h$$ where $\left\vert\left\vert{H\,\to\,G}\right\vert\right\vert$ denotes the number of labeled copies of $H$ in $G$.
Hence, the Key Lemma provides us a theoretical guarantee on the quality of the summary built from an $\epsilon$-regular partition. In particular, for $t=1$, $R(t)=R$, and if the constraint on the edge density $d$ is satisfied, the Key Lemma ensures that every small subgraph of $R$ is also a subgraph of $G$. Thus, we can use the Regularity Lemma to build a summary $R$ of $G$, and then we can infer structural properties of $G$ by studying the same properties on $R$.
#### **Finding Regular Partitions**
The original proof of the Regularity Lemma is not constructive, but during the last decades different constructive versions have been proposed. In this paper, we focus on the Alon et al.’s [@Alon94] work. In particular, they proposed a new formulation of the Regularity Lemma which emphasizes the algorithmic nature of the result.
[(Alon et al., 1994)]{} For every $\varepsilon > 0$ and every positive integer $t$ there is an integer $Q = Q(\varepsilon, t)$ such that every graph with $n > Q$ vertices has an $\varepsilon$-regular partition into $k + 1$ classes, where $t \leq k \leq Q$. For every fixed $\varepsilon > 0$ and $t \geq 1$ such partition can be found in $O(M(n))$ sequential time, where $M(n) = O(n^{2.376})$ is the time for multiplying two $n \times n$ matrices with $0,1$ entries over the integers. It can also be found in time $O(\log{n})$ on an Exclusive Read Exclusive Write Parallel Random Access Machine(EREW PRAM) with a polynomial number of parallel processors.
A sketch of the proof is then presented. Let $H$ be a bipartite graph with color classes $A$ and $B$, with $\left\vert{A}\right\vert = \left\vert{B}\right\vert = n$. Let us define the average degree $\bar{d}$ of $H$ as: $$\bar{d}(A,B)=\frac{1}{2n} \sum_{i \in A \cup B} deg(i)$$ where $deg(i)$ is the degree of vertex $i$.
For two distinct vertices $y_1, y_2 \in B$ the *neighbourhood deviation* of $y_1$ and $y_2$ is defined as: $$\sigma(y_1, y_2) = |N(y_1) \cap N(y_2)| - \frac{\bar{d}^2}{n}$$
where $N(x)$ is the set of neighbours of vertex $x$. For a subset $Y \subset B$ the *deviation* of $Y$ is defined as: $$\sigma(Y)=\frac{\sum_{y_1,y_2 \in Y}\sigma(y_1,y_2)}{|Y|^2}$$\[sigma\]
Let $0 < \varepsilon < 1/16$, it can be proved that, if there exists $Y \subset B,\; |Y| > \varepsilon n$ such that $\sigma(Y) \geq \frac{\varepsilon^3}{2} n$, then at least one of the following cases occurs:
1. $\bar{d} <\varepsilon^3 n$ ($H$ is $\varepsilon$-regular); \[cond1\]
2. there exists in $B$ a set of more than $\frac{1}{8}\varepsilon^4n$ vertices whose degrees deviate from $\bar{d}$ by at least $\varepsilon^4n$ ($H$ is $\varepsilon$-irregular); \[cond2\]
3. there are subsets $A'\subset A, \; B'\subset B, \; |A'|\geq \frac{\varepsilon^4}{n}n, \; |B'|\geq \frac{\varepsilon^4}{n}n$ such that $|\bar{d}(A', B') - \bar{d}(A,B)|\geq \varepsilon^4$ ($H$ is $\varepsilon$-irregular). \[cond3\]
\[alon3\]
Note that one can easily check if \[cond1\] holds in time $O(n^2)$. Similarly, it is trivial to check if \[cond2\] holds in $O(n^2)$ time, and in case it holds to exhibit the required subset of $B$ establishing this fact. If the first two conditions are not verified, the \[cond3\] condition must be checked. To this end, we have to find the subsets $A',B'$, called *certificates*, that witness the irregularity of the bipartite graph $H$. To address this task, we first select a subset of $B$ whose vertex degrees “deviate” the most from the average degree $\bar{d}$ of $H$. More formally: for each $y_0 \in B$ with $|deg(y_0)-\bar{d}| < \varepsilon^4n$ we find the vertex set $B_{y_0} = \{y \in B| \, \sigma(y_0, y) \geq 2\varepsilon^4 n \}$. The proof provided by Alon et al. guarantees the existence of at least one such $y_0$ for which $|B_{y_0}| \geq \frac{\varepsilon^4}{4}n$. Thus, the subsets $B'=B_{y_0}$ and $A'=N(y_0)$ are the required certificates. These two subsets represent the collection of vertices that contribute more to the irregularity of the pair $(A,B)$. The sets $\bar{A^{'}} = A \setminus A^{'}, \, \bar{B^{'}} = B \setminus B^{'}$ are called *complements*. Since the computation of the quantities $\sigma(y, y')$, for $y, \; y' \in B$, can be done by squaring the adjacency matrix of $H$, the overall complexity of this algorithms is $O(M(n)) = O(n^{2.376})$.
Before reporting Alon et al.’s algorithm, we present a measure to assess the goodness of a partition of the vertex set $V$ of a graph $G$, which was introduced by Szemerédi [@Szemeredi75Graphs]. Given a partition $\mathcal{P} = \{C_0,C_1,\dots, C_k\}$ of a graph $G=(V,E)$, the index ($\emph{sze\_idx}$) of the partition $\mathcal{P}$ is defined as follows:
$$sze\_ind(\mathcal{P})=\frac{1}{k^2}\sum_{s=1}^{k}\sum_{t=s+1}^{k}d(C_s,C_t)^2$$\[indexP\]
Since $0\leq d(C_s,C_t) \leq 1$, it can be seen that $sze\_ind(\mathcal{P}) \leq \frac{1}{2}$. Szemerédi proved that if a partition $\mathcal{P}$ violates the regularity condition, then it can be refined by a new partition $\mathcal{P '}$ such that $sze\_ind(\mathcal{P '}) > sze\_ind(\mathcal{P})$. Finally, we can now present Alon et al.’s algorithm, which provides a way to find an $\varepsilon$-regular partition. The procedure is divided into two main steps: in the first step all the constants needed during the next computation are set; in the second one, the partition is iteratively created. An iteration is called *refinement step*, because, at each iteration, the current partition is closer to a regular one.
#### Alon et al.’s Algorithm
1. Create the initial partition: arbitrarily divide the vertices of $G$ into an equitable partition $\mathcal{P}_1$ with classes $C_0,C_1, \dots, C_b$ where $|C_i|=\lfloor \frac{n}{b}\rfloor$. \[step1\]
2. Check Regularity: for every pair $(C_r,C_s)$ of $\mathcal{P}_i$, verify if it is $\varepsilon$-regular or find two certificates $A^{'} \subset C_r, \, B^{'} \subset C_s, \, |A^{'}| \geq \frac{\varepsilon^4}{16}|C_1|, \, |B^{'}| \geq \frac{\varepsilon^4}{16}|C_1| $ such that $|\bar{d}(A^{'},\, B^{'})-\bar{d}(C_s, C_t)| \geq \varepsilon^4$. \[step2\]
3. Count regular pairs: if there are at most $\varepsilon \binom{k_i}{2}$ pairs that are not $\varepsilon$-regular, then stop. $\mathcal{P}_i$ is an $\varepsilon$-regular partition. \[step3\]
4. Refine: apply a refinement algorithm and obtain a partition $\mathcal{P}'$ with $1+k_i4^{k_i}$ classes, such that $sze\_ind(\mathcal{P '}) > sze\_ind(\mathcal{P})$. \[step4\]
5. Go to step 2.
Even if the above mentioned algorithm has polynomial worst case complexity in the size of $G$, there is a hidden tower-type dependence on an accuracy parameter. Unfortunately, Gowers [@Gowers1998] proved that this tower function is necessary in order to guarantee a regular partition for *all* graphs. This implies that, in order to have a faithful approximation, the original graph size should be astronomically big. This has typically discouraged researchers from applying regular partitions to practical problems, thereby confining them to the purely theoretical realm.
To make the algorithm truly applicable, [@Sperotto07], and later [@Sar+12; @Fiorucci17], instead of insisting on provably regular partitions, proposed a few simple heuristics that try to construct an approximately regular partition. In the next section, we present a new heuristic algorithm which is characterized by an improvement of the summary quality both in terms of reconstruction error and of noise filtering.
The Summarization Algorithm {#sumAlg}
===========================
The main limitations which prevent the application of Alon et al.’s algorithm to practical problems concern Step \[step2\] and Step \[step4\]. In particular, in Step \[step2\] the algorithm checks the regularity of all classes pairs by using the three conditions previously described. Given a pair $(C_r,C_s)$, condition \[cond1\] verifies if it is $\varepsilon$-regular, otherwise conditions \[cond2\] and \[cond3\] are used to obtain the *certificates* $C_r^{'}$ and $C_s^{'}$ that witness the irregularity. The main obstacle concerning the implementation of condition \[cond3\] is the necessity to scan over *almost all possible subsets* of $C_s$. To make the implementation of condition \[cond3\] feasible, given a class $C_s$, we select in a *greedy way* \[greedy\] a set $Y^{'} \subseteq C_s$ with the highest deviation $\sigma(Y^{'})$ (the deviation is defined in \[sigma\]). To do so, the nodes of $C_s$ are sorted by bipartite degree, and $Y^{'}$ is built by adding $\frac{\varepsilon^4}{4}n$ nodes with the highest degree. At each iteration of the greedy algorithm, the node with a degree that deviates most from the average degree is added to the candidate certificate $Y^{'}$. This last operation is repeated until the subset $C_s'$, that satisfies condition \[cond3\], is found. This almost guarantees to put in the candidate certificate the nodes that have a *connectivity pattern* that deviates most from the one characterizing the majority of the nodes which belong to $C_s$.
As far Step \[step4\] is concerned, here an irregular partition $\mathcal{P}^{i}_\varepsilon$ is *refined* by a new partition $\mathcal{P}^{i+1}_\varepsilon$, such that the partition index *sze\_idx* is increased. This step poses the main obstacle towards a practical version of Alon et al.’s algorithm involving the creation of an exponentially large number of subclasses at each iteration. Indeed, as we have said, Step \[step2\] finds all possible irregular pairs in the graph. As a consequence, each class may be involved with up to $(k_i-1)$ irregular pairs, $k_i$ being the number of classes in the current partition $\mathcal{P}^{i}_\varepsilon$, thereby leading to an exponential growth. To avoid the problem, for each class, one can limit the number of irregular pairs containing it to at most one, possibly chosen randomly among all irregular pairs. This simple modification allows one to divide the classes into a constant, rather than exponential, number of subclasses $l$ (typically $ 2 \leq l \leq 7$). Despite the crude approximation this seems to work well in practice.
The devised algorithm takes as input two main parameters, the tolerant parameter $\varepsilon$ and the minimum compression rate $c\_min$, that acts as a stopping criterion in the refinement process, and returns an approximated $\epsilon$-regular partition. Its pseudocode is reported in Algorithm \[algorithm:alg1\], where the procedure <span style="font-variant:small-caps;">ApproxAlonCertificates</span>, based on the two heuristics described above, takes as input a partition $\mathcal{P}_\epsilon^{i}$ and returns the number of irregular pairs of $\mathcal{P}_\epsilon^{i}$. In the next paragraph, we describe the <span style="font-variant:small-caps;">Refinement</span> procedure which refines a partition $\mathcal{P}_\epsilon^{i}$ into a partition $\mathcal{P}_\epsilon^{i+1}$. Its pseudocode is reported in Algorithm \[algorithm:newrefinement\]. The overall complexity of our summarization algorithm is $O(M(n)) = O(n^{2.376})$, which is dominated by the verification of condition \[cond3\].
partitions = empty list $\mathcal{P}^1_\varepsilon =$ Create initial random partition from $G$
\#irr\_pairs = $\Call{ApproxAlonCertificates}{\mathcal{P}^i_\varepsilon}$
break $\mathcal{P}^{i+1}_\varepsilon = \Call{Refinement}{\mathcal{P}^i_\varepsilon}$ partitions.add($\mathcal{P}^{i+1}_\varepsilon$) break
Select best partition $\mathcal{P}^*$ with maximum $sze\_idx$ from list partitions
\[algorithm:alg1\]
$C_i =$ $C^1_i,C^2_i = $ Select $C_j$ with most similar internal structure Get certificates $(A^{'}, B^{'})$ and complements $(\bar{A^{'}},\bar{B^{'}})$ of $C_i, C_j$ $C^1_i,C^2_i =$
$C^1_i,C^2_i =$
Perform step 9,10,11,12 for $B^{'}$ Uniformly distribute nodes of $C_0$ between all the classes $(\mathcal{P}^{i+1}_\varepsilon, irregular)$ $(\mathcal{P}^{i+1}_\varepsilon, regular)$
\[algorithm:newrefinement\]
Given a partition $\mathcal{P}_\epsilon^i = \{C_0,C_1,\dots, C_{k_i}\}$, the <span style="font-variant:small-caps;">Refinement</span> procedure starts by randomly selecting a class $C_i$, then iteratively processes all the others.
- If $C_i$ is $\varepsilon$-regular with all the others, the procedure sorts the nodes of $C_i$ by their *internal degree*, i.e. the degree calculated with respect to the nodes of the same class, obtaining the following sorted sequence of nodes $v_1,v_2,v_3,v_4,v_5,v_6, \cdots, v_{|C_i|}$. The next step splits (<span style="font-variant:small-caps;">Unzip</span>) this sequence into two sets $C_i^1 = \{v_1, v_3, v_5, \cdots , n_{|C_i|-1}\}$ and $C_i^2=\{v_2, v_4, v_6, \cdots , v_{|C_i|}\}$. The latter sets are part of the refined partition $\mathcal{P}^{i+1}_\varepsilon$.
- If $C_i$ forms an irregular pair with other classes, the heuristic selects the candidate $C_j$ that shares the most similar internal structure with $C_i$ by maximizing $S = d(C_i, C_j) + (1 - |d(C_i, C_i) - d(C_j, C_j)|)$, where $d(C_i,C_i)=e(C_i,C_i)/|C_i|^2$ is the *internal density*.
After selecting the best matching class $C_j$, we are ready to split the pair $(C_i, C_j)$ in 4 new classes $C_i^1, C_i^2,C_j^1,C_j^2$ based on the internal densities of the certificates $C_i'$ and $C_j'$.
- In particular, a <span style="font-variant:small-caps;">Sparsification</span> procedure is applied when the internal density of a certificate is below a given threshold. This procedure randomly splits the certificate into two new classes. In order to match the equi-cardinality property, the new classes are filled up to $|C_i|/2$ by adding the remaining nodes from the corresponding complement. We choose the nodes that share the minimum number of connections with the new classes.
- On the other hand, if the internal density of a certificate is above a given threshold, then a <span style="font-variant:small-caps;">Densification</span> procedure is applied. In particular, the heuristics sorts the nodes of the certificate by their internal degree and <span style="font-variant:small-caps;">Unzip</span> the set into two new classes. Also in this case, we fill the new sets up to $|C_{i}|/2$ by adding the remaining nodes from the corresponding complement by choosing the nodes which share the major number of connections with the new classes.
Graph Search Using Summaries {#graphSearch}
============================
In this section, we discuss how to use our summarization framework to efficiently address the graph search problem defined under a similarity measure. The aim of graph search is to retrieve from a database the top-$k$ graphs that are most similar to a query graph.
#### **Problem Definition**
We consider a graph database $\mathcal{D}$ containing a high number of simple undirected graphs $g_j \in \mathcal{D}, \; j = 1 \dots |\mathcal{D}|$, and, for the sake of generality, we allow the edges to be weighted.
Given a graph database $\mathcal{D} = \{g_1,g_2,\cdots,g_{|D|}\}$, a query graph $q$, and a positive integer $k$, the graph similarity search problem is to find the top-$k$ graphs in $\mathcal{D}$ that are most similar to $q$ according to a similarity measure.
As far as the similarity measure is concerned, the most used one is the *graph edit distance* ($GED$) due to its generality, broad applicability and noise robustness [@Liang17; @Zhang10]. However, since the $GED$ computation is NP-hard, it is not suited to deal with large graphs. To overcome this limitation, we use the *spectral distance* [@Jin18], which is computed by comparing the eigenvalues of the two graphs being matched. The choice of this measure is motivated by the work of Van Dam and Haemers [@VANDAM03], who show that graphs with similar spectral properties generally share similar structural patterns. In this paper, we introduce a slightly modified version of the spectral distance to increase its range of applicability to pairs of graphs that violate the assumption of the Theorem 1 in [@Jin18], which assumes a precise order between the eigenvalues of the two graphs being matched. To this aim, we simply compute the absolute value of the difference between the $i$-$th$ eigenvalue of the first graph with the $i$-$th$ one of the second graph.
Given two simple undirected weighted\
graphs $G_1 = (V_1,W_1)$ with $|V_1| = n_1$, and $G_2=(V_2,W_2)$ with $|V_2|=n_2$. Let us denote the corresponding spectra as $0=\lambda_1^{(1)} \leq \lambda_2^{(1)}, \leq \cdots \leq \lambda_{n_1}^{(1)} $ and $0=\lambda_1^{(2)} \leq \lambda_2^{(2)}, \leq \cdots \leq\lambda_{n_2}^{(2)}$. We may assume without loss of generality that $n_2 > n_1$. The spectral distance is then defined as follows $$SD(G_1,G_2,l) = \frac{1}{n_1} \left ( \sum_{i=1}^{l}|\lambda_i^{(2)} - \lambda_i^{(1)}| + \sum_{i=l+1}^{n1}|\lambda_{i+n_2-k}^{(2)} - \lambda_i^{(1)}| \right )$$
where, $l$ controls which part of the spectra are being matched. In particular, eigenvalues of $G_1$ are compared with the head and tail eigenvalues of the $G_2$.
#### **Using The Summaries**
In our approach, all the graphs contained in a database are summarized off-line, while the query graph is summarized on-line by means of our summarization framework. Thus, graph search can be performed on graph summaries, and this allows us to speed up the search process and to reduce the storage space. In particular, for each graph $g_j$ of a database $\mathcal{D}$, we store two different quantities: the summary $r_j$ of $g_j$ and the eigenvalues $eig_{r_j}$ of $r_j$. We then summarized on-line the query graph $q$ obtaining its summary $r_q$. Finally, we compute the spectral distance between $r_q$ and each summary $r_j \in \mathcal{D}$. The desired top-$k$ graphs will be obtained by selecting, from $\mathcal{D}$, the $k$ graphs corresponding to the $k$ smallest value of the spectral distance previously computed. The pseudocode of our approach to graph search is reported in Algorithm $3$.
$r =$ Summarize $g$ $eig_r =$ Calculate the eigenvalues of the adj. matrix of $r$ Store $(r, eig_r)$ in $\mathcal{D}$
$r_q =$ Summarize $q$ $eig_{r_q} =$ Calculate the eigenvalues of the adj. matrix of $r_q$ $sd\_array = \emptyset$ $sd =$ Spectral Distance$(r_j, r_q, eig_{r_j}, eig_{r_q})$ Append $sd$ to $sd\_array$ Order $sd\_array$ first $k$ results of $sd\_array$ and their relative graphs.
\[gS\]
Experimental Evaluation {#experiments}
=======================
In this section, we evaluate our summarization algorithm both on synthetic graphs and on real-world networks to assess:
- the ability of the proposed algorithm to separate structure from noise;
- the usefulness of the summaries in retrieving from a database the top-$k$ graphs that are most similar to a query graph.
#### **Experimental Settings**
In our experiments we used both synthetic graphs and real-world networks. We generated synthetic graphs with a cluster structure, where the clusters are perturbed with different levels of noise. In particular, each graph is generated by adding spurious edges between cluster pairs and by dropping edges inside each cluster. Figure \[fig:graphex\] provides a concrete example with a visual explanation. The pseudocode of the algorithm used to generate the synthetic datasets is reported in Algorithm \[algorithm:gen\].
![The adjacency matrix of an undirected synthetic graph of 2000 vertices. The graph is generated by corrupting $5$ cliques as described in Algorithm \[algorithm:gen\]. In particular, the *intra-cluster noise probability* is $0.2$ and the *inter-cluster noise probability* is $0.4$.[]{data-label="fig:graphex"}](imgs/graphex.png){width="0.43\linewidth"}
As far as the real-world networks are concerned, we used two different datasets which have been taken from two famous repositories: the Stanford Large Network Dataset Collection SNAP [@snapnets] and the Konect repository of the University Koblenz-Landau Konect [@konect]. In particular, we used the following networks: Facebook [@facebookNIPS], Email-Eu-core [@Yin2017LocalClustering][@Leskovec07email], Openflights[@konect:2016:openflights], and Reactome [@Joshi-Tope2005Reactome:Pathways]. Our algorithm is implemented in Python $3.6.3$ [^2] and the experiments are performed on an Intel Core i5 @ 2.60GHz HP Pavilion 15 Notebook with 8GB of RAM (DDR3 Synchronous 1600 MHz) running Arch-Linux with kernel version 4.14.4-1.
$G =$ Generate Erdős Rényi graph of size $n$ using $\eta_1$ as edge probability $clust\_dim = n / num\_c$ Select $clust\_dim$ nodes from $G$ and create cluster $c_i$ with them For each edge in $c_i$ drop it with probability $\eta_2$ $G$
\[algorithm:gen\]
Graph Summarization
-------------------
We performed experiments on both synthetic graphs and on real-world networks to assess the ability of the proposed algorithm to separate structure from noise. As evaluation criterion, we used the reconstruction error, which is expressed in terms of normalized $l_p$ norm computed between the similarity matrix of an input graph $G$ and the similarity matrix of the corresponding reconstructed graph $G'$.
Given the similarity matrix of the input graph $A_G$ and the similarity matrix of the reconstructed graph $A_{G'}$, the reconstruction error is defined as follows: $$l_p(\mathbf{A_G}, \mathbf{A_{G'}}) = (\sum_{i=1}^n \sum_{j=1}^n (A_G(i,\,j) - A_{G'}(i,\,j))^p)^\frac{1}{p}$$
We decided to use the reconstruction error in order to compare our results with the ones presented by Riondato et al. [@Riondato2017], who evaluated the summary quality using this measure. This choice is due to the fact that their algorithm summarizes a graph by minimizing the reconstruction error. However, they pointed out that the reconstruction error has some shortcomings. In particular, given an unweighted graph $G$, it is possible to produce an uninteresting summary with only one supenode corresponding to the vertex set and $l_1$ reconstruction error at most $n^2$. On the other hand, if we obtained an useful summary, where each pair of vertices belonging to a supernode share an high number of common neighbors, then we get a low (say $o(n^2)$) $l_1$ reconstruction error: this is a desirable behavior because low values of $l_1$ correspond to high quality summaries. Unfortunately, such low values are often obtained only with summaries having an high number of supernodes. This prevents to adopt the reconstruction error as a general measure to assess the summary quality.
As far the summarization and reconstruction steps are concerned, we proceeded, in all the experiments, in the following way: we applied our summarization algorithm (see Algorithm \[algorithm:alg1\]) to summarize an input graph $G$. We then “blow-up” the summary in order to obtain the reconstructed graph $G'$, which preserves the main structure carried by the input graph (Figure \[fig:summarization\]).
![We summarized an input graph $G$ by using Algorithm \[algorithm:alg1\]. We then “blow up” the summary to obtain the reconstructed graph $G'$.[]{data-label="fig:summarization"}](imgs/summarization.jpg){width="1\linewidth"}
#### **Noise Robustness Evaluation**
We study the ability of the proposed algorithm to separate structure from noise in graphs performing an extensive series of experiments on both synthetic graphs and real-world networks. As far synthetic graph experiments are concerned, we generated a graph $G$ by corrupting the clusters of $GT$ in the following way: we added spurious edges between each cluster pair with probability $\eta_1$, and we dropped edges inside each cluster with probability $\eta_2$ (see algorithm \[algorithm:gen\]). As far as the real-world networks experiments are concerned, we added spurious edges with probability $noise$ $probability$ to a original graph $GT$ obtaining an input graph $G$.
In the ideal case, the distance between $G$ and the corresponding reconstructed graph $G'$ should be only due to the filtered noise, while the distance between $GT$ an $G'$ should be closed to zero. Hence, we computed the reconstruction error $l_2(G',GT)$ to assess the robustness of our summarization framework against noise.
*Experiment 1.* We generated synthetic graphs of different sizes, spanning from $10^3$ up to $10^4$ nodes. We synthesized $250$ graphs by considering, for each of the $10$ different sizes, all the $25$ combinations of the following noise probabilities:
- the probability $\eta_1$ of adding a spurious edge between a pair of clusters, called *inter-cluster noise probability*, which assumes values in $\{0.1,\,0.2,\,0.3,$ $\,0.4,\,0.5\}$;
- the probability $\eta_2$ of dropping an edge inside each cluster, called *intra-cluster noise probability*, which assumes values in $\{0.1,\,0.2,\,0.3,\,0.4,\,0.5\}$.
Let’s consider a synthetic graph $G_{n,(\eta_1,\eta_2)}$, where n is its size, and $(\eta_1,\eta_2)$ corresponds to one of the $25$ pairs of noise probabilities. For each $G_{n,(\eta_1,\eta_2)}$ we obtained the reconstructed graph $G'_{n,(\eta_1,\eta_2)}$, and we then computed the reconstruction error $l_2(G'_{n,(\eta_1,\eta_2)},GT)$. Given a size $n$, we computed the median $m_n$ of $ \{ l_2(G'_{n,(0.1,0.1)},GT), \; l_2(G'_{n,(0.1,0.2)},GT), \cdots ,$\
$ l2_(G'_{n,(0.5,0.5)},GT) \}$. We reported in figure \[fig:l2vsnoise\] the $10$ medians computed using our summarization framework and the corresponding medians obtained by applying Riondato et al.’s algorithm [@Riondato2017]. We can see how our framework outperforms the state-of-the-art summarization algorithm in terms of robustness against noise.
![The plot shows the medians computed, for each size $n$, from the $25$ values of $l_2(G'_{n,(\eta_1,\eta_2)},GT)$, where $(\eta_1,\eta_2)$ corresponds to one of the $25$ pairs of noise probabilities. The curve “rionda” is obtained by using Riondato et al.’s algorithm [@Riondato2017], while the curve “sze” is obtained by applying our summarization framework.[]{data-label="fig:l2vsnoise"}](imgs/l2vsnoise.png){width="0.5\linewidth"}
*Experiment 2.* The aim of this experiment is to study separately the robustness against the inter-cluster and the intra-cluster noise. Let’s consider the probability of dropping an edge inside each cluster $\eta_2$ equals to $0.2$ and the graph size $n$ equals to $10^4$. To asses the inter-cluster noise robustness, we generated synthetic graphs $G_{10^4,(\eta_1,\,0.2)}$, where $\eta_1$ assumes values in $\{0.1,\,0.15,\,0.2,\,0.25,\,0.3,$ $\, 0.35,\,0.4,\,0.45,\,0.5\}$, and we computed the reconstruction errors $l_2(G'_{10^4,(\eta_1,\,0.2)},$ $GT)$. As far the intra-cluster noise is concerned, we chose the probability of adding a spurious edges between each pair of clusters $\eta_1=0.2$, and the graph size $i=10^4$. We then generated synthetic graphs $G_{10^4,(0.2, \, \eta_2)}$, where $\eta_2$ assumes values in $\{0.1,\,0.15,\,0.2,\,0.25,\,0.3,$ $\, 0.35,\,0.4,\,0.45,\,0.5\}$, and we computed the reconstruction errors $l_2(G'_{10^4,(0.2, \, \eta_2)},GT)$.
Figure \[fig:l2\_vs\_inter\] illustrates the comparison between our results with those obtained by applying Riondato et al.’s algorithm [@Riondato2017]. This results are in accord to those presented in figure \[fig:l2vsnoise\], and provides an experimental verification of the ability of our method to separate structure from noise in graphs.
![The plot on the left represents $l_2(G'_{10^4,(0.2,\,\eta_2)},GT)$ versus the intra-noise probability. The plot on the right represents $l_2(G'_{10^4,(\eta_1,\,0.2)},GT)$ versus the inter-noise probability. The curve “rionda” is obtained by using Riondato et al.’s algorithm, while the curve “sze” is obtained by applying our summarization framework.[]{data-label="fig:l2_vs_inter"}](imgs/l2_vs_inter.png "fig:"){width="0.45\linewidth"} ![The plot on the left represents $l_2(G'_{10^4,(0.2,\,\eta_2)},GT)$ versus the intra-noise probability. The plot on the right represents $l_2(G'_{10^4,(\eta_1,\,0.2)},GT)$ versus the inter-noise probability. The curve “rionda” is obtained by using Riondato et al.’s algorithm, while the curve “sze” is obtained by applying our summarization framework.[]{data-label="fig:l2_vs_inter"}](imgs/l2_vs_intra.png "fig:"){width="0.45\linewidth"}
*Experiment 3.* We added spurious edges with probability $noise$ $probability$ to an original real-world network $GT$ obtaining an input graph $G$. The $noise$ $probability$ assumes values in $\{0.01, \, 0.02,\,0.03,\,0.04,$ $\, 0.05,\,0.06,\,0.07,\,0.08, \, 0.09, \,$ $0.1\}$. We applied this procedure on real-word networks, which have been taken from the Stanford Large Network Dataset Collection SNAP [@snapnets] and from the Konect repository of the University Koblenz-Landau Konect [@konect]. Since our framework is based on the Regularity Lemma, which is suited to deal only with dense graphs, we expect to obtain low quality summaries from sparse real-world networks. However, as shown in figure \[fig:l2\_real-wordl\], our framework outperforms the state-of-the-art summarization algorithm in terms of robustness against noise providing good quality summary even on sparse real-world networks. In particular, we can see how the quality increases with the size of the input graph, which is in accord with the assumptions of the Regularity Lemma.
![These plots represent the median of the $l_2(G',GT)$ versus the noise probability. We run $20$ experiments for each value of the noise probability. The curve “rionda” is obtained by using Riondato et al.’s algorithm, while the curve “sze” is obtained by applying our summarization framework.[]{data-label="fig:l2_real-wordl"}](imgs/email-Eu-core.png "fig:"){width="0.45\linewidth"} ![These plots represent the median of the $l_2(G',GT)$ versus the noise probability. We run $20$ experiments for each value of the noise probability. The curve “rionda” is obtained by using Riondato et al.’s algorithm, while the curve “sze” is obtained by applying our summarization framework.[]{data-label="fig:l2_real-wordl"}](imgs/facebook.png "fig:"){width="0.45\linewidth"}\
![These plots represent the median of the $l_2(G',GT)$ versus the noise probability. We run $20$ experiments for each value of the noise probability. The curve “rionda” is obtained by using Riondato et al.’s algorithm, while the curve “sze” is obtained by applying our summarization framework.[]{data-label="fig:l2_real-wordl"}](imgs/openflights.png "fig:"){width="0.45\linewidth"} ![These plots represent the median of the $l_2(G',GT)$ versus the noise probability. We run $20$ experiments for each value of the noise probability. The curve “rionda” is obtained by using Riondato et al.’s algorithm, while the curve “sze” is obtained by applying our summarization framework.[]{data-label="fig:l2_real-wordl"}](imgs/reactome.png "fig:"){width="0.45\linewidth"}
Graph Search
------------
We performed extensive experiments on synthetic datasets to assess the usefulness of the summaries in retrieving, from a database, the top-$k$ graphs that are most similar to a query graph. To this end, we evaluate the quality of the answer in terms of the found top-$k$ similar graphs, and we evaluate the scalability both in the size of the database and in the size of the graphs.
#### **Quality Evaluation** {#exp:quality-eval}
We conducted the following experiment: we compared graph search on the summaries with the baseline approach, in which the spectral distance is computed between no preprocessed graphs. The aim of the experiment is to show that pre-summarizing the graphs in the databases increases the noise robustness of the search process. We created a database $\mathcal{D}$ contained synthetic graphs, which have different structures corrupting with different levels of noise (see algorithm \[algorithm:gen\]). In particular, each graph is generated by combining the following three factors: five different possible number of clusters $\{4, 8, 12, 16, 20\}$, six different possible levels of intra-cluster noise probability $\{0.05, 0.1, 0.15, $ $0.2, 0.25, 0.3\}$ and six different possible levels of inter-cluster noise probability $\{0.05, 0.1, 0.15, 0.2, 0.25,$ $ 0.3\}$. Given a size $n$, we generated $180$ graphs considering all the possible combinations of these three parameters. As described in Algorithm \[gS\], we stored in $\mathcal{D}$ the eigenvalues of the $180$ synthetic graphs, their summaries and the corresponding eigenvalues. Finally, we grouped the database graphs into five groups. Each group $\omega_i$, with $i=1,2,3,4,5$, is composed by 36 graphs that are generated by corrupting the same cluster structure with different combinations of intra-cluster and inter-cluster noise probability. Hence, all the graphs belonging to a given group $\omega_i$ are similar, since they have the same main structure.
More formally, we constructed a set $Q$ of five query graphs by randomly sampling one graph from each group $\omega_i$. Then, we first computed the spectral distance between $q_i \in Q$ and every graph in the database $\mathcal{D}$. We then calculated the $AP@k$ for each query $q_i$ by considering relevant the graphs belonging to $\omega_i$ i.e. the same group of $q_i$. Finally, we computed the $MAP@k$ score by averaging the *average precision* $AP@k(q_i)$ of the five graphs in $Q$.
We repeated the same procedure using the summaries of the $180$ synthetic graphs contained in $\mathcal{D}$. The aim of this experiment is to compare the quality obtained using our approach with that obtained by computing the spectral distance between original graphs. We performed the experiment by considering the following graph sizes $n = 1500, 2000, 3000, 7000$.\
The $AP@k(q_i)$ and the $MAP@k$ are defined as follows.
Given a query $q \in Q$, a set of relevant graphs $\omega_i$ (graphs that share the same structure with $q$). Let us consider the output top-$k$ graphs in a database $\mathcal{D}$ ordered by crescent spectral distance. We define the average precision at $k$ as follows. $$AP@k(q) = \frac{1}{|w_i|} \cdot \sum^{k}_{j=1} Precision(j) \cdot Relevance(j)$$ where $Precision(j)$ is the relevant proportion of the found top-$k$ graphs, while $Relevance(j)$ is $1$ if the considered graph is part of $\omega_i$ and is $0$ otherwise. Finally, $|\omega_i|$ is the number of relevant graphs.
Given a query set $Q$, the mean average precision is defined as follows. $$MAP(Q) = \frac{1}{|Q|} \cdot \sum_{q_i \in Q} AP@k(q_i)$$
In particular, the higher is the value of the $MAP \in [0,1] $, the higher is the quality of the proposed graph search algorithm. Figures \[fig:ranking1\], \[fig:ranking2\], \[fig:ranking3\], \[fig:ranking4\] show that the proposed summarization based approach improved the query quality.
![The MAP@k of the top-$k$ graphs given as output in a database of 180 graphs. The size of the graphs contained in the database is $n = 1500$. Two-stage is referred to the use of our pre-summarization approach to address the graph search problem, while one-stage is referred to the search on the original graphs.[]{data-label="fig:ranking1"}](imgs/onestage_twostage_1.png){width="0.5\linewidth"}
![The MAP@k of the top-$k$ graphs given as output in a database of 180 graphs. The size of the graphs contained in the database is $n = 2000$. Two-stage is referred to the use of our pre-summarization approach to address the graph search problem, while one-stage is referred to the search on the original graphs.[]{data-label="fig:ranking2"}](imgs/onestage_twostage_2.png){width="0.5\linewidth"}
![The MAP@k of the top-$k$ graphs given as output in a database of 180 graphs. The size of the graphs contained in the database is $n = 3000$. Two-stage is referred to the use of our pre-summarization approach to address the graph search problem, while one-stage is referred to the search on the original graphs.[]{data-label="fig:ranking3"}](imgs/onestage_twostage_3.png){width="0.5\linewidth"}
![The MAP@k of the top-$k$ graphs given as output in a database of 180 graphs. The size of the graphs contained in the database is $n = 7000$. Two-stage is referred to the use of our pre-summarization approach to address the graph search problem, while one-stage is referred to the search on the original graphs.[]{data-label="fig:ranking4"}](imgs/onestage_twostage_4.png){width="0.5\linewidth"}
#### **Scalability** {#exp:scalability}
In order to evaluate the scalability of our approach, we conducted two different experiments. In the first one, we investigated the time required to perform a single query as the dimension of the database $\mathcal{D}$ grows. In the second one, we investigated the query time in function of the size of the query graph.
In the first experiment, we fixed the size of all the graphs to be $n=2000$. We then generated the graphs in $\mathcal{D}$ using all the possible combinations of the following factors: three different numbers of clusters $\{4, 12, 20\}$, six different levels of intra-cluster noise probability $\{0.05, 0.1, 0.15, $ $ 0.2, 0.25, 0.3\}$, and six different levels of inter-cluster noise probability $\{0.05, 0.1, 0.15, 0.2, 0.25,$ $ 0.3\}$. The combination of these three parameters allow us to generate 108 graphs. We then copied them enough times to reach a database cardinality spanning from $10^3$ up to $10^4$ graphs.
The query time is calculated as follows: $$t = t\_s(q) + t\_{eig(r_q)} + t\_SD(eig_{r_q}, eig_{r_j}) \;\;\;\;\; j=1,\dots, |\mathcal{D}|.$$
where $t\_s(q)$ is the time required to obtain the summary $r_q$ of the query graph $q$; $t\_{eig}(r_q)$ is the time required to calculate the eigenvectors of $r_q$; and $t\_SD(eig_{r_q}, eig_{r_j})$ is the time required to calculate the spectral distances between $r_q$ and each graph summary $r_j$ contained in $\mathcal{D}$. We reported in Figure \[fig:scalability1\], the computed time $t$ versus the cardinality of the database $\mathcal{D}$.
In the second experiment, we generated different databases $\mathcal{D}_i$, containing $10000$ graphs. All the graphs in $\mathcal{D}_i$ have the same size and have been created analogously as the previous experiment. We then constructed a query graph $q_i$ of the same size of the graphs in $\mathcal{D}_i$, and we measured the query time $t_i$ as we did for the previous experiment. We reported in Figure \[fig:scalability2\], the computed $t_i$ versus the size of the graph query $q_i$. Figures \[fig:scalability1\], \[fig:scalability2\] provide us an experimental verification of the scalability of our approach both in the size of the database and in the size of the query graph.
![This plot shows the time expressed in seconds to perform a query as we increase the size of the database. Two-stage is referred to the use of our pre-summarization approach to address the graph search problem, while one-stage is referred to the search on the original graphs.[]{data-label="fig:scalability1"}](imgs/time_db_query_1.png){width="0.5\linewidth"}
![This plot shows the time (expressed in minutes) for retrieving the top-$k$ graphs in a database composed of 10000 graphs as we increase the dimension of the query graph. Two-stage is referred to the use of our pre-summarization approach to address the graph search problem, while one-stage is referred to the search on the original graphs.[]{data-label="fig:scalability2"}](imgs/time_db_query_2.png){width="0.5\linewidth"}
Conclusions
===========
In this work, we introduced a graph summarization approach based on the Regularity Lemma, which provides us with a principled way to describe the essential structure of large graphs using a small amount of data. We have successfully validated our framework both on synthetic and real-world graphs showing that our algorithm surpasses the state-of-the-art graph summarization methods in terms of noise robustness. In the second part of the paper, we presented an algorithm to address the graph similarity search problem exploiting our summaries. In particular, the proposed method is tailored for efficiently dealing with databases containing a high number of large graphs, and, moreover, it is robust against noise, which is always presented in real-world data. This achievement seems of particular interest since, to the best of our knowledge, we are the first to devise a graph search algorithm which satisfies all the above requirements together. A weak point of our summarization algorithm is related to its time complexity, which prevents the application of our framework to networks of millions of nodes. This demands for the development of efficient approaches suited to deal with large sparse graphs. In this direction, it would be a good idea to develop a heuristic based on the version of the Weak Regularity Lemma introduced by Fox et al. [@Fox2018], as well as designing a distributed version of the regular decomposition algorithm introduced by Reittu et al. [@Reittu2018], who studied the linkage among the Regularity Lemma, the Stochastic Block Model and the Minimum Description Length. We think that the notion of regular partition will allow us to tackle the scalability issue faced by graph-based approaches [@Bunke2011; @Vento2015]. This would pave the way for a principled approach to massive network data analysis by combining modern graph theory and combinatorics with machine learning and pattern recognition.
References {#references .unnumbered}
==========
[^1]: $C_0$ has only a technical purpose: it makes it possible that all other classes have exactly the same number of vertices.
[^2]: The implementation is available from [https://github.com/MarcoFiorucci/graph-summarization-using-regular-partitions]( )
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We consider packing LP’s with $m$ rows where all constraint coefficients are normalized to be in the unit interval. The $n$ columns arrive in random order and the goal is to set the corresponding decision variables irrevocably when they arrive so as to obtain a feasible solution maximizing the expected reward. Previous $(1 - \epsilon)$-competitive algorithms require the right-hand side of the LP to be $\Omega (\frac{m}{\epsilon^2} \log \frac{n}{\epsilon})$, a bound that worsens with the number of columns and rows. However, the dependence on the number of columns is not required in the single-row case and known lower bounds for the general case are also independent of $n$.
Our goal is to understand whether the dependence on $n$ is required in the multi-row case, making it fundamentally harder than the single-row version. We refute this by exhibiting an algorithm which is $(1 - \epsilon)$-competitive as long as the right-hand sides are $\Omega (\frac{m^2}{\epsilon^2} \log \frac{m}{\epsilon})$. Our techniques refine previous PAC-learning based approaches which interpret the online decisions as linear classifications of the columns based on sampled dual prices. The key ingredient of our improvement comes from a non-standard covering argument together with the realization that only when the columns of the LP belong to few 1-d subspaces we can obtain small such covers; bounding the size of the cover constructed also relies on the geometry of linear classifiers. General packing LP’s are handled by perturbing the input columns, which can be seen as making the learning problem more robust.
author:
- |
Marco Molinaro\
Carnegie Mellon
- |
R. Ravi\
Carnegie Mellon
bibliography:
- 'online-lp.bib'
title: Geometry of Online Packing Linear Programs
---
Introduction
============
Traditional optimization models usually assume that the input is known a priori. However, in most applications, the data is either revealed over time or only coarse information about the input is known, often modeled in terms of a probability distribution. Consequently, much effort has been directed towards understanding the quality of solutions that can be obtained without full knowledge of the input, which led to the development of online and stochastic optimization [@borodinbook; @stocprogbook]. Emerging problems such as allocating advertisement slots to advertisers and yield management in the internet are of inherent online nature and have further accelerated this development [@agrawal].
Linear programming is arguably the most important and thus well-studied optimization problem. Therefore, understanding the limitations of solving linear programs when complete data is not available is a fundamental theoretical problem with a slew of applications, including the ad allocation and yield management problems above. Indeed, a simple linear program with one uniform knapsack constraint, the Secretary Problem, was one of the first online problems to be considered and an optimal solution was already obtained by the early 60’s [@Dynkin; @GilbertMosteller]. Although the single knapsack case is currently well-understood under different models of how information is revealed [@BabaioffSurvey], much less is known about problems with multiple knapsacks and only recently algorithms with solution guarantees have been developed [@feldman; @agrawal; @Devanur11].
[*The Model.*]{} We study online packing LP’s in the *random permutation model*. Consider a fixed but unknown LP with $n$ columns $a^1, a^2, \ldots, a^n \in [0,1]^m$, whose associated variables are constrained to be in $[0,1]$, and $m$ packing constraints: $$\begin{aligned}
{\ensuremath{\textrm{OPT}}}= \max \sum_{t = 1}^n \pi_t x_t \notag \\
\sum_{t = 1}^n a^t x_t \le B \label{eq:LP} \tag{LP}\\
x_t \in [0,1] \,. \notag
\end{aligned}$$ Columns are presented in uniformly random order, and when a column is presented we are required to irrevocably choose the value of its corresponding variable. We assume that the number of columns $n$ is known.[^1] The goal is to obtain a feasible solution while maximizing its value. We use ${\ensuremath{\textrm{OPT}}}$ to denote the optimum value of the (offline) LP.
By scaling down rows as necessary, we assume without loss of generality that all entries of $B$ are the same, which we also denote with some overload of notation by $B$. Due to the packing nature of the problem, we also assume without loss of generality that all the $\pi_t$’s are non-negative and all the $a^t$’s are non-zero: we can simply ignore columns which do not satisfy the first property and always set to 1 the variables associated to the remaining columns which do not satisfy the second property. Finally, we assume that the columns $a^t$’s are in *general position*: for all $p \in \mathbb{R}^m$, there are at most $m$ different $t \in [n]$ such that $\pi_t = p a^t$. Notice that perturbing the input randomly by a tiny amount achieves this property with probability one, while the effect of the perturbation is absorbed in our approximation guarantees [@DevanurHayes09; @agrawal].
Applications {#applications .unnumbered}
------------
Write about applications on online revenue management and resource allocation (like ads) to motivate the problem.
[*Related work.*]{} The random permutation model has grown in popularity [@GoelMehta08; @DevanurHayes09; @BabaioffSurvey] since it avoids strong lower bounds of the pessimistic adversarial-order model [@BuchbinderMOR] while still capturing the lack of total information a priori. Different online problems have already been studied in this model, including bin-packing [@kenyon], matchings [@KVV; @GoelMehta08], the AdWords Problem [@DevanurHayes09] and different generalizations of the Secretary Problem [@BabaioffSurvey; @weightsSecretary; @submodularSecretary; @soto; @sungjin]. Closest to our work are packing problems with a single knapsack constraint. In [@kleinberg], Kleinberg considered the $B$-Choice Secretary Problem, where the goal is to select at most $B$ items coming online in random order to maximize profit. The author presented an algorithm with competitive ratio $1 - O(1/\sqrt{B})$ and showed that $1 - \Omega(1/\sqrt{B})$ is best possible. Generalizing the $B$-Choice Secretary Problem, Babaioff et al. [@babaioff] considered the online knapsack problem and presented a $(1/10e)$-competitive algorithm. Notice that in both cases the competitive ratio does not depend on $n$.
Despite all these works, the first result for more general online packing LP’s here was only recently obtained by Feldman et al. [@feldman] and Agrawal et al. [@agrawal]. The first paper presents an algorithm that obtains with high probability a solution of value at least $(1 - \epsilon) {\ensuremath{\textrm{OPT}}}$ whenever $B \ge \Omega(\frac{m \log n}{\epsilon^3})$ and ${\ensuremath{\textrm{OPT}}}\ge \Omega(\frac{\pi_{\max} m \log n}{\epsilon})$, where $\pi_{\max}$ is the largest profit. In the second paper, the authors present an algorithm which obtains a solution of expected value at least $(1 - \epsilon) {\ensuremath{\textrm{OPT}}}$ under the weaker assumptions $B \ge \Omega\left(\frac{m}{\epsilon^2} \log \frac{n}{\epsilon}\right)$ or ${\ensuremath{\textrm{OPT}}}\ge \Omega\left(\frac{\pi_{\max} m^2}{\epsilon^2} \log \frac{n}{\epsilon}\right)$. One other way of stating this result is that the algorithm obtains a solution with competitive ratio $1 - O(\sqrt{\frac{m \log(n) \log B}{B}})$; notice that the guarantee degrades as $n$ increases. The current lower bound on $B$ to allow $(1 - \epsilon)$-competitive algorithms is $B \ge \frac{\log m}{\epsilon^2}$, also presented in [@agrawal]. We remark that these algorithms actually work for more general allocation problems, where a set of columns representing various options arrive at each step and the solution may choose at most one of the options.
Both of the above algorithms use a connection between solving the online LP and PAC-learning [@cucker] a linear classification of its columns, which was initiated by Devanur and Hayes [@DevanurHayes09] in the context of the AdWords problem. Here we further explore this connection and our improved bounds can be seen as a consequence of making the learning algorithm more robust by suitably changing the input LP. Robustness is a topic well-studied in learning theory [@devroye; @partha], although existing results do not seem to apply directly to our problem. We remark that a component of robustness more closely related to the standard PAC-learning literature is used in [@DevanurHayes09].
In recent work, Devanur et al. [@Devanur11] consider the weaker *i.i.d. model* for the general allocation problem. While in the random permutation model one can think that columns are sampled without replacement, in the i.i.d. model they are sampled with replacement. Making use of the independence between samples, Devanur et al. substantially improve requirement on $B$ to $\Omega(\frac{\log (m/\epsilon)}{\epsilon^2})$ while showing that the lower bound $\Omega\left(\frac{\log m}{\epsilon^2}\right)$ still holds in this model. We remark, however, that these models can present very different behaviors: as a simple example, consider an LP with $n$ columns, $m = 1$ constraints and budget $B = 1$, where only one of the columns has $\pi_1 = a^1 = 1$ and all others have $\pi_i = a^i = 0$; in the random permutation model the expected value of the optimal solution is 1, while in the i.i.d. model this value is $1 - (1 - 1/n)^n \rightarrow 1 - 1/e$. The competitiveness of the algorithm of [@Devanur11] under the permutation model is still unknown and was left as an open problem by the authors.
[*Our results.*]{} Our focus is to understand how large $B$ is required to be in order to allow $(1 - \epsilon)$-competitive algorithms. In particular, the requirements for $B$ in the above algorithms degrade as the number of columns in the LP increases, while the the lower bound does not. With the trend of handling LP’s with larger number of columns (e.g. columns correspond to the keywords in the ad allocation problem, which in turn correspond to visits of a search engine’s webpage), this gap is very unsatisfactory from a practical point of view. Furthermore, given that guarantees for the single knapsack case do not depend on the number of columns, it is important to understand if the multi-knapsack case is fundamentally more difficult. In this work, we give a precise indication of why the latter problem was resistant to arguments used in the single knapsack case, and overcome this difficulty to exhibit an algorithm with dimension-independent guarantee.
We show that a modification of the [<span style="font-variant:small-caps;">DPA</span>]{}algorithm from [@agrawal] that we call *Robust [DPA]{}* obtains a $(1 -\epsilon)$-competitive solution for online packing LP’s with $m$ constraints in the random permutation model whenever $B \ge \Omega(\frac{m^2}{\epsilon^2} \log \frac{m}{\epsilon})$. Another way of stating this result is that the algorithm has competitive ratio $1 - O(m \sqrt{\log B}/\sqrt{B})$. Contrasting to previous results, our guarantee does not depend on $n$ and in the case $m = 1$ matches the bounds for the $B$-Choice Secretary Problem up to lower order terms. We finally remark that we can replace the requirement $B \ge \Omega(\frac{m^2}{\epsilon^2} \log \frac{m}{\epsilon})$ by ${\ensuremath{\textrm{OPT}}}\ge \Omega(\frac{\pi_{\max} m^3}{\epsilon^2} \log \frac{m}{\epsilon})$ exactly as done in Section 5.1 of [@agrawal].
[*High-level outline.*]{} As mentioned before, we use the connection between solving an online LP and PAC-learning a good linear classification of its columns; in order to obtain the improved guarantee, we focus on tightening the bounds for the generalization error of the learning problem. More precisely, solving the LP can be seen as classifying the columns into 0/1, which corresponds to setting their associated variable to 0/1. Consider a family $\mathcal{X} \subseteq \{0,1\}^n$ of linear classifications of the columns. Our algorithms sample a set $S$ of columns and learn a classification $x^S \in \mathcal{X}$ which is “good” for the columns $S$ (i.e., obtains large proportional revenue while not filling up the proportionally scaled budget too much). The goal is to upper bound the probability that $x^S$ is not good for the whole LP; this is typically done via a union bound over the classifications in $\mathcal{X}$ [@DevanurHayes09; @agrawal].
To obtain improved guarantees, we refine this bound using an argument akin to covering: we consider *witnesses* (Section \[sec:witness\]), which are representatives of groups of ‘similar’ bad classifications that can be used to bound the probability that *any* classification in the group is learned; for that we need to use a non-standard measure of similarity between classifications which is based on the budget of the LP. The problem is that, when the columns $(\pi_t, a^t)$’s do not lie in a two-dimensional subspace of $\mathbb{R}^{m}$, the set $\mathcal{X}$ may contain a large number of mutually dissimilar bad classifications; this is a roadblock for obtaining a small set of witnesses. In stark contrast, when these columns do lie in a two-dimensional subspace (e.g., $m = 1$), these classifications have a much nicer structure which indeed allows a small set of witnesses. This indicates that the latter learning problem is intrinsically more robust than the former, which seem to precisely capture the increased difficulty in obtained good bounds for the multi-row case.
Motivated by this discussion we first consider LP’s whose columns $a^t$’s lie in *few* one-dimensional subspaces (Section \[sec:otp\]). For each of these subspaces, we are able to approximate the classifications induced in the columns lying in the subspace by considering a small subset of the induced classifications; patching together these partial classifications gives us a witness set for $\mathcal{X}$. However, this strategy as stated does not make use of the fact that the subspaces are embedded in an $m$-dimensional space, and hence leads to large witness sets. By establishing a connection between the “useful” patching possibilities with faces of a hyperplane arrangement in $\mathbb{R}^m$ (Lemma \[lemma:sizeP\]), we are able to make use of the dimension of the host space and exhibit witness sets of much smaller sizes, which leads to improved bounds.
For a general packing LP, we perturb the columns $a^t$’s to make them lie in few one-dimensional subspaces that form an ‘$\epsilon$-net’ of the space, while not altering the feasibility and optimality of the LP by more than a $(1 \pm \epsilon)$ factor (Section \[sec:rotp\]). Finally, we tighten the bound by using the idea of periodically recomputing the classification, following [@agrawal] (Section \[sec:rdpa\]).
OTP for almost 1-dim columns {#sec:otp}
============================
In this section we describe and analyze the algorithm [<span style="font-variant:small-caps;">OTP</span>]{}(One-Time Pricing) over LP’s whose columns are contained in few 1-dimensional subspaces of $\mathbb{R}^{m}$. The overall goal is to find an appropriate dual (perhaps infeasible) solution $p$ for and use it to classify the columns of the LP. More precisely, given $p \in \mathbb{R}^m$, we define $x(p)_t = 1$ if $\pi_t > p a^t$ and $x(p)_t = 0$ otherwise. Thus, $x(p)$ is the result of classifying the columns $(\pi_t, a^t)$’s with the homogeneous hyperplane in $\mathbb{R}^{m+1}$ with normal $(-1, p)$. The motivation behind this classification is that it selects the columns which have positive reduced cost with respect to the dual solution $p$, or alternatively, it solves to optimality the Lagrangian relaxation using $p$ as multipliers.
[*Sampling LP’s.*]{} In order to obtain a good dual solution $p$ we use the (random) LP consisting on the first $s$ columns of with appropriately scaled right-hand side.
$$\begin{aligned}
\tag{$(s, \delta)$-LP} \label{eq:sdLP}
\max & \sum_{t = 1}^{s} \pi_{\sigma(t)} x_{\sigma(t)} \\
&\sum_{t = 1}^{s} a^{\sigma(t)} x_{\sigma(t)} \le \frac{s}{n} \delta B \notag \\
&x_{\sigma(t)} \in [0,1] \ \ \ \ t = 1, \ldots, s \notag .
\end{aligned}$$
$$\begin{aligned}
\tag{$(s, \delta)$-Dual} \label{eq:sdDual}
\min \ & \frac{s}{n} \delta B \sum_{i = 1}^{m} p_i + \sum_{t = 1}^s \alpha_{\sigma(t)} \\
& p a^{\sigma(t)} + \alpha_{\sigma(t)} \ge \pi_{\sigma(t)} \ \ \ \ t = 1, \ldots, s \notag \\
& p \ge 0 \notag \\
& \alpha \ge 0 \notag.
\end{aligned}$$
Here $\sigma$ denotes the random permutation of the columns of the LP. We use ${\ensuremath{\textrm{OPT}}}(s,\delta)$ to denote the optimal value of $(s, \delta)$-LP and ${\ensuremath{\textrm{OPT}}}(s)$ to denote the optimal value of $(s, 1)$-LP.
The static pricing algorithm [<span style="font-variant:small-caps;">OTP</span>]{}of [@agrawal] can then be described as follows.[^2]
1. Wait for the first $\epsilon n$ columns of the LP (indexed by $\sigma(1), \sigma(2), \ldots, \sigma(\epsilon n)$) and solve $(\epsilon n, 1 - \epsilon)$-Dual. Let $(p, \alpha)$ be the obtained dual optimal solution.
2. Use the classification given by $p$ as above by setting $x_{\sigma(t)} = x(p)_{\sigma(t)}$ for $t = \epsilon n + 1, \epsilon n + 2, \ldots$ for as long as the solution obtained remains valid. From this point on set all further variables to zero.
Note that by definition this algorithm outputs a feasible solution with probability one. Our goal is then to analyze the quality of the solution produced, ultimately leading to the following theorem.
\[thm:otp\] Fix $\epsilon \in (0,1]$. Suppose that there are $K \ge m$ 1-dim subspaces of $\mathbb{R}^m$ containing the columns $a^t$’s and that $B \ge \Omega\left(\frac{m}{\epsilon^3} \log \frac{K}{\epsilon}\right)$. Then algorithm [<span style="font-variant:small-caps;">OTP</span>]{}returns a feasible solution with expected value at least $(1 - 5\epsilon) {\ensuremath{\textrm{OPT}}}$.
Let $S = \{\sigma(1), \ldots, \sigma(\epsilon n)\}$ be the (random) index set of the columns sampled by [<span style="font-variant:small-caps;">OTP</span>]{}. We use $p^S$ to denote the optimal dual solution obtained by [<span style="font-variant:small-caps;">OTP</span>]{}; notice that $p^S$ is completely determined by $S$. To simplify the notation, we also use $x^S$ to denote $x(p^S)$.
Notice that, for all the scenarios where $x^S$ is feasible, the solution returned by [<span style="font-variant:small-caps;">OTP</span>]{}is identical to $x^S$ with its components $x^S_{\sigma(1)}, \ldots, x^S_{\sigma(\epsilon n)}$ set to zero. Given this observation and the fact that ${\mathbb{E}}[\sum_{t \leq \epsilon n} \pi_{\sigma(t)} x^S_{\sigma(t)}] \leq \epsilon {\ensuremath{\textrm{OPT}}}$, one can prove that the following lemma implies Theorem \[thm:otp\].
\[lemma:goodOtp\] Fix $\epsilon \in (0,1]$. Suppose that there are $K \ge m$ 1-dim subspaces of $\mathbb{R}^m$ containing the columns $a^t$’s and that $B \ge \Omega\left(\frac{m}{\epsilon^3} \log \frac{K}{\epsilon}\right)$. Then with probability at least $(1 - \epsilon)$, $x^S$ is a feasible solution for with value at least $(1 - 3\epsilon) {\ensuremath{\textrm{OPT}}}$.
Connection to PAC learning
--------------------------
We assume from now on that $B \ge \Omega(\frac{m}{\epsilon^3} \log \frac{K}{\epsilon})$. Let $\mathcal{X} = \{x(p) : p \in \mathbb{R}^m_+\} \subseteq \{0,1\}^n$ denote the set of all possible linear classifications of the LP columns which can be generated by [<span style="font-variant:small-caps;">OTP</span>]{}. With slight overload in the notation, we identify a vector $x \in \{0,1\}^n$ with the subset of $[n]$ corresponding to its support.
Given a scenario, we say that $x^S$ is *bad* if it does not satisfy the properties of Lemma \[lemma:goodOtp\], namely $x^S$ is either infeasible or has value less than $(1 - 3\epsilon) {\ensuremath{\textrm{OPT}}}$. We say that $x^S$ is *good* otherwise.
As noted in previous work, since our decisions are made based on reduced costs it suffices to analyze the *budget occupation* (or complementary slackness) of the solution in order to understand its *value*. To make this precise, given $x \in \{0,1\}^n$ let $a_i(x) = \sum_{t \in x} a_i^t$ be its occupation of the $i$th budget and let $a^S_i(x) = \frac{1}{\epsilon}\sum_{t \in x \cap S} a_i^t$ be its appropriately scaled occupation of $i$th budget in the sampled LP (recall $|S| = \epsilon n$).
\[lemma:approximateCS\] Consider a scenario where $x^S$ satisfies: (i) for all $i \in [m]$, $a_i(x^S) \le B$ and (ii) for all $i \in [m]$ with $p^S_i > 0$, $a_i(x^S) \ge (1 - 3 \epsilon) B$. Then $x^S$ is good.
Moreover, since we are making decisions based on the *optimal* reduced cost for the sampled LP, our solution satisfies the above properties for the sampled LP.
\[lemma:sampleCS\] In every scenario, $x^S$ satisfies the following: (i) for all $i \in [m]$, $a_i^S(x^S) \le (1 - \epsilon)B$ and (ii) for every $i \in [m]$ with $p^S_i > 0$, $a_i^S(x^S) \ge (1 - 2\epsilon) B$.
Given that $a_i(x) = {\mathbb{E}}[a^S_i(x)]$ for all $x$, the idea is to use concentration inequalities to argue that the conditions in Lemma \[lemma:approximateCS\] hold with good probability. Although concentration of $a^S_i(x)$ for *fixed* $x$ can be achieved via Chernoff-type bounds, the quantity $a^S_i(x^S)$ has undesired correlations; obtaining an effective bound is the main technical contribution of this paper.
For a given scenario, we say that $x \in \mathcal{X}$ can be *badly learned for budget $i$* if either (i) $a_i^S(x) \le (1 - \epsilon) B$ and $a_i(x) > B$ or (ii) $a^S_i(x) \ge (1 - 2\epsilon) B$ and $a_i(x) < (1 - 3 \epsilon) B$.
Essentially these are the classifications which look good for the sampled $(\epsilon n, 1- \epsilon)$-LP but are actually bad for . Putting Lemmas \[lemma:approximateCS\] and \[lemma:sampleCS\] together and unraveling the definitions gives that $$\Pr\left(x^S \textrm{ is bad}\right) \le \Pr\left(\bigvee_{i \in [m], x \in \mathcal{X}} x \textrm{ can be badly learned for budget } i\right). \label{eq:badlyLearned}$$ Notice that the right-hand side of this inequality does not depend on $x^S$, it is only a function of how skewed $a_i^S(x)$ is as compared to its expectation $a_i(x)$.
Usually the right-hand side in the previous equation is upper bounded by taking a union bound over all its terms [@agrawal]. Unfortunately this is too wasteful: when $x$ and $x'$ are “similar” there is a large overlap between the scenarios where $a_i^S(x)$ is skewed and those where $a_i^S(x')$ is skewed. In order to obtain improved guarantees, we introduce in the next section a new way of bounding the right-hand side of the above expression.
Similarity via witnesses {#sec:witness}
------------------------
First, we partition the classifications which can be badly learned for budget $i$ into two sets, depending on why they are bad: for $i \in [m]$, let $\mathcal{X}_i^+ = \{x \in \mathcal{X} : a_i(x) > B\}$ and $\mathcal{X}_i^- = \{x \in \mathcal{X} : a_i(x) < (1 - 3 \epsilon) B\}$. In order to simplify the notation, given a set $x$ we define $\operatorname{skewm}_i(\epsilon, x)$ to be the event that $a^S_i(x) \le (1 - \epsilon) B$ and $\operatorname{skewp}_i(\epsilon, x)$ to be the event that $a^S_i(x) \ge (1 - 2\epsilon) B$. Notice that if $x \in \mathcal{X}_i^+$, then $\operatorname{skewm}_i(\epsilon, x)$ is the event that $a_i^S(x)$ is significantly smaller than its expectation (skewed in the minus direction), while for $x \in \mathcal{X}_i^-$ $\operatorname{skewp}_i(\epsilon, x)$ is the event that $a_i^S(x)$ is significantly larger than its expectation (skewed in the plus direction). These definitions directly give the equivalence $$\label{eq:witness1}
\Pr\left(\bigvee_{i,x \in \mathcal{X}} x \textrm{ can be badly learned for budget } i\right) = \Pr\left(\bigvee_{i, x \in \mathcal{X}_i^+} \operatorname{skewm}_i(\epsilon, x) \vee \bigvee_{i, x \in \mathcal{X}_i^-} \operatorname{skewp}_i(\epsilon, x)\right).$$
In order to introduce the concept of witnesses, consider two sets $x,x'$, say, in $\mathcal{X}_i^+$. Take a subset $w \subseteq x \cap x'$; the main observation is that, since $a^t \ge 0$ for all $t$, for all scenarios we have $a_i^S(w) \le a_i^S(x)$ and $a_i^S(w) \le a_i^S(x')$. In particular, the event $\operatorname{skewm}_i(\epsilon, x) \vee \operatorname{skewm}_i(\epsilon,x')$ is contained in $\operatorname{skewm}(\epsilon, w)$. The set $w$ serves as a witness for scenarios which are skewed for either $x$ or $x'$; if additionally $a_i(w)$ reasonably larger than $(1 - \epsilon) B$, we can then use concentration inequalities over $\operatorname{skewm}_i(\epsilon, w)$ in order to bound probability of $\operatorname{skewm}(\epsilon, x) \vee \operatorname{skewm}(\epsilon,x')$. This ability of bounding multiple terms of the right-hand side of simultaneously is what gives an improvement over the naive union bound.
\[def:witness\] We say that $\mathcal{W}_i^+$ is a *witness set* for $\mathcal{X}_i^+$ if: (i) for all $w \in \mathcal{W}_i^+$, $a_i(w) \ge (1 - \epsilon/2) B$ and (ii) for all $x \in \mathcal{X}_i^+$ there is $w \in \mathcal{W}_i^+$ contained in $x$. Similarly, we say that $\mathcal{W}_i^-$ is a *witness set* for $\mathcal{X}_i^-$ if: (i) for all $w \in \mathcal{W}_i^-$, $a_i(w) \le (1 - 3 \epsilon/2) B$ and (ii) for all $x \in \mathcal{X}_i^-$ there is $w \in \mathcal{W}_i^-$ containing $x$.
As indicated by the previous discussion, given witness sets $\mathcal{W}_i^+$ and $\mathcal{W}_i^-$ for $\mathcal{X}_i^+$ and $\mathcal{X}_i^-$, we directly get the bound $$\begin{aligned}
\label{eq:witness2}
\Pr\left(\bigvee_{i, x \in \mathcal{X}_i^+} \operatorname{skewm}(\epsilon, x) \vee \bigvee_{i, x \in \mathcal{X}_i^-} \operatorname{skewp}(\epsilon, x)\right) \le \Pr\left(\bigvee_{i, w \in \mathcal{W}_i^+} \operatorname{skewm}(\epsilon, w) \vee \bigvee_{i, w \in \mathcal{W}_i^-} \operatorname{skewp}(\epsilon, w)\right).
\end{aligned}$$ Putting together the last three displayed equations and using Chernoff-type bounds, we can get an upper estimate on the probability that $x^S$ is bad in terms of the size of witnesses sets.
\[lemma:badWitness\] Suppose that, for all $i \in [m]$, there are witness sets for $\mathcal{X}_i^+$ and $\mathcal{X}_i^-$ of size at most $M$. Then $\Pr(x^S \textrm{ is bad }) \le 8mM \exp\left(-\frac{\epsilon^3 B}{33}\right)$.
One natural choice of a witness set for, say, $\mathcal{X}_i^+$ is the collection of all of its minimal sets; unfortunately this may not give a witness set of small enough size. But notice that a witness set need not be a subset of $\mathcal{X}_i^+$ (or even $\mathcal{X}$). Allowing elements outside $\mathcal{X}_i^+$ gives the flexibility of obtaining witnesses which are associated to multiple “similar” minimal elements of $\mathcal{X}_i^+$, which is effective in reducing the size of witness sets.
Small witness sets for almost 1-dim columns {#sec:goodWitness}
-------------------------------------------
Given the previous lemma, our task is to find small witness sets. Unfortunately, when the $(\pi_t, a^t)$’s lie in a space of dimension at least 3, $\mathcal{X}_i^+$ and $\mathcal{X}_i^-$ may contain many ($\Omega(n)$) disjoint sets (see Figure \[fig:disjoint3d\]), which shows that in general we cannot find small witness sets directly. This sharply contrasts with the case where the $(\pi_t, a^t)$’s lie in a 2-dimensional subspace of $\mathbb{R}^{m + 1}$, where one can show that $\mathcal{X}$ is a union of 2 chains with respect to inclusion. In the special case where the $a^t$’s lie in a 1-dimensional subspace of $\mathbb{R}^m$, we show that $\mathcal{X}$ is actually a single chain (Lemma \[lemma:chain\]) and therefore we can take $\mathcal{W}_i^+$ as *the* minimal set of $\mathcal{X}_i^+$ and $\mathcal{W}_i^-$ as *the* maximal set of $\mathcal{X}_i^-$.
Due to the above observations, we focus on LP’s whose $a^t$’s lie in few 1-dimensional subspaces. In this case, $\mathcal{X}_i^+$ and $\mathcal{X}_i^-$ are sufficiently well-behaved so that we can find small (independent of $n$) witness sets.
\[lemma:witness2Dim\] Suppose that there are $K \ge m$ 1-dimensional subspaces of $\mathbb{R}^m$ which contain the $a^t$’s. Then there are witness sets for $\mathcal{X}_i^+$ and $\mathcal{X}_i^-$ of size at most $(O(\frac{K}{\epsilon} \log \frac{K}{\epsilon}))^m$.
Assuming the hypothesis of the lemma, partition the index set $[n]$ into $C_1, C_2, \ldots, C_K$ such that for all $j \in [K]$ the columns $\{a^t\}_{t \in C_j}$ belong to the same 1-dimensional subspace. Equivalently, for each $j \in [K]$ there is a vector $c^j$ of $\ell_\infty$-norm 1 such that for all $t \in C_j$ we have $a^t = \|a^t\|_{\infty} c^j$. An important observation is that now we can order the columns (locally) by the ratio of profit over budget occupation: without loss of generality assume that for all $j \in [K]$ and $t, t' \in C_j$ with $t < t'$, we have $\frac{\pi_t}{\|a^t\|_{\infty}} \ge \frac{\pi_{t'}}{\|a^{t'}\|}_{\infty}$.[^3]
Given a classification $x$, we use $x|_{C_j}$ to denote its projection onto the coordinates in $C_j$; so $x|_{C_j}$ is the induced classification on columns with indices in $C_j$. Similarly, we define $\mathcal{X}|_{C_j} = \{x|_{C_j} : x \in \mathcal{X}\}$ as the set of all classifications induced in the columns in $C_j$. The most important structure that we get from working with 1-d subspaces, which is implied by the local order of the columns, is the following.
\[lemma:chain\] For each $j \in [K]$, the sets in $\mathcal{X}|_{C_j}$ are prefixes of $C_j$.
To simplify the notation fix $i \in [m]$ for the rest of this section, so we aim at providing witness sets for $\mathcal{X}_i^+$ and $\mathcal{X}_i^-$. The idea is to group the classifications according to their budget occupation caused by the different column classes $C_j$’s. To make this formal, start by covering the interval $[0, B + m]$ with intervals $\{I_\ell\}_{\ell \in L}$, where $I_0 = [0, \frac{\epsilon B}{4K})$ and $I_\ell = [\frac{\epsilon B}{4K} (1 + \frac{\epsilon}{4})^{\ell - 1}, \frac{\epsilon B}{4K} (1 + \frac{\epsilon}{4})^\ell)$ for $\ell > 0$ and $L = \{0, \ldots, \lceil \log_{1 + \epsilon/4} \frac{8K}{\epsilon} \rceil\}$ (note that since $B \geq m$, we have $B + m \leq 2B$). Define $\mathcal{B}_{i,j}^\ell$ as the set of partial classifications $y \in \mathcal{X}|_{C_j}$ whose budget occupation $a_i(y)$ lies in the interval $I_\ell$. For ${v}\in L^K$ define the family of classifications $\mathcal{B}_i^{{v}} = \{(y^1, y^2, \ldots, y^K) : y^j \in \mathcal{B}_{i,j}^{{v}_j}\}$. The $\mathcal{B}_i^{{v}}$’s then provide the desired grouping of the classifications. Note that the $\mathcal{B}_i^{{v}}$’s may include classifications not in $\mathcal{X}$ and may not include classifications in $\mathcal{X}$ which have occupation $a_i(.)$ greater than $B + m$.
Now consider a non-empty $\mathcal{B}_i^{{v}}$. Let $\underline{w}_i^{{v}}$ be the inclusion-wise smallest element in $\mathcal{B}_i^{{v}}$. Notice that such unique smallest element exists: since $\mathcal{X}|_{C_j}$ is a chain, so is $\mathcal{B}_{i,j}^{{v}_j}$, and hence $\underline{w}_i^{{v}}$ is the product (over $j$) of the smallest elements in the sets $\{\mathcal{B}_{i,j}^{{v}_j}\}_j$. Similarly, let $\overline{w}_i^{{v}}$ denote the largest element in $\mathcal{B}_i^{{v}}$. Intuitively, $\underline{w}^{{v}}_i$ and $\overline{w}^{{v}}_i$ will serve as witnesses for all the sets in $\mathcal{B}_i^{{v}}$.
Finally, define the witness sets by adding the $\underline{w}_i^{{v}}$ and $\overline{w}_i^{{v}}$’s of appropriate size corresponding to meaningful $\mathcal{B}_i^{{v}}$’s: set $\mathcal{W}_i^+ = \{\underline{w}_i^{{v}} : {v}\in L^K, \mathcal{B}_i^{{v}} \cap \mathcal{X} \neq \emptyset, a_i(\underline{w}_i^{{v}}) \ge (1 - \epsilon/2) B\}$ and $\mathcal{W}_i^- = \{\overline{w}_i^{{v}} : {v}\in L^K, \mathcal{B}_i^{{v}} \cap \mathcal{X} \neq \emptyset, a_i(\overline{w}_i^{{v}}) \le (1 - 3\epsilon/2) B\}$.
It is not too difficult to see that, say, $\mathcal{W}_i^+$ is a witness set for $\mathcal{X}_i^+$: If $x \in \mathcal{X}_i^+$ belongs to some $\mathcal{B}_i^{{v}}$, then $\underline{w}_i^{{v}}$ belongs to $\mathcal{W}_i^+$ and is easily shown to be a witness for $x$. However, if $x$ does not belong to any $\mathcal{B}_i^{{v}}$, by having too large $a_i(x)$, the idea is to find $x' \subseteq x$ which belongs to some $\mathcal{B}_i^{{v}}$ *and* to $\mathcal{X}$, and then use $\underline{w}_i^{{v}}$ as a witness for $x$. We note that considering $B_i^{{v}}$’s for side lengths at most $B + m$ and only adding witnesses for $B_i^{{v}}$’s which intersect $\mathcal{X}$ are crucially used for bounding the size of $\mathcal{W}_i^+$ and $\mathcal{W}_i^-$.
\[lemma:wWitness\] The sets $\mathcal{W}_i^+$ and $\mathcal{W}_i^-$ are witness sets for $\mathcal{X}_i^+$ and $\mathcal{X}_i^-$.
[*Bounding the size of witness sets.*]{} Clearly the witness sets $\mathcal{W}_i^+$ and $\mathcal{W}_i^-$ have size at most $|L|^K$. Although this size is independent of $n$, it is still unnecessarily large since it only uses locally (for each $C_j$) the fact that $\mathcal{X}$ consists of linear classifications; in particular, it does not use the dimension of the ambient space $\mathbb{R}^m$. Now we sketch the argument for an improved bound, and details are provided in the appendix.
First notice that the partial classification $x(p)|_{C_j}$ is completely defined by the value $pc^j$. Thus, if $J \subseteq [K]$ is such that the directions $\{c^j\}_{j \in J}$ form a basis of $\mathbb{R}^m$ then knowing $pc^j$ for all $j \in J$ completely determines the whole classification $x(p)$. Similarly, if we know that $x(p)|_{C_j} \in \mathcal{B}_i^{{v}_j}$ for all $j \in J$, then for each $j \notin J$ we should have fewer possible $\mathcal{B}_i^{{v}_j}$’s where the partial classification $x(p)|_{C_j}$ can belong to; this indicates that some of the sets $\{\mathcal{B}_i^{{v}}\}_{{v}\in L^K}$ do not contain any element from $\mathcal{X}$, which implies a reduced size for the witness sets.
In order to capture this idea, we focus on the space of dual vectors $p$ and define the sets $P^\ell_j = \{p \in {{\ensuremath{\mathbb{R}}}}_+^m : x(p)|_{C_j} \in \mathcal{B}_{i,j}^{\ell}\}$ and $P^v = \{p \in {{\ensuremath{\mathbb{R}}}}_+^m : x(p) \in \mathcal{B}_i^v\}$. Notice that $P^v = \cap_j P^{v_j}_j$ and that $\mathcal{B}^v_i$ is empty iff $P^v$ is. The main step is to show that each $P^\ell_j$ is a polyhedron with “few” facets, which uses the definition of $x(p)$ and Lemma \[lemma:chain\]. We then consider the arrangement of the hyperplanes which are facet-defining for the $P_j^\ell$’s and conclude that the $P^v$’s are given by unions of the cells in this arrangement; classical bounds on the number of cells in a hyperplane arrangement in ${{\ensuremath{\mathbb{R}}}}^m$ then allow us to upper bound the number of nonempty $P^v$’s. This gives the following.
\[lemma:sizeP\] At most $(O(\frac{K}{\epsilon} \log \frac{K}{\epsilon}))^m$ of the $\mathcal{B}_i^{{v}}$’s contain an element from $\mathcal{X}$.
This lemma implies that $\mathcal{W}_i^+$ and $\mathcal{W}_i^-$ each has size at most $(O(\frac{K}{\epsilon} \log \frac{K}{\epsilon}))^m$, which then proves Lemma \[lemma:witness2Dim\]. Finally, applying Lemma \[lemma:badWitness\] we conclude the proof of Lemma \[lemma:goodOtp\].
Robust [<span style="font-variant:small-caps;">OTP</span>]{} {#sec:robustOTP}
============================================================
\[sec:rotp\] In this section we consider with columns that may not belong to few 1-dimensional subspaces. Given the results of the previous section we would like to perturb the columns of this LP so that it belongs to few 1-dim subspaces, and such that an approximate solution for this perturbed LP is also an approximate solution for the original one. More precisely, we obtain a set of vectors $Q \subseteq \mathbb{R}^m$ and transform each column $a^t$ into a column $\tilde{a}^t$ which is a scaling of a vector in $Q$, and we let the rewards $\pi_t$ remain unchanged. The crucial observation is that the solutions of an LP are robust to slight changes in the the constraint matrix.
\[lemma:robust\] Consider real numbers $\pi_1, \ldots, \pi_n$ and vectors $a^1, \ldots, a^n$ and $\tilde{a}^1, \ldots, \tilde{a}^n$ in $\mathbb{R}^m_+$ such that $\|\tilde{a}^t - a^t\|_{\infty} \le \frac{\epsilon}{m + 1} \|a^t\|_{\infty}$. If $x$ is an $\epsilon$-approximate solution for with columns $(\pi_t, \tilde{a}^t)$ and right-hand side $(1 - \epsilon) B$, then $x$ is a $2\epsilon$-approximate solution for the LP .
[*Perturbing the columns.*]{} To simplify the notation, set $\delta = \frac{\epsilon}{m+1}$; for simplicity of exposition we assume that $1/\delta$ is integral. When constructing $Q$ we want the rays spanned by the each of its vectors to be “uniform” over $\mathbb{R}^m_+$. Using $\ell_\infty$ as normalization, let $Q$ be a $\delta$-net of the unit $\ell_\infty$ sphere, namely let $Q$ be the vectors in $\{0, \delta, 2\delta, 3\delta, \ldots, 1\}^m$ which have $\ell_\infty$ norm 1. Note that $|Q| = (O(\frac{m}{\epsilon}))^m$.
Given a vector $a^t \in {{\ensuremath{\mathbb{R}}}}^m$ we let $\tilde{a}^t = \|a^t\|_\infty q^t$, where $q^t$ is the vector in $Q$ closest (in $\ell_\infty$) to $\frac{a^t}{\|a^t\|_\infty}$. By definition of $Q$, for every vector $v \in \mathbb{R}^m$ with $\|v\|_\infty = 1$ there is a vector $q \in Q$ with $\|v - q\|_{\infty} \le \delta$. It then follows from positive homogeneity of norms that the $\tilde{a}^t$’s satisfy the property required in Lemma \[lemma:robust\]: $\|a^t - \tilde{a}^t\|_{\infty} \le \delta \|a^t\|_{\infty}$.
[*Algorithm Robust [<span style="font-variant:small-caps;">OTP</span>]{}.*]{} One way to think of the algorithm Robust [<span style="font-variant:small-caps;">OTP</span>]{}is that it works in two phases. First, it transforms the vectors $a^t$ into $\tilde{a}^t$ as described above. Then it returns the solution obtained by running the algorithm [<span style="font-variant:small-caps;">OTP</span>]{}over the LP with columns $(\pi_t, \tilde{a}^t)$ and right-hand side $(1 - \epsilon)B$. Notice that this algorithm can indeed be implemented to run in an online fashion.
Putting together the discussion in the previous paragraphs and the guarantee of [<span style="font-variant:small-caps;">OTP</span>]{}for almost 1-dim columns given by Theorem \[thm:otp\] with $K = |Q| = (O(\frac{m}{\epsilon}))^m$, we obtain the following theorem.
\[thm:rotp\] Fix $\epsilon \in (0,1]$ and suppose $B \ge \Omega\left(\frac{m^2}{\epsilon^3} \log \frac{m}{\epsilon}\right)$. Then algorithm Robust [<span style="font-variant:small-caps;">OTP</span>]{}returns a solution to the online with expected value at least $(1 - 10\epsilon) {\ensuremath{\textrm{OPT}}}$.
Robust [<span style="font-variant:small-caps;">DPA</span>]{} {#sec:rdpa}
============================================================
In this section we describe our final algorithm, which has an improved dependence on $1/\epsilon$. Following [@agrawal], the idea is to update the dual vector used in the classification as new columns arrive: we use the first $2^i \epsilon n$ columns to classify columns $2^i \epsilon n + 1, \ldots, 2^{i + 1} \epsilon n$. This leads to improved generalization bounds, which in turn give the reduced dependence on $1/\epsilon$. The algorithm Robust [<span style="font-variant:small-caps;">DPA</span>]{}(as the algorithm [<span style="font-variant:small-caps;">DPA</span>]{}) can be seen as a combination of solutions to multiple sampled LP’s, obtained via a modification of [<span style="font-variant:small-caps;">OTP</span>]{}denoted by $(s,\delta)$-[<span style="font-variant:small-caps;">OTP</span>]{}.
[*Algorithm $(s,\delta)$-[<span style="font-variant:small-caps;">OTP</span>]{}.*]{} This algorithm aims at solving the program $(2s,1)$-LP and can be described as follows: it finds an optimal dual solution $(p, \alpha)$ for $(s, (1 - \delta))$-LP and sets $x_{\sigma(t)} = x(p)_{\sigma(t)}$ for $t = s + 1, s + 2, \ldots, t' \le 2s$ such that $t'$ is the maximum one guaranteeing $\sum_{t = s + 1}^{2s} a^{\sigma(t)} x_{\sigma(t)} \le \frac{s}{n} B$.
The analysis of $(s,\delta)$-[<span style="font-variant:small-caps;">OTP</span>]{}is similar to the one employed for [<span style="font-variant:small-caps;">OTP</span>]{}. The main difference is that this algorithm tries to approximate the value of the *random* LP $(2s,1)$-LP. This requires a partition of the bad classifications which is more refined than simply splitting into $\mathcal{X}_i^+$ and $\mathcal{X}_i^-$, and witness sets need to be redefined appropriately. Nonetheless, using these ideas we can prove the following guarantee for $(s,\delta)$-[<span style="font-variant:small-caps;">OTP</span>]{}. Again let $S = \{\sigma(1), \sigma(2), \ldots, \sigma(s)\}$ be the random index set of the first $s$ columns of the LP, let $T = \{\sigma(s+1), \sigma(s+2), \ldots, \sigma(2s)\}$ and $U = S \cup T$. We use $\pi_U$ to denote the vector $(\pi_t)_{t \in U}$.
\[lemma:sDeltaOtp\] Suppose that there are $K \ge m$ 1-dim subspaces of $\mathbb{R}^m$ containing the columns $a^t$’s. Fix an integer $s$ and a real number $\delta \in (0,1/10)$ such that $\frac{\delta^2 s B}{n} \ge \Omega(m \ln \frac{K}{\delta})$. Then algorithm $(s,\delta)$-[<span style="font-variant:small-caps;">OTP</span>]{}returns a solution $x$ satisfying $a_i^T(x) \le B$ for all $i \in [m]$ with probability 1 and with expected value ${\mathbb{E}}[\pi_U x] \ge (1 - 3 \delta) {\mathbb{E}}[{\ensuremath{\textrm{OPT}}}(2s)] - {\mathbb{E}}[{\ensuremath{\textrm{OPT}}}(s)] - \delta^2 {\ensuremath{\textrm{OPT}}}$.
[*Algorithm Robust [<span style="font-variant:small-caps;">DPA</span>]{}.*]{} In order to simplify the description of the algorithm, we assume in this section that $\log (1/\epsilon)$ is an integer.
Again the algorithm Robust [<span style="font-variant:small-caps;">DPA</span>]{}can be thought as acting in two phases. In the first phase it converts the vectors $a^t$ into $\tilde{a}^t$, just as in the first phase of Robust [<span style="font-variant:small-caps;">OTP</span>]{}. In the second phase, for $i = 0, \ldots, \log (1/\epsilon) - 1$, it runs $(\epsilon 2^i n, \sqrt{\epsilon/2^i})$-[<span style="font-variant:small-caps;">OTP</span>]{}over with columns $(\pi_t, \tilde{a}^t)$ and right-hand side $(1 - \epsilon) B$ to obtain the solution $x^i$. The algorithm finally returns the solution $x$ consisting of the “union” of $x^i$’s: $x = \sum_i x^i$.
Note that the second phase corresponds exactly to using the first $\epsilon 2^i n$ columns to classify the columns $\epsilon 2^i n + 1, \ldots, \epsilon 2^{i+1} n$. This relative increase in the size of the training data for each learning problem allow us to reduce the dependence of $B$ on $\epsilon$ in each of the iterations, while the error from all the iterations telescope and are still bounded as before. Furthermore, notice that Robust [<span style="font-variant:small-caps;">DPA</span>]{}can be implemented to run online.
The analysis of Robust [<span style="font-variant:small-caps;">DPA</span>]{}reduces to that of $(s,\delta)$-[<span style="font-variant:small-caps;">OTP</span>]{}. That is, using the definition of the parameters of $(s,\delta)$-[<span style="font-variant:small-caps;">OTP</span>]{}used in Robust [<span style="font-variant:small-caps;">DPA</span>]{}and Lemma \[lemma:sDeltaOtp\], it is routine to check that the algorithm produces a feasible solution which has expected value $(1 - \epsilon) {\ensuremath{\textrm{OPT}}}$. This is formally stated in the following theorem.
\[thm:expValueDPA\] Fix $\epsilon \in (0,1/100)$ and suppose that $B \ge \Omega(\frac{m^2}{\epsilon^2} \ln \frac{m}{\epsilon})$. Then the algorithm Robust [<span style="font-variant:small-caps;">DPA</span>]{}returns a solution to the online LP with expected value at least $(1 - 50\epsilon) {\ensuremath{\textrm{OPT}}}$.
Open problems
=============
A very interesting open question is whether the techniques introduced in this work can be used to obtain improved algorithms for generalized allocation problems [@feldman]. The difficulty in this problem is that the classifications of the columns are not linear anymore; they essentially come from a conjunction of linear classifiers. Given this additional flexibility, having the columns in few 1-dimensional subspaces does not seem to impose strong enough properties in the classifications. It would be interesting to find the appropriate geometric structure of the columns in this case.
Of course a direct open question is to improve the lower or upper bound on the dependence on the right-hand side $B$ to obtain $(1 - \epsilon)$-competitive algorithms. One possibility is to investigate how much the techniques presented here can be pushed and what are their limitations. Another possibility is to analyze the performance of the algorithm from [@Devanur11] under the random permutation model.
![Case $m = 2$, columns $(\pi_t, a^t)$ equal to $(1,\sin(\frac{\pi}{4} + \delta t), \cos(\frac{\pi}{4} + \delta t))$ for sufficiently small $\delta > 0$, represented by black dots. Each segment $\{t, t + 1, \ldots, t + j\}$ can be linearly classified and hence belongs to $\mathcal{X}$. Furthermore, all segments $\{j 2 B, \ldots, (j + 1) 2 B\}$ belong to $\mathcal{X}_i^+$, which then contains $\Omega(\frac{n}{B})$ disjoint sets. Similar analysis holds for $\mathcal{X}_i^-$.[]{data-label="fig:disjoint3d"}](disjoint3d.pdf)
Bernstein inequality for sampling without replacement
=====================================================
\[lemma:addChernoff\] Let $Y = \{Y_1, \ldots, Y_n\}$ be a set of real numbers in the interval $[0,1]$ and let $0 < \epsilon < 1$. Let $S$ be a random subset of $Y$ of size $s$ and let $Y_S = \sum_{i \in S} Y_i$. Setting $\mu = \frac{1}{n} \sum_i Y_i$ and $\sigma^2 = \frac{1}{n}\sum_i (Y_i - \mu)^2$, we have that for every $\tau > 0$ $$\begin{aligned}
\Pr(|Y_S - s \mu| \ge \tau) \le 2 \exp \left( -\frac{\tau^2}{2 s \sigma^2 + \tau} \right)
\end{aligned}$$
Notice that, since the $Y_i$’s belong to the interval $[0,1]$, we can upper bound the variance by the mean as follows: $$\sigma^2 \le \frac{1}{n} \sum_i |Y_i - \mu| \le \frac{1}{n} \left( \sum_i |Y_i| + \sum_i |\mu| \right) = 2 \mu.$$ This gives the following corollary.
\[cor:multiChernoff\] Consider the conditions of the previous lemma. Then for all $\tau > 0$ $$\begin{aligned}
\Pr(|Y_S - s \mu| \ge \tau) \le 2 \exp \left( -\frac{\tau^2}{4 s \mu + \tau} \right).
\end{aligned}$$
Proof of Lemmas \[lemma:approximateCS\] and \[lemma:sampleCS\]
==============================================================
[Lemma \[lemma:approximateCS\]]{} Fix a scenario $\sigma$ for the duration of the proof. By assumption $x^S$ is feasible for , so it suffices to show that it attains value at least $(1-3\epsilon) {\ensuremath{\textrm{OPT}}}$. For that, consider with a modified right-hand side: $$\begin{aligned}
\max \sum_{t = 1}^n \pi_t x_t \notag\\
\sum_{t = 1}^n a^t_i x_t \le a_i(x^S) \ \ \ \ \forall i \in [m] \tag{modLP} \label{eq:modLP} \\
x \in [0,1]^n. \notag
\end{aligned}$$ Consider the Lagrangian relaxation $L(p, x) = \sum_{t = 1}^{\epsilon n} \pi_t x_t - \sum_{i = 1}^m p_i (\sum_{t = 1}^{\epsilon n} a^t_i x_t - a_i(x^S))$. Notice that $x^S$ is an optimal solution for $\max_{x \in [0,1]^n} L(p^S, x)$, which is at least the [$\textrm{OPT}$]{}(\[eq:modLP\]), the optimum value of LP . Since $x^S$ is clearly feasible for , it follows that $x^S$ is an optimal solution for the latter.
Now let $x^*$ be an optimal solution for . Since $a_i(x^S) \ge (1 - 3 \epsilon) B$ for all $i$, and since $a^t \ge 0$ for all $t$, it follows that $(1 - 3 \epsilon) x^*$ is feasible for . By linearity of the objective function we get that [$\textrm{OPT}$]{}(\[eq:modLP\]) $ \ge (1 - 3\epsilon) \sum_{t = 1}^n \pi_t x^*_t = (1 - 3 \epsilon) {\ensuremath{\textrm{OPT}}}$ and the result follows.
[Lemma \[lemma:sampleCS\]]{} Fix a scenario $\sigma$ for the duration of the proof. Let $x^*$ be an optimal solution for $(\epsilon n, (1 - \epsilon))$-LP in complementary slackness with $p^S$. If $p^S a^t > \pi_t$, the corresponding constraint in the dual is loose and by complementary slackness we get $x^*_t = 0$. If $p^S a^t < \pi_t$, then for dual feasibility we have $\alpha^*_t > 0$ and by complementary slackness we have $x^*_t = 1$.
From the definition of $x^S$ we get that $x^S \le x^*$ and, since the $a^t$’s are non-negative, the feasibility of $x^*$ implies that $a_i^S(x^S) \le (1-\epsilon)B$ for all $i \in [m]$. Moreover, from our assumption that the input is in general position we get that there are at most $m$ values of $t$ such that $p^S a^t = \pi_t$. Therefore, $x^S$ and $x^*$ differ in at most $m$ positions and from primal complementary slackness we get that whenever $p^S > 0$, $a_i^S(x^S) \ge a_i^S(x^*) - m = (1 - \epsilon) B - m \ge (1 - 2\epsilon) B$, where the last inequality follows from the fact that $B \ge \frac{1}{\epsilon}$. This concludes the proof of the lemma.
Proof of Lemma \[lemma:badWitness\]
===================================
The following simple inequalities will be helpful.
\[obs:ratio\] For $\epsilon, \alpha, \beta \ge 0$, $\frac{1-\alpha \epsilon}{1 + \beta \epsilon} \ge 1 - (\alpha + \beta) \epsilon$ and $\frac{1-\alpha \epsilon}{1 - \beta \epsilon} \le 1 - (\alpha - \beta) \epsilon$.
Combining equations , and and union bounding over all terms in the disjunction, we have that $$\begin{aligned}
\Pr\left(x^S \textrm{ is bad}\right) \le \sum_{i, w \in \mathcal{W}_i^+} \Pr\left(\operatorname{skewm}(\epsilon,w)\right) + \sum_{i, w \in \mathcal{W}_i^-} \Pr\left(\operatorname{skewp}(\epsilon,w)\right).
\end{aligned}$$ Thus, it suffices to show that for all $w \in \mathcal{W}_i^+$ (respectively $w \in \mathcal{W}_i^-$), the event $\operatorname{skewm}(\epsilon, w)$ (resp. $\operatorname{skewp}(\epsilon, w)$) occurs with probability at most $2\exp\left(-\frac{\epsilon^3 B}{33}\right)$.
Take $w \in \mathcal{W}_i^+$. By definition of this set, $a_i(w) \ge (1 - \frac{\epsilon}{2})B$, so the event $\operatorname{skewm}(\epsilon,w)$ is contained in the event that $a_i^S(w) \le (1 - \epsilon) a_i(w)/(1 - \frac{\epsilon}{2})$, which is contained in the event $a_i^S(w) \le (1 - \frac{\epsilon}{2}) a_i(w)$. Using Corollary \[cor:multiChernoff\] with $\tau = \epsilon^2 a_i(w) /2$, we obtain that $\Pr(\operatorname{skewm}(\epsilon, w)) \le 2\exp\left(-\frac{\epsilon^3 B}{33}\right)$.
Similarly, take $w \in \mathcal{W}_i^-$, such that $a_i(w) \le (1 - \frac{3\epsilon}{2})B$. It is easy to check that the event $\operatorname{skewp}(\epsilon,w)$ is contained in $a_i^S(w) \ge (1 + \frac{\epsilon}{2})a_i(w)$, so using Corollary \[cor:multiChernoff\] with $\tau = \epsilon^2 B/2$ we get that $\Pr(\operatorname{skewm}(\epsilon, w)) \le 2\exp\left(-\frac{\epsilon^3 B}{33}\right)$. This concludes the proof of the lemma.
Proof of Lemma \[lemma:chain\]
==============================
Fix $j \in [K]$. Consider a set $x \in \mathcal{X}$ and let $p$ be a dual vector such that $x(p) = x$. Let $t'$ be the last index of $C_j$ which belongs to $x|_{C_j}$; this implies that $\pi_{t'} > p a^{t'} = p c^j \|a^{t'}\|_{\infty}$, or alternatively $\frac{\pi_{t'}}{\|a^{t'}\|_{\infty}} > p c^j$. By the ordering of the columns, for all $t \in C_j$ smaller than $t'$ we have $\frac{\pi_t}{\|a^t\|_{\infty}} \ge \frac{\pi_{t'}}{\|a^{t'}\|_{\infty}} > p c^j$ and hence $t \in x|_{C_j}$. By definition of $t'$ it follows that $x|_{C_j} = \{t \in C_j : t \le t'\}$, a prefix of $C_j$; this concludes the proof.
Proof of Lemma \[lemma:wWitness\]
=================================
We prove that $\mathcal{W}_i^+$ is a witness set for $\mathcal{X}_i^+$; the proof that $\mathcal{W}_i^-$ is a witness set for $\mathcal{X}_i^-$ is analogous.
First, we claim that for all $x \in \mathcal{X}_i^+$, there is $x' \in \mathcal{X}$ such that $x' \subseteq x$ and $a_i(x') \in [B, B + m]$. To see this, let $p$ be such that $x = x(p)$. For $\lambda \ge 0$, define $p^\lambda = p + \lambda e_i$, where $e_i$ denotes the $i$th canonical vector. We have that $a_i(x(p^0)) > B$ (since $x(p) \in \mathcal{X}_i^+$) and $a_i(x(p^\infty)) = 0$ (since columns with $a_i^t > 0$ will at have at some point $p^\lambda a^t \ge \pi_t$). Due to the assumption that the input is in general position, whenever $a_i(x(p^\lambda))$ is discontinuous (as a function of $\lambda \ge 0$) the right and the left limits differ by at most $m$. It then follows that there is $\lambda \ge 0$ such that $a_i(x(p^\lambda)) \in [B, B + m]$, and since $x(p^\lambda) \subseteq x$ for all $\lambda \ge 0$ the claim follows.
So take a classification $x \in \mathcal{X}_i^+$ and let $x'$ be as above. The fact that $a_i(x') \le B + m$ and the non-negativity of the $a^t$’s imply that there is an $\ell \in L^K$ such that $x' \in \mathcal{B}_i^\ell$. Since $\underline{w}^\ell$ is the unique smallest set in $\mathcal{B}_i^\ell$, clearly $x' \subseteq \underline{w}^\ell$. To show that $\underline{w}^\ell \in \mathcal{W}_i^+$, it suffices to argue that $a_i(\underline{w}^\ell) \ge (1 - \epsilon/2) B$.
Since $\underline{w}^\ell, x' \in \mathcal{B}_i^\ell$, for all $j$ such that $\ell_j > 0$ we have $a_i(\underline{w}^\ell|_{C_j}) \ge a_i(x'|_{C_j}) / (1 + \frac{\epsilon}{4})$. Moreover, for $j$ such that $\ell = 0$ we have $a_i(x(p)|_{C_j}) < \frac{\epsilon B}{4K}$. Adding over all $j \in [K]$ gives $$\begin{aligned}
a_i(\underline{w}^\ell) \ge \left( \frac{1}{1 + \frac{\epsilon}{4}} \right) \left[ a_i(x(p)) - \sum_{j : \ell_j = 0} a_i(x(p)|_{C_j}) \right] \ge \frac{B}{1 + \frac{\epsilon}{4}} - \frac{\epsilon B}{4} \ge \left(1 - \frac{\epsilon}{2}\right) B,
\end{aligned}$$ where the third inequality follows from Observation \[obs:ratio\]. Thus, $\underline{w}^\ell \in \mathcal{W}_i^+$.
Since this property holds for all $x \in \mathcal{X}_i^+$, we conclude that $\mathcal{W}_i^+$ is a witness set for $\mathcal{X}_i^+$.
Proof of Lemma \[lemma:sizeP\]
==============================
Recall the definitions of $P^{{v}}$ (for ${v}\in L^K$) and $P_j^{\ell}$ (for $j \in [m]$, $\ell \in L$). It suffices to prove that at most $(O(\frac{K}{\epsilon} \log \frac{K}{\epsilon}))^m$ of the families $P^{{v}}$’s are non-empty.
Since $x(p) \in \mathcal{B}_i^{{v}}$ if and only if for all $j \in [K]$ we have $x(p)|_{C_j} \in \mathcal{B}_{i,j}^{{v}_j}$, it follows that $P^{{v}} = \bigcap_j P_j^{{v}_j}$. Let $\tau^\ell_j$ denote the first index in $C_j$ such that the prefix $\{t \in C_j : t \le \tau^\ell_j\}$ occupies the budget $i$ to an extent in $I_\ell$. Using Lemma \[lemma:chain\] and the fact that the $a^t$’s are non-negative, we get that $\mathcal{B}_{i,j}^\ell$ is the set of all prefixes of $C_j$ which contain $\tau_j^\ell$ but do not contain $\tau_j^{\ell + 1}$. Moreover, notice that the set $x(p)|_{C_j}$ contains $\tau_j^\ell$ if and only if $\pi_{\tau_j^\ell} > p a^{\tau_j^\ell}$. It then follows from these observations we can express the set $P_j^\ell$ using linear inequalities: $P_j^\ell = \{p \in \mathbb{R}^m_+ : \pi_{\tau_j^\ell} > p a^{\tau_j^\ell}, \pi_{\tau_j^{\ell + 1}} \le p a^{\tau_j^{\ell + 1}}\}$. Since $P^{{v}} = \bigcap_j P_j^{{v}_j}$, we have that $P^{{v}}$ is given by the intersection of halfspaces defined by hyperplanes of the form $\pi_{\tau_j^\ell} = p a^{\tau_j^\ell}$ and $p_k = 0$ ($k \in [m]$).
So consider the arrangement given by all hyperplanes $\{\pi_{\tau_j^\ell} = p a^{\tau_j^\ell}\}_{j \in [K], \ell \in L}$ and $\{p_i = 0\}_{i = 1}^m$. Given a face $F$ in this arrangement and a set $P^{{v}}$, either $F$ is contained in $P^{{v}}$ or these sets are disjoint. Since the faces of the arrangement cover $\mathbb{R}^m$, it follows that each non-empty $P^{{v}}$ contains at least one of these faces.
Notice that the arrangement is defined by $K|L| + m \le O(\frac{Km}{\epsilon} \log \frac{K}{\epsilon})$ hyperplanes, where the last inequality uses the fact that $\log (1 + \frac{\epsilon}{4}) \ge \epsilon \log (1 + \frac{1}{4})$ holds (by concavity) for $\epsilon \in [0,1]$. It is known that an arrangement with $h \ge m$ hyperplanes in $\mathbb{R}^m$ has at most $\left(\frac{e h}{m}\right)^m$ faces (see Section 6.1 of [@matousek] and page 82 of [@matousekNesetril]). Using the conclusion of the previous paragraph, we get that there are at most $(O(\frac{K}{\epsilon} \log \frac{K}{\epsilon}))^m$ non-empty $P^{{v}}$’s and the result follows.
Proof of Lemma \[lemma:robust\]
===============================
Let LP1 denote the LP with columns $(\pi_t, \tilde{a}^t)$ and right-hand side $(1-\epsilon)B$ and LP2 denote the LP with columns $(\pi_t, a^t)$ and right-hand side $B$.
Let $x$ be an $\epsilon$-approximate solution for LP1. Notice that we can upper bound $\|a^t - \tilde{a}^t\|_{\infty}$ as a function of $\|\tilde{a}^t\|_{\infty}$: $$\begin{aligned}
\|\tilde{a}^t\|_{\infty} \ge \|a^t\|_{\infty} - \|a^t - \tilde{a}^t\|_{\infty} \ge \frac{m}{\epsilon} \|a^t - \tilde{a}^t\|_{\infty},
\end{aligned}$$ where the first inequality follows from triangle inequality. That is, we have $\|a^t - \tilde{a}^t\|_{\infty} \le \frac{\epsilon}{m} \|\tilde{a}^t\|_{\infty}$.
Given this bound, it is easy to see that $x$ is feasible for LP2: $$\begin{aligned}
\sum_t a_i^t x_t \le \sum_t (\tilde{a}_i^t + \|a_i^t - \tilde{a}_i^t\|) x_t \le (1 - \epsilon) B + \sum_t \|a^t - \tilde{a}^t\|_{\infty} x_t \le (1 - \epsilon) B + \frac{\epsilon}{m} \sum_t \|\tilde{a}^t\|_{\infty} x_t \le B,
\end{aligned}$$ where the last inequality uses the fact that $\sum_t \|\tilde{a}^t\|_{\infty} x_t \le \|\tilde{a}^t\|_1 x_t \le mB$, since $x$ is a feasible solution and the $\tilde{a}^t$’s are non-negative.
In order to show that $x$ is a $2\epsilon$-approximate solution for LP2, it suffices to show that the optimum of LP1 is at least $1/(1+\epsilon)$ times the optimum of the LP2, since then $x$ will be within a factor of $(1-\epsilon)/(1+\epsilon) \ge (1 - 2\epsilon)$ the optimum of LP2. So let $x^*$ be an optimal solution for LP2. Using the same argument as before, it is easy to see that $x^*/(1+\epsilon)$ is feasible for LP1; this concludes the proof of the lemma.
Proof of Lemma \[lemma:sDeltaOtp\]
==================================
The proof uses the same ideas used in the analysis of [<span style="font-variant:small-caps;">OTP</span>]{}, although some definitions need to be changed slightly.
Recall that $S = \{\sigma(1), \sigma(2), \ldots, \sigma(s)\}$, $T = \{\sigma(s + 1), \sigma(s + 2), \ldots, \sigma(2s)\}$ and $U = S \cup T$. Again we use $p^S$ to denote the dual vector used by $(s, \delta)$-[<span style="font-variant:small-caps;">OTP</span>]{}for its classification, and set $x^S = x(p^S)$. With slight abuse in the notation, we often see $x^S$ as a (possibly infeasible) solution for $(2s,1)$-LP, which means that we truncate the vector $x^S$ to the first $2s$ coordinates $x^S_{\sigma(1)}, \ldots, x^S_{\sigma(2s)}$.
As before, we focus on proving the following lemma; the proof that this lemma implies Lemma \[lemma:sDeltaOtp\] is presented at the end of this section.
\[lemma:goodSDOTP\] Suppose that there are $K \ge m$ 1-dim subspaces of $\mathbb{R}^m$ containing the columns $a^t$’s. Fix an integer $s$ and a real number $\delta \in (0,1/10)$ such that $\frac{\delta^2 s B}{n} \ge \Omega(m \ln \frac{K}{\delta})$. Then with probability at least $(1 - \delta^2)$, $x^S$ satisfies $a_i^T(x^S) \le B$ for all $i \in [m]$ and has value $\pi_U x^S \ge (1 - 3\delta) {\ensuremath{\textrm{OPT}}}(2s)$.
In a given scenario, we now say that $x^S$ is *bad* if $a_i^T(s^S) > B$ for some $i \in [m]$ or if $\pi_U x^S (1 - 3\delta) {\ensuremath{\textrm{OPT}}}(2s)$. In this scenario, now a classification $x \in \mathcal{X}$ can be *badly learned for budget $i$ due to infeasibility* if $a_i^S(x) \le (1 - \delta) B$ and $a_i^T(x) > B$; $x$ can be *badly learned for budget $i$ due to value* if $a_i^S(x) \ge (1 - 2\delta)B$ and $a_i^U(x) < (1 - 3\delta)B$. Then $x$ can be *badly learned for budget $i$* if it falls into any of the above cases. The following is the appropriate modification of Lemma \[lemma:approximateCS\] for our current setting, and can be proved exactly in the same way.
\[lemma:approximateCS2\] Consider a scenario where $x^S$ satisfies the following: (i) for all $i \in [m]$, $a_i^T(x^S) \le B$ and (ii) for all $i \in [m]$ with $p^S_i > 0$, $a_i^U(x^S) \ge (1 - 3 \delta) B$. Then $x^S$ is good.
Due to our definitions, this lemma implies that inequality still hold.
#### Witness sets.
In the analysis of [<span style="font-variant:small-caps;">OTP</span>]{}, each $x \in \mathcal{X}$ could be badly learned for budget $i$ due to either infeasibility or (exclusively) due to value, which motivated the definitions of $\mathcal{X}_i^+$ and $\mathcal{X}_i^-$. Now the same $x$ can be badly learned for budget $i$ due to both conditions. Therefore, we introduce two different partition of $\mathcal{X}$, which tells *why* a classification is unlikely to be badly learned due to the appropriate condition. That is, we define $\mathcal{X}_i^+ = \{x \in \mathcal{X} : a_i(x) > (1 - \delta) B + \frac{\delta B}{2}\}$ and $\mathcal{Y}_i^+ = \{x \in \mathcal{X} : a_i(x) \le (1 - \delta) B + \frac{\delta B}{2}\}$ as the partition associated to the infeasibility condition and $\mathcal{X}_i^- = \{x \in \mathcal{X} : a_i(x) < (1 - 2\delta) B - \frac{\delta B}{2}\}$ and $\mathcal{Y}_i^- = \{x \in \mathcal{X} : a_i(x) \ge (1 - 2\delta) B - \frac{\delta B}{2}\}$ as the partition associated to the value condition. For example, $\mathcal{X}_i^-$ is the set of classifications which are unlikely to be infeasible because of a small $a_i(.)$ value. Also, note that these classifications are all based on the total budget occupation rather than on the budget occupation in the first $2s$ columns only.
Given this more refined tagging of elements in $\mathcal{X}$, we also need to redefine witness sets. We say that $(\mathcal{W}_i^+, \mathcal{W}_i^-, \mathcal{Z}_i^+, \mathcal{Z}_i^-)$ are *witness sets* for $(\mathcal{X}_i^+, \mathcal{X}_i^-, \mathcal{Y}_i^-, \mathcal{Y}_i^+)$ respectively if they satisfy the following: $$\begin{aligned}
w \in \mathcal{W}_i^+ \Rightarrow a_i(w) \ge (1 - \delta) B + \frac{\delta B}{4}, x \in \mathcal{X}_i^+ \Rightarrow \exists w \in \mathcal{W}_i^+ : w \subseteq x \\
w \in \mathcal{Z}_i^+ \Rightarrow a_i(w) \ge (1 - 2\delta) B - \frac{3\delta B}{4}, x \in \mathcal{Y}_i^- \Rightarrow \exists w \in \mathcal{Z}_i^+ : w \subseteq x \\
w \in \mathcal{W}_i^- \Rightarrow a_i(w) \le (1 - 2\delta) B - \frac{\delta B}{4}, x \in \mathcal{X}_i^- \Rightarrow \exists w \in \mathcal{W}_i^+ : x \subseteq w \\
w \in \mathcal{Z}_i^- \Rightarrow a_i(w) \le (1 - \delta) B + \frac{3 \delta B}{4}, x \in \mathcal{Y}_i^+ \Rightarrow \exists w \in \mathcal{W}_i^+ : x \subseteq w \, .
\end{aligned}$$
Again to simplify the notation, given a set $x$ we define $\operatorname{skewm}_i^S(\delta, x)$ to be the event that $a^S_i(x) \le (1 - \delta) B$, $\operatorname{skewp}^S_i(\delta, x)$ to be the event that $a^S_i(x) \ge (1 - \delta) B$ and similarly replacing the set $S$ by the sets $T$ and $U$. The following expression, which is the analogous to -, establishes the connection between the events where classifications can be badly learned and witness sets: $$\begin{aligned}
\bigvee_{x \in \mathcal{X}} \{\textrm{$x$ can be badly learned for budget $i$}\} \subseteq & \left(\bigvee_{w \in \mathcal{W}_i^+} \operatorname{skewm}^S(\delta, w) \right) \vee \left( \bigvee_{w \in \mathcal{Z}_i^+} \operatorname{skewm}^{U}(3\delta, w) \right) \notag\\
&\vee \left( \bigvee_{w \in \mathcal{W}_i^-} \operatorname{skewp}^S(2\delta, w) \right) \vee \left( \bigvee_{w \in \mathcal{Z}_i^-} \operatorname{skewp}^{T}(0, w)\right). \label{eq:witnessSDOTP}
\end{aligned}$$ To see that this expression holds, take $x \in \mathcal{X}$. Suppose that $x \in \mathcal{X}_i^+$ and let $w \in \mathcal{W}_i^+$ be contained in $x$. Then the event $\{\textrm{$x$ can be badly learned for budget $i$ due to infeasibility}\}$ is contained in $\operatorname{skewm}^S(\delta, w)$. Similarly, if $x \in \mathcal{Y}_i^+$ let $w \in \mathcal{Z}_i^-$ contain $x$; then the event $\{\textrm{$x$ can be badly learned for budget $i$ due to infeasibility}\}$ is contained in $\operatorname{skewm}^T(0, w)$. The reasoning for the event $\{\textrm{$x$ can be badly learned for budget $i$ due to value}\}$ is similar.
The following is analogous to Lemma \[lemma:badWitness\].
\[lemma:badWitnessOTP\] Suppose that, for all $i \in [m]$, there are witness sets for $(\mathcal{X}_i^+, \mathcal{X}_i^-, \mathcal{Y}_i^+, \mathcal{Y}_i^-)$ of size at most $M$. Then $\Pr(x^S \textrm{ is bad }) \le 8mM \exp\left(-\frac{\delta^2 s B}{136 n}\right)$.
#### Good witness sets.
We now construct witness sets of size at most $(O(\frac{K}{\delta} \log \frac{K}{\delta}))^m$, so Lemma \[lemma:goodSDOTP\] will follow directly from Lemma \[lemma:badWitnessOTP\]. The development mirrors that of Section \[sec:goodWitness\]. Let $C_1, C_2, \ldots, C_K$ be a partition of the index set $[n]$ such that for all $j$, the columns $\{a^t\}_{t \in C_j}$ belong to the same 1-dimensional subspace.
Cover the interval $[0, B+m]$ with intervals $\{I_\ell\}_{\ell \in L}$, where $I_0 = [0, \frac{\delta B}{8 K})$ and $I_\ell = [\frac{\delta B}{8K} (1 + \frac{\delta}{8})^{\ell - 1}, \frac{\delta B}{8K} (1 + \frac{\delta}{8})^\ell)$ for $\ell > 0$ and $L = \{0, \ldots, \lceil \log_{1 + \delta/8} \frac{16K}{\delta} \rceil + 1\}$. Define $\mathcal{B}_{i,j}^\ell$ as the set of classifications $x \in \mathcal{X}|_{C_j}$ whose occupation $a_i(x)$ lies in the interval $I_\ell$. Finally, for $\ell \in L^K$, define the family of boxes $\mathcal{B}_i^\ell = \prod_j \mathcal{B}_{i,j}^{\ell_j}$.
Given $\ell \in L$, let $\underline{w}^\ell(j)$ be the smallest set in $\mathcal{X}|_{C_j}$ which has $a_i(\underline{w}^\ell(j)) \in I_\ell$ and for $\ell \in L^K$ define the set $\underline{w}^{\ell}$ as the union of the sets $\underline{w}^{\ell_j}(j)$’s (or equivalently, as the concatenation of the vectors $\underline{w}^{\ell_j}(j)$’s). Similarly, for $\ell \in L$ let $\overline{w}^\ell(j)$ be the largest set in $\mathcal{X}|_{C_j}$ which has $a_i(\overline{w}^\ell(j)) \in I_\ell$ and for $\ell \in L^K$ define the set $\overline{w}^\ell$ as the union of the sets $\overline{w}^{\ell_j}(j)$’s.
Now we construct the witness sets as before. Set $\mathcal{W}_i^+ = \{\underline{w}^\ell : a_i(\underline{w}^\ell) \ge (1 - \delta)B + \frac{\delta B}{4}, \mathcal{B}_i^\ell \cap \mathcal{X} \neq \emptyset\}$, set $\mathcal{Z}_i^+ = \{\underline{w}^\ell : a_i(\underline{w}^\ell) \ge (1 - 2\delta)B - \frac{3\delta B}{4}, \mathcal{B}_i^\ell \cap \mathcal{X} \neq \emptyset\}$, set $\mathcal{W}_i^- = \{\overline{w}^\ell : a_i(\overline{w}^\ell) \le (1 - 2\delta)B - \frac{\delta B}{4}, \mathcal{B}_i^\ell \cap \mathcal{X} \neq \emptyset\}$ and finally set $\mathcal{Z}_i^- = \{\overline{w}^\ell : a_i(\overline{w}^\ell) \le (1 - \delta)B + \frac{3\delta B}{4}, \mathcal{B}_i^\ell \cap \mathcal{X} \neq \emptyset\}$.
Following the same steps as in the proof of Lemma \[lemma:wWitness\], one can check that $(\mathcal{W}_i^+, \mathcal{W}_i^-, \mathcal{Z}_i^+, \mathcal{Z}_i^-)$ are *witness sets* for $(\mathcal{X}_i^+, \mathcal{X}_i^-, \mathcal{Y}_i^+, \mathcal{Y}_i^-)$. Moreover, the proof of Lemma \[lemma:sizeP\] can be used to show that, for a fixed $i \in [m]$, at most $(e \frac{K}{\delta} \log \frac{K}{\delta} )^m$ of the $\mathcal{B}_i^\ell$’s contain an element of $\mathcal{X}$, which then imposes the same upper bound on the size of the witness sets. This concludes the proof of Lemma \[lemma:goodSDOTP\].
[Lemma \[lemma:sDeltaOtp\]]{} Let $x$ be the solution returned by $(s, \delta)$-[<span style="font-variant:small-caps;">OTP</span>]{}and let $\mathcal{E}$ denote the event that $x^S$ is good. For any scenario in $\mathcal{E}$, we have $x_{\sigma(t)} = x^S_{\sigma(t)}$ for all $t = s+1, s+2, \ldots, 2s$. Therefore, we get that $$\begin{aligned}
{\mathbb{E}}\left[\sum_{t = 1}^{2s} \pi_{\sigma(t)} x_{\sigma(t)}\right] &\ge {\mathbb{E}}\left[\sum_{t = 1}^{2s} \pi_{\sigma(t)} x_{\sigma(t)} \mid \mathcal{E} \right] \Pr(\mathcal{E}) \notag \\
& \ge {\mathbb{E}}\left[\sum_{t = 1}^{2s} \pi_{\sigma(t)} x_{\sigma(t)}^S \mid \mathcal{E}\right] \Pr(\mathcal{E}) - {\mathbb{E}}[{\ensuremath{\textrm{OPT}}}(s) \mid \mathcal{E}] \Pr(\mathcal{E}) \notag \\
&\ge {\mathbb{E}}\left[\sum_{t = 1}^{2s} \pi_{\sigma(t)} x_{\sigma(t)}^S \mid \mathcal{E}\right] \Pr(\mathcal{E}) - {\mathbb{E}}[{\ensuremath{\textrm{OPT}}}(s)]. \label{eq:sDeltaOpt1}
\end{aligned}$$
To lower bound the first term in the right hand side we use again the definition of $\mathcal{E}$: $$\begin{aligned}
{\mathbb{E}}\left[\sum_{t = 1}^{2s} \pi_{\sigma(t)} x_{\sigma(t)}^S \mid \mathcal{E} \right] \ge (1 - 3 \delta) {\mathbb{E}}[{\ensuremath{\textrm{OPT}}}(2s) \mid \mathcal{E}] \Pr(\mathcal{E})
\end{aligned}$$ and $$\begin{aligned}
{\mathbb{E}}[{\ensuremath{\textrm{OPT}}}(2s)] = {\mathbb{E}}[{\ensuremath{\textrm{OPT}}}(2s) \mid \mathcal{E}] \Pr(\mathcal{E}) + {\mathbb{E}}[{\ensuremath{\textrm{OPT}}}(2s) \mid \overline{\mathcal{E}}] \Pr(\overline{\mathcal{E}}) \le {\mathbb{E}}[{\ensuremath{\textrm{OPT}}}(2s) \mid \mathcal{E}] \Pr(\mathcal{E}) + \delta^2 {\ensuremath{\textrm{OPT}}},
\end{aligned}$$ where the last inequality uses Lemma \[lemma:goodSDOTP\]. Combining the previous two inequalities give that ${\mathbb{E}}\left[\sum_{t = 1}^{2s} \pi_{\sigma(t)} x_{\sigma(t)}^S \mid \mathcal{E}\right] \ge (1 - 3 \delta) {\mathbb{E}}[{\ensuremath{\textrm{OPT}}}(2s)] - \delta^2 {\ensuremath{\textrm{OPT}}}$, and the result follows from equation .
Proof of Theorem \[thm:expValueDPA\]
====================================
Let LP1 denote the LP with columns $(\pi_t, \tilde{a}^t)$ and right-hand side $\tilde{B} = (1-\epsilon)B$ and LP2 denote the LP with columns $(\pi_t, a^t)$ and right-hand side $B$. We show that Robust [<span style="font-variant:small-caps;">DPA</span>]{}returns a $(1 - 21.5\epsilon)$-approximation for LP1, and the theorem will follow from Lemma \[lemma:robust\].
First we show that the returned solution $x$ is feasible for LP1. By definition of the algorithm, $a_j(x^i) \le \epsilon 2^i \tilde{B}$ for all $i,j$. By linearity, $a_j(x) = \sum_i a_j(x^i) \le \epsilon \tilde{B} \sum_{i = 0}^{\log(1/\epsilon) - 1} 2^i \le \tilde{B}$.
In order to verify the value of the returned solution, we first show that $\frac{\delta^2 s B}{n} \ge \Omega(m \ln \frac{K}{\delta})$ in every call to $(s,\delta)$-[<span style="font-variant:small-caps;">OTP</span>]{}made by Robust [<span style="font-variant:small-caps;">DPA</span>]{}. As in Section \[sec:robustOTP\], the columns $\tilde{a}^t$’s belong to at most $K = O(\frac{m}{\epsilon})^m$ 1-dim subspaces. Since $B \ge \Omega(\frac{m^2}{\epsilon^2} \ln \frac{m}{\epsilon})$, we have that for each $i = 0, \ldots, \log(1/\epsilon)-1$ setting $s = \epsilon 2^i n$ and $\delta = \sqrt{\epsilon/2^i}$ satisfies the expression $\frac{\delta^2 s B}{n} \ge \Omega(m \ln \frac{K}{\delta})$.
Then applying Lemma \[lemma:sDeltaOtp\] we get that for all $i = 0, \ldots, \log(1/\epsilon)-1$, ${\mathbb{E}}[\pi x^i] \ge (1 - 3 \sqrt{\frac{\epsilon}{2^i}}) {\mathbb{E}}[{\ensuremath{\textrm{OPT}}}(\epsilon 2^{i+1} n)] - {\mathbb{E}}[{\ensuremath{\textrm{OPT}}}(\epsilon 2^i n)] - \frac{\epsilon {\ensuremath{\textrm{OPT}}}}{2^i}$. By linearity of the objective value and of expectations $$\begin{aligned}
{\mathbb{E}}[\pi x] = \sum_i {\mathbb{E}}[\pi x^i] \ge - {\mathbb{E}}[{\ensuremath{\textrm{OPT}}}(\epsilon n)] - \sum_{i = 0}^{\log(1/\epsilon) - 2} \left(3 \sqrt{\frac{\epsilon}{2^i}}\right) {\mathbb{E}}[{\ensuremath{\textrm{OPT}}}(\epsilon n 2^{i + 1})] + (1 - 3 \sqrt{2} \epsilon - \epsilon) {\ensuremath{\textrm{OPT}}}.
\end{aligned}$$
Lemma 2.4 of [@agrawal] states that ${\mathbb{E}}[{\ensuremath{\textrm{OPT}}}(s)] \le \frac{s}{n} {\ensuremath{\textrm{OPT}}}$ for all $s \ge 0$. Employing this observation, we get $$\begin{aligned}
{\mathbb{E}}[\pi x] \ge {\ensuremath{\textrm{OPT}}}- \epsilon {\ensuremath{\textrm{OPT}}}\left[3 \sqrt{2} + 2 + 3 \sqrt{\epsilon} \sum_{i = 0}^{\log(1/\epsilon) - 2} 2^{i/2 + 1} \right].
\end{aligned}$$ Since the summation in the expression can be upper bounded by $\frac{2 \sqrt{2}^{\log(1/\epsilon)}}{\sqrt{2} - 1} \le \frac{5}{\sqrt{\epsilon}}$, we get that ${\mathbb{E}}[\tilde{\pi} x] \ge (1 - 21.5 \epsilon) {\ensuremath{\textrm{OPT}}}$. This concludes the proof of the theorem.
[^1]: Actually knowing $n$ up to $(1 \pm \epsilon)$ factor is enough. This assumption is required to allow algorithms with non-trivial competitive ratio [@DevanurHayes09].
[^2]: To simplify the exposition, we assume that $\epsilon n$ is an integer.
[^3]: Notice that this ratio is well-defined since by assumption $a^t \neq 0$ for all $t \in [n]$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We derive a multiconfigurational time-dependent Hartree theory for systems with particle conversion. In such systems particles of one kind can convert to another kind and the total number of particles varies in time. The theory thus extends the scope of the available and successful multiconfigurational time-dependent Hartree methods – which were solely formulated for and applied to systems with a fixed number of particles – to new physical systems and problems. As a guiding example we treat explicitly a system where bosonic atoms can combine to form bosonic molecules and vise versa. In the theory for particle conversion, the time-dependent many-particle wavefunction is written as a sum of configurations made of a different number of particles, and assembled from sets of atomic and molecular orbitals. Both the expansion coefficients and the orbitals forming the configurations are time-dependent quantities that are fully determined according to the Dirac-Frenkel time-dependent variational principle. By employing the Lagrangian formulation of the Dirac-Frenkel variational principle we arrive at two sets of coupled equations of motion, one for the atomic and molecular orbitals and one for the expansion coefficients. The first set is comprised of first-order differential equations in time and nonlinear in-general integrodifferential equations in position space, whereas the second set consists of first-order differential equations with coefficients forming a time-dependent Hermitian matrix. Particular attention is paid to the reduced density matrices of the many-particle wavefunction that appear in the theory and enter the equations of motion. There are two kinds of reduced density matrices: particle-conserving reduced density matrices which directly only couple configurations with the same number of atoms and molecules, and particle non-conserving reduced density matrices which couple configurations with a different number of atoms and molecules. Closed-form and compact equations of motion are derived for contact as well as general two-body interactions, and their properties are analyzed and discussed.'
author:
- 'Ofir E. Alon[^1], Alexej I. Streltsov[^2], and Lorenz S. Cederbaum[^3]'
title: 'Many-body theory for systems with particle conversion: Extending the multiconfigurational time-dependent Hartree method'
---
Introduction {#Intro}
============
The exploration of quantum dynamics of many-particle systems is a fundamental and on-going challenge of many branches in physics [@Book_dynamics1; @Book_dynamics2; @Nuclear_book; @Book_dynamics3; @Pit_Stri_book; @Book_dynamics4]. The equation of motion governing the evolution of quantum particles is, in many cases, the well-known time-dependent Schrödinger equation. Solving the time-dependent Schrödinger equation for many-particle systems can rarely be made analytically or exactly, which renders efficient approximations a must.
The multiconfigurational time-dependent Hartree method (MCTDH) [@CPL; @JCP], which has been developed in the past two decades, is considered at present the most efficient wave-packet propagation approach [@MCTDH_package] and has successfully and routinely been used for multi-dimensional dynamical systems consisting of distinguishable degrees-of-freedom, such as molecular vibrations, see Ref. [@JCP_24a; @JCP_24b; @PR; @Dieter_review; @Manthe_review; @Lenz_CI; @relaxation2; @vib_new1; @vib_new2; @irene]. The main idea behind the MCTDH method is to expand the time-dependent many-body wavefunction of distinguishable particles by [*time-dependent*]{} configurations that are assembled from [*time-dependent*]{} orbitals (one-body functions) and optimized according to the Dirac-Frenkel variational principle [@DF1; @DF2]. In this way, a much larger effective subspace of the many-particle Hilbert space can be spanned in practice in comparison to multiconfigurational expansions with [*stationary*]{} configurations. By grouping several “elementary” degrees-of-freedom together and treating them as “generalized” particles, the efficiency of the MCTDH algorithm increases [@JCP_24a; @JCP_24b]. Choosing to use MCTDH itself to propagate multi-dimensional “generalized” particles has led to the idea of cascading [@Dieter_review]. Finally, expanding the time-dependent orbitals themselves by other time-dependent orbitals, and so on, putting the resulting time-dependent expansion under the Dirac-Frenkel variational principle, leads to the multi-layer formulation of the MCTDH theory [@Multi_L] which further increases the efficiency of the MCTDH method for larger, complex systems. The MCTDH can be applied to systems of identical particles. In this direction, we would like to mention that the MCTDH approach has very successfully been employed to unveiling fundamental physics of few-boson systems [@ZO_st1; @ZO_st2; @ZO_dy1; @ZO_dy2; @ZO_dy3; @axel] on the numerically-exact many-body level.
A new branch of MCTDH-based methods has emerged after it had been realized that, to effectively treat the dynamics of more than a handful [*identical*]{} particles, it is essential to use their quantum statistics, Fermi-Dirac or Bose-Einstein, to eliminate the large amount of redundancies of coefficients in the distinguishable-particle multiconfigurational expansion of the MCTDH wavefunction. First, taking explicitly the antisymmetry of the many-fermion wavefunction to permutations of any two particles into account, the fermionic version of MCTDH – MCTDHF – was independently developed by several groups [@MCTDHF1; @MCTDHF2; @MCTDHF3]. Shortly after, the bosonic version of MCTDH – MCTDHB – was developed in [@MCTDHB1; @MCTDHB2]. This advancement is in particular valuable since very-many bosons can reside in only a small number of orbitals owing to Bose-Einstein statistics, thereby allowing the successful and quantitative attack of the dynamics of a much larger number of bosons with the MCTDHB theory. For applications of MCTDHF to the many-body dynamics of at-present few-fermion systems with or without external laser field see Refs. [@applF1; @applF2; @applF3; @applF4; @applF5; @applF6; @applF7], and for applications of MCTDHB for the many-body dynamics of repulsive and attractive bosonic systems Refs. [@MCTDHB1; @applB1; @applB2; @applB3]. We mention that Ref. [@applB3] has combined optimal control theory with MCTDHB.
Five decades ago, in his seminal paper, Löwdin defined the reduced density matrices of many-fermion wavefunctions [@Lowdin]. Since then, reduced density matrices and, in particular, reduced two-body density matrices is a fruitful and vivid research area including theory and applications in electronic structure of molecules, quantum phase-transitions, and ground-state nuclear motion [@Slava; @MAZZ1; @MAZZ2; @MAZZ3; @MAZZ4; @MAZZ5; @MAZZ6]. In the present context, reduced one- and two-body density matrices were first used to derive the stationary many-body states within the general variational theory with complete self-consistency for trapped bosonic systems – the multiconfigurational Hartree for bosons (MCHB) [@MCHB]. Later on, the MCTDHF and MCTDHB were formulated in a unified manner, making use of the reduced one- and two-body density matrices of the time-dependent many-body wavefunction [@Unified_paper]. Finally, treating mixtures of two kinds of identical particles in a unified manner, and utilizing the reduced one- and two-body density matrices of the mixture’s wavefunction, a multiconfigurational time-dependent Hartree method for Fermi-Fermi (MCTDH-FF), Bose-Bose (MCTDH-BB) and Bose-Fermi (MCTDH-BF) mixtures has been derived [@MIX].
The multiconfigurational time-dependent Hartree method and its versions specified for identical particles and mixtures are [*particle-conserving*]{} many-body propagation theories. Namely, they were solely formulated for and applied to systems with a fixed number of particles. Conceptually, they aim at describing systems of coupled degrees-of-freedom or interacting particles which have first-quantization Hamiltonian. This brings us to the theme of the present work, which is to derive a multiconfigurational time-dependent Hartree theory for systems with [*particle conversion*]{}. In such systems particles of one kind can convert to another kind and the total number of particles varies in time. Hence, they are generally represented by a phenomenological second-quantized Hamiltonian which includes a [*conversion term*]{}. Doing so, we extend the scope of the available and successful multiconfigurational time-dependent Hartree method and its versions specified for identical particles and mixtures to new physical systems and problems. We abbreviate the multiconfigurational time-dependent Hartree theory for systems with particle conversion by MCTDH-[*conversion*]{}.
As a concrete and guiding example for a many-body system with particle conversion and without loss of generality, we consider explicitly the conversion of bosonic atoms ($a$) to bosonic molecules ($m$) via the ‘reaction’ $2a \leftrightharpoons m$, which has been a system of tremendous theoretical and experimental interest in quantum-gas physics [@PD1; @2M1; @Timmermans_review; @Yurovsky; @Timm1; @EXP1; @PD2; @Beyond; @Holland_PRL; @Sadhan; @Vardi; @EXP2; @EXP3; @Jaksch; @G_STOOF; @Tilman; @EXP3h; @EXP4; @EXP5; @Review_STOOF; @QFT2; @Holland_PRA; @PRL_STOOF_Subir; @EXP6; @Meystre_Rev; @2M2_2005; @Band; @Thorsten_Rev; @Vardi_PRL; @EXP7; @NBIG].
An effective quantum-field-theory-based Hamiltonian for atomic and molecular Bose-Einstein condensates (BECs) coupled by conversion was first put forward by Drummond [*et al.*]{} in [@PD1]. In [@2M1], a proposition that a molecular BEC could be produced by coherent photoassociation was made and a phenomenological two-mode Hamiltonian to describe this process was suggested. A microscopic theory to derive the many-body Hamiltonian of bosonic atoms and molecules with conversion and the respective Gross-Pitaevskii theory with conversion were put forward in [@Timmermans_review], also see [@Timm1]. In [@Yurovsky], a coupled system of Gross-Pitaevskii equations with conversion and deactivation-rate (dissipation) terms has been derived. The validity of the two-mode approach for conversion, at least in the homogeneous system, was questioned in [@Beyond], where dissociation of molecules to other than the ground atomic mode signifies that one needs to go beyond the two-mode approximation. The importance of pair correlations in the dynamics of resonantly-coupled atomic and molecular BECs, leading to significant deviations from the respective Gross-Pitaevskii theory, was put forward in [@Holland_PRL]. That even in the perfect two-mode limit the mean-field theory with conversion can fail, because of strong particle-particle entanglement near the dynamically unstable molecular mode, was reported in [@Vardi]. A proposition to create a molecule BEC from an atomic Mott-insulator phase with exactly two bosons per lattice site was made in [@Jaksch]. A full microscopic theory to derive the Hamiltonian of atoms and molecules with the conversion term from the microscopic particle-conserving Hamiltonian of a homogeneous gas of identical bosonic atoms with two internal states was given in [@G_STOOF], also see [@Review_STOOF]. Quantum phase transitions and effects of rotations in homogeneous systems of atomic and molecular BECs with conversion have been discovered in [@QFT2; @PRL_STOOF_Subir] and [@Holland_PRA], respectively. Finally, in a (harmonic) trap, confinement effects on the stimulated dissociation (effective conversion rate) of a molecular to an atomic BEC were recently found in [@Vardi_PRL], and unique phases (vortex configurations) of rotating interacting atomic-molecular BECs in [@NBIG].
Molecules were first produced from and identified in a $^{\mathrm {87}}$Rb BEC by Wynar [*et al.*]{} [@EXP1]. Soon after, photoassociation of ultracold sodium molecules in an atomic BEC was made [@EXP2]. Atomic-molecular coherence in a BEC \[made of $^{\mathrm {85}}$Rb atoms\] was first achieved in [@EXP3]. A pure molecular quantum gas produced from an atomic cesium BEC was reported in [@EXP3h], and a quantum-degenerate gas of sodium molecules in [@EXP4]. More recently, with $^{\mathrm {87}}$Rb atoms in the Mott-phase of optical lattices, state-selective conversion of atoms to molecules [@EXP6], following the theoretical proposition in [@Jaksch], and atom-molecule Rabi oscillations [@EXP7] have been observed.
Finally, systems with particle conversion can involve of course fermions, in the cold-atom world – see the reviews [@Thorsten_Rev; @F_Review] and references therein – and beyond it. In the latter context, it is gratifying to mention the Friedberg-Lee model of superconductivity, describing the conversion of two electrons to a single Cooper-pair and vice versa by a boson-fermion Hamiltonian with a phenomenological conversion term [@SC0; @SC1].
Let us return now to MCTDH-[*conversion*]{}, and put it in the particular context of interacting atomic and molecular BECs with conversion. For the explicit scenario of the conversion ‘reaction’ $2a{\leftrightharpoons}m$ dealt with throughout this work, the theory shall be referred to as MCTDH-\[$2a{\leftrightharpoons}m$\]. MCTDH-\[$2a{\leftrightharpoons}m$\], as its particle-conserving predecessors [@CPL; @JCP; @PR; @MCTDHF1; @MCTDHF2; @MCTDHF3; @MCTDHB1; @MCTDHB2; @Unified_paper; @MIX], is intended for systems with a finite number of interacting particles, typically trapped in an external potential. As a first step, we extend or “merge” two theoretical approaches much in use in the literature: the Gross-Pitaevskii theory with conversion and the two-mode approximation, see, e.g., Refs. [@Timmermans_review; @Timm1; @PD2; @Sadhan; @Band; @NBIG] and Refs. [@2M1; @Vardi; @2M2_2005], respectively. This results in a fully variational theory where the two modes – the atomic and molecular orbitals – and each and every expansion coefficient in the two-mode many-body wavefunction are fully optimized – the orbitals in time and space and the expansion coefficients in time – according to the Dirac-Frenkel variational principle [@DF1; @DF2]. Our main aim is to go beyond any two-mode description of the atomic-molecular coupled system and present a fully-variational multiconfigurational time-dependent many-body theory for bosonic atoms and molecules coupled by conversion – the MCTDH-\[$2a{\leftrightharpoons}m$\] theory.
The structure of the paper is as follows. We open in section \[secII\] with the many-body Hamiltonian of the system of atoms and molecules with conversion. In section \[two\_mode\_AM\] we consider as mentioned above a specific case of interest, the fully-variational theory where there are one atomic and one molecular orbitals. This specific theory will be referred to as [*conversion mean field*]{}. Next, section \[full\_Hilbert\] is devoted to the general theory. Both time-dependent as well as time-independent theories are presented. Finally, in section \[dis\_sum\] we put forward a summary and concluding remarks. Complementary derivations and relevant matrix elements are deferred to and collected in appendices \[Lagrange\] and \[matrix\_Appen\], respectively.
The many-body Hamiltonian of interacting atoms and molecules with conversion {#secII}
============================================================================
As a concrete example for a many-body system with particle conversion and without loss of generality, we consider a system of bosons which will be referred to as atoms ($a$) and their conversion to another type of bosons which will be referred to as molecules ($m$) via the ‘reaction’ $2a \leftrightharpoons m$. The many-body Hamiltonian of the coupled atom–molecule system is taken from the literature of cold-atom physics [@PD1; @Timmermans_review; @G_STOOF; @Holland_PRA] and is written for our needs as a sum of four terms: \[ham\_am\_1\] H\^[(2am)]{} = H\^[(am)]{} + W\^[(2am)]{} = H\^[(a)]{} + H\^[(m)]{} + W\^[(am)]{} + W\^[(2am)]{}. The first three terms are particle-conserving terms and together describe a mixture of two kinds of interacting bosonic particles; $a$ (atoms) and $m$ (molecules): \[ham\_am\_2\] H\^[(a)]{} &=& h\^[(a)]{} + W\^[(a)]{} =\
&=& d,\
H\^[(m)]{} &=& h\^[(m)]{} + W\^[(m)]{} =\
&=& d,\
W\^[(am)]{} &=& dd’ \^\_a() \^\_m(’) W\^[(am)]{}(,’) \_m(’) \_a(). The last term describes the conversion of atoms to molecules and vise versa and is given by [@PD1; @Timmermans_review; @G_STOOF; @Holland_PRA]: \[ham\_am\_3\] W\^[(2am)]{} &=& W\^[(2am)]{} + W\^[(m2a)]{} =\
&=& dd’ ,\
& & W\^[(m2a)]{}(,’) = {W\^[(2am)]{}(,’)}\^. The coordinates entering the field operators in (\[ham\_am\_3\]) represent the annihilation (creation) of two atoms, one at position $\r$ the second at position $\r'$, and the creation (annihilation) of a molecule at the center-of-mass coordinate $\R=\frac{\r+\r'}{2}$. The atomic, molecular field operators satisfy the usual commutation relations for bosons: $\left[\hat{\mathbf \Psi}_a(\r), \hat{\mathbf \Psi}^\dag_a(\r')\right] =
\left[\hat{\mathbf \Psi}_m(\r), \hat{\mathbf \Psi}^\dag_m(\r')\right] = \delta(\r-\r')$ and $\left[\hat{\mathbf \Psi}_a(\r), \hat{\mathbf \Psi}_a(\r')\right] =
\left[\hat{\mathbf \Psi}_m(\r), \hat{\mathbf \Psi}_m(\r')\right] = 0$. Since the atoms and molecules are distinguishable, different particles, their mutual field operators commute, $\left[\hat{\mathbf \Psi}_a(\r), \hat{\mathbf \Psi}^\dag_m(\r')\right] =
\left[\hat{\mathbf \Psi}_a(\r), \hat{\mathbf \Psi}_m(\r')\right] = 0$. Finally, we note that the interaction terms appearing in the Hamiltonian (\[ham\_am\_1\]-\[ham\_am\_3\]) are symmetric, i.e., $\hat W^{(a)}(\r,\r')=\hat W^{(a)}(\r',\r),\ldots,\hat W^{(m\rightharpoondown 2a)}(\r,\r') =
\hat W^{(m\rightharpoondown 2a)}(\r',\r)$, because the Hamiltonian is symmetric to the exchange of position of any two particles of the same kind.
The Hamiltonian (\[ham\_am\_1\]-\[ham\_am\_3\]) commutes with the following particle-number operator \[particle\_number\] N = N\_a + 2N\_m = d, reflecting a conservation law in presence of particle conversion. Accordingly, the Hilbert-space of the problem is a direct sum of Hilbert subspaces with different number of atoms and molecules: {$N$ atoms; $0$ molecules}$\oplus${$N-2$ atoms; $1$ molecule}$\oplus${$N-4$ atoms; $2$ molecules}$\oplus
\ldots \oplus${$N-2\left[\frac{N}{2}\right]$ atoms; $\left[\frac{N}{2}\right]$ molecules}, where $\left[j\right]$ means the greatest integer not exceeding $j$.
The purpose of this work is to treat the many-body Hamiltonian with atom–molecule conversion (\[ham\_am\_1\]-\[ham\_am\_3\]) multiconfigurationally. To this end, we expand the atomic $\hat{\mathbf \Psi}_a(\r)$ and molecular $\hat{\mathbf \Psi}_m(\r)$ field operators by two complete sets of [*time-dependent*]{} orbitals, \[annihilation\_def\] \_a() = \_k b\_k(t)\_k(,t), \_m() = \_[k’]{} c\_[k’]{}(t)\_[k’]{}(,t). The sets of atomic $\left\{\phi_k(\r,t)\right\}$ and molecular $\left\{\psi_{k'}(\r,t)\right\}$ orbitals span the [*time-dependent*]{} Hilbert space in which the system is to be propagated. The advantages of time-dependent multiconfigurational expansions, see the Introduction, is the employment of [*time-dependent*]{} orbitals which change in time according to a time-dependent variational principle. This allows one to use in practical computations a smaller number of time-dependent orbitals than the number of time-independent orbitals that would have been required otherwise. A general multiconfigurational expansion, see section \[full\_Hilbert\], employs $M$ orbitals for the bosonic atoms and $M'$ orbitals for the bosonic molecules. In particular, even if only one orbital is available for the bosonic atoms and another one for the bosonic molecules, the resulting theory [*goes beyond*]{} the Gross-Pitaevskii mean-field theory for this system [@Timmermans_review], see subsequent section \[two\_mode\_AM\].
Finally, it is convenient to derive the relevant results first for the popular contact interaction, \[contact\_potentials\] & & W\^[(a)]{}(,’) = \_a (-’), W\^[(m)]{}(,’) = \_m (-’),\
& & W\^[(am)]{}(,’) = \_[am]{} (-’), W\^[(2am)]{}(,’) = W\^[(m2a)]{}(,’) = \_[con]{} (-’). Thus, substituting Eqs. (\[annihilation\_def\],\[contact\_potentials\]) into the generic Hamiltonian (\[ham\_am\_1\]-\[ham\_am\_3\]) we get \[ham\_contact\_inter\] & & H\^[(2am)]{} = \_[k,q]{} b\_k\^b\_q + \_[k,s,l,q]{} b\_k\^b\_s\^b\_l b\_q +\
& & + \_[k’,q’]{} c\_[k’]{}\^c\_[q’]{} + \_[k’,s’,l’,q’]{} c\_[k’]{}\^c\_[s’]{}\^c\_[l’]{} c\_[q’]{} +\
& & + \_[am]{}\_[k,k’,q,q’]{} b\_k\^b\_q c\_[k’]{}\^c\_[q’]{} + \_[k’,k,q]{} . Here and hereafter, the dependence of quantities on time is not shown explicitly whenever unambiguous. Below, we will work throughout sections \[two\_mode\_AM\] and \[contact\] with the contact-interaction Hamiltonian (\[ham\_contact\_inter\]) and handle the case of general interactions (\[ham\_am\_1\]-\[ham\_am\_3\]) thereafter, in section \[non\_contact\].
The simplest case of atom–molecule conversion: Conversion mean field (Fully-variational two-mode approximation) {#two_mode_AM}
===============================================================================================================
The multiconfigurational ansatz {#two_mode_AM_asnatz}
-------------------------------
To introduce the nomenclature in the first stage of this work and, independently, as an interesting and relevant problem for itself, we consider the resulting theory when there is only one orbital available for the (bosonic) atoms and one orbital available for the (bosonic) molecules. The atomic orbital will be denoted by $\phi_1(\r,t)\equiv\phi_a(\r,t)$ and the molecular orbital by $\psi_1(\r,t)\equiv\psi_m(\r,t)$. The corresponding creation operators are denoted by $\hat b_1^\dag(t)\equiv\hat b_a^\dag(t)$ and $\hat c_1^\dag(t)\equiv\hat c_m^\dag(t)$. The atomic and molecular creation, annihilation operators obey the bosonic commutation relations corresponding to the field operators.
The problem we wish to solve may now be formulated. In the present section we would like to derive a multiconfigurational theory for atom–molecule conversion which is exact in the smallest Hilbert subspace possible for bosonic species, namely, the Hilbert space spanned by the [*single*]{} molecular orbital $\psi_m(\r,t)$ and [*single*]{} atomic orbital $\phi_a(\r,t)$. We term this specific case of the general theory: [*conversion mean field*]{}. More technically, this theory is a fully-variational extension of the literature two-mode approximation [@2M1; @Vardi] and, of course, of the Gross-Pitaevskii theory with conversion [@Timmermans_review].
The multiconfigurational wavefunction takes on the following form: \[2orb\_Phi\] & & |(t)> = \_[p=0]{}\^ C\_p(t) |N-2p,p;t>,\
& & |N-2p,p;t> = (b\_a\^(t))\^[N-2p]{} (c\_m\^(t))\^p |vac>, where $\left|vac\right>$ is a [*common*]{} vacuum of no atoms and no molecules. The index $p$ enumerates the number of bosonic molecules in the system. The corresponding number of atoms is $N-2p$. $N$ is the maximal number of atoms in the system which is obtained when there are no molecules. Obviously, $\left|\Psi(t)\right>$ is an eigenfunction of the particle-number operator $\hat N$, Eq. (\[particle\_number\]), with the eigenvalue $N$. The atomic and molecular number operators in the relevant Hilbert space boil down to $\hat N_a=\hat b_a^\dag \hat b_a$ and $\hat N_m=\hat c_m^\dag \hat c_m$, respectively. The size of this Hilbert space is $\left[\frac{N}{2}\right]+1$.
The functional action $S$ of the time-dependent Schrödinger equation and its evaluation {#two_mode_AM_energy}
---------------------------------------------------------------------------------------
Solving the time-dependent Schrödinger equation with the Hamiltonian (\[ham\_contact\_inter\]) and the multiconfigurational ansatz (\[2orb\_Phi\]) means finding the equations governing the time evolution of the atomic and molecular orbitals, $\phi_a(\r,t)$ and $\psi_m(\r,t)$, and of the expansion coefficients $\{C_p(t)\}$. The derivation of these equations of motion for $\phi_a(\r,t)$, $\psi_m(\r,t)$, and $\{C_p(t)\}$ requires a time-dependent variational principle. We employ the Lagrangian formulation of the (Dirac-Frenkel) time-dependent variational principle [@LF1; @LF2], also see Refs. [@MCTDHB2; @Unified_paper; @MIX], and write the functional action of the time-dependent Schrödinger equation which takes on the form: \[action\_functional\_AM\] & & S= dt {\
& & - \_a(t) - \_m(t)\
& & - (t)}. The time-dependent Lagrange multiplies $\mu_a(t)$, $\mu_m(t)$ and $\varepsilon(t)$ are introduced to ensure normalization of the atomic $\phi_a(\r,t)$ and molecular $\psi_m(\r,t)$ orbitals and of the expansion coefficients $\{C_p(t)\}$ at all times. $\mu_a(t)$ and $\mu_m(t)$ also serve another role. They exactly “compensate” for those terms appearing within the Dirac-Frenkel formulation of the variational principle $\left<\delta \Psi(t)\left| \hat H^{(2a\leftrightharpoons m)} -i\frac{\partial}{\partial t} \right|\Psi(t)\right>$ [@DF1; @DF2], i.e., when the variation of $\Psi(t)$ is performed before the expectation value $\left<\Psi(t)\left| \hat H^{(2a\leftrightharpoons m)} -i\frac{\partial}{\partial t} \right|\Psi(t)\right>$ is evaluated; see in this context [@MCTDHB2; @LF2]. We shall see below and more elaborately in appendix \[Lagrange\_two\_mode\] that these Lagrange multipliers can be eliminated from the resulting equations of motion by making use of the normalization of the orbitals in combination with unitary transformations.
The expectation value of the Hamiltonian in Eq. (\[action\_functional\_AM\]) can be expressed in two equivalent forms. The first form depends explicitly on the orbitals $\phi_a(\r,t)$, $\psi_m(\r,t)$ and the second on the expansion coefficients $\{C_p(t)\}$. The two forms are needed to derive the respective equations of motion for the orbitals and expansion coefficients.
### Orbital-explicit expression of $S$ {#two_mode_AM_energy_orbital}
Utilizing the multiconfigurational expansion (\[2orb\_Phi\]) and the individual terms of the many-body Hamiltonian (\[ham\_contact\_inter\]), the first form of the expectation value of $\hat H^{(a\leftrightharpoons m)} - i\frac{\partial}{\partial t}$ reads: \[Hamiltonian\_matrix\_element\_AM\_a\] & & = +\
& & + + +\
& & + + \_[am]{} +\
& & + - i \_[p=0]{}\^ C\_p\^. In Eq. (\[Hamiltonian\_matrix\_element\_AM\_a\]) and hereafter we use the shorthand notation for expectation values of operators with respect to $\Psi(t)$: $\left<\hat N_a\right> \equiv \left<\Psi(t)\left|\hat N_a\right|\Psi(t)\right>$, $\left<\hat b_a^\dag \hat b_a^\dag \hat c_m\right> \equiv
\left<\Psi(t)\left|\hat b_a^\dag \hat b_a^\dag \hat c_m\right|\Psi(t)\right>$, etc. We can indeed see that expression (\[Hamiltonian\_matrix\_element\_AM\_a\]) depends explicitly on the orbitals $\phi_a(\r,t)$ and $\psi_m(\r,t)$ through integrals over one-body terms, two-body interaction terms, and the conversion term: $\left<\phi_a\left|\hat h^{(a)} - i\frac{\partial}{\partial t} \right|\phi_a\right>$, $\left<\phi_a\psi_m\left|\right.\phi_a\psi_m\right>$, etc.
Making use of the multiconfigurational expansion (\[2orb\_Phi\]), we can express the above expectation values $\left<\Psi(t)\left| \ldots \right|\Psi(t)\right>$ in a closed form. The expectation values of particle-conserving operators, like the number operators $\hat N_a$ and $\hat N_m$, with respect to $\Psi(t)$ read: \[matrix\_elements\_fock1\_conserve\] & & = \_[p=0]{}\^ (N-2p) |C\_p(t)|\^2, = \_[p=0]{}\^ p |C\_p(t)|\^2,\
& & = \_[p=0]{}\^ (N-2p)(N-2p-1) |C\_p(t)|\^2,\
& & = \_[p=0]{}\^ p(p-1) |C\_p(t)|\^2, = \_[p=0]{}\^ p(N-2p) |C\_p(t)|\^2. We see that the dependence of the expectation values (\[matrix\_elements\_fock1\_conserve\]) on the expansion coefficients is only through weighted sums $\sum_{p=0}^{\left[N/2\right]}$ of the terms $\left|C_p(t)\right|^2$, i.e., that configurations of a different number $p$ of molecules are not directly coupled. The expectation values of particle non-conserving operators, originating from the conversion of particles, are given by \[matrix\_elements\_fock1\_non\_conserve\] & & = \_[p=1]{}\^ C\_p\^(t)C\_[p-1]{}(t),\
& & = {}\^, and seen to couple directly configurations with a different number of $p$ and $p-1$ molecules.
### Expansion-coefficient-explicit expression of $S$ {#two_mode_AM_energy_coeff}
Utilizing the multiconfigurational expansion (\[2orb\_Phi\]) and the many-body Hamiltonian (\[ham\_contact\_inter\]) as a whole, we can express the expectation value of $\hat H^{(2a\leftrightharpoons m)} - i\frac{\partial}{\partial t}$ in the functional action (\[action\_functional\_AM\]) as an explicit function of the expansion coefficients $\{C_p(t)\}$. One readily finds, \[Hamiltonian\_matrix\_element\_AM\_b\] & & =\
& & = \_[p=0]{}\^ C\_p\^. Eq. (\[Hamiltonian\_matrix\_element\_AM\_b\]) contains yet another type of matrix elements, which are the representation of $\hat H^{(2a\leftrightharpoons m)} - i\frac{\partial}{\partial t}$ in the subspace of configurations $\left\{\left|N-2p,p;t\right>\right\}$. These matrix elements can be evaluated explicitly. We divide them into two types, recalling that the Hamiltonian is expressed as a sum of particle-conserving and particle non-conserving parts, $\hat H^{(2a\rightleftharpoons m)} = \hat H^{(am)} + \hat W^{(2a\rightleftharpoons m)}$. The diagonal, or particle-conserving matrix elements read: \[matrix\_elements\_Hamil\_diag\] & & = (N-2p) +\
& & + (N-2p)(N-2p-1) + p +\
& & + p(p-1) + \_[am]{} p(N-2p) , p=0,…,\[N/2\]. The off-diagonal, particle non-conserving matrix elements, originating from the conversion term $\hat W^{(2a\rightleftharpoons m)}=\hat W^{(2a\rightharpoonup m)} + \hat W^{(m\rightharpoondown 2a)}$, take on the following form: \[matrix\_elements\_Hamil\_off\_diag\] & & =\
& & = , p=1,…,\[N/2\],\
& & =\
& & = {}\^, p=1,…,\[N/2\]. All other matrix elements of $\hat H^{(2a\leftrightharpoons m)} - i\frac{\partial}{\partial t}$ in the subspace of configurations $\left\{\left|N-2p,p;t\right>\right\}$ vanish. With explicit expressions of the functional action (\[action\_functional\_AM\]) we can now proceed and derive the equations of motion of $\Psi(t)$.
The equations of motion for $\Psi(t)$ {#two_mode_AM_EOM}
-------------------------------------
We perform the variation of the action functional (\[action\_functional\_AM\],\[Hamiltonian\_matrix\_element\_AM\_a\]) with respect to the orbitals and coefficients. Equating the variation of $S\left[\{C_p(t)\},\phi_a(\r,t),\psi_m(\r,t)\right]$ with respect to the orbitals to zero, eliminating the Lagrange multipliers $\mu_a(t)$ and $\mu_m(t)$ from the resulting equations (see appendix \[Lagrange\_two\_mode\] for details), and dividing the result by $\left<\hat N_a\right>$ and $\left<\hat N_m\right>$ respectively, we obtained the following equations of motion for the orbitals: \[simplest\_EOM\_orbitals\] & & \^[(a)]{} i|\_a> = \^[(a)]{} {|\_a> + \_[con]{}(t) \_a\^|\_m>},\
& & \^[(m)]{} i|\_m> = \^[(m)]{} {|\_m> + \_a |\_a>}, where the shorthand notation $\dot \phi_a \equiv \frac{\partial \phi_a}{\partial t}$, $\dot \psi_m \equiv \frac{\partial \psi_m}{\partial t}$ is used here and hereafter. The “interaction strengths” are given by \[time\_dependent\_interaction\_str\] & & \_a(t)=\_a, \_[am]{}(t)=\_[am]{} , \_[con]{}(t) = \_[con]{} ,\
& & \_m(t)=\_m , \_[ma]{}(t) = \_[am]{} , ’\_[con]{}(t) = \_[con]{} and [*vary in time*]{} due to the conversion of atoms to molecules and vise versa.
The quantities appearing on both the right- and left-hand sides of equations of motion (\[simplest\_EOM\_orbitals\]) are projection operators and given by \[projection\_conver\_coherent\] \^[(a)]{} = 1 - |\_a> $.
Then,
we can compensate for transforming the orbitals
by the ``reverse'' transformation
of the expansion coefficients
${C\_p(t)} {\_p(t)}$.
Overall,
we write this
invariance of the many-body wavefunction as follows:
\beq\label{trans_appen_Psi_text}
\left|\Psi(t)\right> = \sum_{p=0}^{\left[N/2\right]} C_p(t) \left|N-2p,p;t\right> =
\sum_{p=0}^{\left[N/2\right]} \overline{C}_p(t) \overline{\left|N-2p,p;t\right>}. \
\eeq
Clearly,
unitary transformations of the orbitals
and the respective
transformation of the expansion coefficients
neither change the size of the Hilbert space
nor couple configurations
with a different number of molecules.
To express these properties
we use the same summation index $p$
in both the middle part and right-hand side
of Eq.~(\ref{trans_appen_Psi_text}).
We can now make use of the invariance relation (\ref{trans_appen_Psi_text})
to simplify Eq.~(\ref{simplest_EOM_orbitals}).
Specifically,
there exists one unitary transformation that
eliminates the projection operators
acting on the time-derivatives (left-hand sides)
in Eq.~(\ref{simplest_EOM_orbitals}),
without introducing any further constraint
into
the equations of motion;
see appendix \ref{Lagrange_two_mode} for more details.
The equations of motion for the
atomic and molecular orbitals thus
finally read:
\beqn\label{simplest_EOM_orbitals_final}
& & i\left|\dot \phi_a\right> =
\hat {\mathbf P}^{(a)}
\left\{\left[\hat h^{(a)} + \Lambda_a(t) |\phi_a|^2 + \Lambda_{am}(t) |\psi_m|^2 \right] \left|\phi_a\right> +
\sqrt{2} \Lambda_{con}(t) \phi_a^\ast \left|\psi_m\right>\right\}, \nonumber \\
& & i\left|\dot \psi_m\right> =
\hat {\mathbf P}^{(m)}
\left\{\left[\hat h^{(m)} + \Lambda_m(t) |\psi_m|^2 + \Lambda_{ma}(t) |\phi_a|^2 \right] \left|\psi_m\right> +
\frac{\Lambda'_{con}(t)}{\sqrt{2}} \phi_a \left|\phi_a\right>\right\}. \
\eeqn
Eq.~(\ref{simplest_EOM_orbitals_final})
has the following property.
Operating from the left with $ = 0, = 0, clearly ensuring that initially-normalized orbitals remain normalized for all times. We can see the meaning of the unitary transformation carrying Eq. (\[simplest\_EOM\_orbitals\]) to Eq. (\[simplest\_EOM\_orbitals\_final\]). This unitary transformation takes normalized time-dependent orbitals, $\left<\phi_a\left|\right.\phi_a\right>=1$ and $\left<\psi_m\left|\right.\psi_m\right>=1$, which therefore satisfy the general relations $\frac{\partial \left<\phi_a\left|\right.\phi_a\right>}{\partial t} =
\left<\dot\phi_a\left|\right.\phi_a\right> + \left<\phi_a\left|\right.\dot\phi_a\right> = 0$ and $\frac{\partial \left<\psi_m\left|\right.\psi_m\right>}{\partial t} =
\left<\dot\psi_m\left|\right.\psi_m\right> + \left<\psi_m\left|\right.\dot\psi_m\right> = 0$, and transforms them to time-dependent orbitals satisfying the specific differential condition (\[MCTDHI\_mix\_conv\_const\_coherent\]).
Before we move to the corresponding working equations for the expansion coefficients $\{C_p(t)\}$, it is instructive to enquire whether we could further simplify the equations of motion (\[simplest\_EOM\_orbitals\_final\]), by eliminating the projection operators $\hat {\mathbf P}^{(a)}$, $\hat {\mathbf P}^{(m)}$ also from the right-hand sides. The answer is in general negative. If we could eliminate the projection operators remaining on the right-hand sides, it means that conditions (\[MCTDHI\_mix\_conv\_const\_coherent\]) are not satisfied any more. What would then guarantee that the atomic and molecular orbitals remain normalized at all times? It turns out that the condition for that is: $\mathrm{Im}\left\{\lambda_{con}
\left<c_m^\dag b_a b_a\right> \left<\psi_m\left|\right.\phi_a\phi_a\right>\right\} = 0$ for all times. In turn, even if this condition is satisfied at $t=0$, it is not in general guaranteed that it remains so for all times. Thus, the [*presence of particle conversion*]{} does not allow one to eliminate the projection operators $\hat {\mathbf P}^{(a)}$, $\hat {\mathbf P}^{(m)}$ also from the right-hand sides of the equations of motion (\[simplest\_EOM\_orbitals\_final\]). Alternatively speaking, in the absence of particle conversion, it is possible to eliminate the projection operators completely from (\[simplest\_EOM\_orbitals\_final\]), see in this context [@TDMF].
To derive the equations of motion of $\{C_p(t)\}$, we equate the variation of the action functional (\[action\_functional\_AM\],\[Hamiltonian\_matrix\_element\_AM\_b\]) with respect to the expansion coefficients to zero and eliminate the Lagrange multiplier $\varepsilon(t)$ (see for details appendix \[Lagrange\_two\_mode\]). The following equations of motion are obtained: \[simplest\_EOM\_coeffieints\] & & \^[(2a m)]{}(t) (t) = i ,\
& & [H]{}\_[p,p’]{}\^[(2a m)]{}(t) = , where the vector $\C(t)$ collects the expansion coefficients $\{C_p(t)\}$. Eq. (\[simplest\_EOM\_coeffieints\]) is a set of coupled first-order differential equations with time-dependent coefficients, and preserves the norm of an initially-normalized vector of coefficients $\C(0)$. The time-dependent coefficients ${\mathcal H}_{p,p'}^{(2a \leftrightharpoons m)}(t)$, being the matrix representation of $\hat H^{(2a\leftrightharpoons m)} - i \frac{\partial}{\partial t}$ in the subspace of configurations $\left\{\left|N-2p,p;t\right>\right\}$ and hence depending on the atomic $\phi_a(\r,t)$ and molecular $\psi_m(\r,t)$ orbitals, are prescribed in the previous subsection \[two\_mode\_AM\_energy\].
Next, we make use of the invariance of the multiconfigurational wavefunction to unitary transformations (\[trans\_appen\_Psi\_text\]). Explicitly, the unitary transformation responsible for transforming Eq. (\[simplest\_EOM\_orbitals\]) for the orbitals to Eq. (\[simplest\_EOM\_orbitals\_final\]), transforms equations of motion (\[simplest\_EOM\_coeffieints\]) for the expansion coefficients to the final form: \[simplest\_EOM\_coeffieints\_final\] & & \^[(2a m)]{}(t) (t) = i ,\
& & H\_[p,p’]{}\^[(2a m)]{}(t) = . Equivalently, Eq. (\[simplest\_EOM\_coeffieints\_final\]) can be obtained from Eq. (\[simplest\_EOM\_coeffieints\]) by substituting into the latter the differential condition (\[MCTDHI\_mix\_conv\_const\_coherent\]).
The coupled sets of equations of motion for the atomic $\phi(\r,t)$ and molecular $\psi_m(\r,t)$ orbitals and expansion coefficients $\{C_p(t)\}$, Eqs. (\[simplest\_EOM\_orbitals\]) and (\[simplest\_EOM\_coeffieints\]) or, respectively, Eqs. (\[simplest\_EOM\_orbitals\_final\]) and (\[simplest\_EOM\_coeffieints\_final\]) constitute the [*conversion mean field*]{} theory (fully-variational two-mode approximation) for the interacting atomic–molecular system with conversion.
The stationary self-consistent coherent mean field (time-independent fully-variational two-mode approximation) {#two_mode_AM_stationary}
--------------------------------------------------------------------------------------------------------------
The theory presented above is a time-dependent many-body theory. It is certainly interesting to enquire what are the corresponding stationary many-body states of the atomic–molecular Hamiltonian (\[ham\_am\_1\]-\[ham\_am\_3\])? In other words, what are the self-consistent solutions that minimize (extremize) the expectation value $\left<\Psi\left|\hat H^{(2a\rightleftharpoons m)}\right|\Psi\right>$ for a given time-independent multiconfigurational ansatz $\left|\Psi\right> = \sum_{p=0}^{\left[N/2\right]} C_p \left|N-2p,p\right>$ assembled from time-independent atomic $\phi_a(\r)$ and molecular orbitals $\psi_m(\r)$? To get this stationary self-consistent coherent mean field (time-independent fully-variational two-mode approximation), we resort to imaginary time-propagation and set $t \to -it$ in the corresponding equations of motion.
Setting $t \to -it$ in (the left-hand side of) either Eq. (\[simplest\_EOM\_orbitals\]) or (\[simplest\_EOM\_orbitals\_final\]), the left-hand side decays to zero in time and the equation becomes time-independent. Then, by multiplying the result, respectively, by $\left<\hat N_a\right>$ and $\left<\hat N_m\right>$, and translating back the projection operators $\hat {\mathbf P}^{(a)}$ and $\hat {\mathbf P}^{(b)}$ to the corresponding Lagrange multipliers $\mu_a$ and $\mu_b$ (see appendix \[Lagrange\_two\_mode\]), we obtain the multiconfigurational self-consistent (time-independent) equations for the atomic and molecular orbitals: \[stationary\_equations\_orbitals\] & & |\_a> +\
& & + \_[con]{} \_a\^|\_m> = \_a |\_a>,\
& & |\_m> +\
& & + \_a |\_a> = \_m |\_m>. Similarly, restoring the Lagrange multiplier $\varepsilon(t)$ into either Eq. (\[simplest\_EOM\_coeffieints\]) or (\[simplest\_EOM\_coeffieints\_final\]), see in this respect appendix \[Lagrange\_two\_mode\], and setting $t \to -it$ therein, we obtain the stationary (self-consistent) eigenvalue equation \[stationary\_equations\_coeff\] & & \^[(2a m)]{} = ,\
& & H\_[p,p’]{}\^[(2a m)]{} = for the expansion coefficients. We see that the (redundant) time-dependent Lagrange multiplier $\varepsilon(t)$ of the time-dependent theory has emerged as the eigenenergy $\varepsilon = \left<\Psi\left|\hat H^{(2a\leftrightharpoons m)}\right|\Psi\right>$ of the stationary theory.
The theory “distilled” into Eqs. (\[stationary\_equations\_orbitals\],\[stationary\_equations\_coeff\]) is a fully-variational stationary theory for the interacting atomic–molecular system in presence of conversion, where a single orbital is allowed for the atoms and a single orbital to the molecules. It is a system of coupled eigenvalue-like equations for the orbitals and eigenvalue equation for the coefficients, thought non-linear and integrodifferential ones.
The general multiconfigurational theory with atom–molecule conversion {#full_Hilbert}
=====================================================================
In this section we develop a general many-body theory for atom–molecule conversion, by allowing the atoms and molecules to occupy more orbitals. Section \[contact\] builds the theory for the popular contact interaction, whereas the case of generic non-contact interactions is presented in section \[non\_contact\].
Formulation for contact interactions {#contact}
------------------------------------
### The multiconfigurational ansatz for the wavefunction {#contact_ansatz}
The multiconfigurational expansion mixes atomic–molecular states with different numbers of particles which are eigenfunctions of the particle-number operator $\hat N \left|\Psi(t)\right>=N\left|\Psi(t)\right>$: \[MCTDH\_AM\_Psi\] |(t)> &=& \_[p=0]{}\^[\[N/2\]]{} \_[\^p,\^p]{} C\_[\^p\^p]{}(t) |\^p,\^p;t>,\
|\^p,\^p;t> &&\
&& (b\_1\^(t))\^[n\_1\^p]{}(b\_M\^(t))\^[n\_M\^p]{} (c\_1\^(t))\^[m\_1\^p]{}(c\_[M’]{}\^(t))\^[m\_[M’]{}\^p]{}|vac>. We collect the individual occupations in the vectors $\vec{n}^p=(n_1^p,\ldots,n_M^p)$, $\vec{m}^p=(m_1,\ldots,m_{M'}^p)$. The number of bosonic atoms $|\vec{n}^p| \equiv n_1^p+\ldots+n_M^p = N-2p$ and molecules $|\vec{m}^p| \equiv m_1^p+\ldots+m_{M'}^p = p$ of each configuration $\left|\vec{n}^p,\vec{m}^p;t\right>$ satisfies the particle-conservation law $|\vec{n}^p|+2|\vec{m}^p|=N$. Observe that the number of molecules $p$ serves as an index to the occupation numbers $\vec{n}^p$ and $\vec{m}^p$. This simply reflects the fact that, for a given number of $N-2p$ atoms and $p$ molecules, the possible occupation numbers which the configurations can assume depend on $p$ itself. The index $p$ together with the occupation numbers $\vec{n}^p$, $\vec{m}^p$ make a unique representation of each configuration. The atomic and molecular number operators in the corresponding Hilbert space boil down to $\hat N_a=\sum_{k=1}^M \hat b_k^\dag \hat b_k$ and $\hat N_m=\sum_{k'=1}^{M'} \hat c_{k'}^\dag \hat c_{k'}$, respectively. The size of the resulting Hilbert space is given by $\sum_{p=0}^{[N/2]}
\begin{pmatrix}
N-2p+M-1 \cr
M-1
\end{pmatrix}
\begin{pmatrix}
p+M'-1 \cr
M'-1
\end{pmatrix}$, i.e., by the sum of products of the sizes of the respective Hilbert subspaces for $N-2p$ bosonic atoms with $M$ orbitals and $p$ bosonic molecules with $M'$ orbitals.
### Reduced density matrices for systems with particle conversion {#contact_reduced}
As part of the variational derivation we will need the expectation value of $\hat H^{(2a\rightleftharpoons m)} - i\frac{\partial}{\partial t}$ with respect to $\Psi(t)$. To this end, it will be proved valuable to define and employ the reduced density matrices of $\Psi(t)$. We remind that Löwdin has introduced the concept of reduced density matrices for systems of a fixed number of particles (identical fermions) [@Lowdin]. Nevertheless and although $\Psi(t)$ [*is not*]{} comprised of a fixed number of atoms or a fixed number of molecules, it is possible to define the reduced density matrices of a mixture of atoms and molecules [*with conversion*]{}.
Having at hand the normalized many-body wavefunction $\Psi(t)$, the reduced one-body density matrices of the atoms and molecules are defined by: \[reduced\_1B\_2mix\] & & \^[(a)]{}(\_1|\_2;t) = = \^M\_[k,q=1]{} \^[(a)]{}\_[kq]{}(t) \^\_k(\_2,t)\_q(\_1,t),\
& & \^[(m)]{}(\_1|\_2;t) = = \^[M’]{}\_[k’,q’=1]{} \^[(m)]{}\_[k’q’]{}(t) \^\_[k’]{}(\_2,t)\_[q’]{}(\_1,t), where the matrix elements $\rho_{kq}^{(a)}(t) = \left<\hat b_k^\dag \hat b_q\right>$ and $\rho_{k'q'}^{(m)}(t) = \left<\hat c_{k'}^\dag \hat c_{q'}\right>$ are prescribed in appendix \[matrix\_Appen\]. We collect these matrix elements as $\brho^{(a)}(t)=\left\{\rho_{kq}^{(a)}(t)\right\}$ and $\brho^{(m)}(t)=\left\{\rho_{k'q'}^{(m)}(t)\right\}$. Similarly, the reduced two-body density matrices of the atoms and molecules are defined by: \[reduced\_2B\_2mix\] & & \^[(a)]{}(\_1,\_2|\_3,\_4;t) = =\
& & = \^M\_[k,s,l,q=1]{} \^[(a)]{}\_[kslq]{}(t) \^\_k(\_3,t)\^\_s(\_4,t)\_l(\_2,t)\_q(\_1,t),\
& & \^[(m)]{}(\_1,\_2|\_3,\_4;t) = =\
& & = \^[M’]{}\_[k’,s’,l’,q’=1]{} \^[(m)]{}\_[k’s’l’q’]{}(t) \^\_[k’]{}(\_3,t)\^\_[s’]{}(\_4,t)\_[l’]{}(\_2,t)\_[q’]{}(\_1,t),\
& & \^[(am)]{}(\_1,\_2|\_3,\_4;t) = =\
& & = \^[M]{}\_[k,q=1]{} \^[M’]{}\_[k’,q’=1]{} \^[(am)]{}\_[kk’qq’]{}(t) \^\_[k]{}(\_3,t)\_[q]{}(\_1,t)\^\_[k’]{}(\_4,t)\_[q’]{}(\_2,t), where the matrix elements $\rho_{kslq}^{(a)}(t) = \left<\hat b_k^\dag \hat b_s^\dag \hat b_l \hat b_q\right>$, $\rho_{k's'l'q'}^{(m)}(t) = \left<\hat c_{k'}^\dag \hat c_{s'}^\dag \hat c_{l'} \hat c_{q'}\right>$, and $\rho_{kk'qq'}^{(am)}(t) = \left<\hat b_k^\dag \hat b_q \hat c_{k'}^\dag \hat c_{q'}\right>$ are prescribed in appendix \[matrix\_Appen\]. Because the reduced density matrices (\[reduced\_1B\_2mix\]) and (\[reduced\_2B\_2mix\]) directly only couple configurations with the same number of atoms and molecules, we will refer to them as [*particle-conserving*]{} reduced density matrices. In this context, $\rho^{(am)}(\r_1,\r_2|\r_3,\r_4;t)$ is the lowest-order [*inter-species*]{} particle-conserving reduced density matrix.
From the above discussion it is anticipated that, due to the conversion term (\[ham\_am\_3\]) in the Hamiltonian, another kind of reduced density matrices appear in the theory. Specifically, we define the [*particle non-conserving*]{} reduced density matrices as follows: \[reduced\_conver\] & & \^[(2am)]{}(\_1,\_2|\_3;t) = = \^[M’]{}\_[k’=1]{} \^[M]{}\_[k,q=1]{} \^[(2am)]{}\_[k’kq]{}(t) \^\_[k’]{}(\_3,t)\_[k]{}(\_2,t)\_[q]{}(\_1,t),\
& & \^[(m2a)]{}(\_3|\_2,\_1;t) = = \^[M’]{}\_[k’=1]{} \^[M]{}\_[k,q=1]{} \^[(m2a)]{}\_[qkk’]{}(t) \^\_q(\_1,t) \^\_k(\_2,t) \_[k’]{}(\_3,t),\
& & \^[(m2a)]{}(\_3|\_2,\_1;t) = {\^[(2am)]{}(\_1,\_2|\_3;t)}\^, \^[(m2a)]{}\_[qkk’]{}(t) = {\^[(2am)]{}\_[k’kq]{}(t)}\^. The matrix elements $\rho^{(2a\rightharpoonup m)}_{k'kq} = \left<\hat c_{k'}^\dag \hat b_k \hat b_q\right>$ are given in appendix \[matrix\_Appen\].
### The functional action $S$ and its evaluation {#contact_func}
We start from the functional action of the time-dependent Schrödinger equation which in the general multiconfigurational case takes on the form: \[action\_functional\_AM\_full\] & & S= dt {\
& & - \_[k,j=1]{}\^[M]{} \^[(a)]{}\_[kj]{}(t)- \_[k’,j’=1]{}\^[M’]{} \^[(m)]{}\_[k’j’]{}(t)\
& & - (t)}. The time-dependent Lagrange multiplies $\{\mu^{(a)}_{kj}(t)\}$, $\{\mu^{(m)}_{k'j'}(t)\}$ and $\varepsilon(t)$ are introduced to ensure orthonormalization of the atomic $\left\{\phi_k(\r,t)\right\}$ and molecular $\left\{\psi_{k'}(\r,t)\right\}$ orbital sets and normalization of the expansion coefficients $\{C_{\vec{n}^p\vec{m}^p}(t)\}$.
To derive the equations of motion for the atomic–molecular multiconfigurational wavefunction (\[MCTDH\_AM\_Psi\]), the expectation value of $\hat H^{(2a\rightleftharpoons m)} - i\frac{\partial}{\partial t}$ with respect to $\Psi(t)$ is needed, where $\hat H^{(2a\rightleftharpoons m)}$ is given in Eq. (\[ham\_contact\_inter\]). The expectation value of $\hat H^{(2a\rightleftharpoons m)} - i\frac{\partial}{\partial t}$ is expressed by two equivalent forms, as done in section \[two\_mode\_AM\_energy\]. The first form, where the dependence of Eq. (\[action\_functional\_AM\_full\]) on the atomic and molecular orbitals is explicit, reads: \[expectation\_ham\_a\] & & = \_[k,q=1]{}\^M \^[(a)]{}\_[kq]{} +\
& & + \_[k,s,l,q=1]{}\^M \^[(a)]{}\_[kslq]{} + \_[k’,q’=1]{}\^[M’]{} \^[(m)]{}\_[k’q’]{} +\
& & + \_[k’,s’,l’,q’=1]{}\^[M’]{} \^[(m)]{}\_[k’s’l’q’]{} + \_[am]{}\_[k,q=1]{}\^M \_[k’,q’=1]{}\^[M’]{} \^[(am)]{}\_[kk’qq’]{} +\
& & + \_[k,q=1]{}\^[M]{} \_[k’=1]{}\^[M’]{} - i \_[p=0]{}\^ \_[\^p,\^p]{} C\_[\^p\^p]{}\^. We see in (\[expectation\_ham\_a\]) the appearance of the particle-conserving and particle non-conserving reduced density matrices introduced in the previous subsection \[contact\_reduced\]. Eq. (\[expectation\_ham\_a\]) is to be used to derive the equations of motion of $\left\{\phi_k(\r,t)\right\}$ and $\left\{\psi_{k'}(\r,t)\right\}$.
The second form of the expectation value of $\hat H^{(2a\rightleftharpoons m)} - i\frac{\partial}{\partial t}$ in the functional action (\[action\_functional\_AM\_full\]), \[expectation\_ham\_b\] & & = \_[p=0]{}\^ \_[\^p,\^p]{} C\_[\^p\^p]{}\^\
& & , displays its explicit dependence on the expansion coefficients, and therefore will be employed to derive the equations of motion of $\{C_{\vec{n}^p\vec{m}^p}(t)\}$. Finally, it is deductive to compare the structure of Eqs. (\[Hamiltonian\_matrix\_element\_AM\_a\],\[Hamiltonian\_matrix\_element\_AM\_b\]) in the conversion mean field (fully-variational two-mode) problem to that of Eqs. (\[expectation\_ham\_a\],\[expectation\_ham\_b\]) of the general problem.
### The equations of motion for $\Psi(t)$ {#contact_equation}
Collecting the above ingredients, we are ready to perform the variation of the functional action $S\left[\{C_{\vec{n}^p\vec{m}^p}(t)\},\left\{\phi_k(\r,t)\right\},\left\{\psi_{k'}(\r,t)\right\}\right]$ and arrive at the equations of motion of $\Psi(t)$. Equating the variation of the action functional (\[action\_functional\_AM\_full\],\[expectation\_ham\_a\]) with respect to the orbitals to zero and eliminating the Lagrange multipliers $\{\mu_{kj}^{(a)}(t)\}$, $\{\mu_{k'j'}^{(m)}(t)\}$ (see appendix \[appen\_C\]), we obtain the following result, $j=1,\ldots,M$, $j'=1,\ldots,M'$: \[MCTDH\_conver\_orb\_eqs\_P\] & & \^[(a)]{} i|\_j> = \^[(a)]{} ,\
& & \^[(m)]{} i|\_[j’]{}> = \^[(m)]{} , where terms with products of reduced two-body density matrices times orbital pairs are collected together and denoted for brevity as \[rho\_phi\] & & {\_2(\^2,\^2)}\^[(a)]{}\_[kq]{} \_a \^M\_[s,l=1]{} \^[(a)]{}\_[kslq]{} (\_s\^\_l) + \_[am]{} \_[k’,q’=1]{}\^[M’]{} \^[(am)]{}\_[kk’qq’]{} (\_[k’]{}\^\_[q’]{}),\
& & {\_2(\^2,\^2)}\^[(m)]{}\_[k’q’]{} \_m \^[M’]{}\_[s’,l’=1]{} \^[(m)]{}\_[k’s’l’q’]{} (\_[s’]{}\^\_[l’]{}) + \_[am]{} \_[k,q=1]{}\^[M]{} \^[(am)]{}\_[kk’qq’]{} (\_[k]{}\^\_[q]{}), and \[projection\_conver\] \^[(a)]{} = 1 - \_[u=1]{}\^[M]{}|\_u> $.
Then,
we can compensate for the transformations of the orbital sets
by the ``reverse'' transformation of the expansion coefficients
${C\_[\^p\^p]{}(t)} {\_[\^p\^p]{}(t)}$.
We represent this invariance by the following equality:
\beq\label{invariance_rotations}
\left|\Psi(t)\right> =
\sum_{p=0}^{[N/2]} \sum_{\vec{n}^p,\vec{m}^p}
C_{\vec{n}^p\vec{m}^p}(t) \left|\vec{n}^p,\vec{m}^p;t\right> =
\sum_{p=0}^{[N/2]} \sum_{\vec{n}^p,\vec{m}^p}
\overline{C}_{\vec{n}^p\vec{m}^p}(t) \overline{\left|\vec{n}^p,\vec{m}^p;t\right>}.
\eeq
The transformations of the
orbital sets
and expansion coefficients do not change the size
of the Hilbert space,
or couple systems with different numbers of atoms, molecules.
We remark that
transformations which inter-mix atomic and molecular
orbitals are not required for our needs.
To represent these properties,
the same occupation numbers
$\^p$, $\^p$ and summation
index $p$ of the number of molecules
are used for both
multiconfigurational expansions of $(t)$
in Eq.~(\ref{invariance_rotations}).
We can now make use of the invariance (\ref{invariance_rotations})
to simplify the equations of motion (\ref{MCTDH_conver_orb_eqs_P}),
without introducing further constraints into the equations of motion.
We utilize
a specific unitary transformation of the many-particle
wavefunction that
eliminates the projection operators
acting on the time-derivatives (left-hand sides)
in Eq.~(\ref{MCTDH_conver_orb_eqs_P});
see appendix \ref{appen_C} for more details.
The final result
for the equations of motion for the
atomic ${\_k(,t)}$ and molecular ${\_[k’]{}(,t)}$
orbitals thus takes on the form,
$j=1,…,M$, $j’=1,…,M’$:
\beqn\label{MCTDH_conver_orb_eqs}
& & \!\!\!\!\!\!\!\! i\left|\dot\phi_j\right> =
\hat{\mathbf P}^{(a)} \Bigg[\hat h^{(a)} \left|\phi_j\right> + \nonumber \\
& & + \sum^M_{k=1} \left\{\brho^{(a)}(t)\right\}^{-1}_{jk} \times \sum^M_{q=1}
\Bigg( \left\{\rho_2(\phi^2,\psi^2)\right\}^{(a)}_{kq} \left|\phi_q\right>
+ \sqrt{2} \lambda_{con} \sum_{k'=1}^{M'}
\rho^{(m\rightharpoondown 2a)}_{qkk'} \phi_q^\ast \left|\psi_{k'}\right> \Bigg) \Bigg], \nonumber \\
& & \!\!\!\!\!\!\!\! i\left|\dot\psi_{j'}\right> =
\hat{\mathbf P}^{(m)} \Bigg[\hat h^{(m)} \left|\psi_{j'}\right> + \nonumber \\
& & + \sum^{M'}_{k'=1} \left\{\brho^{(m)}(t)\right\}^{-1}_{j'k'} \times
\Bigg( \sum^{M'}_{q'=1} \left\{\rho_2(\psi^2,\phi^2)\right\}^{(m)}_{k'q'}
\left|\psi_{q'}\right>
+ \frac{\lambda_{con}}{\sqrt{2}} \sum_{k,q=1}^{M}
\rho^{(2a\rightharpoonup m)}_{kk'q} \phi_k \left|\phi_{q}\right> \Bigg) \Bigg], \
\eeqn
with the projection operators
$\^[(a)]{}$ and
$\^[(m)]{}$
appearing
now on the right-hand sides only.
Now,
taking the respective scalar products of (\ref{MCTDH_conver_orb_eqs})
with ${ = 0, k,q=1,…,M, = 0, k’,q’=1,…,M’. It is instructive to mention that these differential condition have been introduced originally by the MCTDH developers [@CPL; @JCP], and used thereafter in particle-conserving multiconfigurational theories for identical particles and mixtures [@MCTDHF1; @MCTDHF2; @MCTDHF3; @MCTDHB1; @MCTDHB2; @Unified_paper; @MIX].
The differential conditions (\[MCTDHI\_mix\_conv\_const\]) ensure that initially-orthonormalized orbital sets $\left\{\phi_k(\r,t)\right\}$, $\left\{\psi_{k'}(\r,t)\right\}$ remain orthonormalized at all times. The meaning of the unitary transformation carrying equations of motion (\[MCTDH\_conver\_orb\_eqs\_P\]) to (\[MCTDH\_conver\_orb\_eqs\]) can now be seen. This unitary transformation takes orthonormal time-dependent orbitals, $\left<\phi_k\left|\right.\phi_q\right> = \delta_{kq}$ and $\left<\psi_{k'}\left|\right.\psi_{q'}\right> = \delta_{k'q'}$ \[$\delta_{kq}$, $\delta_{k'q'}$ is the Dirac delta-function\], which therefore satisfy the general relations $\frac{\partial \left<\phi_k\left|\right.\phi_q\right>}{\partial t} =
\left<\dot\phi_k\left|\right.\phi_q\right> + \left<\phi_k\left|\right.\dot\phi_q\right> = 0$ and $\frac{\partial \left<\psi_{k'}\left|\right.\psi_{q'}\right>}{\partial t} =
\left<\dot\psi_{k'}\left|\right.\psi_{q'}\right> + \left<\psi_{k'}\left|\right.\dot\psi_{q'}\right> = 0$, and transforms them to time-dependent orbitals satisfying the specific differential conditions (\[MCTDHI\_mix\_conv\_const\]).
Moving to the equations of motion for the coefficients $\{C_{\vec{n}^p\vec{m}^p}(t)\}$, we equate the variation of the functional action (\[action\_functional\_AM\_full\],\[expectation\_ham\_b\]) with respect to the expansion coefficients to zero. Eliminating the Lagrange multiplier $\varepsilon(t)$ by a respective phase transformation of the coefficients, we arrive at the form: \[MCTDH\_conver\_coeff\_eqs\_P\] & & \^[(2a m)]{}(t)(t) = i,\
& & [H]{}\^[(2a m)]{}\_[\^p\^p,’\^[p’]{}’\^[p’]{}]{}(t) = . Eq. (\[MCTDH\_conver\_coeff\_eqs\_P\]) has exactly the same form as Eq. (\[simplest\_EOM\_coeffieints\]) in the specific theory of section \[two\_mode\_AM\_EOM\], and constitutes a set of coupled first-order differential equations with time-dependent coefficients that preserve the norm of an initially-normalized vector of coefficients $\C(0)$. The matrix elements of $\hat H^{(2a\rightleftharpoons m)} - i\frac{\partial}{\partial t}$ with respect to two general configurations, ${\cal H}^{(2a \leftrightharpoons m)}_{\vec{n}^p\vec{m}^p,\vec{n}'^{p'}\vec{m}'^{p'}}(t)$, are prescribed in appendix \[matrix\_Appen\].
Now, to arrive at the final form of the equations of motion for the coefficients, we make use of the invariance of $\Psi(t)$ to unitary transformations (\[invariance\_rotations\]). Specifically, the unitary transformation carrying Eq. (\[MCTDH\_conver\_orb\_eqs\_P\]) for the orbitals to Eq. (\[MCTDH\_conver\_orb\_eqs\]), casts Eq. (\[MCTDH\_conver\_coeff\_eqs\_P\]) for the expansion coefficients into the final result: \[MCTDH\_conver\_coeff\_eqs\] & & \^[(2a m)]{}(t)(t) = i,\
& & H\^[(2a m)]{}\_[\^p\^p,’\^[p’]{}’\^[p’]{}]{}(t) = . Eq. (\[MCTDH\_conver\_coeff\_eqs\]) has exactly the same form as Eq. (\[simplest\_EOM\_coeffieints\_final\]) in the specific case of section \[two\_mode\_AM\_EOM\]. Of course, there are much more expansion coefficients in the general case. Finally and equivalently, we note that the result (\[MCTDH\_conver\_coeff\_eqs\]) can be obtained from Eq. (\[MCTDH\_conver\_coeff\_eqs\_P\]) when the differential condition (\[MCTDHI\_mix\_conv\_const\]) is substituted into the latter.
Let us pause for a moment and summarize. We have started from the functional action (\[action\_functional\_AM\_full\]) and arrived at the equations of motion for $\Psi(t)$. Eq. (\[MCTDH\_conver\_orb\_eqs\_P\]) for the orbitals $\left\{\phi_k(\r,t)\right\}$, $\left\{\psi_{k'}(\r,t)\right\}$ and Eq. (\[MCTDH\_conver\_coeff\_eqs\_P\]) for the expansion coefficients $\{C_{\vec{n}^p\vec{m}^p}(t)\}$, or, respectively, Eqs. (\[MCTDH\_conver\_orb\_eqs\]) and (\[MCTDH\_conver\_coeff\_eqs\]) constitute the multiconfigurational time-dependent Hartree theory for systems with particle conversion; here specifically the theory for bosonic atoms and bosonic molecules with conversion.
### The stationary self-consistent general multiconfigurational theory with conversion {#contact_equation_stationary}
The theory presented above is a time-dependent many-body theory and, as done in the previous section \[two\_mode\_AM\], it is relevant to put forward the corresponding stationary general theory. Consider the multiconfigurational expansion $\left|\Psi\right> = \sum_{p=0}^{[N/2]} \sum_{\vec{n}^p,\vec{m}^p}
C_{\vec{n}^p\vec{m}^p}\left|\vec{n}^p,\vec{m}^p\right>$, where the expansion coefficients $\{C_{\vec{n}^p\vec{m}^p}\}$ and orbitals $\left\{\phi_k(\r)\right\}$, $\left\{\psi_{k'}(\r)\right\}$ assembling the configurations $\left\{\left|\vec{n}^p,\vec{m}^p\right>\right\}$ are time-independent quantities. What are then the self-consistent solutions that minimize (extremize) the expectation value $\left<\Psi\left|\hat H^{(2a\leftrightharpoons m)}\right|\Psi\right>$?
The working equations of the stationary theory can be obtained by resorting to imaginary time-propagation and setting $t \to -it$ in either equations of motion (\[MCTDH\_conver\_orb\_eqs\_P\]) and (\[MCTDH\_conver\_coeff\_eqs\_P\]) or equations of motion (\[MCTDH\_conver\_orb\_eqs\]) and (\[MCTDH\_conver\_coeff\_eqs\]), after the time-dependent Lagrange multiplier $\varepsilon(t)$ has been reinstated. Then, by translating the projection operators $\hat{\mathbf P}^{(a)}$, $\hat{\mathbf P}^{(m)}$ to the respective Lagrange multipliers $\{\mu^{(a)}_{kj}\}$, $\{\mu^{(m)}_{k'j'}\}$, the resulting working equations take on the form: \[MCTDH\_conver\_stationary\_orb\] & & \_[q=1]{}\^M { |\_q> + \_[con]{} \_[k’=1]{}\^[M’]{} \^[(m2a)]{}\_[qkk’]{} \_q\^|\_[k’]{}> } = \_[j=1]{}\^M \^[(a)]{}\_[kj]{} |\_j>,\
& & k=1,…,M,\
& & \_[q’=1]{}\^[M’]{} |\_[q’]{}> + \_[k,q=1]{}\^[M]{} \^[(2am)]{}\_[k’kq]{} \_k |\_[q]{}> = \_[j’=1]{}\^[M’]{} \^[(m)]{}\_[k’j’]{} |\_[j’]{}>,\
& & k’=1,…,M’ for the orbitals, and \[MCTDH\_conver\_stationary\_coeff\] & & \^[(2am)]{} = ,\
& & H\^[(2a m)]{}\_[\^p\^p,’\^[p’]{}’\^[p’]{}]{} = for the expansion coefficients. As seen in section \[two\_mode\_AM\], the time-dependent Lagrange multiplier $\varepsilon(t)$ emerges in the time-independent theory as the eigenenergy of the coupled atom–molecule system with $\varepsilon=\left<\Psi\left|\hat H^{(2a\leftrightharpoons m)}\right|\Psi\right>$.
The stationary equations for the orbitals (\[MCTDH\_conver\_stationary\_orb\]) can be further simplified. The stationary wavefunction $\Psi$, as its time-dependent counterpart, is invariant to independent unitary transformations of the orbital sets $\left\{\phi_k(\r)\right\}$, $\left\{\psi_{k'}(\r)\right\}$ and the “inverse” transformation of the expansion coefficients $\{C_{\vec{n}^p\vec{m}^p}\}$. We can use the unitary matrices which diagonalize the matrices of Lagrange multipliers $\{\mu^{(a)}_{kj}\}$, $\{\mu^{(m)}_{k'j'}\}$. We note that the matrices of Lagrange multipliers are Hermitian matrices for stationary states. As a result of this transformation, we obtain a set of coupled equations for the orbitals that look just as Eq. (\[MCTDH\_conver\_stationary\_orb\]), except for the right-hand sides being diagonal: \[diagonal\_form\] & & \_[q=1]{}\^M { |\_q> + \_[con]{} \_[k’=1]{}\^[M’]{} \^[(m2a)]{}\_[qkk’]{} \_q\^|\_[k’]{}> } = \^[(a)]{}\_[k]{} |\_k>,\
& & k=1,…,M,\
& & \_[q’=1]{}\^[M’]{} |\_[q’]{}> + \_[k,q=1]{}\^[M]{} \^[(2am)]{}\_[k’kq]{} \_k |\_[q]{}> = \^[(m)]{}\_[k’]{}|\_[k’]{}>,\
& & k’=1,…,M’. The form of the equation for the expansion coefficients (\[MCTDH\_conver\_stationary\_coeff\]) does not change. Thus, the final result for the stationary theory, Eqs. (\[MCTDH\_conver\_stationary\_coeff\],\[diagonal\_form\]), is a coupled system of integrodifferential, non-linear equations constituting eigenvalue-like equations for the orbitals and eigenvalue equation for the expansion coefficients; compare to section \[two\_mode\_AM\_stationary\]. Generally, the transformation of the matrices of Lagrange multipliers to diagonal form would make the orbitals delocalized. Hence, in problems where working with localized orbitals is of advantage or relevance, for instance in lattices, it is the form (\[MCTDH\_conver\_stationary\_orb\]) with the in-general non-diagonal Lagrange multipliers which is to be preferred.
Formulation for general interactions {#non_contact}
------------------------------------
The last stage of the theory is to return to the case of generic non-contact interactions in the Hamiltonian (\[ham\_am\_1\]-\[ham\_am\_3\]) and have the respective theory derived. The derivation of the equations of motion follows essentially the same steps taken in the previous subsection \[contact\] and there is obviously no need to repeat it. The only extra care needed is when minimizing the conversion term in the functional action with respect to the molecular orbitals, where exchange of variables is used. This point and related derivations are discussed in appendix \[appen\_C\]. Below, we report the final results of the time-dependent as well as the self-consistent time-independent theories.
The final form of the time-dependent equations of motion of the orbitals reads, $j=1,\ldots,M$, $j'=1,\ldots,M'$: \[MCTDH\_W\_conver\_orb\_eqs\] & & i|\_j> = \^[(a)]{} ,\
& & i|\_[j’]{}> = \^[(m)]{} , where terms with products of reduced two-body density matrices times one-body potentials (see below) are collected together and denoted for brevity as \[rho\_W\] & & {\_2W}\^[(a)]{}\_[kq]{} \^M\_[s,l=1]{} \^[(a)]{}\_[kslq]{} W\^[(a)]{}\_[sl]{} + \_[k’,q’=1]{}\^[M’]{} \^[(am)]{}\_[kk’qq’]{} W\^[(am)]{}\_[k’q’]{},\
& & {\_2W}\^[(m)]{}\_[k’q’]{} \^[M’]{}\_[s’,l’=1]{} \^[(m)]{}\_[k’s’l’q’]{} W\^[(m)]{}\_[s’l’]{} + \_[k,q=1]{}\^[M]{} \^[(am)]{}\_[kk’qq’]{} W\^[(ma)]{}\_[kq]{}. Comparing Eq. (\[MCTDH\_conver\_orb\_eqs\]) of the previous subsection \[contact\] to Eq. (\[MCTDH\_W\_conver\_orb\_eqs\]), we see that in the latter more general, time-dependent local potentials appear which are given explicitly by: \[one\_body\_pot\] & & W\^[(a)]{}\_[sl]{}(,t) = \_s\^(’,t) W\^[(a)]{}(,’) \_l(’,t) d’,\
& & W\^[(m)]{}\_[s’l’]{}(,t) = \_[s’]{}\^(’,t) W\^[(m)]{}(,’) \_[l’]{}(’,t) d’,\
& & W\^[(am)]{}\_[k’q’]{}(,t) = \_[k’]{}\^(’,t) W\^[(am)]{}(,’) \_[q’]{}(’,t) d’,\
& & W\^[(ma)]{}\_[kq]{}(,t) = \_[k]{}\^(’,t) W\^[(am)]{}(,’) \_[q]{}(’,t) d’,\
& & W\^[(2am)]{}\_[kq]{}(,t) = d’ W\^[(2am)]{}(+,-) \_k(+,t) \_q(-,t),\
& & W\^[(m2a)]{}\_[qk’]{}(,t) = d’ \_q\^(’,t) W\^[(m2a)]{}(,’) \_[k’]{}(,t). These potentials derive from the interaction terms and conversion term in the Hamiltonian (\[ham\_am\_1\]-\[ham\_am\_3\]) and, in the specific case of contact particle-particle interactions, boil down to products-of-orbitals, see for comparison Eq. (\[MCTDH\_conver\_orb\_eqs\]). The form of the equations of motion for the corresponding expansion coefficients, \[MCTDH\_W\_conver\_coeff\_eqs\] & & \^[(2a m)]{}(t)(t) = i,\
& & H\^[(2a m)]{}\_[\^p\^p,’\^[p’]{}’\^[p’]{}]{}(t) = , does not change for general interactions. Of course, the matrix elements $H^{(2a \leftrightharpoons m)}_{\vec{n}^p\vec{m}^p,\vec{n}'^{p'}\vec{m}'^{p'}}(t)$ do depend on the specific form of the particle-particle interactions.
Finally, the self-consistent, time-independent general theory is obtained from the time-dependent one by taking $t \to -it$. The stationary self-consistent equations for the orbitals read: \[MCTDH\_conver\_stationary\_orb\_gen\] & & \_[q=1]{}\^M { |\_q> + \_[k’=1]{}\^[M’]{} \^[(m2a)]{}\_[qkk’]{} W\^[(m2a)]{}\_[qk’]{} } = \_[j=1]{}\^M \^[(a)]{}\_[kj]{} |\_j> \^[(a)]{}\_[k]{} |\_k>,\
& & k=1,…,M,\
& & \_[q’=1]{}\^[M’]{} |\_[q’]{}> + \_[k,q=1]{}\^[M]{} \^[(2am)]{}\_[k’kq]{} W\^[(2am)]{}\_[kq]{} = \_[j’=1]{}\^[M’]{} \^[(m)]{}\_[k’j’]{} |\_[j’]{}> \^[(m)]{}\_[k’]{} |\_[k’]{}>,\
& & k’=1,…,M’, where the arrows indicate the Lagrange multipliers in their diagonal form, as done in Eq. (\[diagonal\_form\]). Finally, the self-consistent–eigenvalue form of the equation for the expansion coefficients, \[MCTDH\_conver\_stationary\_coeff\_gen\] & & \^[(2am)]{} =\
& & H\^[(2a m)]{}\_[\^p\^p,’\^[p’]{}’\^[p’]{}]{} = , remains unchanged for general interactions.
Eqs. (\[MCTDH\_W\_conver\_orb\_eqs\]) and (\[MCTDH\_W\_conver\_coeff\_eqs\]) constitute a multiconfigurational [*time-dependent*]{} theory for systems of bosonic atoms and molecules with conversion (particle conversion in the generic case). Furthermore, Eqs. (\[MCTDH\_conver\_stationary\_orb\_gen\]) and (\[MCTDH\_conver\_stationary\_coeff\_gen\]) constitute a multiconfigurational self-consistent [*time-independent*]{} theory for systems of bosonic atoms and molecules with conversion (particle conversion in the generic case). Both theories extend the scope of the successful multiconfigurational time-dependent Hartree method and its versions specified for systems of identical particles and mixtures to new physical systems and problems.
Summary and concluding remarks {#dis_sum}
==============================
In this work we have derived a many-body propagation theory for systems with particle conversion. The theory is intended for systems with a finite number of interacting particles, typically in a trap potential. The theory has been exemplified and working equations have been explicitly derived for systems of interacting structureless bosonic atoms and bosonic molecules undergoing the conversion ‘reaction’ $2a \leftrightharpoons m$. In doing so, we have also extended the scope of the successful multiconfigurational time-dependent Hartree (MCTDH) method and its versions specified for systems of identical particles and mixtures to new physical systems and problems. We note that the MCTDH method is considered at present the most efficient wave-packet propagation approach for in-general distinguishable coupled degrees-of-freedom, with no particle conversion, of course. The general multiconfigurational theory with particle conversion shall be referred to as MCTDH-[*conversion*]{}, whereas the explicit scenario derived throughout this work by MCTDH-\[$2a{\leftrightharpoons}m$\].
To treat systems with particle conversion, one has to work in second quantization formalism, where the Hamiltonian with particle-conversion terms can be represented. The next step is to define the configurations. In presence of particle conversion configurations with different numbers of atoms and different numbers of molecules are coupled. For instance, consider the particular case of the conversion ‘reaction’ $2a \leftrightharpoons m$. In this case, the subspace of coupled configurations can be easily obtained by starting from the configurations made of $N$ atoms only, and operating repeatedly with the conversion operators in the Hamiltonian until configurations made of the maximal number $\left[\frac{N}{2}\right]$ of molecules are reached.
In the multiconfigurational theory for the ’reaction’ $2a \leftrightharpoons m$, there are $M$ [*time-dependent*]{} orbitals $\{\phi_k(\r,t)\}$ available for the atoms and $M'$ [*time-dependent*]{} orbitals $\{\psi_{k'}(\r,t)\}$ for the molecules. The multiconfigurational ansatz for the many-particle wavefunction $\Psi(t)$ is taken as linear combination with [*time-dependent*]{} coefficients $\{C_{\vec{n}^p\vec{m}^p}(t)\}$ of all possible configurations $\left\{\left|\vec{n}^p,\vec{m}^p;t\right>\right\}$ assembled from $p$ molecules and $N-2p$ atoms – distributed over the $M$ and $M'$ respective orbitals – and coupled by the conversion term in the Hamiltonian.
The evolution of $\Psi(t)$ is then determined by the Dirac-Frenkel time-dependent variational principle. Utilizing the Lagrangian formulation of the Dirac-Frenkel variational principle, one arrives at two sets of coupled equations of motion: The first set is for the orbitals $\{\phi_k(\r,t)\}$ and $\{\psi_{k'}(\r,t)\}$, and the second for the expansion coefficients $\{C_{\vec{n}^p\vec{m}^p}(t)\}$. The first set is comprised of first-order differential equations in time and non-linear integrodifferential (for non-contact interactions) in position space. The second set consists of first-order differential equations with coefficients forming a time-dependent Hermitian matrix. Thus, equations of motion (\[MCTDH\_conver\_orb\_eqs\_P\],\[MCTDH\_conver\_coeff\_eqs\_P\]), or Eqs. (\[MCTDH\_conver\_orb\_eqs\],\[MCTDH\_conver\_coeff\_eqs\]), constitute the time-dependent multiconfigurational theory for bosonic atoms and molecules with conversion – MCTDH-\[$2a{\leftrightharpoons}m$\].
The structure of the equations of motion for systems with particle conversion reminds of the structure of the equations of motion in multiconfigurational time-dependent theories for systems of identical particles and mixtures [@MCTDHB1; @MCTDHB2; @Unified_paper; @MIX]: (i) There are projection operators on the right-hand-sides of the equations of motion for the orbitals, ensuring that the respective orbitals remain normalized and orthogonal to one another for all times; (ii) The equations for the expansion coefficients are first-order differential equations with time-dependent coefficients; and (iii) The equations of motion for the orbitals are formulated in terms of reduced density matrices. This resemblance would allow one to transfer the effective numerical techniques that have been developed in the past almost-twenty years for multiconfigurational time-dependent many-body systems [*without*]{} particle conversion [@CPL; @JCP; @PR; @MCTDHF1; @MCTDHF2; @MCTDHF3; @MCTDHB1; @MCTDHB2; @Unified_paper; @MIX] to the present theory for systems [*with*]{} particle conversion.
Particular attention has been paid to the reduced density matrices appearing in the theory. As the multiconfigurational expansion involves configurations with different numbers of atoms and molecules, two types of reduced density matrices are defined. There are [*particle-conserving*]{} reduced density matrices, $\rho^{(a)}(\r_1|\r_2;t)$, $\rho^{(m)}(\r_1|\r_2;t)$, $\rho^{(a)}(\r_1,\r_2|\r_3,\r_4;t)$, $\rho^{(m)}(\r_1,\r_2|\r_3,\r_4;t)$ and $\rho^{(am)}(\r_1,\r_2|\r_3,\r_4;t)$, which directly do not couple configurations with different numbers of atoms and molecules. Despite this property, the particle-conserving reduced density matrices, such as $\rho^{(a)}(\r_1|\r_2;t)$ and $\rho^{(a)}(\r_1,\r_2|\r_3,\r_4;t)$, [*are not*]{} the standard density matrices introduced by Löwdin [@Lowdin] for many-particle systems [*without*]{} conversion. The second type of reduced density matrices that appear in the theory are [*particle non-conserving*]{} reduced density matrices, $\rho^{(a\rightharpoonup m)}(\r_1,\r_2|\r_3;t)$ and $\rho^{(m\rightharpoondown a)}(\r_3|\r_2,\r_1;t)$, and originate from the conversion term in the Hamiltonian. They obviously have no analogs in systems without conversion. Here, it is of interest by itself to study properties of particle-conserving reduced density matrices and certainly of particle non-conserving reduced density matrices in systems [*with*]{} particle conversion.
The time-dependent multiconfigurational theory MCTDH-\[$2a{\leftrightharpoons}m$\] readily admits the corresponding [*stationary*]{} theory. By resorting to imaginary time propagation, the equations of motion of the time-dependent theory boil down to the fully-self-consistent time-independent multiconfigurational theory, Eqs. (\[MCTDH\_conver\_stationary\_orb\_gen\],\[MCTDH\_conver\_stationary\_coeff\_gen\]), for stationary states of the system $2a \leftrightharpoons m$, in presence of all particle-particle interactions, of course. With this result, available self-consistent multiconfigurational theories for systems [*without*]{} particle conversion, noticeably for fermions [@Szabo_book; @elect], distinguishable degrees-of-freedom [@Dieter_review], bosons [@MCHB], and mixtures [@MIX], are taken a step further, to systems [*with*]{} particle conversion.
A specific case of interest for systems of bosonic atoms and molecules with conversion is the case of $M=1$ atomic and $M'=1$ molecular orbitals, which is presented in section \[two\_mode\_AM\] before the general MCTDH-\[$2a{\leftrightharpoons}m$\] theory is developed. For $M=1$ and $M'=1$, the corresponding multiconfigurational theory is the fully-variational theory that results when the shape of the atomic $\phi_a(\r,t)$ and molecular $\psi_m(\r,t)$ orbitals and of each and every expansion coefficient $C_p(t)$ are optimized according to the variational principle. Being fully-variational with respect to the shape of the orbitals $\phi_a(\r,t)$, $\psi_m(\r,t)$ and with respect to the expansion coefficients $\{C_p(t)\}$, the theory generalizes the literature Gross-Pitaevskii equation [@Timmermans_review] and two-mode approximation [@2M1; @Vardi] for bosonic atoms and molecules with conversion. We term this specific case of the general theory [*conversion mean field*]{}, as there is only one orbital available for the bosonic atoms and one for the molecules – the minimal number possible for bosonic species.
At the other end, in the limit where the number $M$ of atomic orbitals and $M'$ of molecular orbitals goes to infinity, the MCTDH-\[$2a{\leftrightharpoons}m$\] theory becomes an exact representation of the time-dependent many-particle Schrödinger equation with the particle-conversion Hamiltonian $\hat H^{(2a\rightleftharpoons m)}$. In practice, one obviously has to limit $M$ and $M'$. Here, the employment of [*time-dependent*]{} orbitals, which has been very successful for the MCTDH approach and its versions specified for identical particles and mixtures, is of great help and advantage. Of course, even with time-dependent orbitals the size of the Hilbert space grows rapidly with the size of the system and the number of orbitals $M$ and $M'$ employed. Consequently, with increasing system size and as the number of orbitals which one has to employ becomes larger, e.g., for stronger interactions, it is instructive to devise truncation schemes beyond the usage of [*time-dependent*]{} multiconfigurational expansions over [*complete*]{} Hilbert subspaces. We mention two such truncation strategies: (i) to truncate time-dependent multiconfigurational expansions to include parts of Hilbert subspaces, i.e., to include not all available configurations for a given system size and number of orbitals $M$,$M'$; and (ii) to concentrate on the reduced density matrices, write equations of motion for them directly, and thereafter truncate the resulting hierarchy of equations of motion for higher-order reduced density matrices at some given order. The development of these truncation schemes for [*time-dependent*]{} multiconfigurational expansions in systems with particle conversion extends beyond the scope of the present work.
Finally, the explicit equations of motion presented in this work are for the specific ‘reaction’ $2a \leftrightharpoons m$ where the atoms and molecules are structureless bosons – the MCTDH-\[$2a{\leftrightharpoons}m$\] theory. Several other systems come to mind: (i) Other ‘reactions’ with bosonic atoms of the same kind, e.g., $3a \leftrightharpoons m$; (ii) ‘Reactions’ with bosonic atoms of a different kind, e.g., $a+a' \leftrightharpoons m$. In this case and for general particle-particle interactions, the center-of-mass coordinate is, of course, $\R=\frac{m_a\r+m_{a'}\r'}{m_a+m_{a'}}$, where $m_a$ and $m_{a'}$ are the masses of the respective species; (iii) ‘Reactions’ including fermionic atoms, e.g., $a_f+a_b \leftrightharpoons m_f$ and $a_f+a_f \leftrightharpoons m_b$ where the subscript $b$, $f$ stands for bosonic, fermionic species. In the latter case, a unified form of the respective equations of motion and those of the present work is anticipated; and (iv) A whole zoo of ‘reactions’ for particles with spin, internal-structure. The extension of MCTDH-\[$2a{\leftrightharpoons}m$\] for the above concrete examples as well as for other systems with particle conversion can be done by following the theory and derivation steps of the present work.
Financial support by the Deutsche Forschungsgemeinschaft is gratefully acknowledged.
Further details of the derivation of the equations of motion {#Lagrange}
============================================================
The specific case of the conversion mean field (fully-variational two-mode approximation) {#Lagrange_two_mode}
-----------------------------------------------------------------------------------------
When the variations of the functional action (\[action\_functional\_AM\]) with respect to the orbitals and expansion coefficients are put to zero, the following equations of motion are obtained: \[general\_EOM\_orbitals\] & & |\_a> +\
& & + \_[con]{} \_a\^|\_m> = \_a(t) |\_a>,\
& & |\_m> +\
& & + \_a |\_a> = \_m(t) |\_m> and \[general\_EOM\_coeffieints\] (t) = i . The three time-dependent Lagrange multipliers $\mu_a(t)$, $\mu_m(t)$ and $\varepsilon(t)$ appear therein. How to eliminate them?
It is straightforward to eliminate $\varepsilon(t)$. This is done by transforming the expansion coefficients as follows: \[C\_coeff\_global\_phase\] (t) = e\^[-i(t’) dt’]{} (t). Substituting Eq. (\[C\_coeff\_global\_phase\]) into (\[general\_EOM\_coeffieints\]) and removing at the end the ’bar’ from all quantities we obtain: \[general\_simplest\_EOM\_coeffieints\] \^[(a m)]{}(t) (t) = i . From Eqs. (\[general\_EOM\_coeffieints\]-\[general\_simplest\_EOM\_coeffieints\]) we see that the role of the Lagrange multiplier $\varepsilon(t)$ is that of a (redundant) global time-dependent phase of the many-particle wavefunction $\overline{\Psi}(t) = e^{-i\int \varepsilon(t') dt'}\Psi(t)$. We note that Eq. (\[general\_EOM\_orbitals\]) is not affected by the transformation (\[C\_coeff\_global\_phase\]) because reduced density matrices are “insensitive” to a global phase of the wavefunction, namely $\left<\overline{\Psi}(t)\left| \, \ldots \, \right|\overline{\Psi}(t)\right> =
\left<\Psi(t)\left| \, \ldots \, \right|\Psi(t)\right>$.
The next step is to eliminate the remaining Lagrange multipliers $\mu_a(t)$ and $\mu_m(t)$. Making use of the orbitals being normalized and taking the respective scalar products of Eq. (\[general\_EOM\_orbitals\]) with $\left<\phi_a\right|$ and $\left<\psi_m\right|$, we obtain: \[Lagrange\_2orbitals\] \_a(t) &=& + +\
&+& \_[am]{} + \_[con]{} ,\
\_m(t) &=& + +\
&+& \_[am]{} + . Substituting Eq. (\[Lagrange\_2orbitals\]) into Eq. (\[general\_EOM\_orbitals\]), employing the identities: \[mu\_identities\] & & |\_a> + \_[con]{} \_a\^|\_m> - \_a(t) |\_a> =\
& & = (1 - |\_a> (h\^[(a)]{} - i) + \_a |\_a|\^2 +\
& & + \_[am]{} |\_m|\^2 \] |\_a> + \_[con]{} \_a\^|\_m>},\
& & |\_m> + \_a |\_a> - \_m(t) |\_m> =\
& & = (1-|\_m> (h\^[(m)]{} - i) + \_m |\_m|\^2 +\
& & + \_[am]{} |\_a|\^2 \] |\_m> + \_a |\_a>}, and dividing the result, respectively, by $\left<\hat N_a\right>$ and $\left<\hat N_m\right>$, the equations of motion (\[simplest\_EOM\_orbitals\]) are obtained.
To eliminate the projection operators in front of the time derivatives in Eq. (\[mu\_identities\]), i.e., $\hat{\mathbf P}^{(a)} = 1 - \left|\phi_a\left>\right<\phi_a\right|$ and $\hat{\mathbf P}^{(m)} = 1 - \left|\psi_m\left>\right<\psi_m\right|$ on the left-hand sides of Eq. (\[simplest\_EOM\_orbitals\]), we exploit the invariance property of the wavefunction $\Psi(t)$. Consider the following phase transformations of the orbitals and coefficients: \[trans\_appen\_orb\_coeff\] & & \_a(,t) = e\^[+i\_a(t)]{} \_a(,t), \_m(,t) = e\^[+i\_m(t)]{} \_m(,t),\
& & \_p(t) = e\^[-i]{} C\_p(t), p=0,…,. Combing these phase transformations, the wavefunction does not change: $\left|\Psi(t)\right> = \sum_{p=0}^{\left[N/2\right]} C_p(t) \left|N-2p,p;t\right> =
\sum_{p=0}^{\left[N/2\right]} \overline{C}_p(t) \overline{\left|N-2p,p;t\right>}$. We should also recall that transforming the orbitals goes along with transforming the corresponding annihilation, creation operators (another way to look at this is that the field operators $\hat{\mathbf \Psi}_a(\r)$ and $\hat{\mathbf \Psi}_m(\r)$ are time-independent, basis-set-independent quantities and, consequently, transforming the orbitals requires the reverse transformation of the annihilation operators). Thus, Eq. (\[trans\_appen\_orb\_coeff\]) implies also the phase transformations \[trans\_appen\_anni\] \_a(t) = e\^[-i\_a(t)]{} b\_a(t), \_m(t) = e\^[-i\_m(t)]{} c\_m(t) for the atomic and molecular annihilation operators.
Now, plugging Eqs. (\[trans\_appen\_orb\_coeff\]-\[trans\_appen\_anni\]) into the equations of motion (\[simplest\_EOM\_orbitals\]) for the orbitals and (\[simplest\_EOM\_coeffieints\]) for the expansion coefficients \[see also Eqs. (\[general\_simplest\_EOM\_coeffieints\],\[mu\_identities\])\], and choosing the phases \[phase\_choice\_1\] \_a(t) = i dt’, \_m(t) = i dt’, equations of motion (\[simplest\_EOM\_orbitals\_final\]) and (\[simplest\_EOM\_coeffieints\_final\]) are found. We note that the phases $\beta_a(t)$ and $\gamma_m(t)$ are real quantities since $\left<\phi_a(t)\left|\right.\phi_a(t)\right>=1$ and $\left<\psi_m(t)\left|\right.\psi_m(t)\right>=1$, respectively, for all times.
The general multiconfigurational theory {#appen_C}
---------------------------------------
When the expressions for the field operators $\hat{\mathbf \Psi}_a(\r)$ and $\hat{\mathbf \Psi}_m(\r)$ in terms of the time-dependent orbitals, Eq. (\[annihilation\_def\]), are substituted into the generic many-body Hamiltonian (\[ham\_am\_1\]-\[ham\_am\_3\]), one obtains: \[ham\_general\_inter\] & & H\^[(2am)]{} = \_[k,q]{} h\^[(a)]{}\_[kq]{} b\_k\^b\_q + \_[k,s,l,q]{} W\^[(a)]{}\_[ksql]{} b\_k\^b\_s\^b\_l b\_q + \_[k’,q’]{} h\^[(m)]{}\_[k’q’]{} c\_[k’]{}\^c\_[q’]{} + \_[k’,s’,l’,q’]{} W\^[(m)]{}\_[k’s’q’l’]{} c\_[k’]{}\^c\_[s’]{}\^c\_[l’]{} c\_[q’]{} +\
& & + \_[k,k’,q,q’]{} W\^[(am)]{}\_[kk’qq’]{} b\_k\^b\_q c\_[k’]{}\^c\_[q’]{} + \_[k’,k,q]{} . The one-body, two-body and conversion matrix elements appearing in Eq. (\[ham\_general\_inter\]) are given by: \[one\_two\_matrix\_elements\] h\^[(a)]{}\_[kq]{} &=& \_k\^(,t) h\^[(a)]{}() \_q(,t) d,\
W\^[(a)]{}\_[ksql]{} &=& \_k\^(,t) \_s\^(’,t) W\^[(a)]{}(,’) \_q(,t) \_l(’,t) dd’,\
h\^[(m)]{}\_[k’q’]{} &=& \_[k’]{}\^(,t) h\^[(m)]{}() \_[q’]{}(,t) d,\
W\^[(m)]{}\_[k’s’q’l’]{} &=& \_[k’]{}\^(,t) \_[s’]{}\^(’,t) W\^[(m)]{}(,’) \_[q’]{}(,t) \_[l’]{}(’,t) dd’,\
W\^[(am)]{}\_[kk’qq’]{} &=& \_k\^(,t) \_[k’]{}\^(’,t) W\^[(am)]{}(,’) \_q(,t) \_[q’]{}(’,t) dd’,\
W\^[(2am)]{}\_[k’kq]{} &=& \^\_[k’]{}(,t) W\^[(2am)]{}(,’) \_k(,t) \_q(’,t) dd’ =\
&=& \^\_[k’]{}(,t) W\^[(2am)]{}(+,-) \_k(+,t) \_q(-,t) dd’,\
W\^[(m2a)]{}\_[qkk’]{} &=& \^\_q(,t) \^\_k(’,t) W\^[(m2a)]{}(,’) \_[k’]{}(,t) dd’ = {W\^[(2am)]{}\_[k’kq]{}}\^. The change of variables used for $W^{(2a\rightharpoonup m)}_{k'kq}$ is needed in order to perform the variation of this term with respect to the molecular orbitals, see below.
Now, the expectation value appearing in the functional action (\[action\_functional\_AM\_full\]) when expressed explicit with respect to the orbitals reads: \[exp\_full\_app\] & & = \_[k,q=1]{}\^M \^[(a)]{}\_[kq]{} + \_[k,s,l,q=1]{}\^M \^[(a)]{}\_[kslq]{} W\^[(a)]{}\_[ksql]{} +\
& & + \_[k’,q’=1]{}\^[M’]{} \^[(m)]{}\_[k’q’]{} + \_[k’,s’,l’,q’=1]{}\^[M’]{} \^[(m)]{}\_[k’s’l’q’]{} W\^[(m)]{}\_[k’s’q’l’]{} + \_[k,q=1]{}\^M \_[k’,q’=1]{}\^[M’]{} \^[(am)]{}\_[kk’qq’]{} W\^[(am)]{}\_[kk’qq’]{} +\
& & + \_[k,q=1]{}\^[M]{} \_[k’=1]{}\^[M’]{} - i \_[p=0]{}\^ \_[\^p,\^p]{} C\_[\^p\^p]{}\^, where (i)\_[kq]{}\^[(a)]{} = i\_k\^(,t) d, (i)\_[k’q’]{}\^[(m)]{} = i\_[k’]{}\^(,t) d. Equating the variation of the functional action (\[action\_functional\_AM\_full\]) with respect to the orbitals to zero, making use of Eq. (\[exp\_full\_app\]), the following equations are obtained: \[general\_variation\_orbitals\_direct\] & & \_[q=1]{}\^M { |\_q> +\
& & + \_[k’=1]{}\^[M’]{} \^[(m2a)]{}\_[qkk’]{} W\^[(m2a)]{}\_[qk’]{} } = \_[j=1]{}\^M \^[(a)]{}\_[kj]{} |\_j>, k=1,…,M,\
& & \_[q’=1]{}\^[M’]{} |\_[q’]{}> +\
& & + \_[k,q=1]{}\^[M]{} \^[(2am)]{}\_[k’kq]{} W\^[(2am)]{}\_[kq]{} = \_[j’=1]{}\^[M’]{} \^[(m)]{}\_[k’j’]{} |\_[j’]{}>, k’=1,…,M’. One delicate point in performing the variation of $W^{(2a\rightharpoonup m)}_{k'kq}$ \[the first term in the last line of the expectation value Eq. (\[exp\_full\_app\])\] with respect to the molecular orbitals $\psi^\ast_{k'}(\r,t)$ is worth mentioning. To perform this variation, a change of the integration variables $\r$, $\r'$ to the center-of-mass $\R=\frac{\r+\r'}{2}$ and relative $\bar\r=\r-\r'$ coordinates is required. Assigning thereafter back $\R \to \r$, $\bar\r \to \r'$, the matrix element $W^{(2a\rightharpoonup m)}_{k'kq}$ is re-written in a form, see Eq. (\[one\_two\_matrix\_elements\]), which is amenable to explicit variation with respect to $\psi^\ast_{k'}(\r,t)$.
The next step it to eliminate the Lagrange multipliers $\{\mu_{kj}^{(a)}(t)\}$ and $\{\mu_{k'j'}^{(m)}(t)\}$. Making use of the orthonormality properties of the atomic and molecular orbitals, $\left<\phi_k(t)\left|\right.\phi_q(t)\right>=\delta_{kq}$ and $\left<\psi_{k'}(t)\left|\right.\psi_{q'}(t)\right>=\delta_{k'q'}$, and taking the corresponding scalar products of Eq. (\[general\_variation\_orbitals\_direct\]) with respect to the orbitals, we obtain explicit expressions for the Lagrange multipliers: \[MCTDH\_conver\_mu\_gener\] & & \_[kj]{}\^[(a)]{}(t) = \^M\_[q=1]{}{\_[kq]{}\^[(a)]{} + \^M\_[s,l=1]{} \_[kslq]{}\^[(a)]{} W\^[(a)]{}\_[jsql]{} +\
& & + \_[k’,q’=1]{}\^[M’]{} \^[(am)]{}\_[kk’qq’]{} W\^[(am)]{}\_[jk’qq’]{} + \_[k’=1]{}\^[M’]{} \^[(m2a)]{}\_[qkk’]{} W\^[(m2a)]{}\_[qjk’]{} },\
& & \_[k’j’]{}\^[(m)]{}(t) = \^[M’]{}\_[q’=1]{}{\_[k’q’]{}\^[(m)]{} + \^[M’]{}\_[s’,l’=1]{} \_[k’s’l’q’]{}\^[(m)]{} W\^[(m)]{}\_[j’s’q’l’]{} +\
& & + \_[k,q=1]{}\^[M]{} \^[(am)]{}\_[kk’qq’]{} W\^[(am)]{}\_[kj’qq’]{} } + \_[k,q=1]{}\^[M]{} \^[(2am)]{}\_[k’kq]{} W\^[(2am)]{}\_[j’kq]{}. Substituting Eq. (\[MCTDH\_conver\_mu\_gener\]) into Eq. (\[general\_variation\_orbitals\_direct\]), making use of the identities: \[general\_identities\] & & \_[q=1]{}\^M { |\_q> +\
& & + \_[k’=1]{}\^[M’]{} \^[(m2a)]{}\_[qkk’]{} W\^[(m2a)]{}\_[qk’]{} } - \_[u=1]{}\^M \^[(a)]{}\_[ku]{} |\_u> =\
& & (1-\_[u=1]{}\^M |\_u> + \_[k’=1]{}\^[M’]{} \^[(m2a)]{}\_[qkk’]{} W\^[(m2a)]{}\_[qk’]{} }, k=1,…,M,\
& & \_[q’=1]{}\^[M’]{} |\_[q’]{}> +\
& & + \_[k,q=1]{}\^[M]{} \^[(2am)]{}\_[k’kq]{} W\^[(2am)]{}\_[kq]{} - \_[u’=1]{}\^[M’]{} \^[(m)]{}\_[k’u’]{} |\_[u’]{}> =\
& & (1-\_[u’=1]{}\^[M’]{} |\_[u’]{}> + \_[k,q=1]{}\^[M]{} \^[(2am)]{}\_[k’kq]{} W\^[(2am)]{}\_[kq]{} }, k’=1,…,M’, and multiplying the result, respectively, by the inverse of the reduced one-body density matrices and summing over $\sum_{k=1}^M \left\{\brho^{(a)}(t)\right\}_{jk}^{-1}$ and $\sum_{k'=1}^{M'} \left\{\brho^{(m)}(t)\right\}_{j'k'}^{-1}$, we obtain equations of motion like (\[MCTDH\_conver\_orb\_eqs\_P\]) with general interactions.
Finally, to eliminate the projection operators $\hat{\mathbf P}^{(a)} = 1 - \sum_{u=1}^M \left|\phi_u\left>\right<\phi_u\right|$ and $\hat{\mathbf P}^{(m)} = 1 - \sum_{u'=1}^{M'} \left|\psi_{u'}\left>\right<\psi_{u'}\right|$ in front of the time derivatives \[see Eqs. (\[general\_identities\]) and (\[MCTDH\_conver\_orb\_eqs\_P\])\], we employ the invariance properties of the wavefunction $\Psi(t)=\overline{\Psi}(t)$, where $\left\{\phi_k(\r,t)\right\} \to \left\{\overline{\phi}_k(\r,t)\right\}$, $\left\{\psi_{k'}(\r,t)\right\} \to \left\{\overline{\psi}_{k'}(\r,t)\right\}$, and $\{C_{\vec{n}^p\vec{m}^p}(t)\} \to \{\overline{C}_{\vec{n}^p\vec{m}^p}(t)\}$. For this, consider the following time-dependent matrices: \[matrix\_U\] \^[(a)]{}(t), D\^[(a)]{}\_[kq]{}=i, \^[(m)]{}(t), D\^[(m)]{}\_[k’q’]{}=i. The matrices $\D^{(a)}(t)$ and $\D^{(m)}(t)$ are Hermitian matrices \[because the respective orbitals are normalized and orthogonal to one another, $\left<\phi_k(t)\left|\right.\phi_q(t)\right>=\delta_{kq}$ and $\left<\psi_{k'}(t)\left|\right.\psi_{q'}(t)\right>=\delta_{k'q'}$\] and hence can be diagonalized \[diagonal\_U\] {\^[(a)]{}(t)}\^\^[(a)]{}(t) \^[(a)]{}(t) = \^[(a)]{}(t), {\^[(m)]{}(t)}\^\^[(m)]{}(t) \^[(m)]{}(t) = \^[(m)]{}(t), where $\d^{(a)}(t)$ and $\d^{(m)}(t)$ are the diagonal matrices of the respective eigenvalues. Now, we define the unitary transformations (which are [*symbolically*]{} integrated): \[equation\_U\] & & iU\^[(a)]{}\_[sq]{}(t)= - \_[k=1]{}\^M D\^[(a)]{}\_[sk]{}(t) U\^[(a)]{}\_[kq]{}(t) \^[(a)]{}(t)=e\^[+i\^t \^[(a)]{}(t’) dt’]{} \^[(a)]{}(0),\
& & iU\^[(m)]{}\_[s’q’]{}(t)= - \_[k’=1]{}\^M D\^[(m)]{}\_[s’k’]{}(t) U\^[(m)]{}\_[k’q’]{}(t) \^[(m)]{}(t)=e\^[+i\^t \^[(m)]{}(t’) dt’]{} \^[(m)]{}(0), with the initial conditions defined in the limes $\tau \to 0$ as (see in this respect Ref. [@MCTDHB2]): \[U0\_initial\_conditions\] \^[(a)]{}() = \^[(a)]{}(0) e\^[+i \^[(a)]{}(0)]{}, \^[(m)]{}() = \^[(m)]{}(0) e\^[+i \^[(m)]{}(0)]{}. Then, the unitary transformations of the orbitals \[orbitals\_U\] & & \_q(,t) = \_[k=1]{}\^M U\^[(a)]{}\_[kq]{}(t) \_k(,t), q=1,…,M,\
& & \_[q’]{}(,t) = \_[k’=1]{}\^[M’]{} U\^[(m)]{}\_[k’q’]{}(t) \_[k’]{}(,t), q’=1,…,M’ lead to the desired result – equations of motion (\[MCTDH\_conver\_orb\_eqs\]) and (\[MCTDH\_W\_conver\_orb\_eqs\]) – where the projection operators $\hat{\mathbf P}^{(a)}$ and $\hat{\mathbf P}^{(m)}$ have been eliminated from the left-hand sides.
The transformation $\{C_{\vec{n}^p\vec{m}^p}(t)\} \to \{\overline{C}_{\vec{n}^p\vec{m}^p}(t)\}$ accompanying Eq. (\[orbitals\_U\]), carries equations of motion (\[MCTDH\_conver\_coeff\_eqs\_P\]) for the expansion coefficients to the respective final result, Eqs. (\[MCTDH\_conver\_coeff\_eqs\]) and (\[MCTDH\_W\_conver\_coeff\_eqs\]). It is instructive to obtain this result by proving that the equations of motion for the expansion coefficients are [*form-invariant*]{}. Namely, if ${\bcalH}^{(2a \leftrightharpoons m)}(t)\C(t) = i\frac{\partial \C(t)}{\partial t}$ are satisfied for the untransformed quantities $\left[\{C_{\vec{n}^p\vec{m}^p}(t)\}, \left\{\phi_k(\r,t)\right\},
\left\{\psi_{k'}(\r,t)\right\}\right]$ then $\overline{{\bcalH}}^{(2a \leftrightharpoons m)}(t)\overline{\C}(t) = i\frac{\partial \overline{\C}(t)}{\partial t}$ are satisfied for the transformed ones $\left[\{\overline{C}_{\vec{n}^p\vec{m}^p}(t)\}, \left\{\overline{\phi}_k(\r,t)\right\},
\left\{\overline{\psi}_{k'}(\r,t)\right\}\right]$. The proof is straightforward. Equating the variation of the functional action (\[action\_functional\_AM\_full\],\[expectation\_ham\_b\]) with respect to the expansion coefficients to zero, the result can be written as follows: $\left<\vec{n}^p,\vec{m}^p;t\left|\hat H^{(2a\leftrightharpoons m)} -
i\frac{\partial}{\partial t}\right|\Psi(t)\right>, \forall p,\vec{n}^p,\vec{m}^p$. Since, the transformed configurations $\left\{\overline{\left<\vec{n}^p,\vec{m}^p;t\right|}\right\}$ are given as linear combinations of the untransformed configurations $\left\{\left<\vec{n}^p,\vec{m}^p;t\right|\right\}$, the operator $\hat H^{(2a\leftrightharpoons m)} - i\frac{\partial}{\partial t}$ does not depend on the orbitals, and $\left|\overline{\Psi}(t)\right>=\left|\Psi(t)\right>$, we immediately get: $\overline{\left<\vec{n}^p,\vec{m}^p;t\right|}\hat H^{(2a\leftrightharpoons m)} -
i\frac{\partial}{\partial t}\left|\overline{\Psi}(t)\right>, \forall p,\vec{n}^p,\vec{m}^p$, which concludes our proof. To our needs, since the transformed orbitals (\[orbitals\_U\]) obey the differential conditions $\left<\overline{\phi}_k\left|\right.\dot{\overline{\phi}}_q\right> = 0; k,q=1,\ldots,M$ and $\left<\overline{\psi}_{k'}\left|\right.\dot{\overline{\psi}}_{q'}\right> = 0; k',q'=1,\ldots,M'$ \[see Eq. (\[MCTDHI\_mix\_conv\_const\])\], the respective equations of motion for the transformed coefficients boil down to $\overline{{\H}}^{(2a \leftrightharpoons m)}(t)\overline{\C}(t) = i\frac{\partial \overline{\C}(t)}{\partial t}$ \[see Eqs. (\[MCTDH\_conver\_coeff\_eqs\]) and (\[MCTDH\_W\_conver\_coeff\_eqs\])\].
Matrix elements with multiconfigurational wavefunctions in systems with particle conversion {#matrix_Appen}
===========================================================================================
There are two types of matrix elements in the theory. The first type are matrix elements of the many-body Hamiltonian with respect to the configurations. These matrix elements are expressed using the matrix elements of the one-body terms, two-body interaction terms, and the conversion term with respect to the atomic and molecular orbitals. The second type of matrix elements are the matrix elements of the reduced density matrices appearing in the theory, which are expressed in terms of the expansion coefficients.
In this appendix we prescribe these matrix elements. It is easy to connect the matrix elements of particle-conserving operators to the corresponding matrix elements appearing in the available multiconfigurational theories for identical particles and mixtures. This assignment will shorten substantially the discussion below. The matrix elements of particle non-conserving operators are new and will be presented in full details.
Matrix elements of the Hamiltonian
----------------------------------
The many-body Hamiltonian (\[ham\_am\_1\]-\[ham\_am\_3\]) is written as a sum of particle-conserving and particle non-conserving parts: $\hat H^{(2a\rightleftharpoons m)} = \hat H^{(am)} + \hat W^{(2a\rightleftharpoons m)}$. The matrix elements of the particle-conserving part $\hat H^{(am)}$, see Eqs. (\[ham\_am\_1\],\[ham\_am\_2\]), between two general configurations derive from the following relation: \[ham\_particle\_consere\_ME\] = \_[p,p’]{} . Thus, it corresponds to a matrix element of a mixture with $N-2p$ atoms and $p$ molecules [*without conversion*]{}. The matrix elements of the Hamiltonian of a mixture of two kinds of bosons with respect to two general configurations have been prescribed within the respective particle-conserving multiconfigurational theory for Bose-Bose mixtures, the MCTDH-BB theory, see Ref. [@MIX].
To evaluate the matrix elements of the particle non-conserving part of the Hamiltonian, $\hat W^{(2a\rightleftharpoons m)}=\hat W^{(2a\rightharpoonup m)} + \hat W^{(m\rightharpoondown 2a)}$, see Eqs. (\[ham\_am\_1\],\[ham\_am\_3\]), between two general configurations we have to introduce a shorthand notation for different configurations. The reference configuration is denoted by $\left|\vec{n}^p,\vec{m}^p;t\right> =
\left|n_1^p,\ldots,n_k^p,\ldots,n_q^p,\ldots,n_M^p:m_1^p,\ldots,m_{k'}^p,\ldots,m_{M'}^p;t\right>$. We remind that the occupation numbers satisfy the relations: $|\vec{n}^p|=n_1^p+\ldots+n_M^p=N-2p$ and $|\vec{m}^p|=m_1^p+\ldots+m_{M'}^p=p$, where $p$ is the number of molecules. Now, the configuration $\left|\vec{n}^{p-1}_{kq},\vec{m}^{p-1}_{k'};t\right> \equiv
\left|n_1^p,\ldots,n_k^p+1,\ldots,n_q^p+1,\ldots,n_M^p:m_1^p,\ldots,m_{k'}^p-1,\ldots,m_{M'}^p;t\right>$ differs from $\left|\vec{n}^p,\vec{m}^p;t\right>$ by having $p-1$ molecules and $N-2p+2$ atoms, where a molecule in the $k'$-th orbital has dissociated to two atoms, one in the $k$-th orbital and the second in the $q$-th orbital; and the configuration $\left|\vec{n}^{p-1}_{kk},\vec{m}^{p-1}_{k'};t\right> \equiv
\left|n_1^p,\ldots,n_k^p+2,\ldots,n_M^p:m_1^p,\ldots,m_{k'}^p-1,\ldots,m_{M'}^p;t\right>$ differs from $\left|\vec{n}^p,\vec{m}^p;t\right>$ by having $p-1$ molecules and $N-2p+2$ atoms, where a molecule in the $k'$-th orbital has dissociated to two atoms, both in the $k$-th orbital. We employ a nomenclature where the same ordering of the orbitals $\phi_1,\ldots,\phi_M$ and $\psi_1,\ldots,\psi_{M'}$ as in Eq. (\[MCTDH\_AM\_Psi\]) is kept for all configurations. In this nomenclature the following states are equivalent: $\left|\vec{n}^{p-1}_{kq},\vec{m}^{p-1}_{k'};t\right> \equiv
\left|\vec{n}^{p-1}_{qk},\vec{m}^{p-1}_{k'};t\right>$.
With this notation, the non-vanishing matrix elements of the particle non-conserving part of the Hamiltonian follow from \[ham\_particle\_NON\_consere\_ME\] & & = W\^[(2am)]{}\_[k’kq]{} , k<q,\
& & = W\^[(2am)]{}\_[k’kk]{} , and the relation $\left<\vec{n}^p,\vec{m}^p;t\left| \hat W^{(m\rightharpoondown 2a)}\right|\vec{n}'^{p'},\vec{m}'^{p'};t\right>=
\left\{\left<\vec{n}'^{p'},\vec{m}'^{p'};t\left| \hat W^{(2a\rightharpoonup m)}\right|\vec{n}^p,\vec{m}^p;t\right>\right\}^\ast$. We note that $W^{(2a\rightharpoonup m)}_{k'kq} = W^{(2a\rightharpoonup m)}_{k'qk}$ because $\hat W^{(2a\rightharpoonup m)}(\r,\r') = \hat W^{(2a\rightharpoonup m)}(\r',\r)$. To summarize, direct coupling in the matrix representation of the Hamiltonian with respect to the configurations exists due to the particle non-conserving part of the Hamiltonian $\hat W^{(2a\rightleftharpoons m)}$ between configurations with $p$ and $p-1$ molecules only, for $p=1,\ldots,\left[N/2\right]$.
Matrix elements of reduced density matrices
-------------------------------------------
The multiconfigurational ansatz (\[MCTDH\_AM\_Psi\]) can be written in the following form, \[multi\_psi\_appen\] |(t)> = \_[p=0]{}\^ |\_p(t)>, where each $\left|\Psi_p(t)\right> = \sum_{\vec{n}^p,\vec{m}^p}
C_{\vec{n}^p\vec{m}^p}(t) \left|\vec{n}^p,\vec{m}^p;t\right>$ is a (non-normalized) many-particle wavefunction with a [*definite*]{} number of $p$ molecules and $N-2p$ atoms, and thus “describes” a bosonic mixture [*without conversion*]{}. Consequently, the matrix elements of the particle-conserving reduced density matrices can be expressed as follows: \[matrix\_elements\_rho\_conserved\] & & \_[kq]{}\^[(a)]{}(t) = = \_[p=0]{}\^ ,\
& & \_[k’q’]{}\^[(m)]{}(t) = = \_[p=0]{}\^ ,\
& & \_[kslq]{}\^[(a)]{}(t) = = \_[p=0]{}\^ ,\
& & \_[k’s’l’q’]{}\^[(m)]{}(t) = = \_[p=0]{}\^ ,\
& & \_[kk’qq’]{}\^[(am)]{}(t) = = \_[p=0]{}\^ . In other words, the matrix elements of the particle-conserving reduced density matrices can be readily read from the reduced density matrices of the respective particle-conserving multiconfigurational theory for Bose-Bose mixtures, the MCTDH-BB theory, see Ref. [@MIX].
The matrix elements of the particle non-conserving reduced density matrices are given explicitly by \[matrix\_ele\_particle\_non\_den\_gen\] & & \^[(2am)]{}\_[k’kq]{} = = \_[p=0]{}\^[\[N/2\]]{} \_[\^p,\^p]{} C\^\_[\^p,\^p]{} C\_[\^[p-1]{}\_[kq]{},\^[p-1]{}\_[k’]{}]{} , k<q,\
& & \^[(2am)]{}\_[k’kk]{} = \_[p=0]{}\^[\[N/2\]]{} \_[\^p,\^p]{} C\^\_[\^p,\^p]{} C\_[\^[p-1]{}\_[kk]{},\^[p-1]{}\_[k’]{}]{} . All other matrix elements are derived from the symmetry of the conversion operator, $\rho^{(2a\rightharpoonup m)}_{k'qk}=\rho^{(2a\rightharpoonup m)}_{k'kq}$, and the Hermiticity relation $\rho^{(m\rightharpoondown 2a)}_{qkk'}(t) =
\left\{\rho^{(2a\rightharpoonup m)}_{k'kq}(t)\right\}^\ast$.
[99]{} , edited by K. C. Kulander (North-Holland, Amsterdam, 1991).
J. E. Bayfield, [*Quantum Evolution: An Introduction to Time-Dependent Quantum Mechanics*]{} (Wiley, New York, 1999).
P. Ring and P. Schuck, [*The Nuclear Many-Body Problem*]{} (Springer, Berlin, 2000).
, edited by J. Ullrich and V. P. Shevelko (Springer, Berlin, 2003).
L. Pitaevskii and S. Stringari, [*Bose-Einstein Condensation*]{} (Oxford University Press, Oxford, 2003).
, edited by D. A. Micha and I. Burghardt, Springer Series in Chemical Physics, Vol. 83 (Springer, Berlin, 2007).
H.-D. Meyer, U. Manthe, and L. S. Cederbaum, Chem. Phys. Lett. [**165**]{}, 73 (1990).
U. Manthe, H.-D. Meyer, and L. S. Cederbaum, J. Chem. Phys. [**97**]{}, 3199 (1992).
G. A. Worth, M. H. Beck, A. Jäckle, and H.-D. Meyer, The MCTDH Package, Version 8.2, (2000); H.-D. Meyer, Version 8.3 (2002); Version 8.4 (2007). See http://mctdh.uni-hd.de/ .
G. A. Worth, H.-D. Meyer, and L. S. Cederbaum, J. Chem. Phys. [**109**]{}, 3518 (1998).
A. Raab, G. Worth, H.-D. Meyer, and L. S. Cederbaum, J. Chem. Phys. [**110**]{}, 936 (1999).
M. H. Beck, A. Jäckle, G. A. Worth, and H.-D. Meyer, Phys. Rep. [**324**]{}, 1 (2000).
H.-D. Meyer and G. A. Worth, Theor. Chem. Acc. [**109**]{} 251 (2003).
R. van Harrevelt and U. Manthe, J. Chem. Phys. [**123**]{}, 064106 (2005).
L. S. Cederbaum, E. Gindensperger, and I. Burghardt, Phys. Rev. Lett. [**94**]{}, 113003 (2005).
H.-D. Meyer, F. Le Quere, C. Leonard, and F. Gatti, Chem. Phys. [**329**]{}, 179 (2006).
O. Vendrell, F. Gatti, D. Lauvergnat, and H.-D. Meyer, J. Chem. Phys. [**127**]{}, 184302 (2007).
O. Vendrell, F. Gatti, and H.-D. Meyer, J. Chem. Phys. [**127**]{}, 184303 (2007).
H. Tamura, J. G. S. Ramon, E. R. Bittner, and I. Burghardt, Phys. Rev. Lett. [**100**]{}, 107402 (2008).
P. A. M. Dirac, Proc. Cambridge Phil. Soc. [**26**]{}, 376 (1930).
J. Frenkel, [*Wave Mechanics*]{} (Oxford University Press, Oxford, 1934).
H. Wang and M. Thoss, J. Chem. Phys. [**119**]{}, 1289 (2003).
S. Zöllner, H.-D. Meyer, and P. Schmelcher, Phys. Rev. A [**74**]{}, 053612 (2006).
S. Zöllner, H.-D. Meyer, and P. Schmelcher, Phys. Rev. A [**74**]{}, 063611 (2006).
C. Matthies, S. Zöllner, H.-D. Meyer, and P. Schmelcher, Phys. Rev. A [**76**]{}, 023602 (2007).
S. Zöllner, H.-D. Meyer, and P. Schmelcher, Phys. Rev. Lett. [**100**]{}, 040401 (2008).
S. Zöllner, H.-D. Meyer, and P. Schmelcher, Phys. Rev. A [**78**]{}, 013621 (2008).
A. U. J. Lode, A. I. Streltsov, O. E. Alon, H.-D. Meyer, and L. S. Cederbaum, J. Phys. B (in press).
J. Zanghellini, M. Kitzler, C. Fabian, T. Brabec, and A. Scrinzi, Laser Phys. [**13**]{}, 1064 (2003).
T. Kato and H. Kono, Chem. Phys. Lett. [**392**]{}, 533 (2004).
M. Nest, T. Klamroth, and P. Saalfrank, J. Chem. Phys. [**122**]{}, 124102 (2005).
A. I. Streltsov, O. E. Alon, and L. S. Cederbaum, Phys. Rev. Lett. [**99**]{}, 030402 (2007).
O. E. Alon, A. I. Streltsov, and L. S. Cederbaum, Phys. Rev. A [**77**]{}, 033613 (2008).
M. Kitzler, J. Zanghellini, Ch. Jungreuthmayer, M. Smits, A. Scrinzi, and T. Brabec, Phys. Rev. A [**70**]{}, 041401(R) (2004).
J. Caillat, J. Zanghellini, M. Kitzler, O. Koch, W. Kreuzer, and A. Scrinzi, Phys. Rev. A [**71**]{}, 012712 (2005).
Z. Zhang, C. F. Destefani, C. McDonald, and T. Brabec, Phys. Rev. B [**72**]{}, 161309(R) (2005).
G. Jordan, J. Caillat, C. Ede, and A. Scrinzi, J. Phys. B [**39**]{}, S341 (2006).
M. Nest, Phys. Rev. A [**73**]{}, 023613 (2006).
M. Nest, R. Padmanaban, and P. Saalfrank, J. Chem. Phys. [**126**]{}, 214106 (2007).
F. Remacle, M. Nest, and R. D. Levine, Phys. Rev. Lett. [**99**]{}, 183902 (2007).
A. I. Streltsov, O. E. Alon, and L. S. Cederbaum, Phys. Rev. Lett [**100**]{}, 130401 (2008).
K. Sakmann, A. I. Streltsov, O. E. Alon, and L. S. Cederbaum, Phys. Rev. A [**78**]{}, 023615 (2008).
J. Grond, J. Schmiedmayer, and U. Hohenester, arXiv:0806.3877v1.
P.-O. Löwdin, Phys. Rev. [**97**]{}, 1474 (1955).
A. J. Coleman and V. I. Yukalov, [*Reduced Density Matrices: Coulson’s Challenge*]{} (Springer-Verlag, New York, 2000).
D. A. Mazziotti, Phys. Rev. Lett. [**93**]{}, 213001 (2004).
D. A. Mazziotti, Phys. Rev. Lett. [**97**]{}, 143002 (2006).
G. Gidofalvi and D. A. Mazziotti, Phys. Rev. A [**74**]{}, 012501 (2006).
, edited by D. A. Mazziotti, Advances in Chemical Physics, Vol. 134 (Wiley, New York, 2007).
D. A. Mazziotti, J. Chem. Phys. [**126**]{}, 184101 (2007).
E. Kamarchik and D. A. Mazziotti, Phys. Rev. A [**75**]{}, 013203 (2007).
A. I. Streltsov, O. E. Alon, and L. S. Cederbaum, Phys. Rev. A [**73**]{}, 063626 (2006).
O. E. Alon, A. I. Streltsov, and L. S. Cederbaum, J. Chem. Phys. [**127**]{}, 154103 (2007).
O. E. Alon, A. I. Streltsov, and L. S. Cederbaum, Phys. Rev. A [**76**]{}, 062501 (2007).
P. D. Drummond, K. V. Kheruntsyan, and H. He, Phys. Rev. Lett. [**81**]{}, 3055 (1998).
J. Javanainen and M. Mackie, Phys. Rev. A [**59**]{}, R3186 (1999).
E. Timmermans, P. Tommasini, M. Hussein, and A. Kerman, Phys. Rep. [**315**]{}, 199 (1999).
V. A. Yurovsky, A. Ben-Reuven, P. S. Julienne, and C. J. Williams, Phys. Rev. A [**60**]{}, R765 (1999).
E. Timmermans, P. Tommasini, R. Côté, M. Hussein, and A. Kerman, Phys. Rev. Lett. [**83**]{}, 2691 (1999).
R. Wynar, R. S. Freeland, D. J. Han, C. Ryu, and D. J. Heinzen, Science [**287**]{}, 1016 (2000).
D. J. Heinzen, R. Wynar, P. D. Drummond, and K. V. Kheruntsyan, Phys. Rev. Lett. [**84**]{}, 5029 (2000).
K. Góral, M. Gajda, and K. Rzażewski, Phys. Rev. Lett. [**86**]{}, 1397 (2001).
M. Holland, J. Park, and R. Walser, Phys. Rev. Lett. [**86**]{}, 1915 (2001).
S. K. Adhikari, J. Phys. B. [**34**]{}, 4231 (2001).
A. Vardi, V. A. Yurovsky, and J. R. Anglin, Phys. Rev. A [**64**]{}, 063611 (2001).
C. McKenzie, J. Hecker Denschlag, H. Häffner, A. Browaeys, L. E. E. de Araujo, F. K. Fatemi, K. M. Jones, J. E. Simsarian, D. Cho, A. Simoni, E. Tiesinga, P. S. Julienne, K. Helmerson, P. D. Lett, S. L. Rolston, and W. D. Phillips, Phys. Rev. Lett. [**88**]{}, 120403 (2002).
E. A. Donley, N. R. Claussen, S. T. Thompson, and D. E. Weiman, Nature (London) [**417**]{}, 529 (2002).
D. Jaksch, V. Venturi, J. I. Cirac, C. J. Williams, and P. Zoller, Phys. Rev. Lett. [**89**]{}, 040402 (2002).
R. A. Duine and H. T. C. Stoof, J. Opt. B [**5**]{}, S212 (2003).
T. Esslinger and K. M[ø]{}lmer, Phys. Rev. Lett. [**90**]{}, 160406 (2003).
J. Herbig, T. Kraemer, M. Mark, T. Weber, C. Chin, H.-C. Nägerl, and R. Grimm, Science [**301**]{}, 1510 (2003).
K. Xu, T. Mukaiyama, J. R. Abo-Shaeer, J. K. Chin, D. E. Miller, and W. Ketterle, Phys. Rev. Lett. [**91**]{}, 210402 (2003).
S. Dürr, T. Volz, A. Marte, and G. Rempe, Phys. Rev. Lett. [**92**]{}, 020406 (2004).
R. A. Duine and H. T. C. Stoof, Phys. Rep. [**396**]{}, 115 (2004).
L. Radzihovsky, J. Park, and P. B. Weichman, Phys. Rev. Lett. [**92**]{}, 160402 (2004).
S. G. Bhongale, J. N. Milstein, and M. J. Holland, Phys. Rev. A [**69**]{}, 053603 (2004).
M. W. J. Romans, R. A. Duine, S. Sachdev, and H. T. C. Stoof, Phys. Rev. Lett. [**93**]{}, 020405 (2004).
T. Rom, T. Best, O. Mandel, A. Widera, M. Greiner, T. W. Hänsch, and I. Bloch, Phys. Rev. Lett. [**93**]{}, 073002 (2004).
P. Meystre, J. Phys. B [**38**]{}, S617 (2005).
G.-R. Jin, C. K. Kim, and K. Nahm, Phys. Rev. A [**72**]{}, 045601 (2005).
R. S. Tasgal, G. Menabde, and Y. B. Band, Phys. Rev. A [**74**]{}, 053613 (2006).
T. Köhler, K. Góral, and P. S. Julienne, Rev. Mod. Phys. [**78**]{}, 1311 (2006).
I. Tikhonenkov and A. Vardi, Phys. Rev. Lett. [**98**]{}, 080403 (2007).
N. Syassen, D. M. Bauer, M. Lettner, D. Dietze, T. Volz, S. Dürr, and G. Rempe, Phys. Rev. Lett. [**99**]{}, 033201 (2007).
S. J. Woo, Q-Han Park, and N. P. Bigelow, Phys. Rev. Lett. [**100**]{}, 120403 (2008).
S. Giorgini, L. P. Pitaevskii, and S. Stringari, Rev. Mod. Phys. [**80**]{}, 1215 (2008).
R. Friedberg and T. D. Lee, Phys. Lett. A [**138**]{}, 423 (1989).
R. Friedberg and T. D. Lee, Phys. Rev. B [**40**]{}, 6745 (1989).
P. Kramer and M. Saracento, [*Geometry of the time-dependent variational principle*]{} (Springer, Berlin, 1981).
H.-J. Kull and D. Pfirsch, Phys. Rev. E [**61**]{}, 5940 (2000).
O. E. Alon, A. I. Streltsov, and L. S. Cederbaum, Phys. Lett. A 362, 453 (2007).
A. Szabo and N. S. Ostlund, [*Modern Quantum Chemistry*]{} (Dover, Mineola, NY, 1996).
, edited by D. R. Yarkony (World Scientific, Singapore, 1995).
[^1]: E-mail: [email protected]
[^2]: E-mail: [email protected]
[^3]: E-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
It is shown that the two-level correlation function $R(s,s')$ in the invariant random matrix ensembles (RME) with soft confinement exhibits a “ghost peak” at $s\approx
-s'$. This lifts the sum rule prohibition for the level number variance to have a Poisson-like term ${\rm var}(n)=\eta n$ that is typical of RME with broken $U(N)$ symmetry. Thus we conclude that the $U(N)$ invariance is broken spontaneously in the RME with soft confinement, $\eta$ playing the role of an order-parameter.
address: |
$^1$ International Centre for Theoretical Physics, 34100 Trieste, Italy\
$^2$Institute of Spectroscopy, Russian Academy of Sciences, 142092 Troitsk, Moscow r-n, Russia
author:
- 'C. M. Canali$^1$, and V. E. Kravtsov$^{1,2}$'
date: 'February 1, 1995'
title: 'Normalization Sum Rule and Spontaneous Breaking of U(N) Invariance in Random Matrix Ensembles.'
---
The statistical description of complex systems by ensembles of random matrices turned out to be a powerful general approach that was successively applied to a great variety of systems in different fields from nuclear physics \[1\] to mesoscopics \[2\] and quantum chaos \[3\].
The classical random matrix theory (RMT) by Wigner, Dyson and Mehta \[1\] describes the statistics of eigenvalues for a Gaussian ensemble of random Hermitian $N\times N$-matrices ${\bf H}$ with the probability distribution $P({\bf H})\propto \exp[-Tr {\bf H}^2]$. By definition, the statistical properties of this ensemble are invariant under unitary transformations $U(N)$ of matrices ${\bf H}$ and thus there is no basis preference in the RMT. This means that the classical RMT can be applied only to quantum systems where all (normalized) linear combinations of eigenfunctions have similar properties.
For disordered electronic systems, it implies that all eigenstates must be extended. In other words, the classical RMT is applicable only for describing the energy level statistics in the metal phase \[4,5\] that exists in the dimensionality $d>2$ at a relatively weak disorder.
With disorder increasing the system goes through the Anderson transition to an insulating phase in which all eigenstates are localized. The level statistics in this phase obviously cannot be described by the $U(N)$-invariant RME, since one can construct an extended state by a linear combination of localized states randomly positioned throughout the sample. Thus the proper probability distribution $P({\bf H})$ must contain a basis preference in order to exclude unitary transformations which would lead to formation of such extended states.
The ensemble of random banded matrices (RBME) \[6,7\] is an example of such a non-invariant RMT. It describes properties of systems belonging to so-called quasi-1d universality class which includes quasi-1d disordered electronic systems with localization \[6\] and certain quantum chaotic systems \[7\]. The corresponding eigenvalue statistics are Poissonic in the $N\rightarrow\infty$ limit (at a fixed bandwidth $b$) and reduce to the Wigner-Dyson form at $b\rightarrow\infty$. Thus changing the parameter $b/N$, one can describe the crossover from Wigner-Dyson to Poisson level statistics which occurs in quasi-1d disordered systems with increasing the ratio $L/\xi$ of the sample length $L$ and the localization radius $\xi$.
While the localization in quasi-1d systems seems to be well described in terms of the RBME, the problem of the random matrix description of the critical region near the Anderson transition and the Anderson insulator phase for $d>2$ remains open. The recent works \[8-10\] where the existence of the universal critical level statistics has been demonstrated, resulted in an intensive search for the proper RME description.
In this connection, two different generalizations of the classical RMT have been recently proposed \[11,12\]. The generalized RME studied in Ref.\[11\] was obtained from the Gaussian invariant ensemble by introducing a symmetry-breaking term: $$\label{ShE}
P({\bf H})\propto e^{- Tr {\bf H}^2}\,e^{-h^2 N^2 Tr([{\bf \Lambda},{\bf
H}][{\bf \Lambda},{\bf H}]^{\dagger})}.$$ The $h$-dependent term breaks the $U(N)$ invariance and tends to align ${\bf H}$ with a symmetry breaking unitary matrix ${\bf \Lambda}$ thus setting the basis preference. It turned out \[11\] that even after averaging over ${\bf
\Lambda}$ the resulting ensemble leads to the eigenvalue statistics that deviate from the Wigner-Dyson form. The difference between the Wigner-Dyson statistics that correspond to $h=0$ and the level statistics for [*any non zero*]{} $h$ turns out to be dramatic in the thermodynamic (TD) limit $N\rightarrow\infty$. Namely, for $h\neq 0$ the variance ${\rm var}(n)$ of the number of levels in an energy window that contains $n$ levels on the average, grows linearly with $n$ at $n\gg 1$: $$\label{var}
{\rm var}(n)=\langle (\delta n)^2\rangle=\eta(h) \,n\sim h\,n, \;\;\;\eta(0)=0.$$ For the classical RMT \[1\], ${\rm var}(n)\propto \ln n$ that is negligible as compared to Eq.(\[var\]) for [*any*]{} nonzero $0<\eta(h)<1$ in the limit $n\rightarrow\infty$. The Poisson-like behavior described by Eq.(\[var\]) is valid also for RBME in the TD limit.
In contrast to Eq.(\[ShE\]), the probability distribution suggested in Ref.\[12\] is explicitly $U(N)$-invariant: $$\label{MutE}
P({\bf H})\propto e^{-Tr V({\bf H}) }.$$ The only singularity in this model is that the (even in $E$) “confining potential” $V(E)$ grows extremely slowly: $$\label{V}
V(E)=\mbox{$\frac{A}{2}$} \ln^2 |E|,\;\;\;\;\;\; |E|\rightarrow\infty.$$ However, the Poisson-like behavior, Eq.(\[var\]), turns out to be valid for this model too, provided that the energy window does not contain the origin $E=0$.
It should be stressed \[13\] that for steeper confining potentials, $V(E)=
|E|^{\alpha}$, no deviation from the Wigner-Dyson statistics was observed in the bulk of the spectrum, so that $\eta_{\alpha}=0$ for [*all*]{} $\alpha>0$. What happens with the invariant RME at the transition from a power-law to logarithmic confinement, looks like the spontaneous breaking of the $U(N)$-symmetry.
In this Letter we present both analytical and numerical arguments showing this novel phenomenon to exist. We show the parameter $\eta$ to play the role of an order-parameter and clarify its connection with the breaking of the normalization sum rule in the TD limit.
The reason why $\eta=0$ for a wide class of invariant RME is connected with the normalization sum rule: $$\label{SR}
\int_{-\infty}^{+\infty} Y^{(N)}_{2}(s,s')ds'=1.$$ Here the cluster function $Y_{2}(s,s')=\delta(s-s')-R(s,s')$ is related to the two-level correlation function $R(s,s')$: $$\label{TL}
R^{(N)}(s,s')=\frac{\langle
\rho(E_{s})\rho(E_{s'})\rangle}{\langle\rho(E_{s})\rangle\,\langle\rho(E_{s'})
\rangle}-1.$$ where $\rho(E)=Tr\{\delta(E-{\bf H})\}$ is the level density, $\langle...\rangle$ denotes ensemble averaging with the probability distribution $P({\bf H})$, and the new variable $s(E)$ is chosen so that the mean level density in this variable, $\langle\tilde{\rho}(s)\rangle=\langle\rho(E_{s})\rangle
dE_{s}/ds=1$ for $-N/2<s<N/2$.
For $N$ finite, Eq.(\[SR\]) is an exact property of any RME that follows simply from the normalization condition $\int
\rho(E) dE=N$. However, the sum rule may be violated \[14\] in the limit $N\rightarrow\infty$, since after the integration in Eq.(\[SR\]) both terms in Eq.(\[TL\]) result in the divergent constants $N$ to be subtracted from each other. Below we will assume this limit to be taken.
The general relationship between the cluster function $Y_{2}(s,s')$ and the coefficient $\eta=\lim_{n\rightarrow\infty}\{d [{\rm var}(n)]/dn\}$ in the level number variance, Eq(\[var\]), for an energy window centered at a point $E_{0}=E(s_{0})$, reads: $$\label{dvdn}
\eta=\lim_{n\rightarrow\infty}\left[1-\int_{a_{-}}^{a_{+}}[Y^{\infty}_{2}
(a_{+},s')+
Y^{\infty}_{2}(a_{-},s')]\,ds'\right],$$ where $a_{\pm}=s_{0}\pm n/2$. In the case where the cluster function in the TD limit $Y^{\infty}_{2}(s,s')=Y^{\infty}_{2}(s-s')$ is translationally invariant, the relationship reduces to: $$\label{dvdnTD}
\eta=1- \int_{-\infty}^{+\infty} Y^{\infty}_{2}(s)\,ds.$$ Now comparing Eq.(\[SR\]) and Eq.(\[dvdnTD\]) we can conclude that the parameter $\eta$ is equal to the deficiency of the sum rule in the $N\rightarrow\infty$ limit \[9\].
Whether the sum rule is broken or not depends on the behavior of the effective level interaction at large distances. Quite generally, the joint probability distribution $P[\{x_{i}\}]$ of eigenvalues $x_{i}$ of a matrix ${\bf H}$ can be represented in the form similar to the Gibbs distribution $P[\{ x_{i}\}]=\exp[-\beta {\cal H}]$ for the one-dimensional plasma of classical particles described by the potential energy functional ${\cal H}[\{x_{i}\}]= \sum_{i}V(x_{i})+W[\{x_{i} \}]$. For the unitary ensembles considered in this Letter, the effective temperature $\beta=2$. In case of the invariant ensembles given by Eq.(\[MutE\]), the many-body interaction $W[\{x_{i} \}]$ reduces to the pair-wise logarithmic effective level interaction \[1\]: $$\label{log}
W[\{x_{i}\}]= -
\sum_{i>j}\ln|x_{i}-x_{j}|.$$ For RME similar to that given by Eq.(\[ShE\]), the corresponding effective level interaction that appears after averaging \[11\] over the symmetry-breaking matrix ${\bf \Lambda}$, contains all the many-body terms and can be rewritten in the form of a determinant $Det_{ij}$, $i,j=1,...N$: $$\label{det}
W[\{x_{i} \}]= - \mbox{$\frac{1}{2}$} \ln Det_{ij}[e^{-h^2 N^2
(x_{i}-x_{j})^2}].$$ In both limiting cases, $|x_{i}-x_{j}|\rightarrow 0$ and $|x_{i}-x_{j}|\rightarrow\infty$, the determinant in Eq.(\[det\]) can be approximated by a pair-wise interaction of the form: $$\label{cut}
W[\{x_{i}\}]=
-\mbox{$\frac{1}{2}$} \sum_{i>j}
\ln(1-e^{-2h^2 N^2 (x_{i}-x_{j})^2})+const.$$
The main difference between Eq.(\[log\]) and Eq.(\[cut\]) is that the symmetry-breaking field $h\neq 0$ results in the cut-off of the effective level interaction at sufficiently large distances \[15\]. This difference is crucial for the fulfillment of the sum rule, Eq.(\[SR\]), in the TD limit.
It can be shown that the normalization sum rule persists also in the TD limit if the effective level interaction is long-range (non-integrable). We will present here a simple proof of this statement based on the mean-field approach \[16\] that is valid exactly for the long-range interactions in the $N\rightarrow\infty$ limit. Suppose that the effective level interaction $f(x_{i}-x_{j})$ is pair-wise and long-range. Then using the relationship $\delta
\langle\rho(E) \rangle/\delta V(E')= -\beta [\langle\rho(E)\rho(E')
\rangle-\langle\rho(E) \rangle\,\langle\rho(E') \rangle]$ and the Dyson mean-field equation \[1\] for $\langle\rho(E)\rangle$ $$\label{MFE}
\int_{-\infty}^{+\infty}\langle\rho(E'')\rangle f(E''-E')+V(E')=const,$$ one can derive the integral equation for the two-level correlation function \[16\]: $$\label{IE}
\int_{-\infty}^{+\infty} ds''
R^{\infty}(s,s'')\,f(E_{s''}-E_{s'})=\beta^{-1}\delta(s-s').$$ Now integrate this equation over all $s$ changing the order of integration in the left-hand side and denote the integral $\int_{-\infty}^{+\infty} R^{\infty}(s,s'')ds=I$: $$\label{div}
I\beta \int_{-\infty}^{+\infty}f(E_{s''}-E_{s})\,ds''=1.$$ From the definition of the function $E_{s}$ just after Eq.(\[TL\]) and a physically obvious property $d\langle\rho(E) \rangle/d|E|<0$ that holds for an even confinement potential $V(E)$, it follows that $E_{s}$ increases linearly or faster with $s$. Therefore for any long-range interaction with $\int f(E)dE$ divergent at infinity, the integral in Eq.(\[div\]) is also divergent. Thus we arrive at the relationship: $$\label{SRp}
I=1-
\int_{-\infty}^{+\infty}Y_{2}^{\infty}(s,s')ds'=0,$$ which proves the validity of Eq.(\[SR\]) in the limit $N\rightarrow\infty$. On the contrary, if the effective interaction $f(E-E')$ is cut at large distances, the quantity $I$ in Eqs.(\[div\]),(\[SRp\]) must be finite and thus the sum rule is violated.
We see that the symmetry-breaking term in Eq.(\[ShE\]) leads to the cut-off of the effective level interaction at large distances and thus to the break-down of the normalization sum rule in the TD limit. This results in a quasi-Poisson level number variance, Eq.(\[var\]), governed by the nonzero parameter $\eta$, Eq.(\[dvdnTD\]).
The situation resembles the appearance of the long-range order in spin systems in the external magnetic field ${\bf h}$ which breaks down the rotational invariance. In this case, the spin-spin correlator $\langle {\bf S}({\bf r})
{\bf S}({\bf r'}) \rangle\rightarrow m^2$ is constant at large distances, the magnetization $|m|$ depending linearly on $|{\bf h}|$. Since the spin-spin correlator is invariant under global rotations of the spins, the order parameter, that is $|m|$, is unchanged after averaging over the direction of the symmetry-breaking field ${\bf h}$.
In our problem, the quantity analogous to $\langle {\bf S}({\bf r}){\bf S}({\bf r'}) \rangle$, which is invariant under global (independent on $E$, $E'$) unitary transformations, is the two-level correlation function $\langle
Tr\{\delta(E-{\bf H})\}Tr\{\delta(E'-{\bf H}) \} \rangle$. The symmetry-breaking field $h{\bf \Lambda}$ plays the role of the magnetic field, and the parameter $\eta$ is similar to the magnetization. Like $|m|$, $\eta$ remains nonzero after averaging over ${\bf \Lambda}$.
Now we return to consider the TD limit of the invariant ensemble with soft confinement, defined by Eqs.(\[MutE\]),(\[V\]). Since the effective level interaction, Eq.(\[log\]), is long-range, the sum rule, Eq.(\[SR\]), must hold for any confining potential $V(E)$. However, for soft confinement this does not necessarily lead to $\eta=0$. The reason is the dramatic break-down of the translational invariance of the cluster function $Y^{\infty}_{2}(s,s')=Y^{\infty}_{2}(s-s')+Y_{a}(s,s')$ which turns out to have an anomalous part $Y_{a}(s,s')$ that exhibits a sharp peak of width $A$ near $s=-s'$.
With the translational invariance broken in such a way, the level number variance ${\rm var}(n)$ becomes dependent on the position of the energy window $s_{0}$, and can oscillate as a function of $n$. The parameter $\eta$ is given by the general Eq.(\[dvdn\]), where $n = 1, 2, ...$. It is easy to show that the integral in Eq.(\[dvdn\]) reduces approximately to that in the sum rule, Eq.(\[SR\]), only provided that $n/2-|s_{0}|\gg A$, that is the origin $E=s=0$ is far within the energy window. This means that the parameter $\eta$ vanishes for a symmetric window $s_{0}=0$ and it is nonzero if $|s_{0}|\gg n/2$.
Indeed, the Monte Carlo simulations on the classical 1d plasma described by Eqs.(\[V\]),(\[log\]) show a dramatic difference in the level number variance in these two cases. A remarkable property of the model is that for a symmetric window, the level number variance ${\rm var}(n)$ is [*constant*]{} for all integers $n\gg1$. Thus the “level rigidity” is even higher than that for the classical RMT where ${\rm var}(n)\propto\ln n$.
Though the anomalous part $Y_{a}(s,s')$ can be extracted from the exact solution \[12\] for the particular confining potential obeying Eq.(\[V\]), its existence and importance was never mentioned before. Below we present a simplified derivation of the cluster function $Y_{2}^{\infty}(s,s')$ for small values of the parameter $q=e^{-\pi^2 A}$.
We start with the representation \[1\] of $Y_{2}^{\infty}(s,s')$ in terms of the orthonormal “wave functions” $\varphi_{i}(E)=p_{i}(E) exp(-V(E))$, where $p_{i}(E)$ are orthogonal polynomials corresponding to the weight function $e^{-2V(E)}$: $$\label{Y}
Y_{2}^{\infty}(s,s')=
\frac{K^{2}(E_{s},E_{s'})}{K(E_{s},E_{s})K(E_{s'},E_{s'})}.$$ Here the kernel $K(E,E')$ is given by \[1\]: $$\label{K}
K(E,E')=\frac{1}{\pi C^2}
\,\frac{\varphi_{o}(E')\varphi_{e}(E)-\varphi_{e}(E')\varphi_{o}(E)}{E'-E},$$ where $C$ is a constant and $\varphi_{o(e)}$ are $N\rightarrow\infty$ limits of wave functions $\varphi_{N}(E)$ of the odd and even order, respectively.
For $q\ll 1$, the wave functions $\varphi_{o(e)}(E)$ have a “semi-classical” form that is a generalization of the $N\rightarrow\infty$ limit of the oscillator wave functions for a nonlinear $s(E)$: $$\label{anz}
\varphi_{o}(E)=C\sin[\pi s(E)],\;\;\; \varphi_{e}(E)=C\cos[\pi s(E)].$$ The arguments in $\sin$ and $\cos$ are chosen so that the density of zeros $\rho_{0}=ds/dE$ coincides with the mean level density $\langle\rho(E)
\rangle$.
The function $s(E)$ can be easily found from the known solution of the Dyson mean field equation (\[MFE\]). Solving Eq.(\[MFE\]) with $f(E-E')=-\ln|E-E'|$ for the confining potential, Eq.(\[V\]), and taking the limit $N\rightarrow\infty$, we have: $$\label{s(E)}
\langle\rho(E)\rangle=ds/dE=\mbox{$\frac{A}{2|E|}$},\;\;\;E_{s}=\lambda
e^{2|s|/A}\,{\rm sign}(s),$$ where $\lambda$ is a constant of integration.
Now using Eqs.(\[Y\])-(\[s(E)\]) and the relationship \[1\] $\langle
\rho(E)\rangle=K(E,E)$ one can obtain an explicit form of the cluster function $Y_{2}(s,s')$ for $q=e^{-\pi^2 A}\ll 1$, where $A$ is a coefficient in Eq.(\[V\]). For $ss'>0$ we arrive at the translationally invariant expression \[12\]: $$\label{Yn}
Y_{2}^{\infty}(s-s')=\frac{1}{\pi^2
A^2}\,\frac{\sin^2[\pi(s-s')]}{\sinh^2[(s-s')/A]},\;\;\;ss'>0.$$ However, there are strong correlations for $ss'<0$ too: $$\label{Ya}
Y_{a}(s,s')=\frac{1}{\pi^2
A^2}\,\frac{\sin^2[\pi(s-s')]}{\cosh^2[(s+s')/A]},\;\;\;ss'<0.$$ This is just the anomalous part of the cluster function discussed above. Its remarkable property is a sharp peak at $s\approx -s'$ with a height that [*does not*]{} decrease when $|s-s'|\sim 2|s|\rightarrow\infty$.
The existence of such a “ghost” peak in the two-level correlation function is confirmed also by the Monte Carlo simulations on the effective plasma model with a soft confinement \[see Fig.1\]. Such anomalous, infinitely long-range level correlations make it possible for the [*invariant*]{} random matrix ensemble defined by Eqs.(\[MutE\]), (\[V\]) to exhibit the Poisson-like level number variance ${\rm var}(n)=\eta n$ which is typical for RME with [*broken*]{} $U(N)$ symmetry. Moreover, after the substitution $\pi^2 A=2/h\gg 1$, the form of the “normal”, translationally invariant part of $Y^{\infty}_{2}(s,s')$ given by Eq.(\[Yn\]) is identical to that found in Ref.\[11\] for the matrix ensemble, Eq.(\[ShE\]), with symmetry breaking.
This is a crucial point in the chain of arguments in favour of the statement that the $U(N)$ symmetry is [*spontaneously broken*]{} in the random matrix ensemble, Eqs.(\[MutE\]),(\[V\]), with soft confinement.
The corresponding order parameter which arises either due to the external symmetry breaking field $h$ or spontaneously for a sufficiently soft confinement, can be defined in the $N\rightarrow\infty$ as follows: $$\label{op}
\eta=\lim_{s\rightarrow+\infty}\left[1-\int_{s'>0}Y^{\infty}_{2}(s,s')\,ds'
\right].$$ It coincides with the coefficient in the quasi-Poisson term in the level number variance, Eq.(\[var\]) provided that the energy window does not contain the origin.
In case of the spontaneously broken $U(N)$ symmetry, a complementary order parameter can be introduced: $$\label{cop}
\tilde{\eta}=\lim_{s\rightarrow+\infty}
\left[-\int_{s'<0}Y^{\infty}_{2}(s,s')\,ds'\right],$$ which involves integration only over the anomalous part of the cluster function. Because of the sum rule, Eq.(\[SR\]), the sum $\eta+\tilde{\eta}=0$ must be zero. Exploiting the spin analogy discussed above, one can say that the spontaneous symmetry breaking is of the AFM type, with $\eta$ and $\tilde{\eta}$ playing a role of the sublattice magnetizations.
We conclude that the anomalous statistics of eigenvalues in RME with soft confinement is a result of the spontaneous breaking of the $U(N)$ symmetry which exhibits itself in the quasi-Poisson term in the level number variance and in the appearance of the “ghost” peak at $s=-s'$ in the two-level correlation function $R(s,s')$.
We are grateful to Yu Lu, A. Mirlin, E. Tosatti and Mats Wallin for stimulating discussions.
See for reviews M. L. Mehta, [*Random matrices*]{} (Academic Press, Boston, 1991; F. Haake, [*Quantum signatures of chaos*]{} (Springer-Verlag, Berlin, 1991). B. L. Altshuler, V. E. Kravtsov, and I. V. Lerner, [in]{} [ *Mesoscopic Phenomena in Solids*]{}, [edited by]{} B. L. Altshuler, P. A. Lee, and R. A. Webb, North-Holland, Amsterdam, 449 (1991). M. C. Gutzwiller, [*Chaos in Classical and Quantum Mechanics*]{} (Springer Verlag, New York, 1990). L. P. Gor’kov and G. M. Eliashberg, [Sov. Phys. JETP]{} [**21**]{}, 940 (1965). K. B. Efetov, Adv. Phys. [**32**]{}, 53 (1983). Y. V. Fyodorov and A. D. Mirlin, Phys. Rev. Let.,[**69**]{}, 1093 (1992); Phys. Rev. Lett. [**71**]{}, 412 (1993). F. M. Izrailev, Phys. Rep. [**196**]{}, 299 (1990); G. Casati, L. Molinari, F. M. Izrailev, Phys. Rev. Lett. [**64**]{}, 1851 (1990). B. I. Shklovskii, B. Shapiro, B. R. Sears, P. Lambrianides, and H. B. Shore, Phys. Rev. B [**47**]{}, 11487 (1993). V. E. Kravtsov, I. V. Lerner, B. L. Altshuler, and A. G. Aronov, Phys. Rev. Lett. [**72**]{}, 888 (1994). A. G. Aronov, V. E. Kravtsov, and I. V. Lerner, Phys. Rev. Lett., [**74**]{}, 1174, (1995). J. -L. Pichard and B. Shapiro, J. Phys. [**4**]{}, 623 (1994); M. Moshe, H. Neuberger, and B. Shapiro, Phys. Rev. Lett. [**73**]{}, 1497 (1994). K. A. Muttalib, Y. Chen, M. E. H. Ismail, and V. N. Nicopoulos, Phys. Rev. Lett. [**71**]{}, 471 (1993); C. Blecken, Y. Chen, and K. A. Muttalib, J. Phys. A, [**27**]{}, L563 (1994). C. M. Canali, Mats Wallin, V. E. Kravtsov, Phys. Rev. B., [**51**]{}, 2831 (1995). V. E. Kravtsov, and I. V. Lerner, Phys. Rev. Lett. [**74**]{}, 2563, (1995); A. G. Aronov, and A. D. Mirlin, Phys. Rev. B [**51**]{}, 6131 (1995). Distances $\Delta x\sim \mbox{$1/hN$}$ correspond to $\Delta s\sim h^{-1/2}$. C. W. J. Beenakker, Phys. Rev. Lett., [**70**]{}, 1155 (1993). R. A. Jalabert, J. -L. Pichard, and C. W. J. Beenakker, Europhys. Lett., [**24**]{}, 1 (1993).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Hysteresis-controlled devices are widely used in industrial applications. For example, cooling devices usually contain a two-point controller, resulting in a nonlinear hybrid system with two discrete states. Dynamic models of systems are essential for optimizing such industrial supply technology. However, conventional system identification approaches cannot handle hysteresis-controlled devices. Thus, the new identification method Sparse Identification of Nonlinear Dynamics (SINDy) is extended to handle hybrid systems. In this new method (SINDyHybrid), tailored basis functions in form of relay hysterons are added to the library which is used by SINDy. Experiments with a hysteresis controlled water basin show that this approach correctly identifies state transitions of hybrid systems and also succeeds in modeling the dynamics of the discrete system states. A novel proximity hysteron achieves the robustness of this method. The impacts of the sampling rate and the signal noise ration of the measurement data are examined accordingly.'
author:
-
-
bibliography:
- 'IEEEabrv.bib'
- 'literature.bib'
title: 'System identification of a hysteresis-controlled chiller plant using SINDy [^1] '
---
SINDyHybrid, system identification, nonlinear systems, hysteresis, hybrid dynamical systems, sparse regression
Introduction
============
Nowadays, data-driven approaches for modeling dynamics without a fixed model structure gain relevance. However, classical grey-box modeling is still popular because it allows to include domain-specific knowledge, constraints, and to achieve transparency. The presented work provides an example of combining explicit knowledge with a flexible data-driven approach in order to estimate hysteresis behavior.
A traditional approach of obtaining a system model combines physical laws and technical parameters to so called white-box-models. Alternatively, grey-box-models can be used, which have a fixed model structure but chosen parameters are tuned using measurement data. Methods from machine learning partially abandon fixed model structures, like in the case of neural networks. These approaches do not provide a clear transparency, which implies a need for empirical validation. Furthermore, pure data-driven methods require for extensive data for all considered situations and operation points. However, due to high costs, efforts or dangers it is not always feasible to conduct real experiments.
This article proposes an extension of the Sparse Identification of Nonlinear Dynamics (SINDy) framework developed and presented by S. Brunton and J.N. Kutz in their prominent publication [@BruntonSINDy.2016]. Adding the robust identification of hysteresis behavior in hybrid systems holds the potential to identify all discrete states of a hysteresis controlled system and build a nonlinear model for the dynamics on all states in one step. The approach aims to handle diverse systems, to include system knowledge and to result in an easily interpretable model.
Section \[sec:stateOfTheArt\] will cover basic theory regarding system identification and current approaches. A new concept is presented in Section \[sec:Concept\]. Part \[sec:practicalEvaluation\] contains the practical experiments and their results while Part \[sec:discussion\] gives a deepening discussion. Section \[sec:conclusion\] summarizes the findings and gives a brief outlook to further applications.
State of the art {#sec:stateOfTheArt}
================
Whenever a certain degree of precision is required for tasks like optimization or anomaly detection, simple dynamic models may deviate too much from the actual system, especially in the critical regions of hysteretic systems, where the state transitions happen. In order to pursue model-based approaches for these tasks, sophisticated models are required.
System Identification for nonlinear hybrid systems
--------------------------------------------------
### Hybrid systems.
The considered type of dynamic systems contains both discrete and continuous parts. Depending on a discrete state, a different behavior is observed. A famous example is the temperature characteristics of water: the behavior changes with reaching the threshold of 100 $^\circ C$ due to the phase transition. Therefore, it makes sense to create two distinct models: one for the discrete state “temperature below 100 $^\circ C$" and one for “temperature above 100 $^\circ C$".
### PWARX models. {#PWARX}
Piece-Wise-Affine (PWA) methods divide the state-space into discrete regions, commonly shaped as polyhedronscite[Ferrari-Trecate.2000]{}. For every region a continuous model is set up, e.g. the well known linear *autoregressive model with exogenous input* (ARX-model). This procedure is called PWARX[@pwarx_paoletti]. If the transition from one state to another is affected by hysteresis behavior, PWA models are not sufficient anymore, because they assume fixed transition borders from one discrete state to another. They are unable to include memory of the past trajectory of the system’s state. [@Fang.2015] present a data-driven approach to model sticky valves using the Preisach model for hysteresis, which originated from describing natural hysteresis in ferromagnetic materials. They succeed in modeling the state transitions, but they separate the identification of the transition model from the identification of dynamic models for the discrete states. Also, the number of discrete states must be known beforehand. This will be overcome by using SINDyHybrid.
### SARX models. {#SARX}
**S**witched **A**ffine Modelle (*SARX*) models can be used if transitions of the state are triggered by external events. It is also possible to use **p**iece**w**ise continuously differentiable **n**onlinear functions (*PWNARX*, *SNARX*). An overview of related techniques is given in[@Hutchison.2008]. Furthermore, the authors present a support-vector regression to separate the domains with hyperplanes of the form $$\label{eq:pwa-svr}
h(x) = \sum_{i=1}^{N}(\beta_i k_c (x,x_i) + b_c) = 0 ,$$ with $N$ samples of p-dimensional data $x = [x_1, ..., x_N]$, kernel function $k_c(.,.)$ and tunable parameters $\beta_i, b_c$. The sign of $h$ gives the classification.
### Black-box-techniques.
An alternative approach would be to use neural networks to learn the system dynamics, but the resulting models are usually difficult to check for plausibility and stability, which poses further problems for industrial applications[@Ljung_BlackBox_2001].
Hysteresis Modeling {#hysteresisModeling}
-------------------
Whenever a system’s behavior is not only depending on its current internal states, but also on its past trajectory, the system is called hysteretic. This nonlinear effect occurs in nature phenomena like in magnetization of ferromagnetic materials, but also in artificial systems like two-point controllers. The Hungarian physicist Ferenc Preisach presented the first suitable model to describe hysteretic behavior of magnetism in 1935[@Preisach.1935; @PreisachInMemorial]. His hysteron-operator allows for approximation of natural hysteresis using weighted stairs. The relay operator $R_{\alpha, \beta}$, also called relay hysteron, is expressed by Eq. and illustrated in Fig.\[fig:preisachRelay\]. $$\label{eq:relayHysteron}
y(x) = \begin{cases}
1, & \text{if $x \geq \beta$.} \\
0, & \text{if $x \leq \alpha$.} \\
y_p, & \text{if $ \alpha < x < \beta $.}
\end{cases}$$
![Relay hysteron [@RelayHyst1][]{data-label="fig:preisachRelay"}](Preisach_RelaySVG){width="0.82\linewidth"}
The value $y_p$ always corresponds to the prior value of $y(t)$. The complex shape of a magnetic hysteresis curve can be achieved by superposition of many elementary relay hysterons.
Modern approaches of hysteresis modeling are discussed in[@Hassani.2014]. The Krasnosel’skii–Pokrovskii (KP) model[@Krasnoselskii.1989] formalizes Preisach’s model in a mathematical way and allows for different operators to be included. Next to operator-based approaches, hysteresis can also be described by differential equations, e.g. using the Bouc–Wen model[@wen1976method] or the extended Bouc-Wen model, which requires 13 parameters to approximate the complex shape of the hysteresis[@Zhang.2001]. The here presented work utilizes the concept of hysterons which is the basis of Preisach’s model.
System Identification
---------------------
The SINDy framework is based upon the concepts of the Koopman operator and of sparsity. This section will give a brief introduction to these concepts, as well as present an overview of the SINDy methodology and its applications. In the following Section \[sec:Concept\], a concept is shown which implements the hysteresis model into the given framework.
### Koopman operator.
The Koopman operator $\mathcal{K}$[@Koopman.1931] is a linear but typically infinite-dimensional operator, which is capable of describing the full dynamics of an underlying system. For this paper we focus on the practical implementation of Koopman operator theory. An in-depth overview about theoretical concepts and applications can be found in [@Budisic.2012].
Each dynamic system of the form with state $x \in \mathbb{R}^n$ and time $t$ has a linear representation of the form $$\frac{d}{dt}g(x) = \mathcal{K} g(x),$$
where $g = [h_1(x), ...]^T$ is a function of possibly infinite observables $h_i(x)$ of the state. The Koopman operator $\mathcal{K}$ evolves observable functions $h_i(x)$ instead of direct traces of the states, which can be arbitrary functions. An example is shown in Eq. . $$\label{eq:observables_example}
g(x) = [h_1(x), h_2(x)...]^T = [1, x, x^2, x^3, \sin(x), \sqrt{x}, e^x, ...]^T$$
These observable functions span a subspace of the Hilbert space[@BruntonKoopmanInvariant.2016b]. In general, the resulting system is infinite dimensional but linear. This enables tools like spectral analysis without any inaccuracy caused by linearization[@Budisic.2012]. In 2005, Igor Mezic presented the application of the Koopman operator for spectral analysis of high-dimensional, nonlinear systems[@Mezic.2005]. The Dynamic Mode Decomposition (DMD) extracts eigenvalues and modes of the Koopman operator from measurement data[@Schmid2010], i.e. it finds the most probable system matrix $A$ for The extended-DMD (e-DMD) additionally approximates eigenfunctions of the Koopman operator[@WilliamsKernel; @WilliamsEDMD.2015].
### Sparsity.
The concept of sparsity is useful to find solutions of underdetermined systems of linear equations. A solution $x \in \mathbb{C}^n$ is considered *sparse*, when most of the entries of $x$ vanish,[@Cands2014MathematicsOS].
Next to offering numerical advantages, a solution with a limited number of terms facilitates the manual interpretation of the results. The problem of finding sparse solutions can be connected to model reduction techniques, as discussed in [@loiseau_brunton_2018].
[@Donoho.2006] showed that an underdetermined system of equations can be converted to a convex problem using a sparsity promoting condition. Consider solving $$y = \theta s^{\prime}$$ for $s^{\prime}$, with $y \in \mathbb{R}^n$, $\theta \in \mathbb{R}^{n \times m}$, $s^{\prime} \in \mathbb{R}^m$ and $m>n$. The parameter vector $ s^{\prime}$ needs to be determined in order to fit the measurement vector $y$ using the product of the observation matrix $\theta$ and $ s^{\prime}$.
The common $\text{L}_2$-regularized regression promotes a high number of involved terms. Contrariwise, by using the $\text{L}_1$-norm as regularization term, terms of minor impact are neglected from the solution. The $\text{L}_1$-regularized regression (see Eq. ) with weight $\lambda$ can be solved e.g. by the LASSO described by [@TibshiraniLasso]. The solution $s$ is optimal in $\text{L}_1$-sense and sparse within the space spanned by the columns of $\theta$, i.e. the basis functions. $$\label{eq:l1regression}
s = \operatorname*{arg\,min}_{s^{\prime}} \| \theta s^{\prime} - y \|_2 + \lambda \| s^{\prime} \|_1.$$
### Sparse Identification of Nonlinear Dynamics (SINDy).
Since 2015, Nathan Kutz, Steven Brunton and Joshua Proctor have developed the Sparse Identification of Nonlinear Dynamics method[@BruntonSINDy.2016], which is the basis for our data-driven approach to identify hysteretic behavior. First, a library $\Theta(X)$ with possible basis functions is built. SINDy aims to find the minimal amount of those functions, with which a given signal can be approximated. In contrast to DMD, which results in a linear model, SINDy can include arbitrary nonlinearities. As a further development, SINDYc can include feedback and control[@BRUNTON_SINDYc]. A successful application of SINDy in model predictive control is reported in [@Kaiser.].
The main principle of SINDy is the approximation of signals using only few basis functions, i.e. *sparse* solutions. The vector $x(t) = \begin{bmatrix}
x_1(t) & x_2(t) & \dots & x_n(t)
\end{bmatrix}^T \in \mathbb{R}^n $ represents all states of the system at time $t$. Measurement data of these states $x(t)$ and their derivatives $\dot{x}(t)$ at the time-steps $t_1, t_2, \dots , t_m$ form two data matrices $X$ and $\dot{X}$, see Eq. , . $$\begin{aligned}
\label{eq:dataMatricesX}
X &= \begin{bmatrix}
x^T(t_1) \\
x^T(t_2) \\
\vdots \\
x^T(t_m)
\end{bmatrix}
= \newcommand{\mymatrix}[1]{\ensuremath{\overset{\xrightarrow[\hphantom{#1}]{\text{\scriptsize States}}}{#1}\left\downarrow\vphantom{#1}\right.}}
\mymatrix{\begin{bmatrix}
x_1(t_1) & x_2(t_1) & \dots & x_n(t_1) \\
x_1(t_2) & x_2(t_2) & \dots & x_n(t_2) \\
\vdots & \vdots & \ddots& \vdots \\
x_1(t_m) & x_2(t_m) & \dots & x_n(t_m) \\
\end{bmatrix}}
\text{\scriptsize Time}
\\
\dot{X} &=
\begin{bmatrix}
\dot{x}^T (t_1) \\
\dot{x}^T (t_2) \\
\vdots \\
\dot{x}^T (t_m)
\end{bmatrix}
= \begin{bmatrix}
\dot{x}_1(t_1) & \dot{x}_2(t_1) & \dots & \dot{x}_n(t_1) \\
\dot{x}_1(t_2) & \dot{x}_2(t_2) & \dots & \dot{x}_n(t_2) \\
\vdots & \vdots & \ddots& \vdots \\
\dot{x}_1(t_m) & \dot{x}_2(t_m) & \dots & \dot{x}_n(t_m)
\end{bmatrix}
\label{eq:dataMatricesXdot}\end{aligned}$$
As a next step, a library $\Theta(X)$ is built from candidate functions, in which $X$ is evaluated. Typical candidates are polynomials, trigonometric functions and constants, but arbitrary functions can be included as well. An example is given in Eq. . $$\begin{aligned}
\Theta(X)\!= \!
\begin{bmatrix}
\vline &\vline & \vline &\vline & & \vline & \vline & & \\
1 & X & X^{P_2} & X^{P_3} & \!\dots\! & \sin(X) & \cos(X) & \!\dots\! & e^X \\
\vline &\vline & \vline &\vline & & \vline & \vline & &
\end{bmatrix}
\label{eq:bibliothekStruktur}\end{aligned}$$
The terms $ X^{P_n} $ are polynomials of degree $n$. This includes cross terms of the single states. For the case of quadratic polynomials, this yields to Eq. . $$\delimitershortfall=0pt
\begin{footnotesize}
X^{P_2} \!= \!
\begin{bmatrix}
x_1^2(t_1) \!&\! x_1(t_1)x_2(t_1) \!& \!\dots\! &\! x_2^2(t_1) &\! x_2(t_1)x_3(t_1) & \!\dots\! &\! x_n^2(t_1) \\
x_1^2(t_2) \!&\! x_1(t_2)x_2(t_2) \!& \!\dots\! &\! x_2^2(t_2) &\! x_2(t_2)x_3(t_2) & \!\dots\! &\! x_n^2(t_2) \\
\vdots \!&\! \vdots \!& \!\ddots\!&\! \vdots &\! \vdots & \!\ddots\!&\!\vdots \\
x_1^2(t_m) \!&\! x_1(t_m)x_2(t_m) \!& \!\dots\! &\! x_2^2(t_m) &\! x_2(t_m)x_3(t_m) & \!\dots\! &\! x_n^2(t_m)
\end{bmatrix}\label{eq:polynomialLib}
\end{footnotesize}$$
So, the dynamic can be expressed by $\dot{X} = \Theta(X) \Xi$, where $\Theta(X)$ is the evaluated library and $\Xi = \begin{bmatrix}
\xi_1 & \xi_2 & \dots & \xi_n
\end{bmatrix}$ is the coefficient matrix. The coefficients can be determined using sparse regression techniques, e.g. sequential least-squares regression, as has been done in [@BruntonSINDy.2016].
For time-discrete systems, the equation $$\label{eq:sindyDiscrete}
X_{k+1} = \Theta(X_k) \Xi.$$ can be solved accordingly. SINDy allows for easy identification of few dominant functions out of a basis library. Choosing an appropriate library, which contains sufficient functions to describe the dynamics, is essential for a successful identification.
### Criteria for model quality.
There are several methods to compare different models and to decide which one to use. Parameter tuning often uses cross-validation[@kohavi1995study], which bases on the separation of data-sets for training and for validation. The fewer parameters a model has, the less flexible it is, and the less likely it is that overfitting will take place. Therefore the *Akaike Information Criterion* (AIC)[@akaike1974new] was defined to include the number of parameters into the quality of the model: $$\label{eq:aic_definition}
AIC = 2k - 2\ln(RSS).$$ The integer $k$ gives the number of parameters and the residual sum of squares (RSS) quantifies the prediction error[@ManganModelSelection.2017]. Therefore, the AIC is a trade-off between prediction and model complexity.\
In order to be able to compare errors between different time series, the Mean Absolute Scaled Error *MASE*[@Hyndman.2006] can be used, which relates the error to the overall change of the time series. For a small number of samples, there is a special correction term for the AIC. This so called AICc is defined in Eq. following[@Burnham.2002 Chap. 6.8]. $$\label{eq:aicc}
AICc = AIC + \frac{2k^2 + 2k}{n-k-1}.$$ The best model achieves the smallest AIC, or AICc respectively. The term $\Delta_j = AIC_i - AIC_{min}$ is a measurement for the difference between models. For $\Delta_j < 2$, models are similar, whereas for $\Delta_j > 10$, the models are significantly different[@KennethBurnham2004]. In context of SINDy, the AIC was applied successfully for model selection[@ManganModelSelection.2017].
Concept for identification of hysteresis-controlled systems {#sec:Concept}
===========================================================
This section provides a technical concept for system identification of hysteretic systems using SINDy, named SINDyHybrid. Based upon theoretical considerations of SINDy and hybrid systems, we developed an approach to incorporate tailored basis functions into the function library. To assess the results, we define criteria for model quality. SINDy has been used by [@BruntonSINDy.2016] to model the dynamics of systems with a bifurcation parameter $\mu$, which changes the qualitative behavior of the system’s dynamics. In order to incorporate $\mu$ into the observables, the equation $\dot{x} = f(x)$ was extended by $\mu$, such that $$\begin{aligned}
\dot{x} &= f(x, \mu), \label{eq:sindyBifurcation}\\
\dot{\mu} &= 0. \label{eq:MuDerivative}\end{aligned}$$ The function library $\Theta$ then consists of functions of $x$ and $\mu$. The value of $\mu$ however remains unchanged, as indicated by Eq. . A more general formulation has been used in [@BRUNTON_SINDYc], where feedback control signals were included by choosing $$\dot{\mu} = g(x). \label{eq:MuDerivativeSindyC}$$
Expanding the theory for hybrid systems
---------------------------------------
We now further generalize the approach of Eq. by choosing $$\dot{\mu} = g(x, \mu), \label{eq:MuDerivativeGeneral}$$ which connects changes of $\mu$ to current values of $\mu$. This represents the behavior of a hybrid system. Special cases of Eq. can be linked to the previously presented methods of modeling hybrid systems. Choosing $$\dot{\mu} = g(t)$$ results in the formulation of SARX models (see \[SARX\]). In cases where the change of the qualitative behavior of the system depends on only current values of the state $x$, the change of $\mu$ can be represented by $$\dot{\mu} = g(x(t)), \label{eq:MuDerivativePWARX}$$ which is the premise of PieceWise-Affine (PWA) models (see \[PWARX\]). For a linear function $g$, the regions for each state have a linear border. For a nonlinear $g$, complex borders can be modeled between regions, as is the case for NPWA models. If the function $f$ in Eq. is linear, the formulation is equivalent to PWARX models, otherwise it results in PWNARX models.\
Hysteresis models not only depend on the current state of the system, but also on past values up to a horizon $q$. Let $$\dot{\mu} = g(x(t-q), x(t-q+1), ..., x(t-1), x(t), \mu) = g(x_q, \mu).$$ This way, $\mu(t)$ serves as a storage for information of past values of the system. The general formulation Eq. is equivalent to $$\dot{\mu} = g(x(t), \mu(t)). \label{eq:MuDerivativeSimple}$$
Therefore it is sufficient to model the change of $\mu$ by current values of $x$ and $\mu$, even for hysteretic systems. The question of how to choose a proper function $g$ will be addressed next.
Modeling discrete states {#modelingDiscreteStates}
------------------------
In order to find a suited function $g$ from Eq. to model changes of discrete systems, one has to consider the way the change of states takes place. The transition can for example be either abrupt or smooth, following a certain shape. In the case of two-point controlled systems the state transition is abrupt: As soon as a certain threshold gets passed, the system behavior changes more or less instantaneously. This assumption holds for many hybrid systems.
In \[hysteresisModeling\] we discussed the Preisach model for identification of hysteresis, which uses the relay operator $R_{\alpha, \beta}$. Based upon the idea that state transitions in many hybrid systems happen abruptly, the relay hysteron seems to be a suited candidate for the function $g$.
Preparing the data
------------------
The first step of data processing is affine scaling of all data to a sensible and comparable range, e.g. from -1 to 1. This promotes an equal treatment of all time series, not depending on the absolute values of the data.
Before identification with SINDy begins, the possible state indicators (relay hysterons, Section \[modelingDiscreteStates\]) have to be computed from the data. This is done in three steps: Building indicator functions, pairwise matching of indicator functions and evaluating the relay hysterons.
We assume a change of the system behavior as soon as a value $x$ rises above $\beta \in \mathbb{R} $ or falls below $\alpha \in \mathbb{R} $. We also assume $\alpha$ and $\beta$ are included in our data. The first step is to build differences between all time series in our data. If the units of the time series are known, it is helpful only to build differences between commensurable data, e.g. time series which share the same unit, as exemplarily done in Eq. . $$\begin{aligned}
\label{eq:differences}
\tilde{x}_1 &= x-\alpha \\
\tilde{x}_2 &= x-\beta \\
\tilde{x}_3 &= \alpha - \beta
\end{aligned}$$
We then create indicator functions masking these differences depending on the sign of the value. $I(\tilde{x}_2)$ masks ranges, where $x \geq \beta$. $\bar{I}(\tilde{x}_1)$ on the other hand masks ranges, where $x < \alpha$. $$\begin{aligned}
\begin{split}
I(x) &= \begin{cases}
1, & \text{if $x \geq 0 $} \\
0, & \text{if $x < 0 $} \\
\end{cases} \label{eq:indicator1} \\
\bar{I}(x) &= \begin{cases}
1, & \text{if $x < 0 $} \\
0, & \text{if $x \geq 0 $} \\
\end{cases}
\end{split}\end{aligned}$$
A hysteron can be formed out of two indicator functions, which are not both true at the same time. For all matching indicator pairs $I_\alpha, I_\beta$, the hysteron $H_{\alpha, \beta}$ is built. This way, $I(\tilde{x}_2)$ and $\bar{I}(\tilde{x}_1)$ would also be combined. For each hysteron, a complementary hysteron $\bar{H}_{\alpha, \beta}$ will also be built. The evaluation is done according to
$$\label{eq:hysteron}
H(k)_{\alpha, \beta} = \begin{cases}
1, & \text{if $I_\beta = 1$.} \\
0, & \text{if $I_\alpha = 1$.} \\
H(k-1), & \text{for $I_\beta = I_\alpha = 0$.}
\end{cases}$$
for each step $k>1$.
This is equivalent to the formulation of the relay hysteron in Eq. . Presuming alternating switches between the states, $H(1)_{\alpha, \beta}$ can be initialized with the opposite state of the first occuring switch. This presumption has to be reviewed depending on the use case.
Robust handling of transitions – the proximity hysteron {#sec:proximity}
-------------------------------------------------------
A problem in the design of hysterons is the fact, that the critical threshold values $\alpha$ and $\beta$ from Eq. are only reached for a short moment. Due to a slow sampling frequency or noisy data it may happen that these critical points are not even included in the data at all. In order to improve robustness of the relay hysterons, we developed a “proximity-hysteron", which utilizes an $\epsilon$-range around the actual threshold value for switches between states. The evaluation is done according to $$\label{eq:proxHysteron}
H_{\epsilon}(k)_{\alpha, \beta} = \begin{cases}
1, & \text{if $I_{\epsilon, \beta} = 1$.} \\
0, & \text{if $I_{\epsilon, \alpha} = 1$.} \\
H_{\epsilon}(k-1), & \text{for $I_{\epsilon, \beta} = I_{\epsilon, \alpha} = 0$.}
\end{cases}$$
with the proximity indicator functions $$\begin{aligned}
\begin{split}
I_{\epsilon_\beta}(x) &= \begin{cases}
1, & \text{if $x = \min\{\max( Z_{x_\beta}(t) ), \beta\} $.} \\
0, & \text{otherwise.}
\end{cases} \\
I_{\epsilon_\alpha}(x) &= \begin{cases}
1, & \text{if $x = \max\{\min( Z_{x_\alpha}(t) ), \alpha\} $.} \\
0, & \text{otherwise.}
\end{cases}
\end{split}\end{aligned}$$
and the connected sets and\
$Z_{x_\alpha}(t) = \{x \mid x < \alpha + \epsilon_\alpha\}$, with $\epsilon_\beta,\epsilon_\alpha \in \mathbb{R} > 0$, which include $x(t)$.
The proximity indicator functions define an $\epsilon$-range, which moves the threshold for the transition in the corresponding direction (see Fig.\[fig:proximity\_hysteron\]). However, the transition between states does not happen as soon as the shifted threshold is reached, but only when the extremum within the connected set of points within the $\epsilon$-range is reached. An exception occurs, when the actual threshold $\alpha$ or $\beta$ is reached. In this case, the switch is done regardless of whether this is the extremum within this set.
![Switches of the proximity hysteron[]{data-label="fig:proximity_hysteron"}](proximityHysteronSVG){width="\linewidth"}
Due to the introduction of the proximity hysteron, state transitions can even be localized under harsh conditions – if the threshold points do not occur in the measurement data. The connection between signal, threshold values, indicator functions and hysterons is visualized in Fig.\[fig:buildingOfHysterons\].
![Interaction between signal, threshold values, indicator functions and hysterons[]{data-label="fig:buildingOfHysterons"}](simSignal "fig:"){width="\linewidth"}\
![Interaction between signal, threshold values, indicator functions and hysterons[]{data-label="fig:buildingOfHysterons"}](simIndicatorsSVG "fig:"){width="\linewidth"}\
![Interaction between signal, threshold values, indicator functions and hysterons[]{data-label="fig:buildingOfHysterons"}](simHysteronSVG.pdf "fig:"){width="\linewidth"}
System Identification with SINDy
--------------------------------
Once the hysterons are built and other preprocessing measures have been taken, the SINDy algorithm can start. The schematics are shown in \[fig:sindyFlow\]. Creating the data matrices includes choosing the right state vector $X$. The past values $H(k-1)_i$ for $i = 1, ..., m$ (see Eq. ) of all $m$ hysterons have to be included, holding information about the system’s current discrete state. Past measurement values $x(k-1), ..., x(k-q) $ up to a defined horizon $q$ can also be considered, allowing a representation resembling an FIR-filter. The state at time $k$ can be expressed by
$$\begin{aligned}
\begin{split}
X(k) =& [x_1(k), ... , x_n(k), H_1(k-1), \bar{H}_1(k-1), ...
\\& \bar{H}_m(k-1), x_1(k-1), ... , \bar{H}_m(k-q)]
\end{split}\end{aligned}$$
![The SINDy algorithm with tailored basis functions for identification of hysteretic systems[]{data-label="fig:sindyFlow"}](sindyFlowSVG){width="0.6\linewidth"}
As for the design of the library $\Theta$, polynomial basis functions are useful and easy to include. The function $\varphi_{poly}(x_1, x_2, ..., x_n)$ evaluates all polynomials between $x_1$ to $x_n$ up to a predefined polynomial degree. Hysterons are evolved by $\varphi_{relay}(x_1, x_2, ..., x_n, H_i)$. After evaluating all basis functions $\phi$, all basis functions that are unaffected by hysterons are multiplied with the updated hysterons. The finished library $\Theta$ consists of three parts: One without hysterons, the second with cross terms of hysterons and non hysteretic terms and the third of only the updated hysterons. Once the library is evaluated, a sequential least squares regression with a tuning parameter $\lambda$ is used to solve $ X(k+1) = \Theta(X(k)) \Xi$ for the sparse coefficients $\Xi$, as proposed in[@BruntonSINDy.2016].
For a library consisting of only polynomial basis functions and tailored basis functions for the propagation of hysterons, the evaluation and calculation of the next step is shown in Fig. \[fig:sindyLib\].
![Building and evaluation of the library[]{data-label="fig:sindyLib"}](bibGrafikSVG){width="\linewidth"}
Practical evaluation and results {#sec:practicalEvaluation}
================================
In order to evaluate the presented concept, we used different model parameters and differing data quality to identify a simple system. Our aim was to determine under which conditions the proposed identification procedure succeeds in finding a suited model, which parameters influence model quality, and where the limitations lay. The influence of the size of the library is studied in the first experiment by varying the degree of polynomials used. The robustness to sample rate and measurement noise are further aspects under test.
Example: Hysteresis-controlled tank system
------------------------------------------
The level of the basin in Fig. \[fig:Wassertank\] is controlled to a certain height by a *two-point controller* with an upper set point $h_{max}$ and a lower set point $h_{min}$. A leak causes a constant drain $Q_{out} \in \mathbb{R} < 0$, which neglects the influence of the current water level. The controller with control signal $u(t)$ turns the pump on whenever $h_{min}$ is reached, resulting in a steady inflow $Q_{in} \in \mathbb{R} > 0$. The pump is turned off when the level raises over $h_{max}$.
\[tank\_model\]
The length of the supply pipe $l_P$ can be used to analyze the effect of delay.
The simulation is implemented in Simulink^^. For this example, 20 different settings were processed, whereof 16 served as training- and four as validation set. The dynamics of the tank system is given in Eq. . $$\begin{aligned}
\begin{split}
\label{eq:tank_dynamics}
h(t+1) = h(t) + u(t-l_p) + Q_{out} \text{, with } \\
u(t) = \begin{cases}
Q_{in}, & \text{if pump on.} \\
0, & \text{if pump off.}
\end{cases}
\end{split}\end{aligned}$$ As a result, the pump works in rectangular pulses and the water level changes in a saw-tooth curve.
Influence of polynomial degree
------------------------------
The library prepared for SINDy consists of polynomials and hysterons. In a first experiment, we vary the degree of the polynomials. Figure \[fig:exp2\_sim\_h\] contains the simulated prediction of the filling level. Apparently, polynomials of low degree lead to the best results. For polynomials of degree one, the identified model is shown in Eq. . It consists of 4 out of 55 functions of the library, meaning a high degree of sparsity. Due to the hybrid form of hysteron $H_1$, it has the same structure as the correct model in Eq. . The models of higher degree manage to reproduce state transitions at the correct critical points, but they induce an inaccuracy for the slope.\
$$\begin{aligned}
\tilde{h}_{poly1}(k+1) &= h(k) + 1.56 \times 10^{-4}(4.98 Q_{in} (1 - H_1) + Q_{out} ) \label{eq:poly1_model}, \nonumber \\
H_{1}(k) &= \begin{cases}
1, & \text{if } h > h_{max}. \\
0, & \text{if } h < h_{min}. \\
H_{1}(k-1), & \text{else}.
\end{cases} \end{aligned}$$
{width="\linewidth"} \[fig:exp2\_sim\_h\]
Influence of sample rate
------------------------
As a further objective, we analyzed the effect of the sampling rate. As explained in Part \[sec:proximity\], this method cannot localize transitions if the critical points are not included in the given data, which may happen for slow sampling rates. Figure \[fig:exp3\_normal\_sim\_h\] shows, that state transitions are not detected using standard hysterons for slower sampling than $0.02$ s.
{width="\linewidth"} \[fig:exp3\_normal\_sim\_h\]
As a remedy, we presented the proximity hysteron in Section \[sec:proximity\]. The effect can be seen in Fig. \[fig:exp3\_proxi\_sim\_h\]. With this extension, the algorithm achieves good results for all sampling rates. The best model is shown in Eq. and was built from a library of 23 basis functions. It is equivalent to the model in Eq. , as $1 - H_1 = \bar{H}_1$.\
$$\begin{aligned}
\tilde{h}_{prox, 0.01}(k+1) &= h(k) + 1.56 \times 10^{-4}(4.98 Q_{in} \ \bar{H}_1 + Q_{out} ) \label{eq:snr_model}, \nonumber \\
H_{1}(k) &= \begin{cases}
1, & \text{if } h > h_{max}. \\
0, & \text{if } h < h_{min}. \\
H_{1}(k-1), & \text{else}.
\end{cases} \end{aligned}$$
{width="\linewidth"} \[fig:exp3\_proxi\_sim\_h\]
Effect of noisy data
--------------------
Beside an insufficient sampling rate, noise is a typical problem for data-driven approaches. In this experiment we investigate its influence with the SNR values 1000, 100, 50 and 10. The results are depicted in Fig.\[fig:exp4\_sim\_h\].
{width="\linewidth"} \[fig:exp4\_sim\_h\]
It can be seen that noise influences the identifcation quality very harshly. The approach works properly for an SNR of 1000. For an SNR of 100, the slope of the dynamics is not modeled accurately, resulting in a phase shift. The model for SNR 50 apparently identified two discrete states, but the state transitions as well as the dynamics within the states are faulty. For SNR 10, the measurements are too noisy to identify discrete states. As such, the model tries to satisfy all measurements by converging to the average.
Discussion {#sec:discussion}
==========
The results show, that SINDy with taylored basis functions in form of hysterons (SINDyHybrid) is a suitable method for the identification of hysteresis-controlled systems. The continuous system dynamics and the discrete states do not have do be modeled separately. With given signals and guesses for the thresholds, the hysteresis relation is found automatically. Adequate models were identified under varying sample rate and signal noise ratio. Designing a suited library remains the main challenge for the user. It was shown that systemic knowledge can be included in the function library, and that specific adaptations, e.g. with regard to poor data quality, can be integrated. The library should not be larger than necessary, as this complicates the regression and hinders identification of the correct basis functions. The resulting models are sparse and thus easily interpretable. This enables testing for plausibility, which is required for many industrial applications. However, this method may show problems for systems with many discrete states. For each state, a hysteron will be identified, which is multiplied with functions describing the dynamics for that particular state. This results in a solution with a high number of overall terms, which is not according to the idea behind SINDy that the optimal solution is sparse.
It must be noted, that creating the hysteron requires knowledge about the actual threshold value of the state transition. For two-point controlled systems, those threshold values can usually be read from data. The $\epsilon$-range around the threshold values used by the proximity hysteron can compensate inaccuracies of the threshold values, and this way even compensate aging effects.
Conclusion {#sec:conclusion}
==========
The presented work proposes a novel data-driven approach for identification of hybrid systems, SINDyHybrid. This method is exemplarily applied to a hysteresis-controlled system. The flexibility of the SINDy framework enables the integration of tailored basis functions, which can incorporate domain-specific knowledge. Based upon the relay-operator of the Preisach model, we developed the proximity hysteron, which allows for robust identification of state transitions and dynamics.
[^1]: The presented work was partially performed in the project *Reinforcement Learning for complex automation technology*, supported by Investitionsbank Berlin (IBB), co-financed by the European Regional Development Fund.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Continuing the project described by Kato et al. (2009, PASJ, 61, S395), we collected times of superhump maxima for 102 SU UMa-type dwarf novae observed mainly during the 2017 season and characterized these objects. WZ Sge-type stars identified in this study are PT And, ASASSN-17ei, ASASSN-17el, ASASSN-17es, ASASSN-17fn, ASASSN-17fz, ASASSN-17hw, ASASSN-17kd, ASASSN-17la, PNV J20205397$+$2508145 and TCP J00332502$-$3518565. We obtained new mass ratios for 7 objects using growing superhumps (stage A). ASASSN-17gf is an EI Psc-type object below the period minimum. CRTS J080941.3$+$171528 and DDE 51 are objects in the period gap and both showed long-lasting phase of stage A superhumps. We also summarized the recent advances in understanding of SU UMa-type and WZ Sge-type dwarf novae.'
author:
- 'Taichi <span style="font-variant:small-caps;">Kato</span>, Keisuke <span style="font-variant:small-caps;">Isogai</span>,$^,$ Yasuyuki <span style="font-variant:small-caps;">Wakamatsu</span>, Franz-Josef <span style="font-variant:small-caps;">Hambsch</span>,$^,$$^,$ Hiroshi <span style="font-variant:small-caps;">Itoh</span>, Tamás <span style="font-variant:small-caps;">Tordai</span>, Tonny <span style="font-variant:small-caps;">Vanmunster</span>, Pavol A. <span style="font-variant:small-caps;">Dubovsky</span>, Igor <span style="font-variant:small-caps;">Kudzej</span>, Tomáš <span style="font-variant:small-caps;">Medulka</span>, Mariko <span style="font-variant:small-caps;">Kimura</span>, Ryuhei <span style="font-variant:small-caps;">Ohnishi</span>, Berto <span style="font-variant:small-caps;">Monard</span>,$^,$ Elena P. <span style="font-variant:small-caps;">Pavlenko</span>, Kirill A. <span style="font-variant:small-caps;">Antonyuk</span>, Nikolaj V. <span style="font-variant:small-caps;">Pit</span>, Oksana I. <span style="font-variant:small-caps;">Antonyuk</span>, Julia V. <span style="font-variant:small-caps;">Babina</span>, Aleksei V. <span style="font-variant:small-caps;">Baklanov</span>, Aleksei A. <span style="font-variant:small-caps;">Sosnovskij</span>, Roger D. <span style="font-variant:small-caps;">Pickard</span>,$^,$ Ian <span style="font-variant:small-caps;">Miller</span>, Yutaka <span style="font-variant:small-caps;">Maeda</span>, Enrique <span style="font-variant:small-caps;">de Miguel</span>,$^,$ Stephen M. <span style="font-variant:small-caps;">Brincat</span>, Domenico <span style="font-variant:small-caps;">Licchelli</span>,$^,$ Lewis M. <span style="font-variant:small-caps;">Cook</span>, Sergey Yu. <span style="font-variant:small-caps;">Shugarov</span>,$^,$ Anna M. <span style="font-variant:small-caps;">Zaostrojnykh</span>, Drahomir <span style="font-variant:small-caps;">Chochol</span>, Polina <span style="font-variant:small-caps;">Golysheva</span>, Natalia <span style="font-variant:small-caps;">Katysheva</span>, Alexandra M. <span style="font-variant:small-caps;">Zubareva</span>,$^,$ Geoff <span style="font-variant:small-caps;">Stone</span>, Kiyoshi <span style="font-variant:small-caps;">Kasai</span>, Peter <span style="font-variant:small-caps;">Starr</span>, Colin <span style="font-variant:small-caps;">Littlefield</span>, Seiichiro <span style="font-variant:small-caps;">Kiyota</span>, Maksim V. <span style="font-variant:small-caps;">Andreev</span>,$^,$ Alexandr V. <span style="font-variant:small-caps;">Sergeev</span>,$^,$ Javier <span style="font-variant:small-caps;">Ruiz</span>,$^,$$^,$ Gordon <span style="font-variant:small-caps;">Myers</span>, Andrii O. <span style="font-variant:small-caps;">Simon</span>, Volodymyr V. <span style="font-variant:small-caps;">Vasylenko</span>, Francisco <span style="font-variant:small-caps;">Soldán</span>, Yenal <span style="font-variant:small-caps;">Ögmen</span>, Kazuhiro <span style="font-variant:small-caps;">Nakajima</span>, Peter <span style="font-variant:small-caps;">Nelson</span>, Gianluca <span style="font-variant:small-caps;">Masi</span>, Kenneth <span style="font-variant:small-caps;">Menzies</span>, Richard <span style="font-variant:small-caps;">Sabo</span>, Greg <span style="font-variant:small-caps;">Bolt</span>, Shawn <span style="font-variant:small-caps;">Dvorak</span>, Krzysztof Z. <span style="font-variant:small-caps;">Stanek</span>, Joseph V. <span style="font-variant:small-caps;">Shields</span>, Christopher S. <span style="font-variant:small-caps;">Kochanek</span>, Thomas W.-S. <span style="font-variant:small-caps;">Holoien</span>, Benjamin <span style="font-variant:small-caps;">Shappee</span>, José L. <span style="font-variant:small-caps;">Prieto</span>,$^,$ Tadashi <span style="font-variant:small-caps;">Kojima</span>, Hideo <span style="font-variant:small-caps;">Nishimura</span>, Shizuo <span style="font-variant:small-caps;">Kaneko</span>, Shigehisa <span style="font-variant:small-caps;">Fujikawa</span>, Rod <span style="font-variant:small-caps;">Stubbings</span>, Eddy <span style="font-variant:small-caps;">Muyllaert</span>, Gary <span style="font-variant:small-caps;">Poyner</span>, Masayuki <span style="font-variant:small-caps;">Moriyama</span>, Hiroyuki <span style="font-variant:small-caps;">Maehara</span>, Patrick <span style="font-variant:small-caps;">Schmeer</span>, Denis <span style="font-variant:small-caps;">Denisenko</span>,'
title: 'Survey of Period Variations of Superhumps in SU UMa-Type Dwarf Novae. X: The Tenth Year (2017)'
---
Introduction
============
This is a continuation of series of papers @Pdot, @Pdot2, @Pdot3, @Pdot4, @Pdot5, @Pdot6, @Pdot7, @Pdot8 and @Pdot9 reporting new observations of superhumps in SU UMa-type dwarf novae. \[see e.g. @war95book for SU UMa-type dwarf novae and CVs in general\].
Upon recommendation from the previous reviewer and the PASJ office, we provide the result in a concise form: presenting the results in the Supporting Information (SI) and only the list the objects (table \[tab:outobs\]), the obtained parameters (table \[tab:perlist\]) and the references (section \[sec:list\]) are given in the main paper. For the details of the analysis, terminology and definitions see @Pdot and for the initial and current aims of this survey, see @Pdot and @Pdot9, respectively. For superhump stages, see @Pdot and a concise version in e-section 1 in SI. A short description of the data analysis is given in e-section 2 in SI. In table \[tab:perlist\], $P_1$ and $P_2$ represent periods in stage B and C, respectively ($P_1$ is averaged during the entire course of the observed segment of stage B), and $E_1$ and $E_2$ represent intervals (in cycle numbers) to determine $P_1$ and $P_2$, respectively.
Data source {#sec:data}
===========
The CCD time-series observations were were obtained under campaigns led by the VSNET Collaboration [@VSNET]. We also used the public data from the AAVSO International Database[^1].
Outburst detections of many new and known objects relied on the ASAS-SN CV patrol [@dav15ASASSNCVAAS][^2], the MASTER network [@MASTER], and Catalina Real-time Transient Survey (CRTS; [@CRTS])[^3]. There were also outburst detections reported to VSNET, AAVSO[^4], BAAVSS alert[^5] and cvnet-outburst.[^6]
[cccl]{} Object & Year & Observers or references&\
PT And & 2017 & DPV, Ioh &\
DH Aql & 2017 & Ioh &\
V1047 Aql & 2017 & SGE, deM, Trt &\
NN Cam & 2017 & DPV &\
V391 Cam & 2017 & @Pdot9 &\
KP Cas & 2017 & Ioh &\
VW CrB & 2017 & COO, Kis &\
V503 Cyg & 2017 & IMi &\
V632 Cyg & 2017 & deM &\
GP CVn & 2017 & Trt &\
GQ CVn & 2017 & deM, Mdy, COO &\
HO Del & 2017 & BSM &\
MN Dra & 2017 & COO &\
OV Dra & 2015 & Trt &\
& 2017 & DPV, Lis, Ioh, Trt &\
V1454 Cyg & 2006 & @Pdot &\
BE Oct & 2017 & HaC &\
V521 Peg & 2017 & Ioh, BSM, RPc &\
V368 Per & 2017 & CRI, Ter, Ioh, Trt, IMi, RPc, DPV &\
XY Psc & 2017 & KU, HaC, Ioh &\
V701 Tau & 2017 & BSM &\
V1208 Tau & 2017 & MZK, CRI, Trt &\
TU Tri & 2017 & Trt, Ioh, RPc, DPV &\
SU UMa & 2017 & DPV &\
\
\
\
Object Year Observers or references
-------------- ------ ---------------------------------------- -------------------------
HS Vir 2007 Njh
2017 HaC, DPV, Mdy
V406 Vir 2017 MLF, MGW, HaC, Nel
NSV 35 2017 MGW, HaC
1RXS J161659 2017 deM, IMi 1RXS J161659.5$+$620014
ASASSN-13ce 2017 Van, Ioh
ASASSN-13dh 2017 SGE, DPV, IMi, BSM
ASASSN-14ca 2017 Ter, Trt, Lis, DPV
ASASSN-14cr 2017 DPV
ASASSN-14kb 2017 HaC
ASASSN-14lk 2017 MLF
ASASSN-15fu 2017 HaC
ASASSN-15fv 2017 Van
ASASSN-15qu 2017 MLF, HaC
ASASSN-17ei 2017 MLF, HaC, SPE
ASASSN-17el 2017 MLF, HaC
ASASSN-17eq 2017 Van, Ioh
ASASSN-17es 2017 HaC, Van, Ioh
ASASSN-17et 2017 MLF, HaC
ASASSN-17ew 2017 HaC
ASASSN-17ex 2017 HaC
ASASSN-17fh 2017 Van
ASASSN-17fi 2017 Van
ASASSN-17fj 2017 HaC
ASASSN-17fl 2017 HaC
ASASSN-17fn 2017 Van, Ioh, DPV, Trt, Mdy, Shu, Lic, CRI
ASASSN-17fo 2017 Mdy, Kis, HaC, Lic, COO, RPc, Ioh, CRI
ASASSN-17fp 2017 MLF, HaC
ASASSN-17fz 2017 MLF, HaC, SPE
ASASSN-17gf 2017 MLF, HaC
ASASSN-17gh 2017 Ioh, Van
ASASSN-17gv 2017 MLF, HaC
Object Year Observers or references
-------------- ------ ----------------------------------- -------------------------
ASASSN-17hm 2017 HaC
ASASSN-17hw 2017 MLF, HaC, BSM, Ioh, SPE, Shu, Van
ASASSN-17hy 2017 HaC
ASASSN-17id 2017 HaC
ASASSN-17if 2017 HaC
ASASSN-17ig 2017 GBo, HaC
ASASSN-17il 2017 Van
ASASSN-17iv 2017 HaC
ASASSN-17iw 2017 HaC
ASASSN-17ix 2017 HaC
ASASSN-17ji 2017 IMi, Trt, RPc
ASASSN-17jr 2017 HaC
ASASSN-17kc 2017 HaC
ASASSN-17kd 2017 HaC
ASASSN-17kg 2017 HaC, RPc, Van, Trt
ASASSN-17kp 2017 Trt, Van, RPc
ASASSN-17la 2017 COO, Van, DPV, IMi, Trt, NKa, KU
ASASSN-17lr 2017 IMi
ASASSN-17me 2017 LCO, CRI
ASASSN-17np 2017 MLF, HaC
ASASSN-17nr 2017 HaC
ASASSN-17of 2017 Van, Ioh, KU, IMi, CRI
ASASSN-17oo 2017 KU, HaC
ASASSN-17ou 2017 Shu, KU, HaC, Trt
ASASSN-17pb 2017 Van, CRI, IMi, KU
CRTS J044027 2017 HaC, Van CRTS J044027.1$+$023301
CRTS J080941 2017 Van, HaC, CRI, Trt CRTS J080941.3$+$171528
CRTS J120052 2017 Mdy CRTS J120052.9$-$152620
CRTS J122221 2017 @neu17j1222 CRTS J122221.6$-$311524
CRTS J162806 2017 Trt CRTS J162806.2$+$065316
CRTS J214934 2017 HaC, Ioh CRTS J214934.1$-$121908
Object Year Observers or references
---------------- ------- ------------------------------------------- ---------------------------------
CRTS J223235 2017 IMi, Van CRTS J223235.4$+$304105
CTCV J1940 2017 HaC CTCV J1940$-$4724
DDE 51 2017 Mdy, Trt, RPc, Rui, CRI, IMi
MASTER J132501 2017 Kai, Lic, deM, Van MASTER OT J132501.00$+$431846.1
MASTER J174305 2017 Mdy, Kai, DPV, Lic, Trt MASTER OT J174305.70$+$231107.8
MASTER J192757 2017 Van MASTER OT J192757.03$+$404042.8
MASTER J200904 2017 KU, deM, Lic MASTER OT J200904.69$+$825153.6
MASTER J205110 2017 LCO, Lic, Ioh, KU, Trt MASTER OT J205110.36$+$044842.2
MASTER J212624 2017 Shu, DPV, BSM, Trt, RPc, Ioh MASTER OT J212624.16+253827.2
OT J182142 2017 DPV, Ioh OT J182142.8$+$212154
OT J204222 2017 Ioh, LCO, Mas, RPc, Trt, Mdy OT J204222.3$+$271211
PNV J202053 2017 deM, CRI, AAVSO, COO, SGE, Lic, Ioh, Trt, PNV J20205397$+$2508145
Van, OYE, Sol, DPV, Rui, RPc, Kis
SDSS J152857 2017 Van SDSS J152857.86$+$034911.7
SDSS 153015 2017b KU, Trt, CRI SDSS J153015.04$+$094946.3
SDSS J204817 2017 BSM, Ioh SDSS J204817.85$-$061044.8
TCP J003325 2017 MLF, HaC TCP J00332502$-$3518565
TCP J201005 2017 Van, Kai, deM, HaC, SRI, SGE, Trt, DKS, TCP J20100517$+$1303006
Ioh, Kai, BSM, DPV
[c@c@c@c@c@c@c@c@c@c@c@c@c@c]{} Object & Year & $P_1$ (d) & err & & $P_{\rm dot}$& err& $P_2$ (d) & err & & $P_{\rm orb}$ (d)& Q\
V1047 Aql & 2017 & 0.073914 & 0.000098 & 0 & 19 & – & – & – & – & – & – & – & C\
V391 Cam & 2017 & – & – & – & – & – & – & 0.056728 & 0.000012 & 209 & 263 & 0.05620 & C\
KP Cas & 2017 & – & – & – & – & – & – & 0.085143 & 0.000242 & 0 & 13 & – & C\
VW CrB & 2017 & 0.071985 & 0.000528 & 0 & 11 & – & – & – & – & – & – & – & C\
V632 Cyg & 2017 & 0.0655 & 0.0003 & 0 & 2 & – & – & – & – & – & – & – & C\
OV Dra & 2017 & 0.060398 & 0.000033 & 0 & 98 & 14.5 & 2.4 & 0.060032 & 0.000057 & 94 & 150 & 0.058736 & B\
GQ CVn & 2017 & 0.089476 & 0.000091 & 0 & 37 & – & – & – & – & – & – & – & C\
BE Oct & 2017 & 0.077115 & 0.000132 & 0 & 40 & – & – & – & – & – & – & – & C\
V521 Peg & 2017 & 0.061646 & 0.000065 & 0 & 29 & – & – & – & – & – & – & – & C\
V368 Per & 2017 & 0.079224 & 0.000028 & 0 & 41 & – & – & 0.078602 & 0.000166 & 63 & 79 & – & B\
XY Psc & 2017 & 0.060675 & 0.000045 & 0 & 83 & 13.7 & 2.3 & 0.060230 & 0.000053 & 82 & 99 & – & C\
V701 Tau & 2017 & 0.069026 & 0.000037 & 0 & 31 & – & – & – & – & – & – & – & C\
V1208 Tau & 2017 & 0.0698 & 0.0040 & 0 & 3 & – & – & – & – & – & – & 0.0681 & C\
TU Tri & 2017 & 0.076246 & 0.000080 & 0 & 20 & – & – & – & – & – & – & – & C\
SU UMa & 2017b & 0.078924 & 0.000123 & 0 & 64 & – & – & – & – & – & – & 0.07635 & C\
HS Vir & 2017 & 0.080313 & 0.000063 & 0 & 103 & 3.7 & 4.9 & – & – & – & – & 0.0769 & CG\
V406 Vir & 2017 & 0.056960 & 0.000016 & 0 & 88 & 8.1 & 1.5 & – & – & – & – & 0.05592 & B\
1RXS J161659 & 2017 & 0.071028 & 0.000032 & 0 & 70 & $-$10.6 & 3.3 & – & – & – & – & – & CG\
ASASSN-13dh & 2017 & – & – & – & – & – & – & 0.091322 & 0.000056 & 38 & 100 & – & B\
\
\
\
\
Object Year $P_1$ err $P_{\rm dot}$ err $P_2$ err $P_{\rm orb}$ Q
------------- ------ ---------- ---------- ---- --------------- -------- ------- ---------- ---------- --------------- ----- ---------- ----
ASASSN-14ca 2017 0.067036 0.000014 0 45 $-$4.7 3.0 – – – – – C
ASASSN-14cr 2017 – – – – – – 0.068698 0.000055 0 45 – C
ASASSN-14kb 2017 0.070420 0.000030 0 86 3.0 3.9 – – – – 0.068106 CG
ASASSN-14lk 2017 – – – – – – 0.061054 0.000094 0 34 – C
ASASSN-15fu 2017 0.074592 0.000071 0 28 – – – – – – – CG
ASASSN-15fv 2017 0.0682 0.0040 0 1 – – – – – – – C
ASASSN-15qu 2017 0.080449 0.000038 0 78 $-$7.5 3.8 – – – – – CG
ASASSN-17ei 2017 0.057257 0.000011 34 247 3.4 0.4 – – – – 0.05646 BE
ASASSN-17el 2017 0.055183 0.000013 48 213 5.1 0.3 0.054911 0.000184 230 271 0.05434 BE
ASASSN-17eq 2017 0.072197 0.000069 0 28 – – – – – – – C
ASASSN-17es 2017 0.057858 0.000023 33 105 0.6 4.4 – – – – 0.05719 BE
ASASSN-17et 2017 – – – – – – 0.095636 0.000060 0 63 – C
ASASSN-17ew 2017 – – – – – – 0.078497 0.000027 0 65 – C
ASASSN-17ex 2017 – – – – – – 0.068306 0.000096 0 31 – C
ASASSN-17fh 2017 0.064 0.001 0 1 – – – – – – – C
ASASSN-17fi 2017 0.058833 0.000011 0 52 – – – – – – – C
ASASSN-17fj 2017 0.066266 0.000021 0 77 8.4 2.3 0.065950 0.000044 75 135 – B
ASASSN-17fl 2017 0.062632 0.000123 0 18 – – – – – – – C
ASASSN-17fn 2017 0.061584 0.000014 37 169 $-$2.8 1.4 – – – – 0.06096 BE
ASASSN-17fo 2017 0.063240 0.000028 8 80 7.3 3.8 – – – – 0.061548 B
ASASSN-17fz 2017 0.054404 0.000025 41 152 7.0 2.0 – – – – – B
ASASSN-17gf 2017 0.052551 0.000010 31 129 5.2 1.0 – – – – – B
ASASSN-17gh 2017 0.061394 0.000348 0 9 – – – – – – – C
ASASSN-17gv 2017 0.060897 0.000039 0 88 – – – – – – – CG
ASASSN-17hm 2017 0.088586 0.000073 0 37 – – 0.088140 0.000059 34 59 – C
ASASSN-17hw 2017 0.059717 0.000013 29 218 0.3 0.9 – – – – 0.05886 BE
ASASSN-17hy 2017 0.071475 0.000048 0 72 16.3 4.3 – – – – – C
ASASSN-17id 2017 0.078613 0.000074 0 39 – – – – – – – C2
ASASSN-17if 2017 0.058827 0.000031 0 154 8.2 0.8 0.058568 0.000041 153 223 – B
Object Year $P_1$ err $P_{\rm dot}$ err $P_2$ err $P_{\rm orb}$ Q
---------------- ------- ---------- ---------- ---- --------------- -------- ------- ---------- ---------- --------------- ----- ---------- ----
ASASSN-17ig 2017 0.094947 0.000084 0 25 – – 0.094393 0.000024 25 96 – C
ASASSN-17iv 2017 – – – – – – 0.070237 0.000044 15 87 – C
ASASSN-17iw 2017 0.055906 0.000047 0 90 11.3 5.9 – – – – – C
ASASSN-17ix 2017 0.062449 0.000048 0 82 18.2 4.0 – – – – – C
ASASSN-17ji 2017 0.0589 0.0001 0 18 – – – – – – – C
ASASSN-17jr 2017 0.061706 0.000038 0 98 8.0 3.0 – – – – – C
ASASSN-17kc 2017 0.063764 0.000028 0 81 12.8 1.4 0.063320 0.000024 80 160 – B
ASASSN-17kd 2017 0.060919 0.000016 33 213 2.7 0.8 – – – – – B
ASASSN-17kg 2017 0.057620 0.000017 36 228 5.4 0.5 0.057427 0.000025 242 297 – A
ASASSN-17kp 2017 0.057957 0.000030 0 51 9.3 5.7 – – – – – C
ASASSN-17la 2017 0.061571 0.000021 27 175 7.9 0.5 – – – – 0.06039 BE
ASASSN-17lr 2017 0.058635 0.000057 0 102 $-$8.8 3.2 – – – – – CG
ASASSN-17me 2017 0.0614 0.0004 0 1 – – – – – – – C
ASASSN-17np 2017 0.089227 0.000047 0 26 – – 0.088730 0.000032 25 82 – C
ASASSN-17nr 2017 0.056376 0.000027 0 107 5.8 1.6 – – – – – CU
ASASSN-17of 2017 0.064175 0.000067 0 74 – – 0.063567 0.000030 74 109 – C
ASASSN-17oo 2017 0.06781 0.00005 – – – – – – – – – C2
ASASSN-17ou 2017 0.057128 0.000045 0 70 – – – – – – – C
ASASSN-17pb 2017 0.076092 0.000049 47 101 $-$1.0 8.6 – – – – – C
CRTS J044027 2017 – – – – – – 0.064361 0.000034 49 97 – C
CRTS J080941 2017 0.100467 0.000122 20 62 – – – – – – – B
CRTS J214934 2017 0.071482 0.000005 0 65 – – 0.071222 0.000041 64 107 – C
CRTS J223235 2017 0.062994 0.000136 0 32 – – – – – – – C
CTCV J1940 2017 0.076668 0.000027 0 79 $-$3.7 3.2 – – – – – CU
DDE 51 2017 0.100277 0.000020 49 108 $-$0.5 2.1 – – – – – B
MASTER J174305 2017 0.069949 0.000079 0 14 – – 0.069425 0.000074 27 44 – C
MASTER J192757 2017 0.08161 0.00005 0 12 – – – – – – – C
MASTER J200904 2017 0.073646 0.000115 0 20 – – – – – – – C
MASTER J205110 2017 0.080710 0.000044 0 59 7.6 4.7 – – – – – C
MASTER J212624 2017 0.090888 0.000074 43 75 – – – – – – – B
NSV 35 2017 0.081034 0.000039 0 112 $-$1.1 2.4 – – – – – BG
OT J182142 2017 0.082140 0.000095 0 40 – – – – – – – C2
OT J204222 2017 0.056152 0.000045 65 167 $-$1.1 6.6 – – – – – C
PNV J202053 2017 0.057392 0.000010 53 250 4.3 0.4 0.056443 0.000153 246 263 0.056509 AE
SDSS J152857 2017 0.06319 0.00024 0 3 – – – – – – – C
SDSS J153015 2017b 0.075310 0.000134 0 32 – – – – – – – C
TCP J003325 2017 0.055222 0.000019 91 256 4.6 0.3 – – – – 0.05485 BE
TCP J201005 2017 0.081030 0.000046 0 44 – – – – – – – B2
Major findings in objects in this paper {#sec:findings}
=======================================
We list major findings in this paper.
- Suspected WZ Sge-type dwarf novae XY Psc and V406 Vir underwent long-awaited superoutbursts, but neither of them showed WZ Sge-type characteristics.
- ASASSN-17fo is a deeply eclipsing SU UMa-type dwarf nova.
- ASASSN-17gf is an EI Psc-type object below the period minimum.
- ASASSN-17kg showed a dip before the termination of the superoutburst.
- ASASSN-17la is a WZ Sge-type dwarf nova with an intermediate mass ratio \[0.084(5)\] and a medium long orbital period \[0.06039(3) d\].
- CRTS J080941 and DDE 51 are in the period gap and had a long-lasting stage A.
- MASTER J212624 is a long-period system with a long-lasting stage A.
- WZ Sge-type stars identified in this study are PT And, ASASSN-17ei, ASASSN-17el, ASASSN-17es, ASASSN-17fn, ASASSN-17fz, ASASSN-17hw, ASASSN-17kd, ASASSN-17la, PNV J202053 and TCP J003325.
- New mass ratios from stage A superhumps (using [@kat13qfromstageA]) are: ASASSN-17ei 0.074(3), ASASSN-17el 0.071(3), ASASSN-17es 0.095(9), ASASSN-17fn 0.097(1), ASASSN-17hw 0.078(1), CRTS J122221 0.032(2), PNV J202053 0.090(3)
Summary of recent progress in understanding of SU UMa-type dwarf novae {#sec:progress}
======================================================================
In this section, we provide brief descriptions of recent progress in understanding of SU UMa-type dwarf novae based on this series of papers and other published papers upon the request from the reviewer.
SU UMa-type dwarf novae and superhump stages
--------------------------------------------
For SU UMa-type dwarf novae in general, we have verified that the relation between the period derivative ($P_{\rm dot}$) for stage B versus the oribital period ($P_{\rm orb}$) that we found in @Pdot essentially applies to most of ordinary superoutbursts. The refined relation was shown in @Pdot8 and @Pdot9 \[we consider that @Pdot9 to be the final regular summary of the statistics\]. The stage A, B and C are now well-established and used in many publications by various authors: @kat14wzsgestarsproc, @bak14czev404, @skl16asassn14cv, @neu17j1222, @bak17mndra, @neu18j1222gwlib, @skl18nyser, @lit18j0359, @pal18qzlib, @pal19j1238, @pav19asassn18fk, @mca19DNeclipse, @cou19zcha (this work also illustrates the difficulty in determining superhump times in a deeply eclipsing system). The rapid growth of papers referring to our superhump stages indicates that this concept and application are now widely accepted in this field.
SU UMa-type/WZ Sge-type relation and period bouncers
----------------------------------------------------
The SU UMa-type/WZ Sge-type relation and the nature of period bouncers would be one of the most intriguing subjects for many readers. We have already give a conclusion to this subject as a review [@kat15wzsge]. The distinction between SU UMa-type and WZ Sge-type dwarf novae is the manifestaion of the 2:1 resonance in the latter, and this classification is now widely accepted (such as in AAVSO VSX[^7]). After the release of @kat15wzsge, there have been increasing number of WZ Sge-type dwarf novae mainly thanks to the ASAS-SN survey. The major advance since then has been the increase of examples of type-E outbursts. The objects with type-E outbursts have an initial superoutburst corresponding to the 2:1 resonance (high-inclination systems show early superhumps) and the second superoutburst showing the development of ordinary superhumps. They are considered to be the best candidates for the still elusive population of period bouncers. The papers dealing with type-E outbursts are @kim16asassn15jd (ASASSN-15jd), @kim18asassn16dtasassn16hg (ASASSN-16dt and ASASSN-16hg), @iso19nsv1440 (NSV 1440, AM CVn star). Among them ASASSN-15jd and ASASSN-16hg showed a transitional feature between single superoutburst and the type-E outburst. These observations suggest that type-E outbursts can be understood as a smooth extension of WZ Sge-type dwarf novae toward a lower mass ratio (i.e. period bouncers). The examples are still increasing and the results are pending publication.
Systems near stability border of 3:1 resonance
----------------------------------------------
The major recent advance in SU UMa-type dwarf novae is around the stability borderline of the 3:1 resonance. When @Pdot was published, it was a mystery why some long-$P_{\rm orb}$ systems show a strong decrease of the superhump periods \[cf. MN Dra and UV Gem, see subsection 4.10 in @Pdot\]. An idea to solve this issue required five years to appear and @Pdot6 gave a working hypothesis that the 3:1 resonance grows slowly in systems near the stability border of the 3:1 resonance. This idea has been reinforced by subsequent observations [@kat16v1006cyg]. @Pdot8 and @Pdot9 increased the number of candidate systems showing this feature. Some of these objects are known to show post-superoutburst rebrightenings, which had been usually considered to be a feature unique to WZ Sge-type dwarf novae (cf. [@kat15wzsge]). With the increasing number of long-$P_{\rm orb}$ object showing rebrightenings \[V1006 Cyg, @kat16v1006cyg; ASASSN-14ho @kat19asassn14ho\], it is now considered that the weak 3:1 resonance could cause the decoupling of the tidal and thermal instabilities, leading to premature quenching of the superoutburst. This idea was originally proposed for extremely low mass-ratio systems such as WZ Sge-type dwarf novae [@hel01eruma]. Recent findings suggest that the same mechanism could work in systems near the stability border of the 3:1 resonance and that such systems can mimic WZ Sge-type outbursts. A long precursor followed by a dip and an ordinary superoutburst in CS Ind [@kat19csind] also strengthens this interpretation. Theoretical supports are still lacking and a futher advance would be expected in this regime.
SU UMa-type dwarf nova showing standstills
------------------------------------------
Currently there is only one known SU UMa-type dwarf nova (NY Ser) which showed standstills in 2018 [@kat19nyser]. This is a single known bona-fide a hybrid SU UMa + Z Cam-type dwarf nova. It was shown that superoutbursts arose from standstills in NY Ser, and the disk should grow in radius to reach the 3:1 resonance during standstills.
List of references {#sec:list}
==================
The references cited in SI are: @alk00ptand, @ant96vwcrb, @ant02var73dra, @aug10CTCVCV2, @avi10j1238, @bal13j1616, @bal14j1927atel6024, @bal12j1743atel4022, @bal17j1325atel10470, @bal14j2009atel5974, @boy10kpcas, @can25dhaql, @car17asassn17fpatel10334, @LOWESS, @dav14asassn14caatel6211, @den17j1325atel10480, @den13j2126atel5111, @dil08SDSSCV, @dra14CRTSCVs, @era73v701tau, @fer89error, @gre82PGsurveyCV, @PGsurvey, @gru58ptand, @har95v503cyg, @hen01xypsc, @hof49newvar, @hof49newvar1, @hof57VSchart, @hof57MVS245, @hof63VSS61, @hof64an28849, @hof67an29043, @ima17qzvir, @kat02v503cyg, @kat15wzsge, @kat16j0333, @Pdot6, @Pdot7, @Pdot4, @Pdot5, @Pdot8, @Pdot, @Pdot9, @Pdot3, @kat12DNSDSS, @Pdot2, @kat13j1222, @kat98hsvir, @kat95hsvir, @kat16v1006cyg, @kat01hvvir, @kat01hsvir, @kat17j0026, @khr05nsv1485, @kim16alcom, @kin00VSID4896, @lit13sbs1108, @liu00CVspec3, @liu99CVspec1, @luy38propermotion2, @mar17asassn17fpatel10354, @mas03faintCV, @men99hsvir, @mot96CVROSAT, @mro15OGLEDNe, @nak13j2112j2037, @nam17asassn15po, @neu17j1222, @nog95dhaql, @nog03var73dra, @nov97vwcrb, @osa13v1504cygKepler, @osa13v344lyrv1504cyg, @ohn19ovboo, @ohs12eruma, @osm85hsvir, @pat05SH, @pat03suumas, @pat08j1507, @pav12v503cyg, @ASAS3, @pri14asassn14kbatel6688, @ric69v1233aql, @rin93thesis, @rod05hs2219, @rom78tutri, @ros72xypsciauc1, @ros72xypsciauc2, @sha91tutri, @sha89ptand, @sha92tutri, @she07v701tau, @she08j1227, @she07CVspec, @shu17j2051atel10790, @sta13asassn13aoatel5118, @PDM, @szk09SDSSCV7, @szk03SDSSCV2, @szk06SDSSCV5, @tho02j2329, @uem02j2329letter, @waa17asassn17fpaan580, @wak17asassn16eg, @wen89v632cygv630cyg, @wil10j1924, @woo11v344lyr, @wou10CVperiod, @wou12SDSSCRTSCVs, @wyr14asassn14kbatel6690, @zem13eruma, @zha06j1238, @zhe10ptandcbet2574, @zlo04tutri.
Acknowledgements {#acknowledgements .unnumbered}
================
This work was also partially supported by Grant VEGA 2/0008/17 (by Shugarov, Chochol) and APVV-15-0458 (by Shugarov, Chochol, Dubovsky, Kudzej, Medulka), RSF-14-12-00146 (Golysheva for processing observation data from Slovak Observatory). ASAS-SN is supported by the Gordon and Betty Moore Foundation through grant GBMF5490 to the Ohio State University and NSF grant AST-1515927. The authors are grateful to observers of VSNET Collaboration and VSOLJ observers who supplied vital data. We acknowledge with thanks the variable star observations from the AAVSO International Database contributed by observers worldwide and used in this research. We are also grateful to the VSOLJ database. This work is helped by outburst detections and announcement by a number of variable star observers worldwide, including participants of CVNET and BAA VSS alert. The CCD operation of the Bronberg Observatory is partly sponsored by the Center for Backyard Astrophysics. We are grateful to the Catalina Real-time Transient Survey team for making their real-time detection of transient objects and the past photometric database available to the public. We are also grateful to the ASAS-3 team for making the past photometric database available to the public. This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France. This research has made use of the International Variable Star Index (VSX) database, operated at AAVSO, Cambridge, Massachusetts, USA.
Supporting information {#supporting-information .unnumbered}
======================
For reader’s convenience, supporting information (sections, figures and tables starting with E-) is combined in this arXiv version.
In the final form of PASJ publication, supporting information is separated in the online version.
addtoreset[equation]{}[section]{}
Superhump Stages
================
It has become evident since @Pdot that the superhump periods systematically vary in a way common to many objects. @Pdot introduced superhump stages (stages A, B and C): initial growing stage with a long period (stage A) and fully developed stage with a systematically varying period (stage B) and later stage C with a shorter, almost constant period (see figure \[fig:stagerev\]). \[This part is an excerpt from @Pdot9\].
(80mm,110mm)[stagerev.eps]{}
Data Analysis
=============
This part includes an excerpt from @Pdot9.
The data analysis was performed in the same way described in @Pdot and @Pdot6 and we mainly used R software[^8] for data analysis.
In de-trending the data, we mainly used locally-weighted polynomial regression (LOWESS: [@LOWESS]) and sometimes lower (1–3rd order) polynomial fitting when the observation baseline was short. The times of superhumps maxima were determined by the template fitting method as described in @Pdot. The times of all observations are expressed in barycentric Julian days (BJD).
We used phase dispersion minimization (PDM; [@PDM]) for period analysis and 1$\sigma$ errors for the PDM analysis was estimated by the methods of @fer89error and @Pdot2. We have used a variety of bootstrapping in estimating the robustness of the result of the PDM analysis since @Pdot3.
We used @kat13qfromstageA to determine the mass ratio ($q$) from the superhump period ($P_{\rm SH}$) and the orbital period ($P_{\rm orb}$, when early superhumps were observed, we assumed them to have the same period as the orbital one). The fractional superhump excess (in frequency) $\epsilon^* \equiv 1-P_{\rm orb}/P_{\rm SH}$ is equal to the dynamical precession rate when the pressure effect can be neglected as in stage A superhumps [@kat13qfromstageA].
Individual Objects {#sec:individual}
==================
PT Andromedae {#obj:ptand}
-------------
PT And was originally discovered as a nova in M31 (R15 = M31N 1957-10b in [@gru58ptand]). @sha89ptand reported a short outburst in 1983 and a long one in 1986. @sha89ptand suggested this object to be an SU UMa-type dwarf nova as judged from the outburst behavior. @alk00ptand studied the 1998 outburst and past ones and suggested that this object is more likely a recurrent nova in M31 based on the lack of plateau phase in SU UMa-type superoutbursts.
There was a long outburst in 2010 December (originally reported as M31N 2010-12a on 2010 December 1 by K. Nishiyama and F. Kabashima, cf. [@zhe10ptandcbet2574]). During this outburst, likely superhumps were detected (vsnet-alert 12484, 12497, 12527).[^9] Due to the short observational runs and interference by the moonlight, it was difficult to determine the superhump period. A spectrum by A. Arai showed no prominent lines, confirming the dwarf nova-type nature of this object (vsnet-alert 12528). The outburst showed a plateau phase followed by rapid fading on 2010 December 24–25 (vsnet-alert 12530). This feature also supported the SU UMa-type interpretation contrary to what was stated in @alk00ptand.
The 2017 outburst was detected on 2017 August 15 at an unfiltered CCD magnitude of 16.02 by E. Muyllaert (cvnet-outburst 7623). Time-resolved CCD photometry recorded double-wave modulations (vsnet-alert 21356). They were most likely early superhumps. Due to the lack of observations, we could not select the alias. The two most likely periods were 0.06063(7) d or 0.05893(7) d. These periods are equally acceptable and phase-averaged profiles are in e-figures \[fig:ptandeshpdm\] and \[fig:ptandeshpdm2\]. Although there were observations after these two nights, the data quality was not sufficient and we could not determine the superhump period. The object is thus a WZ Sge-type dwarf nova. The rapidly fading light curves resembling those of fast novae were probably caused by viscous decay at the start of the superoutburst (cf. [@kat15wzsge]).
(85mm,110mm)[ptandeshpdm.eps]{}
(85mm,110mm)[ptandeshpdm2.eps]{}
DH Aquilae {#obj:dhaql}
----------
DH Aql was discovered as a Mira-type variable (=HV 3899) with a range of 12.5 to fainter than 16 in photographic range [@can25dhaql]. The SU UMa-type nature of this object was clarified by @nog95dhaql. Refer to @Pdot6 for more history.
The 2017 superoutburst was detected by M. Noriyama at an unfiltered CCD magnitude of 14.5 while the object was still rising. The object was observed to be at peak magnitude of 12.4 on the next night. Two superhump maxima were obtained from single-night observations: BJD 2458036.9216(4) ($N$=106) and 2458037.0013(2) ($N$=152).
V1047 Aquilae {#obj:v1047aql}
-------------
V1047 Aql was discovered as a dwarf nova (S 8191) by @hof64an28849. The object was identified to be an SU UMa-type dwarf nova in 2005 by Greg Bolt. Observations of the 2016 superoutburst were reported in @Pdot9, which obtained only two superhump maxima.
The 2017 superoutburst was visually detected by R. Stubbings at a magnitude of 15.2 on 2017 June 26. The ASAS-SN data recorded it at $V$=16.2 on 2017 June 20, which further brightened to $V$=15.5 on 2017 June 24. The times of superhump maxima are listed in e-table \[tab:v1047aqloc2017\]. The superhump stage is unknown.
As reported in @Pdot9, this object shows regular superoutbursts with a short supercycle. We extracted superoutbursts in the ASAS-SN data (e-table \[tab:v1047aqlout\]). These superoutbursts can be well expressed by a supercycle of 89.1(3) d with maximum $|O-C|$ of 4 d. There were also apparently frequent normal outbursts expected for this short supercycle.
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57931.4690 & 0.0026 & $-$0&0030 & 29\
1 & 57931.5464 & 0.0006 & 0&0006 & 52\
4 & 57931.7697 & 0.0005 & 0&0021 & 90\
5 & 57931.8408 & 0.0005 & $-$0&0007 & 95\
6 & 57931.9159 & 0.0008 & 0&0004 & 48\
13 & 57932.4346 & 0.0005 & 0&0018 & 67\
14 & 57932.5088 & 0.0005 & 0&0021 & 72\
18 & 57932.8002 & 0.0011 & $-$0&0022 & 85\
19 & 57932.8752 & 0.0009 & $-$0&0011 & 92\
\
\
\
[ccccc]{} Year & Month & Day & max& $V$ mag\
2015 & 7 & 19 & 57222 & 15.0\
2015 & 10 & 13 & 57308 & 15.1\
2016 & 4 & 10 & 57489 & 15.2\
2016 & 7 & 8 & 57578 & 15.0\
2016 & 10 & 9 & 57671 & 15.2\
2017 & 4 & 3 & 57847 & 15.1\
2017 & 6 & 27 & 57932 & 15.1\
2017 & 9 & 26 & 58022 & 15.1\
\
NN Camelopardalis {#obj:nncam}
-----------------
NN Cam = NSV 1485 was identified as a dwarf nova by @khr05nsv1485. For more history, see @Pdot7. The 2017 superoutburst was detected by the ASAS-SN team at $V$=13.14 on 2017 August 30. Only single superhump maximum was recorded at BJD 2458004.4408(9) ($N$=76).
V391 Camelopardalis {#obj:v391cam}
-------------------
The material is the same as in @Pdot9. A re-analysis of the data yielded a positive detection of post-superoutburst superhumps whose period is consistent with the past observations of stage C superhumps (e-table \[tab:v391camoc2017\]). Note that we used a narrower range than in @Pdot9 to determine the maximum on BJD 2457829, resulting a slightly different value.
A comparison of $O-C$ diagrams of V391 Cam between different superoutbursts (e-figure \[fig:v391camcomp2\]) suggests that there was a separate precursor outburst in the 2017 superoutburst and the initial superhump detection referred to a stage A superhump, when the object was still rising toward the full maximum of the superoutburst.
(85mm,70mm)[v391camcomp2.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57829.3181 & 0.0002 & $-$0&0035 & 185\
209 & 57841.2621 & 0.0005 & 0&0102 & 41\
210 & 57841.3182 & 0.0007 & 0&0093 & 59\
245 & 57843.3037 & 0.0019 & $-$0&0031 & 58\
246 & 57843.3599 & 0.0011 & $-$0&0040 & 58\
263 & 57844.3255 & 0.0015 & $-$0&0088 & 59\
\
\
\
KP Cassiopeiae {#obj:kpcas}
--------------
This object (S 3865) was discovered by @hof49newvar. A finding chart was provided in @hof57MVS245. @kin00VSID4896 provided correct identification. The object, however, has not been regularly monitored before the chance detection of a bright (13.0 mag) outburst by Y. Sano on 2008 October 25 (cf. vsnet-alert 10629). The 2008 outburst was well observed and the object was confirmed to be an SU UMa-type dwarf nova ([@Pdot]; [@boy10kpcas]). Although several outbursts were recorded since 2008, all of them were likely normal outbursts.
The 2017 superoutburst was recorded by Y. Maeda at an unfiltered CCD magnitude of 13.5 and H. Maehara at a visual magnitude of 14.2 on 2017 November 11 (cf. vsnet-alert 21581). Observations starting on 2017 November 18 detected superhumps. The times of maxima are listed in e-table \[tab:kpcasoc2017\]. The observations covered the relatively late phase of the superoutburst, and these superhumps were likely stage C ones. The resultant period agrees with the period of stage C superhumps during the 2008 superoutburst.
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 58076.0720 & 0.0008 & 0&0014 & 81\
1 & 58076.1541 & 0.0007 & $-$0&0016 & 92\
10 & 58076.9245 & 0.0012 & 0&0025 & 96\
11 & 58077.0035 & 0.0012 & $-$0&0036 & 92\
13 & 58077.1786 & 0.0010 & 0&0012 & 94\
\
\
\
VW Coronae Borealis {#obj:vwcrb}
-------------------
VW CrB was discovered as a dwarf nova (Antipin Var 21) by @ant96vwcrb. @nov97vwcrb established the SU UMa-type nature of this object. For more information, see @Pdot8. The 2017 superoutburst was detected by the ASAS-SN team at $V$=14.4 on 2017 May 11 (we later knew that an AAVSO observer detected a rising phase at $V$=15.69 on May 10). The outburst was also detected by M. Hiraga at an unfiltered CCD magnitude of 14.2 on May 13.
We observed this superoutburst on two nights and detected superhumps (e-table \[tab:vwcrboc2017\]). Although our observations were carried out relatively late, they were likely in the middle of stage B (see e-figure \[fig:vwcrbcomp3\]) since the duration of superoutbursts in VW CrB is long.
(85mm,70mm)[vwcrbcomp3.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57892.0529 & 0.0053 & 0&0053 & 84\
1 & 57892.1146 & 0.0013 & $-$0&0050 & 117\
9 & 57892.6903 & 0.0035 & $-$0&0053 & 37\
10 & 57892.7699 & 0.0005 & 0&0024 & 106\
11 & 57892.8421 & 0.0005 & 0&0026 & 108\
\
\
\
GP Canum Venaticorum {#obj:gpcvn}
--------------------
This object was originally selected as a CV (SDSS J122740.83$+$513925.0) during the course of the SDSS [@szk06SDSSCV5]. The object is an eclipsing SU UMa-type dwarf nova. The SU UMa-type nature was confirmed during the 2017 superoutburst ([@she08j1227]; [@Pdot]). For more history, see @Pdot9.
The 2017 superoutburst was detected by the ASAS-SN team at $V$=14.8 on 2017 July 5. Subsequent single-night observations detected two superhumps: BJD 2457941.3788(3) ($N$=40) and 2457941.4457(3) ($N$=50).
GQ Canum Venaticorum {#obj:gqcvn}
--------------------
This object was discovered as ASASSN-13ao by the ASAS-SN team on 2013 June 8 [@sta13asassn13aoatel5118]. The 2013 superoutburst was studied in @Pdot5 yielding only two superhump maxima.
The 2017 superoutburst was detected by the ASAS-SN team at $V$=14.7 on 2017 April 8. Superhumps were better observed than in the 2013 superoutburst (vsnet-alert 20899, 20913; e-figure \[fig:gqcvnshpdm\]). The times of superhump maxima are listed in e-table \[tab:gqcvnoc2017\]. Although there was some tendency of a period decrease, we could not determine the superhump stages.
Three superoutbursts have been known in the ASAS-SN data (2013 June 7, $V$=15.0; 2016 January 27, $V$=14.8 and the present one). The shortest interval between superoutbursts was 437 d.
(85mm,110mm)[gqcvnshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57854.4146 & 0.0005 & $-$0&0022 & 97\
1 & 57854.5026 & 0.0009 & $-$0&0037 & 81\
11 & 57855.4046 & 0.0006 & 0&0035 & 97\
12 & 57855.4919 & 0.0006 & 0&0013 & 94\
13 & 57855.5820 & 0.0007 & 0&0020 & 78\
14 & 57855.6713 & 0.0010 & 0&0018 & 68\
29 & 57857.0064 & 0.0044 & $-$0&0052 & 71\
30 & 57857.1038 & 0.0006 & 0&0027 & 181\
31 & 57857.1941 & 0.0015 & 0&0035 & 53\
37 & 57857.7236 & 0.0004 & $-$0&0038 & 169\
\
\
\
V503 Cygni {#obj:v503cyg}
----------
For this famous SU UMa-type dwarf nova with a short supercycle and negative superhumps [@har95v503cyg]. @kat02v503cyg reported a dramatic variation in the number of normal outbursts, and this finding led to the discovery of the state with negative superhumps suppressing the number of normal outbursts in other objects ([@ohs12eruma]; [@zem13eruma]; [@osa13v1504cygKepler]; [@osa13v344lyrv1504cyg]). @pav12v503cyg indeed confirmed temporal disappearance of negative superhumps in 2010 and shortening of the outburst cycle. The 2017 July superoutburst was detected at a visual magnitude of 14.1 by Alain Klotz on 2017 July 11. Single-night observations detected two superhumps: BJD 2457951.4358(28) ($N$=54) and 2457951.5257(3) ($N$=121).
V632 Cygni {#obj:v632cyg}
----------
This object was discovered by @hof49newvar1, who recorded three outbursts on 1938 December 11 (13.9 mag), 1939 November 3 (12.8 mag) and 1940 June 13 (magnitude unknown). @hof57VSchart provided a finding chart. Although this object had long been introduced in monitoring programs by the AAVSO and AFOEV, the position of the object was not correctly marked [@wen89v632cygv630cyg]. Outbursts started to be detected by visual observers since 1988–1989 when the chart error was corrected (the old AAVSO chart marked the object at a 16 mag unrelated star). Unpublished $I$-band photometry by one of the authors (T.K) in 1991 suggested a large outburst amplitude and the object was suspected to be an SU UMa-type dwarf nova. Based on this information, VSOLJ observers monitored the object and obtained some time-resolved CCD photometry, but it did not lead to a successful detection of superhumps. @liu99CVspec1 reported a spectrum and suggested that the orbital period is likely short. @she07CVspec determined its orbital period to be 0.06377(8) d. The SU UMa-type nature was finally established during the 2008 superoutburst [@Pdot].
The 2017 superoutburst was visually detected by L. Kocsmaros at a magnitude of 13.8 on 2017 June 16 (cvnet-outburst 7520). Only single-night observations were obtained. The times of superhump maxima are listed in e-table \[tab:v632cygoc2017\]. The superhump period of 0.0655(3) d was determined by the PDM method.
We give a list of recent superoutburst in e-table \[tab:v632cygout\]. The supercycle is around 210 d.
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57926.3774 & 0.0012 & $-$0&0002 & 56\
1 & 57926.4438 & 0.0009 & 0&0004 & 68\
2 & 57926.5089 & 0.0006 & $-$0&0002 & 57\
\
\
\
[cccccc]{} Year & Month & Day & max& mag & source\
2013 & 10 & 22 & 56588 & 14.0V & AAVSO\
2016 & 5 & 5 & 57513 & 13.7V & ASAS-SN\
2016 & 11 & 28 & 57721 & 13.8v & AAVSO\
2017 & 6 & 16 & 57920 & 14.1V & ASAS-SN\
\
V1454 Cygni {#obj:v1454cyg}
-----------
Since it turned out the alias selection in @Pdot was wrong for the 2006 superoutburst, we list a corrected $O-C$ table (e-table \[tab:v1454cygoc2006\]). The $P_{\rm dot}$ for stage B has been corrected to be $P_{\rm dot}$ of $+7.4(1.4) \times 10^{-5}$ (120$\le E \le$277).
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 54063.9540 & 0.0017 & $-$0&0013 & 77\
120 & 54070.8827 & 0.0008 & 0&0082 & 84\
121 & 54070.9396 & 0.0011 & 0&0075 & 85\
173 & 54073.9205 & 0.0046 & $-$0&0099 & 81\
190 & 54074.9047 & 0.0015 & $-$0&0059 & 83\
207 & 54075.8866 & 0.0025 & $-$0&0042 & 63\
208 & 54075.9447 & 0.0032 & $-$0&0038 & 64\
248 & 54078.2566 & 0.0022 & 0&0017 & 32\
277 & 54079.9323 & 0.0010 & 0&0052 & 70\
294 & 54080.9099 & 0.0017 & 0&0026 & 84\
\
\
\
HO Delphini {#obj:hodel}
-----------
HO Del (=S 10066) was discovered as a dwarf nova by @hof67an29043. The SU UMa-type nature was confirmed during the 1994 superoutburst. See @Pdot8 for more history. The 2017 superoutburst was detected by the ASAS-SN team at $V$=14.75 on 2017 July 30. The outburst was also visually detected at 14.0 mag on 2017 July 31 by R. Stubbings. Single-night observations detected three superhumps: BJD 2457969.4926(15) ($N$=13), 2457969.5558(10) ($N$=23) and 2457969.6212(17) ($N$=19).
MN Draconis {#obj:mndra}
-----------
This object was discovered as a dwarf nova [@ant02var73dra]. The object was identified as an SU UMa-type dwarf nova in the period gap [@nog03var73dra]. The object has both a short supercycle and negative superhumps in quiescence citep[pav10mndra]{}. It was suggested that the large negative $P_{\rm dot}$ for superhumps reflected stage A-B transition [@Pdot6]. For more information, see @Pdot6.
The 2017 June superoutburst was detected by G. Poyner at an unfiltered CCD magnitude of 16.43 on 2017 June 19 (vsnet-alert 21143). Only one superhump maximum was measured: BJD 2457926.8051(11) ($N$=208).
OV Draconis {#obj:ovdra}
-----------
This object (=SDSS J125023.85$+$665525.5) is a CV selected during the course of the Sloan Digital Sky Survey (SDSS) [@szk03SDSSCV2]. @dil08SDSSCV confirmed the deeply eclipsing nature. The 2008 and 2009 superoutbursts were reported in @Pdot2 and another one in 2011 was reported in @Pdot3. The 2013 superoutburst was reported in @Pdot5.
The 2015 superoutburst was detected by the ASAS-SN team at $V$=15.9 on 2015 February 11. Although time-resolved observations were reported on two nights, we could not detect convincing superhumps. These observations are used in refining the eclipse ephemeris.
The 2017 superoutburst was detected by the ASAS-SN team at $V$=15.56 on 2017 May 26. We updated the eclipse ephemeris using our 2008–2017 observations using the MCMC analysis [@Pdot4]: $${\rm Min(BJD)} = 2456305.98940(7) + 0.0587356736(13) E
\label{equ:ovdraecl}.$$ The epoch corresponds to the center of all the observations. The times of superhump maxima are listed in e-table \[tab:ovdraoc2017\]. Stages B and C are clearly seen (e-figure \[fig:ovdracomp2\]). It is the first time to show a positive $P_{\rm dot}$ and transition to stage C so clearly in a deeply eclipsing system.
(85mm,70mm)[ovdracomp2.eps]{}
[rp[50pt]{}p[30pt]{}r@[.]{}lcr]{} & & & & &\
0 & 57901.4366 & 0.0003 & 0&0024 & 0.17 & 69\
1 & 57901.4969 & 0.0004 & 0&0024 & 0.20 & 65\
2 & 57901.5546 & 0.0008 & $-$0&0003 & 0.18 & 35\
16 & 57902.3969 & 0.0009 & $-$0&0022 & 0.52 & 36\
17 & 57902.4566 & 0.0006 & $-$0&0028 & 0.54 & 65\
18 & 57902.5177 & 0.0010 & $-$0&0020 & 0.58 & 27\
32 & 57903.3593 & 0.0011 & $-$0&0048 & 0.91 & 25\
33 & 57903.4238 & 0.0010 & $-$0&0005 & 0.01 & 26\
34 & 57903.4830 & 0.0023 & $-$0&0016 & 0.01 & 16\
50 & 57904.4450 & 0.0020 & $-$0&0046 & 0.39 & 35\
66 & 57905.4107 & 0.0017 & $-$0&0038 & 0.83 & 39\
83 & 57906.4383 & 0.0016 & $-$0&0014 & 0.33 & 27\
84 & 57906.5055 & 0.0030 & 0&0054 & 0.47 & 21\
94 & 57907.1154 & 0.0019 & 0&0123 & 0.86 & 58\
95 & 57907.1723 & 0.0050 & 0&0089 & 0.83 & 37\
98 & 57907.3557 & 0.0020 & 0&0114 & 0.95 & 27\
115 & 57908.3690 & 0.0018 & $-$0&0006 & 0.20 & 28\
116 & 57908.4290 & 0.0021 & $-$0&0009 & 0.22 & 27\
117 & 57908.4875 & 0.0024 & $-$0&0027 & 0.22 & 25\
148 & 57910.3519 & 0.0034 & $-$0&0078 & 0.96 & 21\
149 & 57910.4167 & 0.0031 & $-$0&0033 & 0.06 & 28\
150 & 57910.4769 & 0.0036 & $-$0&0034 & 0.09 & 27\
\
\
\
\
BE Octantis {#obj:beoct}
-----------
BE Oct was discovered as a possible dwarf nova (S 6633) [@hof63VSS61]. J. Kemp and J. Patterson obtained a superhump period of 0.07712(13) d from observations on 1996 August 17 and 18 (vsnet-obs 3461). Although outbursts have been rather regularly recorded, no further observations of superhumps were reported. @mas03faintCV reported a typical dwarf nova-type spectrum in quiescence.
The 2017 superoutburst was detected at a visual magnitude of 15.4 by R. Stubbings and at $V$=15.97 by the ASAS-SN team on 2017 July 1. Subsequent observations detected superhumps (figure \[fig:beoctshpdm\]). The times of superhump maxima are listed in e-table \[tab:beoctoc2017\]. Although there were observations after BJD 2457941, the object became too faint to measure individual superhump maxima. The light curve indicated brightening on July 7 (BJD 2457941), suggesting that there was stage B-C transition around here. The best superhump period based on the first four nights (figure \[fig:beoctshpdm\]) was determined to be 0.07715(7) d by the PDM method.
(85mm,110mm)[beoctshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57938.7720 & 0.0058 & $-$0&0062 & 12\
1 & 57938.8606 & 0.0026 & 0&0053 & 23\
14 & 57939.8594 & 0.0020 & 0&0016 & 23\
39 & 57941.7878 & 0.0038 & 0&0021 & 19\
40 & 57941.8600 & 0.0032 & $-$0&0028 & 23\
\
\
\
V521 Pegasi {#obj:v521peg}
-----------
This object (=HS 2219$+$1824) is a dwarf nova reported in @rod05hs2219. The SU UMa-type nature was confirmed by @rod05hs2219. For more information, see @Pdot5 and @Pdot6.
The 2017 outburst was detected by the ASAS-SN team at $V$=12.8 on 2017 August 24 and was found to be fading rapidly on the same night by K. Wenzel and E. Muyllaert. This outburst turned out to be a precursor outburst and the true superoutburst occurred 7 d after (cf. vsnet-alert 21384: detection by the ASAS-SN team at $V$=12.0 and visually by H. Maehara on 2017 August 31). The times of superhump maxima are listed in e-table \[tab:v521pegoc2017\]. Although there were observations during the rapidly fading part, we could not determine superhump maxima.
A comparison of $O-C$ diagrams suggests that the 2017 observations recorded the final part of stage B and superhump started to develop 52 cycles (3.2 d) before the detection of the superoutburst. This suggests that superhumps started to develop several days after the precursor outburst \[the actual growth time may have been longer, see @kat16j0333, @ima17qzvir\].
(88mm,70mm)[v521pegcomp2.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57997.5283 & 0.0004 & $-$0&0011 & 64\
13 & 57998.3319 & 0.0003 & 0&0011 & 42\
14 & 57998.3931 & 0.0004 & 0&0007 & 45\
15 & 57998.4539 & 0.0004 & $-$0&0002 & 41\
16 & 57998.5157 & 0.0003 & $-$0&0001 & 47\
17 & 57998.5774 & 0.0004 & 0&0000 & 46\
18 & 57998.6404 & 0.0018 & 0&0014 & 12\
26 & 57999.1281 & 0.0051 & $-$0&0040 & 22\
27 & 57999.1964 & 0.0003 & 0&0026 & 128\
28 & 57999.2556 & 0.0003 & 0&0001 & 129\
29 & 57999.3167 & 0.0009 & $-$0&0005 & 67\
\
\
\
V368 Persei {#obj:v368per}
-----------
V368 Per was discovered by @ric69v1233aql. The SU UMa-type nature was identified by I. Miller in 2012 (BAAVSS alert 3113). For more information see @Pdot5.
The 2017 superoutburst was detected by the ASAS-SN team at $V$=15.43 on 2017 September 26. Subsequent observations detected superhumps (vsnet-alert 21477, 21485, 21500). In contrast to the 2012 superoutburst, when only stage C superhumps were observed, we could observe both stages B and C (e-figure \[fig:v368percomp\]; the maxima for $E \le$ may be stage A superhumps). The $P_{\rm dot}$ for stage B was not determined due to the shortness of stage B for this relatively long-$P_{\rm SH}$ object.
Although this field has been monitored by the ASAS-SN team since 2012 January, the present outburst was the first well-recorded superoutburst. Although the 2012 superoutburst was detected on a single night, it was impossible to recognize it to be a superoutburst by the ASAS-SN data only.
(88mm,70mm)[v368percomp.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 58024.3453 & 0.0023 & $-$0&0081 & 26\
1 & 58024.4292 & 0.0003 & $-$0&0032 & 163\
2 & 58024.5056 & 0.0004 & $-$0&0059 & 86\
3 & 58024.5869 & 0.0002 & $-$0&0036 & 83\
12 & 58025.3031 & 0.0004 & 0&0015 & 79\
13 & 58025.3812 & 0.0003 & 0&0006 & 86\
14 & 58025.4602 & 0.0003 & 0&0006 & 160\
15 & 58025.5394 & 0.0002 & 0&0007 & 263\
16 & 58025.6183 & 0.0004 & 0&0006 & 84\
24 & 58026.2505 & 0.0005 & 0&0007 & 135\
25 & 58026.3298 & 0.0004 & 0&0010 & 85\
26 & 58026.4088 & 0.0003 & 0&0010 & 151\
27 & 58026.4878 & 0.0002 & 0&0010 & 235\
28 & 58026.5674 & 0.0003 & 0&0015 & 209\
29 & 58026.6485 & 0.0009 & 0&0036 & 37\
36 & 58027.2021 & 0.0007 & 0&0041 & 104\
37 & 58027.2803 & 0.0005 & 0&0033 & 191\
38 & 58027.3583 & 0.0007 & 0&0022 & 95\
39 & 58027.4411 & 0.0004 & 0&0060 & 86\
40 & 58027.5167 & 0.0003 & 0&0026 & 86\
41 & 58027.5961 & 0.0004 & 0&0030 & 74\
63 & 58029.3343 & 0.0017 & 0&0028 & 27\
65 & 58029.4920 & 0.0010 & 0&0025 & 83\
66 & 58029.5677 & 0.0012 & $-$0&0008 & 83\
75 & 58030.2783 & 0.0014 & $-$0&0013 & 44\
76 & 58030.3513 & 0.0016 & $-$0&0074 & 44\
77 & 58030.4382 & 0.0015 & 0&0005 & 44\
78 & 58030.5102 & 0.0014 & $-$0&0065 & 29\
79 & 58030.5927 & 0.0015 & $-$0&0030 & 33\
\
\
\
XY Piscium {#obj:xypsc}
----------
XY Psc was discovered as a transient located close to the galaxy UGC 729 on 1972 October 5 at a photographic magnitude of 13.0 and was detected at a magnitude of 15.0 on 1972 October 17 [@ros72xypsciauc1]. Although the object was suspected to be a supernova of this galaxy, it was suggested to be a dwarf nova based on the rapid rise and decline [@ros72xypsciauc2].
VSOLJ members started monitoring this object in 1984 and S. Fujino recorded a possible outburst at a photographic magnitude (using a filter adjusted to reproduce the $V$ band) of 15.0 on 1984 January 21. There was only one positive record and, unfortunately, a negative film was lost and the identity of the object remained unclear. Since the 1990s, the object started to be monitored more regularly by observers worldwide. No outburst, however, was recorded.
In the meantime, @hen01xypsc obtained deep CCD images of this field and identified a quiescent blue counterpart of $V$=21.1. @hen01xypsc also provided some more details of historical observations of this object. @kat01hvvir listed this object as a candidate WZ Sge-type dwarf nova.
The 2017 outburst was detected by the ASAS-SN team at $V$=13.1 on 2017 June 2 (vsnet-alert 21085). Two night before, the object was fainter than $V$=16.3. Since the object was still low in the morning sky, amateur observers had not yet started monitoring. Due to the short observing windows, it took several days to detect superhumps (vsnet-alert 21111). Later observations pinned down the superhump period as observations accumulated (vsnet-alert 21119, 21131; e-figure \[fig:xypscshpdm\]).
The times of superhump maxima are listed in e-table \[tab:xypscoc2017\]. Despite the limited number of observations, stage B with a clearly positive $P_{\rm dot}$ and transition to stage C were recorded.
The initial sign of superhumps was recorded on 2017 June 10 (8 d after the outburst detection). Although there were earlier observations, it was impossible to detect modulation due to the shortness of observations. The relatively early appearance of ordinary superhumps and a strongly positive $P_{\rm dot}$ would suggest either an ordinary SU UMa-type dwarf nova or a WZ Sge/SU UMa-type borderline object. Since the object is not suited for observation around the solar conjunction, past outbursts near solar conjunctions may have been easily missed. We probably need to wait for a superoutburst occurring in the favorable season of the year to possibly detect early superhumps.
The object showed a rebrightening at $V$=15.9 on 2017 June 27 (ASAS-SN detection, vsnet-alert 21171). According to the ASAS-SN observations, there was no evidence of multiple rebrightenings. Regular monitoring by the ASAS-SN team started in 2012 October without any previous detection.
(85mm,110mm)[xypscshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57915.8730 & 0.0006 & 0&0043 & 17\
17 & 57916.8986 & 0.0011 & $-$0&0013 & 17\
33 & 57917.8665 & 0.0010 & $-$0&0040 & 24\
50 & 57918.9012 & 0.0012 & $-$0&0004 & 16\
56 & 57919.2630 & 0.0006 & $-$0&0026 & 53\
66 & 57919.8703 & 0.0013 & $-$0&0018 & 22\
82 & 57920.8461 & 0.0011 & 0&0034 & 17\
83 & 57920.9077 & 0.0012 & 0&0044 & 11\
89 & 57921.2686 & 0.0009 & 0&0014 & 48\
99 & 57921.8706 & 0.0013 & $-$0&0032 & 21\
\
\
\
V701 Tauri {#obj:v701tau}
----------
V701 Tau was discovered by @era73v701tau as an eruptive object. The SU UMa-type nature was confirmed in 1995 (reported in [@Pdot]). @she07v701tau further reported the 2005 superoutburst. For more details, see @Pdot7.
The 2017 superoutburst was detected by the ASAS-SN team at $V$=14.88 on 2017 August 17. Time-resolved photometric observations were carried out on two nights and the times of superhump maxima are listed in e-table \[tab:v701tauoc2017\]. A comparison of $O-C$ diagrams suggests that we observed the late part of stage B superhumps (e-figure \[fig:v701taucomp3\]).
(88mm,70mm)[v701taucomp3.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57987.4742 & 0.0193 & $-$0&0006 & 11\
1 & 57987.5440 & 0.0014 & 0&0002 & 31\
2 & 57987.6132 & 0.0014 & 0&0004 & 29\
30 & 57989.5469 & 0.0031 & 0&0014 & 21\
31 & 57989.6132 & 0.0039 & $-$0&0014 & 31\
\
\
\
V1208 Tauri {#obj:v1208tau}
-----------
V1208 Tau was originally identified as a CV during the course of identification of ROSAT sources (=1RXS J045942.9$+$192625, [@mot96CVROSAT]). P. Schmeer detected the first-ever recorded outburst in 2000 (vsnet-alert 4118). The SU UMa-type nature was confirmed during this outburst. Although @pat05SH gave an orbital period of 0.0681(2) d, the source was unclear. The 2000 and 2002 superoutbursts were reported in @Pdot. The 2011 superoutburst was reported in @Pdot4. The entire course of the superoutburst, however, has not yet been well observed.
The 2017 superoutburst was detected at an unfiltered CCD magnitude of 14.8 on 2017 November 8 by Y. Maeda. This superoutburst was also recorded in the ASAS-SN data ($V$=15.90 on 2017 November 9, vsnet-alert 21576). Superhumps were detected on 2017 November 11 (vsnet-alert 21576). The times of superhump maxima are listed in e-table \[tab:v1208tauoc2017\]. Although there were observations on three nights after these observations, no clear superhumps were detected. It may have been that we only observed the terminal portion of the superoutburst, and that there could have been a rebrightening on 2017 November 17. The data were too sparse to depict the outburst behavior unambiguously. This object still need better observations to obtain precise values of superhump and orbital periods.
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 58068.7076 & 0.0009 & $-$0&0010 & 40\
1 & 58068.7797 & 0.0016 & 0&0017 & 39\
2 & 58068.8471 & 0.0012 & $-$0&0003 & 40\
3 & 58068.9165 & 0.0011 & $-$0&0003 & 39\
\
\
\
TU Trianguli {#obj:tutri}
------------
TU Tri was discovered as a dwarf nova (GR 287) with a photographic range of 14.8 to fainter than 18.0 by @rom78tutri. The coordinates given in this paper, however, was incorrect [@sha92tutri] and @sha91tutri independently discovered this dwarf nova. The observation by @sha91tutri recorded a long outburst starting on 1982 November 9 and lasted at least up to 1982 November 14. @liu00CVspec3 obtained a spectrum in quiescence without emission lines.
There was an outburst on 2013 January 1 at a visual magnitude of 14.6 detected by M. Simonsen. Although K. Torii’s observations could not detect superhumps (vsnet-campaign-dn 3237, 3262), @zlo04tutri recorded superhumps with a period of 0.0745 d from single-night observations.
Although an outburst on 2017 January 29 was observed (P. Dubovsky), it faded quickly and must have been a normal outburst. The 2017 superoutburst was detected by the ASAS-SN team at $V$=14.98 on 2017 October 19. Subsequent observations detected superhumps (vsnet-alert 21540, 21544). The resultant period of 0.07602(2) d (e-figure \[fig:tutrishpdm\]) was significantly longer than the measurement by @zlo04tutri. The times of superhump maxima are listed in e-table \[tab:tutrioc2017\]. There looks like to have been stage B-C transition between $E$=20 and $E$=59.
(85mm,110mm)[tutrishpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 58050.0377 & 0.0023 & $-$0&0038 & 42\
1 & 58050.1166 & 0.0009 & $-$0&0009 & 64\
16 & 58051.2561 & 0.0018 & $-$0&0017 & 51\
17 & 58051.3355 & 0.0005 & 0&0017 & 87\
18 & 58051.4131 & 0.0005 & 0&0032 & 105\
19 & 58051.4875 & 0.0006 & 0&0017 & 44\
20 & 58051.5639 & 0.0006 & 0&0020 & 24\
59 & 58054.5242 & 0.0009 & $-$0&0022 & 84\
\
\
\
SU Ursae Majoris {#obj:suuma}
----------------
This object is the prototype of SU UMa-type dwarf novae. See @Pdot7 for the history. The second superoutburst in 2017 was detected by P. Schmeer at a visual magnitude of 12.3 on 2017 October 16. The object further brightened to 11.2 on 2017 October 18 (J. Toone, baavss-alert 4810). Superhumps were observed on two nights and the times of superhump maxima are listed in e-table \[tab:suumaoc2017b\].
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 58046.5183 & 0.0003 & 0&0027 & 111\
1 & 58046.5920 & 0.0006 & $-$0&0026 & 140\
63 & 58051.4806 & 0.0015 & $-$0&0073 & 149\
64 & 58051.5740 & 0.0009 & 0&0072 & 140\
\
\
\
HS Virginis {#obj:hsvir}
-----------
HS Vir was discovered as an ultraviolet excess object PG 1341$-$079, and was confirmed by spectroscopy to be a cataclysmic variable ([@gre82PGsurveyCV]; [@PGsurvey]). @osm85hsvir reported from photographic observations that this object shows relatively abundant short, faint outbursts, together with a bright ($\sim$12.8 mag) one. @rin93thesis suggested an orbital period of 0.0836 day (with possible aliasing problems) from radial-velocity measurements. Since this observation, HS Vir has been considered to be a candidate for an SU UMa-type dwarf nova. @kat95hsvir reported frequent occurrence of short outbursts. Although the bright outburst in @osm85hsvir was suggestive of a superoutburst, it was only in 1996 March when superhumps with a mean period of 0.08059(3) d were indeed detected, confirming the SU UMa-type nature [@kat98hsvir]. @pat03suumas observed the same superoutburst and reported a superhump period of 0.08045(19) d. @men99hsvir obtained an orbital period of 0.07692(3) d from radial-velocity measurements. @kat01hsvir suggested a supercycle of 186 d or 371 d. @Pdot analyzed the 1996 data again, and yielded a slightly shorter superhump period of 0.08003(3) d.
Although this object was observed several times during later superoutbursts \[2007 March, 2008 June [@Pdot], and 2014 February\], these observations were performed only for a short time and there have been no new measurement of the superhump period.
The 2017 superoutburst was detected by H. Maehara at a visual magnitude of 13.5 on 2017 April 22. Superhumps were observed (vsnet-alert 20958). The times of superhump maxima are listed in e-table \[tab:hsviroc2017\]. There was a gap in the observation following the initial superhump detection and maxima from later observations were not of very good quality due to the low sampling rate and the complex profile of superhumps. The derived mean superhump period was 0.08031(6) d. Since observations were made during the relatively late phase, this period was probably affected by stage C superhumps, which were not apparent on the $O-C$ diagram due to scatter.
We also provide yet unpublished maxima of the 2007 superoutburst: BJD 2454188.0930(5) ($N$=57) and 2454190.0962(6) ($N$=89). A comparison of the $O-C$ diagram (e-figure \[fig:hsvircomp\]) cannot tell much other than the presence of stage A in the 1996 superoutburst.
(85mm,70mm)[hsvircomp.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57868.3646 & 0.0003 & $-$0&0007 & 75\
1 & 57868.4468 & 0.0002 & 0&0012 & 73\
2 & 57868.5274 & 0.0006 & 0&0014 & 57\
29 & 57870.6971 & 0.0010 & 0&0027 & 14\
33 & 57871.0157 & 0.0013 & 0&0001 & 83\
34 & 57871.0962 & 0.0008 & 0&0002 & 99\
41 & 57871.6563 & 0.0009 & $-$0&0019 & 22\
42 & 57871.7364 & 0.0037 & $-$0&0021 & 11\
53 & 57872.6343 & 0.0024 & 0&0123 & 25\
54 & 57872.7008 & 0.0011 & $-$0&0014 & 22\
66 & 57873.6572 & 0.0027 & $-$0&0088 & 19\
67 & 57873.7340 & 0.0030 & $-$0&0123 & 15\
78 & 57874.6413 & 0.0040 & 0&0115 & 19\
79 & 57874.7004 & 0.0021 & $-$0&0097 & 21\
103 & 57876.6450 & 0.0043 & 0&0074 & 22\
\
\
\
V406 Virginis {#obj:v406vir}
-------------
V406 Vir was originally selected as a CV (SDSS J123813.73$-$033933.0) during the course of the SDSS [@szk03SDSSCV2]. @szk03SDSSCV2 suggested a high inclination and an orbital period of 76 min. @zha06j1238 performed time-resolved photometric and spectroscopic observations and obtained an orbital period of 0.05592(35) d. @zha06j1238 classified the object a WZ Sge-like one but with cyclic brightening up to 0.4 mag with periods of the order of 8–12 hr in quiescence. @avi10j1238 performed infrared $JHK$ photometry and optical spectroscopy and indicated that the system has a L4-type brown dwarf. The Doppler mapping of the system showed the permanent presence of a spiral arm pattern in the accretion disk and @avi10j1238 suggested that they can be a result of the 2:1 resonance. @avi10j1238 classified this object to be a period bouncer.
Despite that the object has long been suspected to be a WZ Sge-type dwarf nova, no outburst was recorded until 2017. The 2017 outburst was detected by the ASAS-SN team at $V$=11.86 on 2017 July 31 (cf. vsnet-alert 21308). Despite poor seasonal location in the sky, the outburst was observed and superhumps were detected (vsnet-alert 21325, 21328, 21330; e-figure \[fig:v406virshpdm\]). These superhumps were detected already on 2017 August 4 (they were already stage B superhumps), indicating that the waiting time for the appearance of ordinary superhumps was short. The times of superhump maxima are listed in e-table \[tab:v406viroc2017\]. All superhumps were stage B ones and $P_{\rm dot}$ was relatively large $+8.1(1.5) \times 10^{-5}$
The short waiting time for ordinary superhumps, large amplitude of superhumps (e-figure \[fig:v406virshpdm\]) and the relatively large $P_{\rm dot}$ for stage B superhumps indicate that this object is not an extreme WZ Sge-type dwarf nova as expected from the conclusion by @avi10j1238. By using the relation between $q$ and $P_{\rm dot}$ for WZ Sge-type dwarf novae (equation 6 in [@kat15wzsge]), the $q$ value is expected to be 0.095(6) (the error corresponds to the measurement error only). This value is not particularly small for WZ Sge-type dwarf novae (cf. figure 17 in [@kat15wzsge]), particularly considering the high-mass white dwarf claimed by @avi10j1238. The ASAS-SN observations in 2017 only consisted of three nights and there was 7 d gap before the outburst detection. The waiting time for the appearance of ordinary superhumps may have been longer if the true maximum was missed. In any case, the conclusion by @avi10j1238 would be worth revisiting using higher quality observations.
(85mm,110mm)[v406virshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57970.8436 & 0.0024 & 0&0026 & 18\
1 & 57970.8993 & 0.0005 & 0&0013 & 34\
7 & 57971.2403 & 0.0002 & 0&0005 & 131\
12 & 57971.5255 & 0.0006 & 0&0009 & 13\
18 & 57971.8666 & 0.0003 & 0&0003 & 129\
24 & 57972.2081 & 0.0002 & 0&0000 & 92\
25 & 57972.2652 & 0.0005 & 0&0002 & 87\
29 & 57972.4925 & 0.0004 & $-$0&0004 & 31\
36 & 57972.8893 & 0.0003 & $-$0&0024 & 55\
42 & 57973.2308 & 0.0002 & $-$0&0026 & 131\
47 & 57973.5153 & 0.0007 & $-$0&0029 & 22\
53 & 57973.8572 & 0.0005 & $-$0&0027 & 70\
60 & 57974.2562 & 0.0005 & $-$0&0024 & 86\
64 & 57974.4852 & 0.0005 & $-$0&0013 & 38\
70 & 57974.8316 & 0.0035 & 0&0033 & 14\
71 & 57974.8862 & 0.0004 & 0&0009 & 85\
77 & 57975.2273 & 0.0003 & 0&0003 & 131\
82 & 57975.5133 & 0.0019 & 0&0016 & 14\
88 & 57975.8562 & 0.0006 & 0&0027 & 68\
\
\
\
NSV 35 {#obj:nsv35}
------
NSV 35 was discovered as a variable (HV 8001 = AN 97.1933) with a photographic range of 14.5–16.5 during a proper-motion survey [@luy38propermotion2]. The object is located in the region of the Small Magellanic Cloud and is also given a variable star name of SMC V2. @aug10CTCVCV2 found a cataclysmic variable (CTCV J0006$-$6900) at this location and gave a possible identification with NSV 35. @aug10CTCVCV2 obtained an orbital period of 0.0790(12) d by a radial-velocity study with rather limited baselines. @aug10CTCVCV2 also noted the presence of outbursts in the ASAS-3 data [@ASAS3].
A bright outburst was detected on 2017 October 15 by S. Hovell at a visual magnitude of 13.3 (vsnet-alert 21533). G. Myers detected superhumps, confirming the SU UMa-type nature of this object. This superoutburst reached $V$=12.6 on 2017 October 18. The times of superhump maxima are listed in e-table \[tab:nsv35oc2017\]. Although there were observations after BJD 2458054 (rapid fading from the superoutburst), individual superhump maxima could not be measured.
The mean profile of the superhumps is given in e-figure \[fig:nsv35shpdm\]. There is a prominent secondary maximum. The profile is similar to the one in the late phase of V344 Lyr in Kepler data (cf. [@woo11v344lyr]). @woo11v344lyr interpreted the secondary maximum as the accretion stream bright spot sweeping around the rim of the non-axisymmetric disk. This signal corresponds to the traditional late superhumps, and it is associated with a high mass-transfer rate from the secondary. Both V344 Lyr and NSV 35 have frequent normal outbursts, and the origin of the secondary maximum in these systems is likely the same.
This object has been monitored by the ASAS-SN team since 2014 May (e-figure \[fig:nsv35lc\]). There were no definite superoutburst in the past data. It may have been either that this object rarely showed superoutbursts despite frequent normal outbursts or that past superoutbursts escaped detection around solar conjunctions. The mean brightness (in quiescence) in 2014 gradually rose from 16 mag to 15 mag, suggesting long-term variation of the mass-transfer rate. There was also an interval between 2015 November and 2016 January, when no outbursts were recorded with gradually varying quiescent brightness from 16 mag to 15 mag. This interval was difficult to interpret as a standstill since the brightness was similar to that in 2017. Outbursts in this system may have been somehow suppressed in certain epochs, and this system deserves a further detailed study.
(85mm,110mm)[nsv35shpdm.eps]{}
(160mm,100mm)[nsv35lc.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 58044.9118 & 0.0004 & $-$0&0017 & 94\
37 & 58047.9148 & 0.0012 & 0&0030 & 61\
58 & 58049.6117 & 0.0022 & $-$0&0018 & 24\
59 & 58049.6872 & 0.0057 & $-$0&0073 & 13\
61 & 58049.8509 & 0.0126 & $-$0&0058 & 39\
62 & 58049.9398 & 0.0007 & 0&0022 & 236\
63 & 58050.0220 & 0.0009 & 0&0033 & 230\
64 & 58050.1011 & 0.0008 & 0&0013 & 236\
65 & 58050.1844 & 0.0010 & 0&0037 & 235\
66 & 58050.2644 & 0.0012 & 0&0026 & 133\
70 & 58050.5848 & 0.0040 & $-$0&0011 & 17\
71 & 58050.6723 & 0.0017 & 0&0054 & 16\
72 & 58050.7474 & 0.0044 & $-$0&0005 & 13\
83 & 58051.6379 & 0.0028 & $-$0&0014 & 21\
84 & 58051.7272 & 0.0050 & 0&0068 & 13\
87 & 58051.9587 & 0.0006 & $-$0&0048 & 228\
95 & 58052.6119 & 0.0026 & 0&0001 & 24\
96 & 58052.6874 & 0.0068 & $-$0&0054 & 13\
100 & 58053.0213 & 0.0007 & 0&0043 & 231\
109 & 58053.7392 & 0.0049 & $-$0&0070 & 13\
112 & 58053.9936 & 0.0033 & 0&0043 & 232\
\
\
\
1RXS J161659.5$+$620014 {#obj:j1616}
-----------------------
This object (hereafter 1RXS J161659) was initially identified as an X-ray selected variable (also known as MASTER OT J161700.81$+$620024.9), which was first detected in bright state on 2012 September 11 at an unfiltered CCD magnitude of 14.4 [@bal13j1616]. The SU UMa-type nature was confirmed during the 2016 outburst [@Pdot9]. For more information, see @Pdot9.
The 2017 superoutburst was detected by the ASAS-SN team at $V$=14.9 on 2017 August 31. Subsequent observations detected superhump (vsnet-alert 21429). The time of superhump maxima are listed in e-table \[tab:j1616oc2017\]. A comparison of the $O-C$ diagram suggests that the 2017 observation recorded mostly stage C. The 2017 superoutburst apparently was not detected early enough, since it started fading rapidly on August 6, only 6 d since the initial outburst detection.
(88mm,70mm)[j1616comp2.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57998.4361 & 0.0007 & $-$0&0034 & 75\
1 & 57998.5087 & 0.0014 & $-$0&0018 & 52\
2 & 57998.5820 & 0.0006 & 0&0005 & 72\
13 & 57999.3632 & 0.0008 & 0&0004 & 60\
14 & 57999.4365 & 0.0008 & 0&0026 & 74\
27 & 58000.3583 & 0.0008 & 0&0010 & 50\
28 & 58000.4285 & 0.0006 & 0&0002 & 59\
41 & 58001.3529 & 0.0011 & 0&0013 & 55\
42 & 58001.4229 & 0.0010 & 0&0003 & 62\
55 & 58002.3490 & 0.0028 & 0&0029 & 51\
56 & 58002.4183 & 0.0013 & 0&0013 & 75\
70 & 58003.4062 & 0.0018 & $-$0&0053 & 58\
\
\
\
ASASSN-13ce {#obj:asassn13ce}
-----------
This object was detected as a transient at $V$=16.27 on 2013 August 19 by the ASAS-SN team. The 2017 outburst was detected by the ASAS-SN team at $V$=15.55 on 2017 September 21 and announced after the observation of $V$=15.68 on 2017 September 24. The object had a blue SDSS counterpart and was suggested to be an SU UMa-type dwarf nova (vsnet-alert 21466). Subsequent observations indeed detected superhumps (vsnet-alert 21471; e-figure \[fig:asassn13ceshpdm\]). Although only two superhump maxima were measured, we could reasonably choose the superhump period since observations were relatively well spaced. The maxima were BJD 2458023.4409(16) ($N$=43) and 2458023.6647(11) ($N$=62). The period determined by the PDM method was 0.07511(3) d.
(85mm,110mm)[asassn13ceshpdm.eps]{}
ASASSN-13dh {#obj:asassn13dh}
-----------
This object was detected as a transient at $V$=15.61 on 2013 October 2 by the ASAS-SN team. This outburst was retrospectively detected on MASTER-Amur images on 2013 September 30 (13.1–13.3 unfiltered CCD magnitudes, vsnet-alert 16504). This outburst appears to be a normal outburst.
The 2017 outburst was detected at an unfiltered CCD magnitude of 13.62 by E. Muyllaert on 2017 September 8 (cvnet-outburst 7688). The outburst was immediately confirmed (vsnet-alert 21417) and time-resolved photometry detected superhumps (vsnet-alert 21419, 21432; figure \[fig:asassn13dhshpdm\]). There was a post-superoutburst rebrightening at $V$=15.87 on 2017 September 25 (ASAS-SN data).
The times of superhump maxima are listed in e-table \[tab:asassn13dhoc2017\]. Although initial two maxima may have been stage B superhumps, we could not determine the period. We consider that the later maxima correspond to stage C superhumps since they were recorded before termination of the superoutburst.
This object have undergone superoutbursts relatively regularly (e-table \[tab:asassn13dhout\]). These superoutburst can be expressed by a supercycle of 450(8) d.
(85mm,110mm)[asassn13dhshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 58005.4910 & 0.0004 & $-$0&0020 & 85\
1 & 58005.5821 & 0.0004 & $-$0&0022 & 94\
38 & 58008.9676 & 0.0003 & 0&0013 & 91\
49 & 58009.9730 & 0.0004 & 0&0013 & 95\
59 & 58010.8929 & 0.0034 & 0&0072 & 39\
60 & 58010.9771 & 0.0004 & $-$0&0001 & 88\
75 & 58012.3493 & 0.0006 & 0&0011 & 81\
76 & 58012.4403 & 0.0009 & 0&0007 & 92\
92 & 58013.8953 & 0.0073 & $-$0&0068 & 38\
93 & 58013.9920 & 0.0006 & $-$0&0015 & 82\
100 & 58014.6342 & 0.0019 & 0&0009 & 62\
\
\
\
[ccccc]{} Year & Month & Day & max& $V$ mag\
2012 & 10 & 3 & 56204 & 13.4\
2013 & 12 & 17 & 56644 & 13.3\
2015 & 2 & 11 & 57065 & 14.1\
2017 & 9 & 8 & 58005 & 13.3\
\
\
ASASSN-14ca {#obj:asassn14ca}
-----------
This object was detected as a transient at $V$=15.5 on 2014 June 7 by the ASAS-SN team [@dav14asassn14caatel6211]. The object was confirmed to be an SU UMa-type dwarf nova during the 2015 superoutburst, but the details were unknown [@Pdot8]. Refer to @Pdot8 for more history.
The 2017 superoutburst was detected by the ASAS-SN team at $V$=14.87 on 2017 October 17 and was announced after observation at $V$=14.86 on 2017 October 18. Subsequent observations detected superhumps (vsnet-alert 21529, 21538; e-figure \[fig:asassn14cashpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn14caoc2017\].
The object was recorded in superoutburst three times in the ASAS-SN data (e-table \[tab:asassn14caout\]). Two of them showed a separate precursor outburst. The supercycle is estimated to be 391(20) d.
(85mm,110mm)[asassn14cashpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 58046.3695 & 0.0003 & $-$0&0009 & 74\
1 & 58046.4376 & 0.0004 & 0&0001 & 72\
13 & 58047.2422 & 0.0002 & 0&0002 & 173\
14 & 58047.3094 & 0.0002 & 0&0004 & 171\
15 & 58047.3762 & 0.0002 & 0&0001 & 150\
16 & 58047.4435 & 0.0003 & 0&0004 & 124\
44 & 58049.3191 & 0.0006 & $-$0&0009 & 48\
45 & 58049.3876 & 0.0005 & 0&0005 & 66\
\
\
\
[ccccc]{} Year & Month & Day & max& $V$ mag\
2014 & 8 & 12 & 56882 & 15.4\
2015 & 7 & 11 & 57215 & 15.4\
2017 & 10 & 16 & 58043 & 14.9\
\
\
ASASSN-14cr {#obj:asassn14cr}
-----------
This object was detected as a transient at $V$=15.20 on 2014 June 19 by the ASAS-SN team. Although there were time-resolved observations during long outbursts on 2016 August 5 (T. Vanmunster), observations were too short to detect superhumps (there were also observations on 2015 August 22 by T. Vanmunster during a normal outburst). The 2017 outburst was detected by the ASAS-SN team at $V$=15.08 on 2017 July 26 (ASAS-SN data indicated that the object was already in outburst 8 d before). Subsequent observations detected superhumps (vsnet-alert 21301). The times of superhump maxima are listed in e-table \[tab:asassn14croc2017\]. Although a PDM analysis prefers a period around 0.0646 d (as in vsnet-alert 21301; e-figure \[fig:asassn14crshpdm\]), an alias of 0.0687 d gives much smaller $O-C$ values and we adopted it. The superhump stage is likely C since the observation covered the final part of the superoutburst.
The object showed relatively regular superoutbursts in the past (e-table \[tab:asassn14crout\]). The detection 2014 December 13 looks likely a superoutburst as judged from the brightness. If this was indeed a superoutburst, the supercycle is 190(3) d.
(85mm,110mm)[asassn14crshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57964.3638 & 0.0010 & $-$0&0003 & 34\
1 & 57964.4313 & 0.0013 & $-$0&0015 & 37\
2 & 57964.5014 & 0.0035 & $-$0&0001 & 36\
3 & 57964.5707 & 0.0007 & 0&0006 & 10\
14 & 57965.3267 & 0.0055 & 0&0009 & 23\
15 & 57965.3957 & 0.0017 & 0&0012 & 36\
16 & 57965.4647 & 0.0038 & 0&0015 & 37\
32 & 57966.5577 & 0.0013 & $-$0&0047 & 26\
44 & 57967.3920 & 0.0015 & 0&0052 & 37\
45 & 57967.4525 & 0.0010 & $-$0&0030 & 37\
\
\
\
[ccccc]{} Year & Month & Day & max& $V$ mag\
2014 & 6 & 18 & 56827 & 14.9\
2014 & 12 & 13 & 57005 & 15.0\
2015 & 6 & 7 & 57181 & 14.9\
2016 & 7 & 26 & 57596 & 14.9\
2017 & 7 & 17 & 57952 & 15.1\
\
\
ASASSN-14kb {#obj:asassn14kb}
-----------
This object was detected as a transient at $V$=15.4 on 2014 December 11 by the ASAS-SN team [@pri14asassn14kbatel6688]. @wyr14asassn14kbatel6690 reported from OGLE-IV Magellanic System monitoring program observations that this object (=OGLE-LMC529.30.114) is an eclipsing SU UMa-type dwarf nova with an orbital period of 0.0681057 d. Further details of the OGLE-IV observations were reported in @mro15OGLEDNe, which provides all photometric observations.[^10] Although the OGLE-IV observations recorded the outburst pattern characteristic to an SU UMa-type dwarf nova, no time-resolved photometry was made by the OGLE-IV team.
The 2017 superoutburst was detected by the ASAS-SN team at $V$=15.72 on 2017 April 14. Subsequent observations detected superhumps (vsnet-alert 20935; e-figure \[fig:asassn14kbshpdm\]).
By using the OGLE-IV CVOM data (outside outbursts) and our 2017 observations, we refined the eclipse ephemeris using the MCMC analysis [@Pdot4] as follows: $${\rm Min(BJD)} = 2457865.24389(2) + 0.0681057201(10) E .
\label{equ:asassn14kbecl}$$ The epoch corresponds to the center of all the observations.
The times of superhump maxima are listed in e-table \[tab:asassn14kboc2017\]. Since stages were unclear due to the limited observations (low sampling rate and low signal-to-noise ratio when the object faded), we provided a globally averaged period.
The object shows superoutburst relatively regularly (e-table \[tab:asassn14kbout\]). These superoutburst can be expressed by a mean supercycle of 150.4(9) d with maximum residuals of 42 d. There appear to have been variations of supercycles: 162(1) d for the interval JD 2455276–2456255 and 144.4(5) d for the interval JD 2456255–2457699. The supercycle lengthened again until now.
(85mm,110mm)[asassn14kbshpdm.eps]{}
[rp[50pt]{}p[30pt]{}r@[.]{}lcr]{} & & & & &\
0 & 57860.5214 & 0.0005 & $-$0&0003 & 0.66 & 20\
1 & 57860.5915 & 0.0007 & $-$0&0006 & 0.69 & 14\
15 & 57861.5805 & 0.0021 & 0&0025 & 0.21 & 14\
43 & 57863.5502 & 0.0017 & 0&0005 & 0.13 & 29\
57 & 57864.5366 & 0.0008 & 0&0009 & 0.61 & 29\
58 & 57864.6051 & 0.0009 & $-$0&0010 & 0.62 & 14\
71 & 57865.5174 & 0.0012 & $-$0&0041 & 0.02 & 22\
72 & 57865.5873 & 0.0012 & $-$0&0046 & 0.04 & 13\
85 & 57866.5112 & 0.0032 & 0&0038 & 0.61 & 24\
86 & 57866.5806 & 0.0031 & 0&0028 & 0.63 & 15\
\
\
\
\
[cccccc]{} Year & Month & Day & max& mag & Source\
2010 & 3 & 22 & 55276 & 15.7I & OGLE-IV\
2010 & 9 & 15 & 55455 & 15.3I & OGLE-IV\
2011 & 2 & 14 & 55607 & 15.4I & OGLE-IV\
2012 & 1 & 3 & 55930 & 15.3I & OGLE-IV\
2012 & 11 & 23 & 56255 & 15.9I & OGLE-IV\
2014 & 11 & 10 & 56972 & 15.2V & ASAS-SN\
2015 & 4 & 13 & 57126 & 15.8V & ASAS-SN, OGLE-IV\
2015 & 9 & 1 & 57267 & 15.1V & ASAS-SN, OGLE-IV\
2016 & 1 & 18 & 57406 & 15.7I & OGLE-IV\
2016 & 11 & 6 & 57699 & 15.5V & ASAS-SN, OGLE-IV\
2017 & 4 & 14 & 57858 & 15.6V & ASAS-SN\
2017 & 9 & 24 & 58021 & 15.5V & ASAS-SN\
\
ASASSN-14lk {#obj:asassn14lk}
-----------
This object was detected as a transient at $V$=13.48 on 2014 December 1 by ASAS-SN team (vsnet-alert 18032). It may be identical to NSV 12802 = HV 9672 (see [@Pdot7]). The 2014 observations detected superhumps [@Pdot7].
The 2017 superoutburst was detected by the ASAS-SN team at $V$=14.16 on 2017 October 30 and was visually observed at 14.2 mag on 2017 November 3 by R. Stubbings (vsnet-alert 21566). Subsequent observations starting on 2017 November 9 detected superhumps (vsnet-alert 21575). The times of superhump maxima are listed in e-table \[tab:asassn14lkoc2017\]. The short superhump period and low amplitudes (initially 0.11 mag and later decreased to 0.09 mag) suggest that these superhumps were stage C ones. The object started fading rapidly on 2017 November 11, supporting the stage C nature of the superhumps. The $O-C$ values also support this interpretation (e-figure \[fig:asassn14lkcomp\]).
A list of past superoutbursts recorded in the ASAS-SN data are listed in e-table \[tab:asassn14lkout\]. The supercycle is around 540 d or its $N$-th.
(85mm,70mm)[asassn14lkcomp.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 58067.2468 & 0.0090 & 0&0031 & 86\
1 & 58067.3024 & 0.0009 & $-$0&0024 & 142\
2 & 58067.3631 & 0.0016 & $-$0&0027 & 108\
18 & 58068.3447 & 0.0012 & 0&0021 & 141\
19 & 58068.4053 & 0.0021 & 0&0016 & 73\
34 & 58069.3177 & 0.0015 & $-$0&0018 & 94\
\
\
\
[ccccc]{} Year & Month & Day & max& $V$ mag\
2014 & 11 & 30 & 56992 & 13.6\
2016 & 5 & 14 & 57523 & 14.2\
2017 & 10 & 30 & 58056 & 14.2\
\
ASASSN-15fu {#obj:asassn15fu}
-----------
This object was detected as a transient at $V$=15.6 on 2015 March 27 by the ASAS-SN team. Observations of superhumps during the 2015 superoutburst were reported in @Pdot8.
The 2017 superoutburst was detected by the ASAS-SN team at $V$=15.4 on 2017 May 29. The times of superhump maxima are listed in e-table \[tab:asassn15fuoc2017\]. Since the resultant period was between those of stages B and C in 2015 [@Pdot8], the 2017 observations were likely performed near stage B-C transition. A comparison of $O-C$ diagrams is also consistent with this interpretation (e-figure \[fig:asassn15fucomp\]).
There were three known superoutbursts: 2015 March 26 (JD 2457108), 2015 December 28 (JD 2457385) and the present one. The supercycle is around 260–270 d.
(85mm,70mm)[asassn15fucomp.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57906.4982 & 0.0027 & $-$0&0010 & 19\
1 & 57906.5736 & 0.0017 & $-$0&0002 & 11\
14 & 57907.5461 & 0.0014 & 0&0026 & 13\
27 & 57908.5115 & 0.0017 & $-$0&0017 & 23\
28 & 57908.5882 & 0.0054 & 0&0004 & 8\
\
\
\
ASASSN-15fv {#obj:asassn15fv}
-----------
This object was detected as a transient at $V$=15.7 on 2015 March 27 by the ASAS-SN team. The 2017 outburst was detected by the ASAS-SN team at $V$=15.76 on 2017 May 19. Single-night observations detected superhumps (vsnet-alert 21073; e-figure \[fig:asassn15fvshpdm\]). The times of superhump maxima were BJD 57898.4260(6) ($N$=50) and 2457898.4967(13) ($N$=49).
There was also a superoutburst in the ASAS-SN data on 2016 September 26. The interval between the 2016 and 2017 superoutbursts suggests a supercycle of 235 d.
(85mm,110mm)[asassn15fvshpdm.eps]{}
ASASSN-15qu {#obj:asassn15qu}
-----------
This object was detected as a transient at $V$=15.9 on 2015 October 8 by the ASAS-SN team. The 2017 outburst was detected by the ASAS-SN team at $V$=14.37 on 2017 August 30. Subsequent observations detected well-developed superhumps (vsnet-alert 21390, 21398; e-figure \[fig:asassn15qushpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn15quoc2017\]. The period apparently increase between $E$=0 and $E$=28. We, however, did not adopt $P_{\rm dot}$ from this segment since the baseline was too short and instead gave a global $P_{\rm dot}$.
Although this field was monitored by the ASAS-SN team since 2014 May, no other outbursts were recorded. The 2015 outburst was much fainter and it must have been a normal outburst. ASAS-3 probably recorded a superoutburst in 2005 November ($V$=15.14 on 2005 November 3 and $V$=14.66 on 2015 November 6).
(85mm,110mm)[asassn15qushpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57998.4052 & 0.0005 & $-$0&0025 & 185\
1 & 57998.4880 & 0.0006 & $-$0&0001 & 184\
2 & 57998.5677 & 0.0005 & $-$0&0009 & 185\
15 & 57999.6124 & 0.0016 & $-$0&0021 & 19\
16 & 57999.6925 & 0.0016 & $-$0&0024 & 19\
26 & 58000.5036 & 0.0019 & 0&0043 & 163\
27 & 58000.5817 & 0.0023 & 0&0019 & 188\
28 & 58000.6636 & 0.0024 & 0&0033 & 51\
40 & 58001.6261 & 0.0021 & 0&0004 & 23\
65 & 58003.6410 & 0.0023 & 0&0041 & 24\
77 & 58004.6038 & 0.0016 & 0&0015 & 19\
78 & 58004.6752 & 0.0023 & $-$0&0075 & 20\
\
\
\
ASASSN-17ei {#obj:asassn17ei}
-----------
This object was detected as a transient at $V$=13.0 on 2017 April 1 by the ASAS-SN team (cf. vsnet-alert 20849). Subsequent observations detected low-amplitude modulations (vsnet-alert 20857), which were later identified to be early superhumps (vsnet-alert 20867; e-figure \[fig:asassn17eieshpdm\]). The object started to show ordinary superhumps on 2017 April 9–10 (vsnet-alert 20894, 20900; e-figure \[fig:asassn17eishpdm\]).
The times of superhump maxima are listed in e-table \[tab:asassn17eioc2017\]. The data between $E$=53 and $E$=90 were not very good and the times of superhump maxima had relatively large uncertainties. It is, however, apparent that the superhump stage was already B at $E$=34–36, when the amplitudes of the superhumps may have not yet reached the maximum. An $O-C$ analysis has shown that times of low-amplitude maxima on April 9 were on the smooth extension of the supposed stage A before $E$=34 and we consider they were already stage A superhump rather than early superhumps (cf. e-figure \[fig:asassn17eihumpall\]). This identification appears to be supported by the duration of the phase of stage A superhumps (2 d), which is a normal value for a WZ Sge-type dwarf nova.
The period of early superhumps with the PDM method was 0.056460(9) d. The value for $\epsilon^*$ of stage A superhumps was 0.0275(9), which corresponds to $q$=0.074(3). This small $q$ is consistent with the small $P_{\rm dot}$ for stage B superhumps (cf. [@kat15wzsge]).
The entire light curve (e-figure \[fig:asassn17eihumpall\]) indicates that there was a plateau-type rebrightening. Note that original ASAS-SN measurements gave brighter values than our CCD measurements when the object was faint. This was due to the contamination since the aperture size is large in ASAS-SN measurements. We corrected this difference by subtracting a constant contribution of $V$=16.64 from neighboring stars (this value was determined to make the best match to our CCD measurements). There was some indication that the object faded slightly after the initial rise to the rebrightening phase, particularly at around BJD 2457873.5. This may be evidence of damping oscillation as seen in some WZ Sge-type dwarf novae (ASASSN-15po: [@nam17asassn15po], ASASSN-17el: in this paper).
The ASAS-SN started to observe this field on 2014 April 29 and no past outburst was detected. ASAS-3 recorded no outburst between 2001 and 2009.
(85mm,110mm)[asassn17eieshpdm.eps]{}
(85mm,110mm)[asassn17eishpdm.eps]{}
(160mm,200mm)[asassn17eihumpall.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lrrp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & & & & & & &\
0 & 57852.7066 & 0.0097 & $-$0&0153 & 20 & 140 & 57860.7367 & 0.0013 & $-$0&0045 & 11\
1 & 57852.7652 & 0.0028 & $-$0&0139 & 22 & 141 & 57860.7949 & 0.0014 & $-$0&0036 & 21\
2 & 57852.8263 & 0.0020 & $-$0&0101 & 22 & 142 & 57860.8527 & 0.0014 & $-$0&0031 & 24\
18 & 57853.7520 & 0.0008 & $-$0&0009 & 22 & 151 & 57861.3669 & 0.0010 & $-$0&0043 & 128\
34 & 57854.6820 & 0.0012 & 0&0126 & 18 & 152 & 57861.4237 & 0.0007 & $-$0&0048 & 132\
35 & 57854.7392 & 0.0010 & 0&0125 & 22 & 157 & 57861.7123 & 0.0043 & $-$0&0027 & 11\
36 & 57854.7962 & 0.0009 & 0&0122 & 29 & 158 & 57861.7678 & 0.0016 & $-$0&0044 & 14\
47 & 57855.4222 & 0.0006 & 0&0082 & 131 & 159 & 57861.8253 & 0.0013 & $-$0&0042 & 23\
48 & 57855.4804 & 0.0005 & 0&0090 & 132 & 160 & 57861.8845 & 0.0026 & $-$0&0023 & 14\
49 & 57855.5348 & 0.0006 & 0&0062 & 132 & 168 & 57862.3452 & 0.0010 & 0&0001 & 111\
50 & 57855.5936 & 0.0007 & 0&0077 & 133 & 169 & 57862.4016 & 0.0009 & $-$0&0007 & 129\
51 & 57855.6532 & 0.0009 & 0&0101 & 113 & 170 & 57862.4581 & 0.0010 & $-$0&0015 & 132\
53 & 57855.7604 & 0.0018 & 0&0027 & 21 & 171 & 57862.5161 & 0.0012 & $-$0&0008 & 132\
54 & 57855.8140 & 0.0025 & $-$0&0010 & 36 & 186 & 57863.3728 & 0.0009 & $-$0&0033 & 130\
72 & 57856.8387 & 0.0032 & $-$0&0073 & 36 & 187 & 57863.4334 & 0.0008 & 0&0000 & 132\
82 & 57857.4214 & 0.0034 & 0&0026 & 48 & 188 & 57863.4881 & 0.0009 & $-$0&0026 & 116\
89 & 57857.8257 & 0.0029 & 0&0059 & 36 & 191 & 57863.6579 & 0.0013 & $-$0&0046 & 19\
90 & 57857.8884 & 0.0030 & 0&0113 & 22 & 192 & 57863.7226 & 0.0018 & 0&0028 & 10\
104 & 57858.6792 & 0.0015 & 0&0001 & 17 & 193 & 57863.7724 & 0.0020 & $-$0&0047 & 14\
105 & 57858.7379 & 0.0014 & 0&0016 & 11 & 194 & 57863.8331 & 0.0028 & $-$0&0012 & 17\
106 & 57858.7922 & 0.0013 & $-$0&0014 & 15 & 203 & 57864.3520 & 0.0021 & 0&0021 & 86\
107 & 57858.8509 & 0.0011 & 0&0000 & 18 & 204 & 57864.4056 & 0.0011 & $-$0&0016 & 132\
116 & 57859.3597 & 0.0040 & $-$0&0067 & 64 & 205 & 57864.4624 & 0.0010 & $-$0&0020 & 132\
117 & 57859.4228 & 0.0005 & $-$0&0009 & 132 & 209 & 57864.6926 & 0.0041 & $-$0&0010 & 9\
118 & 57859.4798 & 0.0007 & $-$0&0012 & 131 & 211 & 57864.8076 & 0.0019 & $-$0&0005 & 15\
119 & 57859.5364 & 0.0008 & $-$0&0019 & 132 & 212 & 57864.8628 & 0.0023 & $-$0&0026 & 16\
120 & 57859.5945 & 0.0011 & $-$0&0010 & 112 & 221 & 57865.3853 & 0.0019 & 0&0044 & 131\
121 & 57859.6591 & 0.0009 & 0&0063 & 49 & 222 & 57865.4390 & 0.0013 & 0&0008 & 132\
122 & 57859.7005 & 0.0029 & $-$0&0096 & 11 & 223 & 57865.4932 & 0.0015 & $-$0&0023 & 132\
123 & 57859.7650 & 0.0009 & $-$0&0024 & 11 & 224 & 57865.5577 & 0.0022 & 0&0050 & 129\
124 & 57859.8239 & 0.0010 & $-$0&0007 & 17 & 225 & 57865.6120 & 0.0057 & 0&0020 & 82\
125 & 57859.8811 & 0.0019 & $-$0&0009 & 12 & 228 & 57865.7878 & 0.0031 & 0&0059 & 15\
134 & 57860.3945 & 0.0006 & $-$0&0030 & 132 & 244 & 57866.7173 & 0.0032 & 0&0189 & 15\
135 & 57860.4505 & 0.0006 & $-$0&0043 & 132 & 245 & 57866.7579 & 0.0025 & 0&0022 & 17\
136 & 57860.5094 & 0.0008 & $-$0&0027 & 132 & 246 & 57866.8194 & 0.0027 & 0&0065 & 21\
137 & 57860.5645 & 0.0007 & $-$0&0048 & 131 & 247 & 57866.8663 & 0.0046 & $-$0&0040 & 23\
139 & 57860.6812 & 0.0019 & $-$0&0027 & 14 & & & & &\
\
\
\
ASASSN-17el {#obj:asassn17el}
-----------
This object was detected as a transient at $V$=13.7 on 2017 April 1 by the ASAS-SN team (cf. vsnet-alert 20849). The object further brightened to an unfiltered CCD magnitude of 11.1 on April 2. The object immediately showed likely early superhumps (vsnet-alert 20866; e-figure \[fig:asassn17eleshpdm\]). These modulations were confirmed to be early superhumps by the detection of ordinary superhumps (vsnet-alert 20901, 20936; e-figure \[fig:asassn17elshpdm\]). The object was thereby confirmed to be a WZ Sge-type dwarf nova. The times of superhump maxima are listed in e-table \[tab:asassn17eloc2017\], which also includes post-superoutburst superhumps. There was typical stage B with a positive $P_{\rm dot}$ between $E$=48 and $E$=213 (e-figure \[fig:asassn17elhumpall\]). It looks like that the epoch $E$=36 was obtained during the final phase of stage A. Considering the typical duration of stage A of WZ Sge-type dwarf novae, we identified a hump maximum on BJD 2457852 as appearance of stage A superhumps. As the superoutburst plateau terminated, there was a apparent phase jump after $E$=213. We listed a period derived from $E$=230–271 as stage C superhumps. There were also superhumps during the rebrightening phase (vsnet-alert 20961, 20966, 20974) and the mean period was 0.05509(3) d ($E$=357–466).
The period of early superhumps was determined to be 0.05434(3) d by the PDM method. The period of stage A superhumps gave $\epsilon^*$=0.0265(10), which corresponds to $q$=0.071(3).
The light curve was composed of the main superoutburst, the main dip (BJD 2457865–2457867), a short rebrightening (BJD 2457868–2457869), a smaller dip (BJD 2457870.5) and a plateau-type rebrightening (BJD 2457871–2457878) (e-figure \[fig:asassn17elhumpall\]). This type of phenomenon (damping oscillation when entering the plateau-type rebrightening) was seen in the WZ Sge-type dwarf nova ASASSN-15po [@nam17asassn15po].
The object was recorded in outburst in 2006 by ASAS-3 (D. Denisenko, vsnet-alert 20853). It was also a superoutburst and reached at least $V$=11.6 (the peak may have been missed). This outburst was not accompanied by a rebrightening as far as the ASAS-3 data could tell. There were no other outbursts in the 9-year coverage in the ASAS-3 data and ASAS-SN observations since 2014 April. The outburst cycle length of 11 yr is typical for a WZ Sge-type dwarf nova [@kat15wzsge].
(85mm,110mm)[asassn17eleshpdm.eps]{}
(85mm,110mm)[asassn17elshpdm.eps]{}
(160mm,200mm)[asassn17elhumpall.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57852.5344 & 0.0021 & $-$0&0231 & 29\
36 & 57854.5439 & 0.0006 & 0&0017 & 18\
48 & 57855.2077 & 0.0008 & 0&0039 & 67\
49 & 57855.2626 & 0.0001 & 0&0037 & 127\
50 & 57855.3166 & 0.0002 & 0&0025 & 126\
54 & 57855.5362 & 0.0005 & 0&0016 & 17\
72 & 57856.5280 & 0.0004 & 0&0010 & 15\
85 & 57857.2438 & 0.0022 & 0&0002 & 39\
86 & 57857.2975 & 0.0002 & $-$0&0013 & 127\
87 & 57857.3514 & 0.0005 & $-$0&0026 & 114\
90 & 57857.5184 & 0.0005 & $-$0&0010 & 17\
108 & 57858.5097 & 0.0011 & $-$0&0020 & 14\
109 & 57858.5631 & 0.0022 & $-$0&0038 & 12\
121 & 57859.2262 & 0.0003 & $-$0&0022 & 78\
122 & 57859.2814 & 0.0003 & $-$0&0022 & 126\
123 & 57859.3366 & 0.0004 & $-$0&0021 & 98\
126 & 57859.5049 & 0.0023 & 0&0008 & 12\
127 & 57859.5595 & 0.0029 & 0&0003 & 12\
139 & 57860.2189 & 0.0004 & $-$0&0019 & 75\
140 & 57860.2754 & 0.0002 & $-$0&0005 & 127\
141 & 57860.3304 & 0.0005 & $-$0&0007 & 78\
145 & 57860.5505 & 0.0011 & $-$0&0011 & 17\
157 & 57861.2155 & 0.0004 & 0&0023 & 97\
158 & 57861.2691 & 0.0003 & 0&0008 & 127\
159 & 57861.3243 & 0.0004 & 0&0009 & 101\
163 & 57861.5457 & 0.0012 & 0&0017 & 18\
175 & 57862.2099 & 0.0005 & 0&0043 & 96\
176 & 57862.2658 & 0.0003 & 0&0052 & 126\
177 & 57862.3212 & 0.0014 & 0&0054 & 53\
193 & 57863.2025 & 0.0027 & 0&0046 & 61\
194 & 57863.2633 & 0.0005 & 0&0103 & 127\
195 & 57863.3168 & 0.0007 & 0&0087 & 114\
212 & 57864.2588 & 0.0006 & 0&0134 & 127\
213 & 57864.3132 & 0.0006 & 0&0127 & 89\
230 & 57865.2369 & 0.0014 & $-$0&0009 & 127\
231 & 57865.2887 & 0.0015 & $-$0&0042 & 117\
235 & 57865.5188 & 0.0034 & 0&0054 & 26\
248 & 57866.2222 & 0.0009 & $-$0&0080 & 115\
266 & 57867.2012 & 0.0056 & $-$0&0213 & 61\
267 & 57867.2707 & 0.0009 & $-$0&0070 & 127\
271 & 57867.4982 & 0.0038 & 0&0001 & 15\
357 & 57872.2431 & 0.0007 & 0&0036 & 126\
358 & 57872.2946 & 0.0017 & $-$0&0001 & 79\
375 & 57873.2298 & 0.0005 & $-$0&0021 & 127\
376 & 57873.2848 & 0.0007 & $-$0&0022 & 108\
430 & 57876.2653 & 0.0006 & 0&0012 & 124\
466 & 57878.2427 & 0.0065 & $-$0&0061 & 127\
\
\
\
ASASSN-17eq {#obj:asassn17eq}
-----------
This object was detected as a transient at $V$=13.7 on 2017 April 11 by the ASAS-SN team. Subsequent observations detected superhumps (vsnet-alert 20917, 20920; e-figure \[fig:asassn17eqshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17eqoc2017\]. Well-developed superhumps suggest that they were most likely stage B ones.
There was a normal outburst in the ASAS-SN data at $V$=14.3 on 2016 January 20, which faded to $V$=15.2 on the next night. There was no indication of a past superoutburst in the ASAS-SN data.
(85mm,110mm)[asassn17eqshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57857.5031 & 0.0002 & 0&0013 & 111\
10 & 57858.2251 & 0.0012 & 0&0013 & 52\
11 & 57858.2930 & 0.0029 & $-$0&0030 & 37\
12 & 57858.3679 & 0.0005 & $-$0&0002 & 29\
25 & 57859.3053 & 0.0024 & $-$0&0015 & 26\
26 & 57859.3783 & 0.0002 & $-$0&0006 & 95\
27 & 57859.4512 & 0.0003 & 0&0001 & 112\
28 & 57859.5259 & 0.0007 & 0&0026 & 44\
\
\
\
ASASSN-17es {#obj:asassn17es}
-----------
This object was detected as a transient at $V$=14.8 on 2017 April 11 by the ASAS-SN team. Subsequent observations detected early superhumps (vsnet-alert 20929, 20945; e-figure \[fig:asassn17eseshpdm\]), qualifying this object to be a WZ Sge-type dwarf nova. The object then showed ordinary superhumps (vsnet-alert 20972; e-figure \[fig:asassn17esshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17esoc2017\]. There was a rather smooth transition from stage A to B between $E$=0 and $E$=33. Assuming that the interval between $E$=0 and $E$=18 reflected stage A, we determined the period to be 0.05924(16) d. Using the best period of early superhumps \[0.05719(3) d, PDM method\], $\epsilon^*$ for stage A superhump is 0.037(4), which corresponds to $q$=0.095(9).
According to the ASAS-SN data, there was no past outburst and the 2017 outburst was detected up to May 20 (39 d after the initial rise) at $V$=15.7. Since the object was close to the detection limit in late May to June, the actual termination of the superoutburst was somewhat uncertain. Although the long duration of the outbursting stage suggested the combination of the superoutburst and a plateau-type rebrightening, there was no strong suggestion of a dip in the ASAS-SN data. Time-resolved photometry terminated on April 30 and these observations did not constrain the duration of the superoutburst.
(85mm,110mm)[asassn17eseshpdm.eps]{}
(85mm,110mm)[asassn17esshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57864.7608 & 0.0027 & $-$0&0213 & 12\
16 & 57865.7118 & 0.0027 & 0&0021 & 9\
17 & 57865.7675 & 0.0035 & $-$0&0001 & 13\
18 & 57865.8257 & 0.0017 & 0&0002 & 16\
33 & 57866.7008 & 0.0023 & 0&0058 & 15\
34 & 57866.7592 & 0.0021 & 0&0062 & 18\
35 & 57866.8137 & 0.0020 & 0&0028 & 21\
36 & 57866.8730 & 0.0021 & 0&0041 & 17\
46 & 57867.4518 & 0.0010 & 0&0032 & 19\
50 & 57867.6838 & 0.0022 & 0&0034 & 14\
51 & 57867.7464 & 0.0033 & 0&0080 & 16\
52 & 57867.7944 & 0.0017 & $-$0&0020 & 21\
53 & 57867.8555 & 0.0014 & 0&0012 & 24\
69 & 57868.7815 & 0.0032 & $-$0&0003 & 15\
70 & 57868.8390 & 0.0023 & $-$0&0008 & 16\
85 & 57869.7111 & 0.0025 & 0&0019 & 12\
86 & 57869.7671 & 0.0036 & $-$0&0001 & 13\
87 & 57869.8212 & 0.0025 & $-$0&0039 & 15\
88 & 57869.8853 & 0.0181 & 0&0022 & 5\
102 & 57870.6927 & 0.0029 & $-$0&0019 & 10\
103 & 57870.7468 & 0.0026 & $-$0&0057 & 10\
104 & 57870.8058 & 0.0054 & $-$0&0047 & 14\
105 & 57870.8683 & 0.0038 & $-$0&0003 & 10\
\
\
\
ASASSN-17et {#obj:asassn17et}
-----------
This object was detected as a transient at $V$=14.9 on 2017 April 10 by the ASAS-SN team. The outburst was announced after observation of $V$=15.1 on 2017 April 13. There is an X-ray counterpart 2XMM J175924.9$-$231503. Superhumps were immediately detected (vsnet-alert 20919; e-figure \[fig:asassn17etshpdm\]). Their period qualified this object to be an SU UMa-type dwarf nova in the period gap. The times of superhump maxima are listed in table \[tab:asassn17etoc2017\]. Since we observed the final part of the superoutburst, we tentatively classified the observed superhumps to be stage C ones. Since period variations of superhumps in long-period systems may not be similar to those in short-period systems (cf. [@Pdot3]; [@Pdot6]; [@kat16v1006cyg]), interpretation of superhumps in such a system may not be straightforward.
There was no past outburst detection in the ASAS-SN data starting on 2015 February 24.
(85mm,110mm)[asassn17etshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57858.7842 & 0.0015 & $-$0&0001 & 23\
1 & 57858.8814 & 0.0027 & 0&0015 & 15\
40 & 57862.6077 & 0.0015 & $-$0&0021 & 221\
50 & 57863.5650 & 0.0019 & $-$0&0011 & 220\
51 & 57863.6580 & 0.0029 & $-$0&0037 & 171\
52 & 57863.7583 & 0.0020 & 0&0009 & 21\
53 & 57863.8503 & 0.0032 & $-$0&0027 & 28\
60 & 57864.5318 & 0.0029 & 0&0094 & 203\
61 & 57864.6216 & 0.0042 & 0&0035 & 115\
62 & 57864.7144 & 0.0032 & 0&0007 & 14\
63 & 57864.8031 & 0.0037 & $-$0&0063 & 25\
\
\
\
ASASSN-17ew {#obj:asassn17ew}
-----------
This object was detected as a transient at $V$=14.9 on 2017 April 13 by the ASAS-SN team. Superhumps were immediately detected (vsnet-alert 20925; e-figure \[fig:asassn17ewshpdm\]). The times of superhump maxima are listed in table \[tab:asassn17ewoc2017\]. Since we observed the final part of the superoutburst, we tentatively classified the observed superhumps to be stage C ones.
There were two past outbursts (most likely normal ones) in the ASAS-SN data starting on 2016 February 19: 2016 June 30 ($V$=15.5) and 2017 January 7 ($V$=15.2).
(85mm,110mm)[asassn17ewshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57858.7842 & 0.0015 & $-$0&0001 & 23\
1 & 57858.8814 & 0.0027 & 0&0015 & 15\
40 & 57862.6077 & 0.0015 & $-$0&0021 & 221\
50 & 57863.5650 & 0.0019 & $-$0&0011 & 220\
51 & 57863.6580 & 0.0029 & $-$0&0037 & 171\
52 & 57863.7583 & 0.0020 & 0&0009 & 21\
53 & 57863.8503 & 0.0032 & $-$0&0027 & 28\
60 & 57864.5318 & 0.0029 & 0&0094 & 203\
61 & 57864.6216 & 0.0042 & 0&0035 & 115\
62 & 57864.7144 & 0.0032 & 0&0007 & 14\
63 & 57864.8031 & 0.0037 & $-$0&0063 & 25\
\
\
\
ASASSN-17ex {#obj:asassn17ex}
-----------
This object was detected as a transient at $V$=15.8 on 2017 April 10 by the ASAS-SN team. The outburst was announced after the observation at $V$=15.6 on 2017 April 13. Subsequent observations detected superhumps (vsnet-alert 20926, 20948; e-figure \[fig:asassn17exshpdm\]). The times of superhump maxima are listed in table \[tab:asassn17exoc2017\]. Since we observed the final part of the superoutburst, we tentatively classified the observed superhumps to be stage C ones.
There were no definite past outburst in the ASAS-SN data starting on 2015 September 18. There may have been a borderline detection at $V$=16.3 on 2017 January 28.
(85mm,110mm)[asassn17exshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57859.6447 & 0.0020 & $-$0&0003 & 20\
1 & 57859.7121 & 0.0016 & $-$0&0012 & 12\
15 & 57860.6704 & 0.0014 & 0&0008 & 17\
16 & 57860.7399 & 0.0026 & 0&0020 & 13\
29 & 57861.6252 & 0.0036 & $-$0&0007 & 14\
30 & 57861.6985 & 0.0041 & 0&0043 & 15\
31 & 57861.7576 & 0.0035 & $-$0&0049 & 11\
\
\
\
ASASSN-17fh {#obj:asassn17fh}
-----------
This object was detected as a transient at $V$=16.0 on 2017 April 20 by the ASAS-SN team. Only two superhumps were recorded in single-night observations: BJD 2457869.4704(12) ($N$=55) and 2457869.5343(8) ($N$=66) (vsnet-alert 20963; e-figure \[fig:asassn17fhshlc\]). The superhump stage was unknown.
The ASAS-SN data since 2013 June 4 did not detect other superoutbursts. There was possible single positive detection at $V$=16.2 on 2015 April 1, but it may have been a noise.
(85mm,110mm)[asassn17fhshlc.eps]{}
ASASSN-17fi {#obj:asassn17fi}
-----------
This object was detected as a transient at $V$=14.5 on 2017 April 20 by the ASAS-SN team. Subsequent observations detected superhumps (vsnet-alert 20952; e-figure \[fig:asassn17fishpdm\]). According to the ASAS-SN data, the outburst lasted at least up to 2017 April 27.
The ASAS-SN data covered this region since 2013 October 28. There is nearby $\sim$14.4 mag (Gaia G magnitude) star which affected ASAS-SN photometry. Although the detection limit was shallow due to this companion star, there was no indication of an outburst as bright as in the 2017 one.
(85mm,110mm)[asassn17fishpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57867.4925 & 0.0005 & $-$0&0000 & 62\
1 & 57867.5513 & 0.0004 & $-$0&0001 & 58\
2 & 57867.6102 & 0.0006 & 0&0001 & 55\
51 & 57870.4937 & 0.0004 & 0&0007 & 62\
52 & 57870.5511 & 0.0009 & $-$0&0007 & 32\
\
\
\
ASASSN-17fj {#obj:asassn17fj}
-----------
This object was detected as a transient at $V$=15.7 on 2017 April 21 by the ASAS-SN team. The outburst was announced after the observation at $V$=15.7 on April 22. Subsequent observations detected superhumps (vsnet-alert 20965; e-figure \[fig:asassn17fjshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17fjoc2017\]. Stages B and C can be identified.
This field has been observed by the ASAS-SN team since 2016 March 10. There was an apparent superoutburst on 2016 March 10 at $V$=15.7 and probably lasted at least up to March 15. The supercycle looks likely to be an order of a year.
(85mm,110mm)[asassn17fjshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57868.7539 & 0.0009 & $-$0&0016 & 15\
1 & 57868.8200 & 0.0010 & $-$0&0017 & 17\
2 & 57868.8903 & 0.0026 & 0&0025 & 9\
14 & 57869.6791 & 0.0012 & $-$0&0030 & 11\
15 & 57869.7466 & 0.0018 & $-$0&0017 & 13\
16 & 57869.8118 & 0.0014 & $-$0&0026 & 17\
17 & 57869.8797 & 0.0026 & $-$0&0009 & 10\
30 & 57870.7369 & 0.0012 & $-$0&0041 & 10\
31 & 57870.8060 & 0.0013 & $-$0&0011 & 16\
32 & 57870.8718 & 0.0046 & $-$0&0016 & 12\
46 & 57871.8011 & 0.0023 & 0&0011 & 17\
47 & 57871.8628 & 0.0036 & $-$0&0033 & 16\
59 & 57872.6630 & 0.0037 & 0&0027 & 12\
61 & 57872.7966 & 0.0017 & 0&0039 & 23\
62 & 57872.8598 & 0.0032 & 0&0010 & 25\
75 & 57873.7256 & 0.0017 & 0&0063 & 17\
76 & 57873.7914 & 0.0016 & 0&0059 & 23\
77 & 57873.8550 & 0.0028 & 0&0033 & 26\
90 & 57874.7175 & 0.0014 & 0&0055 & 18\
91 & 57874.7842 & 0.0017 & 0&0060 & 23\
92 & 57874.8479 & 0.0016 & 0&0035 & 25\
105 & 57875.7017 & 0.0037 & $-$0&0030 & 18\
106 & 57875.7775 & 0.0026 & 0&0065 & 24\
107 & 57875.8349 & 0.0030 & $-$0&0022 & 25\
120 & 57876.6957 & 0.0015 & $-$0&0018 & 15\
121 & 57876.7561 & 0.0022 & $-$0&0076 & 19\
122 & 57876.8256 & 0.0021 & $-$0&0043 & 25\
135 & 57877.6826 & 0.0026 & $-$0&0077 & 15\
\
\
\
ASASSN-17fl {#obj:asassn17fl}
-----------
This object was detected as a transient at $V$=16.3 on 2017 April 21 by the ASAS-SN team. The outburst was announced after several confirmatory observations up to 2017 April 26 ($V$=17.3). Although the object was very faint, superhumps were detected (vsnet-alert 20973; e-figure \[fig:asassn17flshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17floc2017\].
Although ASAS-SN observations of this field started on 2016 March 9, no past outburst was detected. Since the object is very faint, past outbursts may have escaped detection even if they existed.
(85mm,110mm)[asassn17flshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57870.7141 & 0.0079 & 0&0014 & 10\
1 & 57870.7761 & 0.0018 & 0&0008 & 13\
2 & 57870.8353 & 0.0019 & $-$0&0027 & 15\
16 & 57871.7179 & 0.0043 & 0&0031 & 11\
17 & 57871.7756 & 0.0017 & $-$0&0018 & 14\
18 & 57871.8392 & 0.0022 & $-$0&0009 & 17\
\
\
\
ASASSN-17fn {#obj:asassn17fn}
-----------
This object was detected as a transient at $V$=17.4 on 2017 April 22 by the ASAS-SN team. The outburst was announced after the observation at $V$=14.8 on 2017 April 26. There is also an X-ray counterpart 2RXP J103528.7$+$541910 (=2RXP J103527.8$+$541852). Subsequent observations detected likely early superhumps (vsnet-alert 20971, 20975; e-figure \[fig:asassn17fneshpdm\]). Ordinary superhumps appeared on May 7 and developed further (vsnet-alert 20997; e-figure \[fig:asassn17fnshpdm\]). Although initially detected superhumps were singly-peaked, the development of ordinary superhumps confirmed the identification as early superhumps. The object is confirmed to be a WZ Sge-type dwarf nova. The times of superhump maxima are listed in e-table \[tab:asassn17fnoc2017\].
Although stage A superhumps were only marginally recorded, the fractional excess $\epsilon^*$ was estimated to be 0.0352(3), which corresponds to $q$=0.097(1). The object appears to have a rather high $q$ despite the long orbital period. The relatively large $q$ appears to be consistent with the large superhump amplitudes (0.20 mag, e-figure \[fig:asassn17fnshpdm\]). WZ Sge-type dwarf novae with such features tend to show multiple rebrightenings [@nak13j2112j2037]. In this object, its faintness prevented us from detecting such rebrightenings.
There were no definite past outburst in the ASAS-SN data starting on 2013 October 9.
(85mm,110mm)[asassn17fneshpdm.eps]{}
(85mm,110mm)[asassn17fnshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57881.0463 & 0.0028 & $-$0&0353 & 85\
21 & 57882.3772 & 0.0018 & 0&0010 & 18\
22 & 57882.4362 & 0.0009 & $-$0&0016 & 51\
23 & 57882.4982 & 0.0010 & $-$0&0013 & 63\
24 & 57882.5620 & 0.0006 & 0&0008 & 63\
37 & 57883.3655 & 0.0013 & 0&0029 & 49\
38 & 57883.4289 & 0.0008 & 0&0047 & 72\
39 & 57883.4915 & 0.0005 & 0&0056 & 76\
40 & 57883.5531 & 0.0010 & 0&0056 & 61\
41 & 57883.6130 & 0.0100 & 0&0039 & 15\
53 & 57884.3553 & 0.0013 & 0&0063 & 27\
54 & 57884.4152 & 0.0006 & 0&0047 & 92\
55 & 57884.4751 & 0.0008 & 0&0028 & 61\
56 & 57884.5389 & 0.0008 & 0&0051 & 55\
65 & 57885.0902 & 0.0027 & 0&0015 & 77\
103 & 57887.4350 & 0.0008 & 0&0037 & 56\
104 & 57887.4966 & 0.0013 & 0&0037 & 40\
118 & 57888.3592 & 0.0011 & 0&0032 & 31\
119 & 57888.4223 & 0.0012 & 0&0046 & 30\
120 & 57888.4781 & 0.0009 & $-$0&0013 & 71\
121 & 57888.5398 & 0.0012 & $-$0&0012 & 44\
122 & 57888.6043 & 0.0055 & 0&0017 & 17\
135 & 57889.4043 & 0.0015 & 0&0003 & 31\
136 & 57889.4781 & 0.0036 & 0&0124 & 11\
150 & 57890.3236 & 0.0034 & $-$0&0052 & 19\
151 & 57890.3861 & 0.0013 & $-$0&0043 & 30\
152 & 57890.4492 & 0.0009 & $-$0&0028 & 73\
153 & 57890.5096 & 0.0032 & $-$0&0041 & 33\
154 & 57890.5690 & 0.0065 & $-$0&0063 & 13\
167 & 57891.3711 & 0.0031 & $-$0&0057 & 32\
168 & 57891.4329 & 0.0015 & $-$0&0055 & 31\
169 & 57891.5001 & 0.0022 & 0&0000 & 31\
\
\
\
ASASSN-17fo {#obj:asassn17fo}
-----------
This object was detected as a transient at $V$=16.0 on 2017 April 27 by the ASAS-SN team. Two previous outbursts were detected in the CRTS data. Subsequent observations detected deep eclipses and superhumps (vsnet-alert 20977; e-figures \[fig:asassn17foshlc\], \[fig:asassn17foshpdm\]).
The times of superhump maxima are listed in e-table \[tab:asassn17fooc2017\]. Except for $E$=0 all superhumps were probably observed during stage B.
We obtained the following eclipse ephemeris using the MCMC analysis [@Pdot4] on the quiescent observations in the CRTS data: $${\rm Min(BJD)} = 2457877.05448(13) + 0.061548044(3) E .
\label{equ:asassn17foecl}$$ The epoch corresponds to the center of our observations during the outburst. This period was uniquely determined both by the PDM method and the MCMC method, and is consistent with the period \[0.061536 d\] determined from observations during the current outburst. The orbital light curve (e-figure \[fig:asassn17foporb\]) indicates the presence of double-wave orbital variations and an orbital hump. The depth of eclipsed was probably underestimated since the object becomes too faint for CRTS at eclipse centers. The quiescent orbital light curve suggests a low-mass transfer rate, which is compatible with the relatively rare outbursts in the past data. Since the object appears to have a very high orbital inclination as judged from high amplitudes of superhumps and low outburst amplitudes, detailed analysis of the quiescent light curve is desired.
(85mm,110mm)[asassn17foshlc.eps]{}
(85mm,110mm)[asassn17foshpdm.eps]{}
(85mm,70mm)[asassn17foporb.eps]{}
[rp[50pt]{}p[30pt]{}r@[.]{}lcr]{} & & & & &\
0 & 57873.5174 & 0.0006 & 0&0018 & 0.52 & 16\
8 & 57874.0230 & 0.0003 & 0&0015 & 0.74 & 61\
9 & 57874.0859 & 0.0004 & 0&0012 & 0.76 & 84\
10 & 57874.1502 & 0.0003 & 0&0023 & 0.80 & 79\
16 & 57874.5253 & 0.0013 & $-$0&0020 & 0.90 & 17\
17 & 57874.5910 & 0.0014 & 0&0005 & 0.97 & 14\
32 & 57875.5386 & 0.0005 & $-$0&0004 & 0.37 & 21\
33 & 57875.6016 & 0.0005 & $-$0&0006 & 0.39 & 19\
39 & 57875.9846 & 0.0009 & 0&0030 & 0.61 & 46\
40 & 57876.0414 & 0.0007 & $-$0&0034 & 0.54 & 83\
41 & 57876.1054 & 0.0009 & $-$0&0027 & 0.58 & 99\
45 & 57876.3583 & 0.0006 & $-$0&0027 & 0.69 & 46\
46 & 57876.4215 & 0.0007 & $-$0&0027 & 0.71 & 27\
48 & 57876.5494 & 0.0008 & $-$0&0013 & 0.79 & 19\
49 & 57876.6113 & 0.0012 & $-$0&0027 & 0.80 & 17\
50 & 57876.6764 & 0.0004 & $-$0&0008 & 0.85 & 45\
51 & 57876.7473 & 0.0026 & 0&0069 & 0.01 & 38\
56 & 57877.0539 & 0.0005 & $-$0&0027 & 0.99 & 75\
57 & 57877.1174 & 0.0013 & $-$0&0024 & 0.02 & 72\
61 & 57877.3691 & 0.0034 & $-$0&0036 & 0.11 & 18\
63 & 57877.5021 & 0.0025 & 0&0029 & 0.27 & 20\
64 & 57877.5621 & 0.0022 & $-$0&0003 & 0.25 & 13\
72 & 57878.0675 & 0.0027 & $-$0&0008 & 0.46 & 62\
73 & 57878.1391 & 0.0028 & 0&0076 & 0.62 & 50\
79 & 57878.5120 & 0.0019 & 0&0011 & 0.68 & 19\
80 & 57878.5747 & 0.0023 & 0&0006 & 0.70 & 14\
\
\
\
\
ASASSN-17fp {#obj:asassn17fp}
-----------
This object was detected as a transient at $V$=16.2 on 2017 April 28 by the ASAS-SN team. The outburst was announced after the observation on 2017 April 28 ($V$=15.7). The object was spectroscopically studied on 2017 April 28 and found to show He I lines but lack hydrogen [@car17asassn17fpatel10334]. The object was then suggested to be a candidate AM CVn-type object. @mar17asassn17fpatel10354 reported 51.0(1)-min modulations on 2017 May 6 and suggested that the object is located at the long period end of the AM CVn period distribution. The object faded quickly after this, but it showed at least one further short outburst [@waa17asassn17fpaan580].
Our observations during the outburst were performed on 2017 May 4–5. A PDM analysis yielded a period of 0.0365(5) d \[52.6(7) min\], close to the value obtained by @mar17asassn17fpatel10354 (e-figure \[fig:asassn17fpshpdm\]).
This field has been monitored by the ASAS-SN team since 2014 April 29 and no past outbursts were recorded. An outburst at $V$=16.1 on 2017 May 16 was the only positive detection by the ASAS-SN team after the main outburst. This outburst is thus considered to be a post-superoutburst rebrightening.
(85mm,110mm)[asassn17fpshpdm.eps]{}
ASASSN-17fz {#obj:asassn17fz}
-----------
This object was detected as a transient at $V$=14.2 on 2017 May 5 by the ASAS-SN team. The large outburst amplitude received attention. Although superhump-like features were suggested (vsnet-alert 20992), they were not confirmed by later observations (vsnet-alert 21003). The object developed evolving ordinary superhumps on May 12 (7 d after the initial outburst detection; vsnet-alert 21017, 21026; e-figure \[fig:asassn17fzshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17fzoc2017\]. The maxima for $E \le$41 were clearly stage A superhumps.
Although a possible period of early superhumps was reported in vsnet-alert 21026, we could not confirm it by further analysis. We consider that the object is likely a WZ Sge-type dwarf nova based on long duration of stage A (suggesting a small $q$). The waiting time (7 d) of the appearance of ordinary superhump is slight shorter than in most WZ Sge-type dwarf novae. This may have been a result of the lack of observations on three nights preceding the initial ASAS-SN detection and the true waiting time may have been longer.
The object was still in superoutburst plateau on 2017 May 21. ASAS-SN observations on subsequent three nights were not deep enough to see whether the outburst still continued. The object was found to have faded to 19 mag on 2017 May 28. There was no indication of a post-superoutburst rebrightening although our CCD observations were not sufficient and ASAS-SN data were not deep enough to detect a rebrightening. ASAS-SN started regularly observing this field on 2016 February 2 and there were no previous outbursts.
(85mm,110mm)[asassn17fzshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57886.2577 & 0.0062 & $-$0&0176 & 95\
1 & 57886.3163 & 0.0021 & $-$0&0135 & 125\
2 & 57886.3822 & 0.0076 & $-$0&0021 & 106\
17 & 57887.2023 & 0.0007 & 0&0007 & 91\
18 & 57887.2579 & 0.0005 & 0&0018 & 125\
19 & 57887.3101 & 0.0007 & $-$0&0005 & 125\
20 & 57887.3665 & 0.0008 & 0&0014 & 125\
23 & 57887.5312 & 0.0013 & 0&0027 & 14\
36 & 57888.2445 & 0.0003 & 0&0075 & 124\
37 & 57888.2989 & 0.0004 & 0&0075 & 124\
38 & 57888.3545 & 0.0004 & 0&0086 & 123\
41 & 57888.5185 & 0.0009 & 0&0091 & 13\
60 & 57889.5480 & 0.0008 & 0&0033 & 12\
72 & 57890.2037 & 0.0006 & 0&0051 & 103\
73 & 57890.2557 & 0.0006 & 0&0027 & 124\
74 & 57890.3085 & 0.0005 & 0&0010 & 124\
75 & 57890.3622 & 0.0007 & 0&0002 & 117\
78 & 57890.5287 & 0.0018 & 0&0032 & 12\
91 & 57891.2335 & 0.0006 & $-$0&0004 & 125\
92 & 57891.2885 & 0.0005 & 0&0001 & 125\
93 & 57891.3429 & 0.0010 & $-$0&0000 & 96\
109 & 57892.2123 & 0.0005 & $-$0&0024 & 118\
110 & 57892.2668 & 0.0005 & $-$0&0024 & 125\
111 & 57892.3228 & 0.0007 & $-$0&0009 & 125\
112 & 57892.3707 & 0.0033 & $-$0&0074 & 63\
115 & 57892.5393 & 0.0015 & $-$0&0024 & 11\
133 & 57893.5174 & 0.0023 & $-$0&0051 & 12\
151 & 57894.4994 & 0.0013 & $-$0&0038 & 17\
152 & 57894.5615 & 0.0048 & 0&0037 & 8\
\
\
\
ASASSN-17gf {#obj:asassn17gf}
-----------
This object was detected as a transient at $V$=13.05 on 2017 May 14 by the ASAS-SN team. Superhumps soon grew and the object was identified as an SU UMa-type dwarf nova below the period minimum (vsnet-alert 21021; e-figure \[fig:asassn17gfshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17gfoc2017\]. The presence of stages A and B is clearly visible.
Although our CCD observations of the superoutburst ended on 2017 May 21, the object was confirmed to be bright at least up to May 26 at a visual magnitude of 14.0 according to the AAVSO observations. The object was found to be fading rapidly on 2017 May 27. There was no indication of post-superoutburst rebrightening. The object was still in quiescence on 2017 May 10. The duration of the superoutburst was between 13 d and 17 d. This duration is typical for an SU UMa-type dwarf nova, and is unlike WZ Sge-type objects below the period minimum, such as ASASSN-15po [@nam17asassn15po] and OV Boo, which are possibly population II dwarf novae (e.g. [@pat08j1507]; [@nam17asassn15po]; [@ohn19ovboo]). ASASSN-17gf is thus most likely an EI Psc-like object. EI Psc-like systems are CVs below the period minimum showing hydrogen (likely somewhat reduced in abundance) in their spectra (cf. [@tho02j2329]; [@uem02j2329letter]; [@lit13sbs1108]) and are considered to be evolving towards AM CVn-type objects.
This object has been monitored by the ASAS-SN team since 2016 February 2 and there has been no other outburst in the ASAS-SN data. There were, however, rather numerous outburst detections in the ASAS-3 data (e-table \[tab:asassn17gfout\]; since there is a possible nearby contaminating object, only outbursts with more than two positive detections are listed). The listed outbursts were all likely superoutbursts and the frequency of superoutbursts appears to be comparable to or even higher than EI Psc ().
(85mm,110mm)[asassn17gfshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57888.5567 & 0.0013 & $-$0&0124 & 10\
12 & 57889.2046 & 0.0011 & 0&0046 & 51\
13 & 57889.2498 & 0.0001 & $-$0&0027 & 121\
14 & 57889.3025 & 0.0002 & $-$0&0026 & 121\
15 & 57889.3569 & 0.0002 & $-$0&0008 & 121\
16 & 57889.4083 & 0.0002 & $-$0&0020 & 121\
18 & 57889.5174 & 0.0005 & 0&0020 & 11\
19 & 57889.5717 & 0.0010 & 0&0037 & 9\
20 & 57889.6224 & 0.0003 & 0&0019 & 10\
31 & 57890.2035 & 0.0002 & 0&0047 & 92\
32 & 57890.2549 & 0.0002 & 0&0035 & 121\
33 & 57890.3068 & 0.0001 & 0&0028 & 121\
34 & 57890.3588 & 0.0002 & 0&0022 & 121\
35 & 57890.4076 & 0.0034 & $-$0&0015 & 32\
37 & 57890.5159 & 0.0006 & 0&0016 & 12\
38 & 57890.5689 & 0.0009 & 0&0020 & 10\
39 & 57890.6219 & 0.0008 & 0&0024 & 10\
50 & 57891.1983 & 0.0004 & 0&0005 & 74\
51 & 57891.2516 & 0.0002 & 0&0013 & 114\
52 & 57891.3030 & 0.0003 & 0&0001 & 106\
53 & 57891.3564 & 0.0002 & 0&0009 & 100\
54 & 57891.4083 & 0.0003 & 0&0002 & 79\
56 & 57891.5126 & 0.0015 & $-$0&0006 & 12\
57 & 57891.5649 & 0.0010 & $-$0&0009 & 10\
58 & 57891.6162 & 0.0011 & $-$0&0021 & 10\
69 & 57892.1970 & 0.0007 & 0&0003 & 62\
70 & 57892.2486 & 0.0002 & $-$0&0006 & 121\
71 & 57892.3014 & 0.0002 & $-$0&0004 & 121\
72 & 57892.3534 & 0.0002 & $-$0&0010 & 121\
75 & 57892.5106 & 0.0006 & $-$0&0016 & 12\
76 & 57892.5662 & 0.0008 & 0&0015 & 10\
77 & 57892.6176 & 0.0006 & 0&0004 & 9\
89 & 57893.2461 & 0.0002 & $-$0&0021 & 121\
90 & 57893.2995 & 0.0002 & $-$0&0012 & 121\
91 & 57893.3518 & 0.0003 & $-$0&0015 & 121\
92 & 57893.4036 & 0.0003 & $-$0&0022 & 121\
94 & 57893.5108 & 0.0015 & $-$0&0002 & 11\
95 & 57893.5597 & 0.0020 & $-$0&0039 & 10\
96 & 57893.6183 & 0.0009 & 0&0021 & 9\
108 & 57894.2454 & 0.0006 & $-$0&0017 & 68\
109 & 57894.2985 & 0.0005 & $-$0&0011 & 84\
113 & 57894.5092 & 0.0006 & $-$0&0007 & 18\
127 & 57895.2468 & 0.0003 & 0&0009 & 118\
128 & 57895.3005 & 0.0005 & 0&0020 & 106\
129 & 57895.3534 & 0.0007 & 0&0023 & 56\
\
\
\
[ccccc]{} Year & Month & Day & max& $V$-mag\
2002 & 10 & 11 & 52559 & 13.00\
2003 & 7 & 10 & 52830 & 13.27\
2004 & 1 & 3 & 53008 & 13.61\
2004 & 5 & 19 & 53145 & 13.56\
2005 & 2 & 11 & 53412 & 13.04\
2005 & 1 & 21 & 53656 & 12.76\
2008 & 2 & 5 & 54502 & 13.27\
\
ASASSN-17gh {#obj:asassn17gh}
-----------
This object was detected as a transient at $V$=16.2 on 2017 May 16 by the ASAS-SN team. The outburst was announced after the observation at $V$=15.9 on 2017 May 16. Subsequent observations detected superhumps (vsnet-alert 21048; e-figure \[fig:asassn17ghshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17ghoc2017\]. This table does not include observations on the second night with less quality to determine the times of maxima. The night, however, was included in the PDM analysis giving a period of 0.0608(1) d, which is considered to be more reliable than the one from single-night $O-C$ analysis.
(85mm,110mm)[asassn17ghshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57894.0326 & 0.0009 & 0&0019 & 91\
1 & 57894.0949 & 0.0013 & 0&0028 & 97\
2 & 57894.1480 & 0.0037 & $-$0&0054 & 52\
6 & 57894.3972 & 0.0009 & $-$0&0018 & 31\
7 & 57894.4623 & 0.0004 & 0&0018 & 63\
8 & 57894.5218 & 0.0005 & $-$0&0000 & 62\
9 & 57894.5839 & 0.0009 & 0&0007 & 31\
\
\
\
ASASSN-17gv {#obj:asassn17gv}
-----------
This object was detected as a transient at $V$=14.2 on 2017 May 28 by the ASAS-SN team. Subsequent observations detected superhumps (vsnet-alert 21065, 21080; e-figure \[fig:asassn17gvshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17gvoc2017\]. Although superhumps apparently had not yet fully developed on the first night and a period of stage A superhumps was reported in vsnet-alert 21080, this value was not confirmed by this $O-C$ analysis. The difference between periods of stage A and B superhumps was too large (more than 4%) and we consider that this period was a spurious detection. Although a positive $P_{\rm dot}$ is expected for this superhump period, this $O-C$ analysis did not give such a tendency. The data were probably of limited quality and the data may have also been contaminated by stage C superhumps.
The outburst lasted at least up to June 4, when the ASAS-SN data apparently showed rapid fading. The object faded to $\sim$18.5 mag on June 9 as recorded by our CCD observations. The region of this object has been covered by the ASAS-SN team since 2016 March 8 and no other outburst was recorded. Although ASAS-3 did not detect any outburst, it may have been due to crowding in this region.
(85mm,110mm)[asassn17gvshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57903.2021 & 0.0056 & $-$0&0117 & 45\
2 & 57903.3428 & 0.0021 & 0&0072 & 139\
3 & 57903.3959 & 0.0005 & $-$0&0006 & 139\
6 & 57903.5845 & 0.0008 & 0&0053 & 16\
25 & 57904.7355 & 0.0022 & $-$0&0007 & 16\
26 & 57904.7965 & 0.0014 & $-$0&0007 & 17\
33 & 57905.2243 & 0.0006 & 0&0009 & 109\
34 & 57905.2812 & 0.0003 & $-$0&0031 & 129\
35 & 57905.3522 & 0.0010 & 0&0070 & 38\
41 & 57905.7167 & 0.0036 & 0&0061 & 12\
42 & 57905.7702 & 0.0020 & $-$0&0013 & 13\
55 & 57906.5626 & 0.0023 & $-$0&0005 & 10\
57 & 57906.6801 & 0.0023 & $-$0&0048 & 13\
58 & 57906.7395 & 0.0030 & $-$0&0063 & 12\
59 & 57906.8069 & 0.0009 & 0&0002 & 13\
66 & 57907.2332 & 0.0014 & 0&0002 & 141\
67 & 57907.2969 & 0.0009 & 0&0030 & 140\
68 & 57907.3536 & 0.0009 & $-$0&0012 & 102\
73 & 57907.6543 & 0.0034 & $-$0&0050 & 11\
74 & 57907.7217 & 0.0031 & 0&0015 & 12\
75 & 57907.7802 & 0.0042 & $-$0&0009 & 13\
88 & 57908.5780 & 0.0023 & 0&0053 & 9\
\
\
\
ASASSN-17hm {#obj:asassn17hm}
-----------
This object was detected as a transient at $V$=14.2 on 2017 June 10 by the ASAS-SN team. The outburst was announced after observations on 2017 June 11 ($V$=14.3) and 2017 June 12 ($V$=14.6). The object was not yet in outburst on June 9. Subsequent observations detected superhumps (vsnet-alert 21117; e-figure \[fig:asassn17hmshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17hmoc2017\]. The epochs for $E \ge$68 refer to post-superoutburst superhumps. Although there was an apparent change in the period around the termination of the superoutburst, no phase jump was recorded.
This field has been monitored by the ASAS-SN team since 2016 March 9 and there was another outburst at $V$=14.1 on 2016 July 16. There were gaps before and after this observation and the outburst type is unknown.
(85mm,110mm)[asassn17hmshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57918.5083 & 0.0011 & $-$0&0021 & 22\
2 & 57918.6894 & 0.0008 & 0&0022 & 18\
3 & 57918.7739 & 0.0014 & $-$0&0018 & 16\
11 & 57919.4738 & 0.0056 & $-$0&0091 & 10\
12 & 57919.5709 & 0.0017 & $-$0&0005 & 10\
13 & 57919.6589 & 0.0027 & $-$0&0009 & 18\
14 & 57919.7482 & 0.0009 & 0&0000 & 19\
23 & 57920.5496 & 0.0021 & 0&0058 & 14\
24 & 57920.6349 & 0.0030 & 0&0027 & 20\
25 & 57920.7225 & 0.0012 & 0&0019 & 21\
34 & 57921.5183 & 0.0018 & 0&0020 & 21\
36 & 57921.6985 & 0.0012 & 0&0054 & 20\
37 & 57921.7850 & 0.0027 & 0&0034 & 12\
45 & 57922.4888 & 0.0040 & 0&0000 & 12\
48 & 57922.7553 & 0.0022 & 0&0012 & 16\
58 & 57923.6347 & 0.0015 & $-$0&0035 & 17\
59 & 57923.7243 & 0.0020 & $-$0&0022 & 16\
68 & 57924.5175 & 0.0016 & $-$0&0047 & 20\
69 & 57924.6084 & 0.0016 & $-$0&0023 & 13\
70 & 57924.6962 & 0.0007 & $-$0&0028 & 16\
81 & 57925.6702 & 0.0044 & $-$0&0014 & 16\
82 & 57925.7666 & 0.0051 & 0&0066 & 12\
\
\
\
ASASSN-17hw {#obj:asassn17hw}
-----------
This object was detected as a transient at $V$=13.2 on 2017 June 20 by the ASAS-SN team. The object started to show superhumps on 2017 July 2 (vsnet-alert 21194, 21208, 21209, 21219; e-figure \[fig:asassn17hwshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17hwoc2017\]. There are unambiguous stages A and B. It is remarkable that $P_{\rm dot}$ is almost zero during stage B (cf. vsnet-alert 21251).
An analysis of the early part of the outburst yielded small-amplitude early superhumps (e-figure \[fig:asassn17hweshpdm\]), confirming the WZ Sge-type nature of this object. The period of early superhumps determined by the PDM method was 0.05886(2) d. Combined with the period of stage A superhumps \[0.060617(5) d\], the value of $\epsilon^*$ for stage A is 0.0290(4), which corresponds to $q$=0.078(1).
Although the object was well monitored after the termination of the outburst, no rebrightening was recorded.
(85mm,110mm)[asassn17hwshpdm.eps]{}
(85mm,110mm)[asassn17hweshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57936.7251 & 0.0012 & $-$0&0170 & 13\
1 & 57936.7829 & 0.0007 & $-$0&0189 & 10\
11 & 57937.3901 & 0.0014 & $-$0&0093 & 43\
16 & 57937.6948 & 0.0004 & $-$0&0033 & 11\
17 & 57937.7565 & 0.0007 & $-$0&0014 & 13\
27 & 57938.3594 & 0.0006 & 0&0039 & 159\
28 & 57938.4200 & 0.0006 & 0&0048 & 175\
29 & 57938.4830 & 0.0011 & 0&0080 & 81\
33 & 57938.7196 & 0.0006 & 0&0056 & 13\
34 & 57938.7779 & 0.0010 & 0&0041 & 10\
44 & 57939.3763 & 0.0020 & 0&0050 & 23\
45 & 57939.4353 & 0.0008 & 0&0042 & 36\
46 & 57939.4966 & 0.0026 & 0&0057 & 14\
60 & 57940.3362 & 0.0031 & 0&0088 & 16\
61 & 57940.3925 & 0.0016 & 0&0052 & 18\
62 & 57940.4410 & 0.0061 & $-$0&0059 & 22\
63 & 57940.4977 & 0.0022 & $-$0&0090 & 13\
66 & 57940.6895 & 0.0011 & 0&0035 & 12\
67 & 57940.7509 & 0.0017 & 0&0052 & 12\
83 & 57941.7080 & 0.0018 & 0&0061 & 12\
84 & 57941.7617 & 0.0040 & 0&0000 & 10\
93 & 57942.3023 & 0.0028 & 0&0029 & 137\
94 & 57942.3632 & 0.0021 & 0&0040 & 137\
95 & 57942.4198 & 0.0021 & 0&0008 & 105\
100 & 57942.7209 & 0.0022 & 0&0031 & 13\
117 & 57943.7331 & 0.0023 & $-$0&0006 & 12\
133 & 57944.6956 & 0.0024 & 0&0059 & 12\
134 & 57944.7522 & 0.0015 & 0&0027 & 11\
143 & 57945.2867 & 0.0013 & $-$0&0007 & 137\
144 & 57945.3468 & 0.0009 & $-$0&0003 & 138\
145 & 57945.4042 & 0.0016 & $-$0&0026 & 137\
146 & 57945.4630 & 0.0042 & $-$0&0037 & 79\
150 & 57945.7034 & 0.0019 & $-$0&0022 & 13\
176 & 57947.2585 & 0.0010 & $-$0&0009 & 129\
177 & 57947.3197 & 0.0010 & 0&0006 & 136\
178 & 57947.3807 & 0.0011 & 0&0018 & 134\
179 & 57947.4415 & 0.0010 & 0&0029 & 128\
198 & 57948.5587 & 0.0023 & $-$0&0153 & 15\
199 & 57948.6312 & 0.0039 & $-$0&0026 & 17\
201 & 57948.7551 & 0.0057 & 0&0018 & 15\
214 & 57949.5249 & 0.0025 & $-$0&0052 & 16\
218 & 57949.7717 & 0.0046 & 0&0025 & 10\
\
\
\
ASASSN-17hy {#obj:asassn17hy}
-----------
This object was detected as a transient at $V$=16.1 on 2017 June 18 by the ASAS-SN team. The outburst was announced after observations on 2017 June 19 ($V$=16.2). Subsequent observations detected superhumps (vsnet-alert 21157; e-figure \[fig:asassn17hyshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17hyoc2017\]. The $O-C$ data suggest a positive $P_{\rm dot}$ although the data were relatively well sampled only on four nights.
(85mm,110mm)[asassn17hyshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57929.4948 & 0.0018 & 0&0019 & 22\
1 & 57929.5704 & 0.0012 & 0&0060 & 20\
2 & 57929.6395 & 0.0013 & 0&0036 & 21\
28 & 57931.4910 & 0.0012 & $-$0&0038 & 17\
29 & 57931.5633 & 0.0009 & $-$0&0031 & 15\
30 & 57931.6346 & 0.0018 & $-$0&0033 & 15\
42 & 57932.4927 & 0.0012 & $-$0&0031 & 16\
43 & 57932.5663 & 0.0013 & $-$0&0011 & 14\
56 & 57933.4965 & 0.0054 & $-$0&0004 & 16\
57 & 57933.5645 & 0.0021 & $-$0&0039 & 15\
58 & 57933.6342 & 0.0029 & $-$0&0057 & 14\
70 & 57934.4953 & 0.0023 & $-$0&0026 & 16\
71 & 57934.5719 & 0.0020 & 0&0025 & 17\
72 & 57934.6502 & 0.0054 & 0&0093 & 10\
100 & 57936.6466 & 0.0042 & 0&0037 & 18\
\
\
\
ASASSN-17id {#obj:asassn17id}
-----------
This object was detected as a transient at $V$=16.8 on 2017 June 21 by the ASAS-SN team. The outburst was announced after observations on 2017 June 23 ($V$=16.8). Subsequent observations detected superhumps (vsnet-alert 21158, 21172; e-figure \[fig:asassn17idshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17idoc2017\]. Although there were observations after BJD 2457933, the object became too faint to detect superhumps. The object faded to 19.6 mag on 2017 July 2.
(85mm,110mm)[asassn17idshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57929.4795 & 0.0025 & $-$0&0019 & 22\
1 & 57929.5587 & 0.0027 & $-$0&0013 & 23\
2 & 57929.6406 & 0.0034 & 0&0020 & 22\
26 & 57931.5270 & 0.0023 & 0&0017 & 18\
27 & 57931.6065 & 0.0039 & 0&0026 & 18\
38 & 57932.4634 & 0.0039 & $-$0&0052 & 12\
39 & 57932.5493 & 0.0019 & 0&0021 & 17\
\
\
\
ASASSN-17if {#obj:asassn17if}
-----------
This object was detected as a transient at $V$=14.0 on 2017 June 25 by the ASAS-SN team. The outburst was announced after observations on 2017 June 26 ($V$=14.1). Subsequent observations detected superhumps (vsnet-alert 21183; e-figure \[fig:asassn17ifshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17ifoc2017\]. The stages B and C were clearly recorded despite relatively low sampling rates. The positive $P_{\rm dot}$ for stage B superhumps was typical for this $P_{\rm SH}$.
This field has been monitored by the ASAS-SN team since 2013 June 27 and no past outbursts were recorded. It may be that outbursts were relatively rare or past outbursts occurred when the object was invisible near the Sun.
(85mm,110mm)[asassn17ifshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57931.8329 & 0.0004 & 0&0074 & 13\
17 & 57932.8273 & 0.0004 & 0&0022 & 16\
34 & 57933.8234 & 0.0003 & $-$0&0014 & 16\
35 & 57933.8810 & 0.0003 & $-$0&0026 & 18\
68 & 57935.8179 & 0.0016 & $-$0&0063 & 21\
69 & 57935.8776 & 0.0020 & $-$0&0054 & 17\
86 & 57936.8767 & 0.0008 & $-$0&0059 & 21\
102 & 57937.8201 & 0.0012 & $-$0&0034 & 15\
103 & 57937.8789 & 0.0008 & $-$0&0034 & 17\
119 & 57938.8203 & 0.0011 & $-$0&0029 & 16\
120 & 57938.8808 & 0.0011 & $-$0&0012 & 18\
121 & 57938.9408 & 0.0025 & $-$0&0000 & 11\
137 & 57939.8838 & 0.0034 & 0&0021 & 11\
153 & 57940.8364 & 0.0034 & 0&0138 & 9\
154 & 57940.8868 & 0.0018 & 0&0055 & 11\
170 & 57941.8285 & 0.0041 & 0&0063 & 9\
171 & 57941.8877 & 0.0013 & 0&0067 & 11\
187 & 57942.8249 & 0.0027 & 0&0030 & 9\
188 & 57942.8854 & 0.0024 & 0&0048 & 11\
204 & 57943.8218 & 0.0041 & 0&0003 & 8\
205 & 57943.8810 & 0.0024 & 0&0006 & 10\
221 & 57944.8093 & 0.0030 & $-$0&0119 & 9\
222 & 57944.8782 & 0.0017 & $-$0&0018 & 14\
223 & 57944.9320 & 0.0022 & $-$0&0068 & 10\
\
\
\
ASASSN-17ig {#obj:asassn17ig}
-----------
This object was detected as a transient at $V$=14.8 on 2017 June 24 by the ASAS-SN team. Subsequent observations detected long-period superhumps (vsnet-alert 21168, 21184; e-figure \[fig:asassn17igshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17igoc2017\]. The superhump period was initially longer, suggesting that they were stage B superhumps (less likely stage A ones, since the amplitudes were already large).
This field has been monitored by the ASAS-SN team since 2015 February 1 and no past outburst was detected. Since the object is an SU UMa-type dwarf nova in the period gap, further observations would be interesting.
(85mm,110mm)[asassn17igshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57931.6531 & 0.0024 & $-$0&0084 & 12\
1 & 57931.7519 & 0.0006 & $-$0&0042 & 30\
12 & 57932.7977 & 0.0016 & 0&0021 & 19\
14 & 57932.9861 & 0.0008 & 0&0015 & 123\
15 & 57933.0785 & 0.0005 & $-$0&0006 & 136\
22 & 57933.7444 & 0.0007 & 0&0037 & 30\
25 & 57934.0283 & 0.0005 & 0&0041 & 190\
32 & 57934.6898 & 0.0007 & 0&0040 & 29\
54 & 57936.7688 & 0.0009 & 0&0039 & 17\
64 & 57937.7102 & 0.0010 & 0&0002 & 20\
75 & 57938.7479 & 0.0013 & $-$0&0016 & 20\
85 & 57939.6946 & 0.0022 & $-$0&0001 & 19\
96 & 57940.7296 & 0.0043 & $-$0&0046 & 19\
\
\
\
ASASSN-17il {#obj:asassn17il}
-----------
This object was detected as a transient at $V$=15.3 on 2017 June 30 by the ASAS-SN team. Although observations detected superhump-like variations (vsnet-alert 21202; e-figure \[fig:asassn17illc\]), we have not been able to determine a unique period using the data on two nights. Observations on July 6 could not detect similar variations. Although the SDSS colors in quiescence suggest an object below the period gap (cf. [@kat12DNSDSS]), the variations detected on the first night may have been large-amplitude random variations. Further observations are needed to clarify the nature of this object.
(85mm,80mm)[asassn17illc.eps]{}
ASASSN-17iv {#obj:asassn17iv}
-----------
This object was detected as a transient at $V$=16.8 on 2017 June 30 by the ASAS-SN team. The outburst was announced after observations on 2017 July 2 ($V$=16.4) and 2017 July 3 ($V$=16.4). Subsequent observations starting on July 5 detected superhumps (vsnet-alert 21245; e-figure \[fig:asassn17ivshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17ivoc2017\]. Maxima for $E \le$1 had negative $O-C$ values and there appears to have been a stage transition between $E$=1 and $E$=15. Since the observation started relatively late after the initial outburst detection, the majority of observations apparently recorded stage C superhumps. We adopted a period for $E \ge$15 following this interpretation.
(85mm,110mm)[asassn17ivshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57939.8022 & 0.0014 & $-$0&0061 & 9\
1 & 57939.8725 & 0.0011 & $-$0&0061 & 12\
15 & 57940.8666 & 0.0030 & 0&0036 & 12\
16 & 57940.9377 & 0.0036 & 0&0044 & 8\
29 & 57941.8466 & 0.0045 & $-$0&0008 & 11\
30 & 57941.9190 & 0.0031 & 0&0013 & 13\
43 & 57942.8348 & 0.0017 & 0&0030 & 11\
44 & 57942.9078 & 0.0069 & 0&0057 & 14\
57 & 57943.8225 & 0.0015 & 0&0063 & 10\
58 & 57943.8829 & 0.0014 & $-$0&0036 & 14\
71 & 57944.8045 & 0.0026 & 0&0038 & 13\
72 & 57944.8646 & 0.0062 & $-$0&0064 & 16\
73 & 57944.9431 & 0.0044 & 0&0018 & 12\
85 & 57945.7878 & 0.0038 & 0&0027 & 11\
86 & 57945.8479 & 0.0023 & $-$0&0075 & 14\
87 & 57945.9234 & 0.0041 & $-$0&0023 & 15\
\
\
\
ASASSN-17iw {#obj:asassn17iw}
-----------
This object was detected as a transient at $V$=16.6 on 2017 June 28 by the ASAS-SN team. The outburst was announced after observations on 2017 June 30 ($V$=16.5) and 2017 July 1 ($V$=17.0). Subsequent observations starting on July 4 detected superhumps (vsnet-alert 21246; e-figure \[fig:asassn17iwshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17iwoc2017\]. Although the $O-C$ diagram is noisy due to the faintness of the object, there was apparently stage B-C transition around $E$=90.
(85mm,110mm)[asassn17iwshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57939.5829 & 0.0031 & 0&0077 & 14\
1 & 57939.6286 & 0.0022 & $-$0&0025 & 13\
18 & 57940.5839 & 0.0020 & 0&0028 & 14\
19 & 57940.6338 & 0.0022 & $-$0&0031 & 13\
35 & 57941.5252 & 0.0019 & $-$0&0058 & 17\
52 & 57942.4720 & 0.0021 & $-$0&0090 & 15\
53 & 57942.5387 & 0.0038 & 0&0018 & 17\
53 & 57942.5394 & 0.0071 & 0&0026 & 17\
55 & 57942.6448 & 0.0024 & $-$0&0038 & 12\
70 & 57943.4820 & 0.0026 & $-$0&0049 & 17\
71 & 57943.5418 & 0.0027 & $-$0&0009 & 17\
72 & 57943.5997 & 0.0026 & 0&0011 & 13\
73 & 57943.6646 & 0.0040 & 0&0101 & 6\
88 & 57944.4920 & 0.0018 & $-$0&0007 & 17\
89 & 57944.5504 & 0.0034 & 0&0018 & 16\
90 & 57944.6134 & 0.0024 & 0&0089 & 13\
108 & 57945.6045 & 0.0043 & $-$0&0059 & 13\
\
\
\
ASASSN-17ix {#obj:asassn17ix}
-----------
This object was detected as a transient at $V$=15.2 on 2017 June 29 by the ASAS-SN team. The outburst was announced after observations between 2017 June 30 ($V$=15.5) and 2017 July 2 ($V$=15.6). Subsequent observations starting on 2017 July 4 detected superhumps (vsnet-alert 21247; e-figure \[fig:asassn17ixshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17ixoc2017\]. Although the $O-C$ diagram is noisy due to the faintness of the object, there was apparently stage B-C transition around $E$=82.
(85mm,110mm)[asassn17ixshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57939.4779 & 0.0013 & 0&0043 & 19\
1 & 57939.5362 & 0.0024 & 0&0001 & 19\
2 & 57939.6021 & 0.0030 & 0&0036 & 16\
3 & 57939.6608 & 0.0020 & $-$0&0001 & 10\
16 & 57940.4733 & 0.0016 & 0&0007 & 17\
17 & 57940.5332 & 0.0020 & $-$0&0018 & 19\
32 & 57941.4713 & 0.0024 & $-$0&0002 & 16\
33 & 57941.5276 & 0.0017 & $-$0&0062 & 19\
34 & 57941.5930 & 0.0032 & $-$0&0033 & 16\
49 & 57942.5306 & 0.0077 & $-$0&0021 & 19\
65 & 57943.5294 & 0.0044 & $-$0&0023 & 19\
66 & 57943.5919 & 0.0043 & $-$0&0021 & 14\
82 & 57944.6047 & 0.0031 & 0&0117 & 14\
96 & 57945.4739 & 0.0016 & 0&0069 & 16\
97 & 57945.5235 & 0.0022 & $-$0&0059 & 19\
98 & 57945.5885 & 0.0021 & $-$0&0034 & 15\
\
\
\
ASASSN-17ji {#obj:asassn17ji}
-----------
This object was detected as a transient at $V$=15.2 on 2017 July 14 by the ASAS-SN team. The outburst was announced after the observation on 2017 July 14 ($V$=15.0). Observations on 2017 July 17 did not show superhumps. Superhumps were later found to have developed at least on 2017 July 24 (vsnet-alert 21275; e-figure \[fig:asassn17jishpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17jioc2017\]. Since the observations on 2017 July 25 were not of good quality, the maximum at $E$=18 was relatively uncertain. The cycle count, however, appears to be certain from a continuous run on July 24. A PDM analysis of the July 24 data yielded 0.0576(8) d. An $O-C$ analysis favored a longer period of 0.061(1) d. It is likely that the true period lies between 0.057 d and 0.061 d. The combined data of July 24 and 25 yielded 0.0589(1) d. The one-day alias of 0.0558(1) d appears to be ruled out (see e-figure \[fig:asassn17jishpdm\]). The superhump stage is unknown.
This field has been monitored by the ASAS-SN team since 2013 June 26 and no past outbursts were detected.
(85mm,110mm)[asassn17jishpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57959.4454 & 0.0019 & $-$0&0016 & 56\
1 & 57959.5048 & 0.0010 & $-$0&0010 & 61\
2 & 57959.5675 & 0.0039 & 0&0029 & 23\
18 & 57960.5051 & 0.0019 & $-$0&0003 & 55\
\
\
\
ASASSN-17jr {#obj:asassn17jr}
-----------
This object was detected as a transient at $V$=15.9 on 2017 July 25 by the ASAS-SN team. Subsequent observations detected superhumps (vsnet-alert 21280, 21297). The times of superhump maxima are listed in e-table \[tab:asassn17jroc2017\]. Only stage B with a positive $P_{\rm dot}$ was observed.
(85mm,110mm)[asassn17jrshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57960.5348 & 0.0012 & 0&0034 & 15\
1 & 57960.5951 & 0.0009 & 0&0020 & 14\
16 & 57961.5198 & 0.0009 & 0&0011 & 16\
17 & 57961.5816 & 0.0013 & 0&0012 & 16\
32 & 57962.5017 & 0.0017 & $-$0&0043 & 15\
33 & 57962.5669 & 0.0015 & $-$0&0008 & 16\
48 & 57963.4925 & 0.0018 & $-$0&0008 & 12\
49 & 57963.5505 & 0.0024 & $-$0&0045 & 15\
51 & 57963.6727 & 0.0034 & $-$0&0057 & 13\
65 & 57964.5416 & 0.0028 & $-$0&0007 & 15\
67 & 57964.6713 & 0.0085 & 0&0056 & 13\
98 & 57966.5821 & 0.0021 & 0&0035 & 13\
\
\
\
ASASSN-17kc {#obj:asassn17kc}
-----------
This object was detected as a transient at $V$=13.5 on 2017 July 30 by the ASAS-SN team. There were no past outbursts in the CRTS data and ASAS-3 data. Subsequent observations detected superhumps (vsnet-alert 21316, 21340; e-figure \[fig:asassn17kcshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17kcoc2017\]. The maxima for $E \ge$158 refer to post-superoutburst observations. Stages B and C can be clearly recognized and the large positive $P_{\rm dot}$ for stage B is apparent.
(85mm,110mm)[asassn17kcshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57968.6996 & 0.0007 & $-$0&0018 & 12\
1 & 57968.7626 & 0.0005 & $-$0&0025 & 15\
2 & 57968.8260 & 0.0004 & $-$0&0027 & 24\
3 & 57968.8896 & 0.0004 & $-$0&0027 & 19\
16 & 57969.7159 & 0.0008 & $-$0&0032 & 14\
17 & 57969.7788 & 0.0008 & $-$0&0039 & 17\
18 & 57969.8423 & 0.0007 & $-$0&0040 & 22\
19 & 57969.9068 & 0.0011 & $-$0&0031 & 13\
33 & 57970.7958 & 0.0010 & $-$0&0046 & 18\
34 & 57970.8607 & 0.0012 & $-$0&0033 & 21\
47 & 57971.6924 & 0.0013 & 0&0015 & 12\
48 & 57971.7510 & 0.0013 & $-$0&0035 & 15\
49 & 57971.8172 & 0.0006 & $-$0&0008 & 24\
50 & 57971.8830 & 0.0014 & 0&0014 & 20\
79 & 57973.7367 & 0.0018 & 0&0105 & 14\
80 & 57973.8003 & 0.0013 & 0&0104 & 23\
81 & 57973.8643 & 0.0013 & 0&0109 & 19\
94 & 57974.6885 & 0.0012 & 0&0083 & 13\
95 & 57974.7515 & 0.0012 & 0&0076 & 16\
96 & 57974.8150 & 0.0009 & 0&0075 & 23\
97 & 57974.8779 & 0.0010 & 0&0069 & 20\
126 & 57976.7174 & 0.0014 & 0&0018 & 15\
127 & 57976.7811 & 0.0020 & 0&0019 & 21\
128 & 57976.8466 & 0.0018 & 0&0038 & 20\
158 & 57978.7398 & 0.0033 & $-$0&0112 & 15\
159 & 57978.8029 & 0.0020 & $-$0&0117 & 24\
160 & 57978.8647 & 0.0038 & $-$0&0135 & 20\
\
\
\
ASASSN-17kd {#obj:asassn17kd}
-----------
This object was detected as a transient at $V$=12.4 on 2017 July 29 by the ASAS-SN team. The outburst was announced after the observation of $V$=12.8 on 2017 July 30. The object started to show superhumps on 2017 August 10 (vsnet-alert 21339; e-figure \[fig:asassn17kdshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17kdoc2017\]. Although stage A was recorded ($E \le$33), observations in the stage was not long enough to measure the period accurately.
We could not detect significant early superhumps to an upper limit of 0.01 mag. Although early superhumps were not observationally confirmed, we consider that the object is a WZ Sge-type dwarf nova since the waiting time to appear ordinary superhumps was long (at least 12 d), which is comparable to typical WZ Sge-type dwarf novae (cf. [@kat15wzsge]). At the time of our initial observation on 2017 August 3, the object already faded to 13.5 mag, indicating that the object faded rapidly since the peak brightness. Such a rapid fading is also characteristic to a WZ Sge-type superoutburst [@kat15wzsge].
No past outbursts are known. Our observations indicated that the object faded close to 18 mag on August 28. According to ASAS-SN observations, the object showed a rebrightening to $V$=15.6 on 2017 September 3–4. Although there was an observational gap between 2017 August 18 and 31 in the ASAS-SN data, the rebrightening was likely a short one.
(85mm,110mm)[asassn17kdshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57974.8537 & 0.0033 & $-$0&0129 & 17\
17 & 57975.8959 & 0.0012 & $-$0&0070 & 16\
33 & 57976.8877 & 0.0013 & 0&0095 & 16\
65 & 57978.8350 & 0.0015 & 0&0063 & 16\
66 & 57978.8946 & 0.0007 & 0&0049 & 13\
82 & 57979.8690 & 0.0008 & 0&0039 & 25\
98 & 57980.8423 & 0.0010 & 0&0020 & 25\
99 & 57980.9029 & 0.0022 & 0&0016 & 14\
115 & 57981.8770 & 0.0011 & 0&0004 & 25\
131 & 57982.8481 & 0.0014 & $-$0&0038 & 25\
147 & 57983.8250 & 0.0013 & $-$0&0022 & 23\
148 & 57983.8866 & 0.0023 & $-$0&0016 & 21\
164 & 57984.8694 & 0.0025 & 0&0060 & 25\
180 & 57985.8377 & 0.0038 & $-$0&0010 & 25\
181 & 57985.9001 & 0.0030 & 0&0004 & 13\
196 & 57986.8129 & 0.0025 & $-$0&0011 & 21\
197 & 57986.8729 & 0.0025 & $-$0&0020 & 25\
213 & 57987.8557 & 0.0030 & 0&0054 & 20\
229 & 57988.8236 & 0.0030 & $-$0&0020 & 20\
230 & 57988.8799 & 0.0029 & $-$0&0067 & 18\
\
\
\
ASASSN-17kg {#obj:asassn17kg}
-----------
This object was detected as a transient at $V$=13.0 on 2017 August 1 by the ASAS-SN team. The object has an X-ray counterpart of 1RXP J003152$+$0841.1. Subsequent observations detected superhumps (vsnet-alert 21317, 21326; e-figure \[fig:asassn17kgshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17kgoc2017\]. There were “textbook” stage A-B-C superhumps (e-figure \[fig:asassn17kghumpall\]).
Before the end of the superoutburst, the object showed a dip of $\sim$1.5 mag on 2017 August 21 (e-figure \[fig:asassn17kghumpall\]). This behavior is very similar to the WZ Sge-type dwarf nova KK Cnc (=OT J080714.2$+$113812, [@Pdot]). Early superhumps, however, were not detected in the case of ASASSN-17kg. The outburst light curve and a relatively small $P_{\rm dot}$ resemble those of WZ Sge-type dwarf novae. ASASSN-17kg may be a WZ Sge-type dwarf nova which failed to develop early superhumps when the stored mass in the disk was not sufficient (e.g. AL Com, [@kim16alcom]). There was no further post-superoutburst rebrightening in the ASAS-SN data.
(85mm,110mm)[asassn17kgshpdm.eps]{}
(160mm,200mm)[asassn17kghumpall.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lrrp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & & & & & & &\
0 & 57968.7404 & 0.0027 & $-$0&0033 & 13 & 139 & 57976.7538 & 0.0032 & $-$0&0007 & 15\
1 & 57968.7941 & 0.0012 & $-$0&0072 & 17 & 140 & 57976.8118 & 0.0017 & $-$0&0003 & 20\
2 & 57968.8521 & 0.0009 & $-$0&0068 & 20 & 141 & 57976.8632 & 0.0027 & $-$0&0065 & 18\
3 & 57968.9148 & 0.0024 & $-$0&0017 & 10 & 171 & 57978.5971 & 0.0010 & $-$0&0016 & 52\
17 & 57969.7312 & 0.0006 & 0&0078 & 13 & 172 & 57978.6508 & 0.0022 & $-$0&0055 & 43\
18 & 57969.7896 & 0.0006 & 0&0086 & 17 & 173 & 57978.7127 & 0.0037 & $-$0&0012 & 13\
19 & 57969.8461 & 0.0003 & 0&0075 & 20 & 174 & 57978.7680 & 0.0028 & $-$0&0036 & 17\
20 & 57969.9053 & 0.0006 & 0&0090 & 13 & 175 & 57978.8262 & 0.0016 & $-$0&0030 & 18\
36 & 57970.8277 & 0.0007 & 0&0092 & 21 & 176 & 57978.8748 & 0.0029 & $-$0&0120 & 18\
37 & 57970.8847 & 0.0009 & 0&0086 & 18 & 191 & 57979.7530 & 0.0021 & 0&0017 & 21\
49 & 57971.5729 & 0.0002 & 0&0052 & 140 & 209 & 57980.7970 & 0.0015 & 0&0083 & 35\
50 & 57971.6301 & 0.0003 & 0&0049 & 86 & 223 & 57981.5971 & 0.0012 & 0&0015 & 27\
52 & 57971.7453 & 0.0008 & 0&0048 & 14 & 227 & 57981.8355 & 0.0012 & 0&0095 & 24\
53 & 57971.8018 & 0.0004 & 0&0036 & 20 & 228 & 57981.8923 & 0.0011 & 0&0086 & 20\
54 & 57971.8587 & 0.0006 & 0&0030 & 18 & 242 & 57982.6979 & 0.0012 & 0&0074 & 16\
55 & 57971.9186 & 0.0006 & 0&0052 & 8 & 243 & 57982.7526 & 0.0010 & 0&0045 & 25\
87 & 57973.7542 & 0.0009 & $-$0&0035 & 14 & 244 & 57982.8135 & 0.0013 & 0&0078 & 30\
88 & 57973.8120 & 0.0006 & $-$0&0032 & 21 & 245 & 57982.8691 & 0.0010 & 0&0057 & 24\
89 & 57973.8686 & 0.0009 & $-$0&0043 & 18 & 259 & 57983.6768 & 0.0018 & 0&0066 & 15\
101 & 57974.5584 & 0.0004 & $-$0&0060 & 46 & 260 & 57983.7310 & 0.0016 & 0&0031 & 17\
104 & 57974.7310 & 0.0010 & $-$0&0064 & 13 & 261 & 57983.7915 & 0.0013 & 0&0059 & 35\
105 & 57974.7887 & 0.0011 & $-$0&0063 & 19 & 262 & 57983.8462 & 0.0017 & 0&0031 & 24\
106 & 57974.8468 & 0.0006 & $-$0&0058 & 18 & 263 & 57983.9008 & 0.0014 & 0&0000 & 15\
107 & 57974.9058 & 0.0015 & $-$0&0044 & 11 & 277 & 57984.7079 & 0.0018 & 0&0003 & 17\
121 & 57975.7134 & 0.0040 & $-$0&0037 & 13 & 278 & 57984.7642 & 0.0020 & $-$0&0011 & 31\
122 & 57975.7690 & 0.0013 & $-$0&0057 & 16 & 279 & 57984.8244 & 0.0013 & 0&0015 & 25\
123 & 57975.8240 & 0.0023 & $-$0&0084 & 19 & 280 & 57984.8768 & 0.0014 & $-$0&0037 & 23\
124 & 57975.8861 & 0.0032 & $-$0&0039 & 18 & 294 & 57985.6847 & 0.0024 & $-$0&0027 & 16\
136 & 57976.5749 & 0.0013 & $-$0&0067 & 50 & 295 & 57985.7394 & 0.0012 & $-$0&0056 & 23\
137 & 57976.6328 & 0.0026 & $-$0&0064 & 31 & 296 & 57985.7970 & 0.0014 & $-$0&0056 & 31\
138 & 57976.6926 & 0.0031 & $-$0&0043 & 12 & 297 & 57985.8588 & 0.0024 & $-$0&0015 & 24\
\
\
\
ASASSN-17kp {#obj:asassn17kp}
-----------
This object was detected as a transient at $V$=14.8 on 2017 August 6 by the ASAS-SN team. The outburst was announced after observation of $V$=15.2 on 2017 August 9. Subsequent observations detected superhumps (vsnet-alert 21338; e-figure \[fig:asassn17kpshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17kpoc2017\]. The fading on 2017 August 9 may suggest that the outburst on 2017 August 6 was a precursor one.
(85mm,110mm)[asassn17kpshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57978.5360 & 0.0007 & 0&0012 & 48\
15 & 57979.4042 & 0.0010 & 0&0001 & 63\
16 & 57979.4596 & 0.0026 & $-$0&0026 & 24\
17 & 57979.5217 & 0.0010 & 0&0016 & 62\
32 & 57980.3889 & 0.0005 & $-$0&0005 & 102\
33 & 57980.4468 & 0.0007 & $-$0&0006 & 119\
34 & 57980.5045 & 0.0006 & $-$0&0009 & 60\
35 & 57980.5639 & 0.0017 & 0&0006 & 38\
36 & 57980.6211 & 0.0017 & $-$0&0001 & 37\
51 & 57981.4919 & 0.0013 & 0&0012 & 41\
\
\
\
ASASSN-17la {#obj:asassn17la}
-----------
This object was detected as a transient at $V$=14.5 on 2017 August 17 by the ASAS-SN team. Early superhumps were detected, indicating that this object is a WZ Sge-type dwarf nova (vsnet-alert 21349; e-figure \[fig:asassn17laeshpdm\]). Ordinary superhump developed on 2017 October 23 (vsnet-alert 21362, 6 d after the outburst detection; e-figure \[fig:asassn17lashpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17laoc2017\]. This object showed “textbook” stages A and B.
The period of early superhump determined by the PDM method was 0.06039(3) d. The $\epsilon^*$ for stage A superhumps was 0.031(2), corresponding to $q$=0.084(5). This value is somewhat smaller than those of ordinary dwarf novae with the corresponding orbital period (cf. figure 17 in [@kat15wzsge]), but somewhat higher than those of typical period bouncers. The resultant parameters were similar to WZ Sge-type dwarf novae with multiple rebrightenings \[see @nak13j2112j2037; @kat15wzsge\]. The $P_{\rm dot}$ for stage B superhumps was not unusually small as in period bouncers. Due to the faintness of the object, we did not have observations after the superoutburst.
(85mm,110mm)[asassn17laeshpdm.eps]{}
(85mm,110mm)[asassn17lashpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57988.7336 & 0.0010 & $-$0&0064 & 115\
1 & 57988.7917 & 0.0012 & $-$0&0098 & 115\
15 & 57989.6640 & 0.0012 & 0&0004 & 76\
16 & 57989.7291 & 0.0005 & 0&0039 & 116\
17 & 57989.7909 & 0.0005 & 0&0041 & 114\
18 & 57989.8546 & 0.0005 & 0&0062 & 92\
27 & 57990.4106 & 0.0004 & 0&0081 & 85\
28 & 57990.4712 & 0.0003 & 0&0071 & 149\
29 & 57990.5328 & 0.0003 & 0&0071 & 121\
30 & 57990.5938 & 0.0012 & 0&0065 & 46\
31 & 57990.6547 & 0.0023 & 0&0059 & 39\
32 & 57990.7172 & 0.0004 & 0&0068 & 83\
33 & 57990.7775 & 0.0004 & 0&0055 & 92\
34 & 57990.8395 & 0.0003 & 0&0059 & 81\
65 & 57992.7391 & 0.0005 & $-$0&0033 & 90\
66 & 57992.8011 & 0.0006 & $-$0&0029 & 80\
77 & 57993.4765 & 0.0005 & $-$0&0048 & 91\
78 & 57993.5392 & 0.0006 & $-$0&0037 & 102\
79 & 57993.6030 & 0.0007 & $-$0&0016 & 98\
90 & 57994.2747 & 0.0006 & $-$0&0072 & 26\
91 & 57994.3367 & 0.0007 & $-$0&0067 & 33\
92 & 57994.4005 & 0.0006 & $-$0&0045 & 34\
93 & 57994.4616 & 0.0008 & $-$0&0050 & 33\
94 & 57994.5231 & 0.0007 & $-$0&0051 & 33\
95 & 57994.5876 & 0.0010 & $-$0&0022 & 28\
97 & 57994.7084 & 0.0008 & $-$0&0045 & 63\
98 & 57994.7713 & 0.0009 & $-$0&0032 & 59\
99 & 57994.8322 & 0.0013 & $-$0&0039 & 46\
106 & 57995.2674 & 0.0041 & 0&0003 & 18\
107 & 57995.3243 & 0.0009 & $-$0&0044 & 33\
108 & 57995.3867 & 0.0015 & $-$0&0036 & 32\
109 & 57995.4481 & 0.0009 & $-$0&0037 & 33\
110 & 57995.5133 & 0.0016 & $-$0&0001 & 33\
111 & 57995.5716 & 0.0013 & $-$0&0034 & 29\
113 & 57995.6961 & 0.0010 & $-$0&0021 & 63\
114 & 57995.7573 & 0.0014 & $-$0&0024 & 64\
115 & 57995.8207 & 0.0015 & $-$0&0006 & 60\
116 & 57995.8775 & 0.0012 & $-$0&0054 & 64\
123 & 57996.3116 & 0.0016 & $-$0&0023 & 33\
124 & 57996.3726 & 0.0013 & $-$0&0029 & 34\
125 & 57996.4345 & 0.0024 & $-$0&0025 & 33\
126 & 57996.4985 & 0.0022 & $-$0&0001 & 33\
127 & 57996.5630 & 0.0035 & 0&0028 & 33\
141 & 57997.4170 & 0.0019 & $-$0&0053 & 37\
172 & 57999.3422 & 0.0013 & 0&0110 & 39\
173 & 57999.4026 & 0.0011 & 0&0099 & 43\
174 & 57999.4644 & 0.0011 & 0&0101 & 42\
175 & 57999.5281 & 0.0017 & 0&0122 & 44\
\
\
\
ASASSN-17lr {#obj:asassn17lr}
-----------
This object was detected as a transient at $V$=14.6 on 2017 September 4 by the ASAS-SN team. The outburst was announced after the observation of $V$=14.9 on 2017 Semtember 5. Observations on 2017 September 9 and 11 did not show superhump-like modulations. Superhumps were observed since 2017 September 15 (vsnet-alert 21443; e-figure \[fig:asassn17lrshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17lroc2017\]. Although a one-day alias 0.06047(3) d was stronger by the PDM method, we adopted this alias based on $O-C$ analysis of the continuous observations on 2017 September 17–18. Although the object was initially suggested to be a WZ Sge-type dwarf nova (vsnet-alert 21443), this was based on the absence of superhumps for a week after the discovery and not based on detection of early superhumps.
(85mm,110mm)[asassn17lrshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 58012.3909 & 0.0010 & $-$0&0046 & 37\
34 & 58014.3925 & 0.0019 & 0&0034 & 42\
36 & 58014.5103 & 0.0015 & 0&0039 & 38\
37 & 58014.5647 & 0.0014 & $-$0&0003 & 48\
102 & 58018.3738 & 0.0037 & $-$0&0024 & 57\
\
\
\
ASASSN-17me {#obj:asassn17me}
-----------
This object was detected as a transient at $V$=15.2 on 2017 September 13 by the ASAS-SN team. The outburst was announced after the observation of $V$=15.4 on 2017 Semtember 15. There is a likely quiescent counterpart $g$=21.45 mag object in Pan-STARRS (vsnet-alert 21440). Subsequent observations detected superhumps (vsnet-alert 21447, e-figure \[fig:asassn17meshpdm\]). Although there were observations on two additional nights, the object already started fading rapidly (2017 September 22 and 23) and superhumps were not detected. The superhump period based on single-nights observations was 0.0614(4) d (PDM method). The times of superhump maxima were BJD 2458013.6486(5) ($N$=162) and 2458013.7130(4) ($N$=186).
(85mm,110mm)[asassn17meshpdm.eps]{}
ASASSN-17np {#obj:asassn17np}
-----------
This object was detected as a transient at $g$=15.0 on 2017 October 18 by the ASAS-SN team. The outburst was announced after the observation of $g$=15.2 on 2017 October 20. Subsequent observations detected superhumps (vsnet-alert 21534, 21535, 21545; e-figure \[fig:asassn17npshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17np2017\]. Although observations were rather sparse, stages B and C can be recognized. The object called attention due to its absence of a known quiescent counterpart and hence a likely large outburst amplitude (cf. vsnet-alert 21531), the resultant superhump period suggests a rather ordinary SU UMa-type dwarf nova. The outburst amplitude may not be as large as initially suspected, or the object may be indeed unusual with a large outburst amplitude despite the long superhump period. The large amplitude of superhumps suggests an ordinary object rather than a WZ Sge-like object with a long superhump period such as ASASSN-16eg [@wak17asassn16eg].
(85mm,110mm)[asassn17npshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 58047.4968 & 0.0005 & $-$0&0055 & 199\
1 & 58047.5872 & 0.0004 & $-$0&0039 & 205\
10 & 58048.3899 & 0.0003 & $-$0&0009 & 205\
11 & 58048.4791 & 0.0003 & $-$0&0006 & 205\
12 & 58048.5701 & 0.0004 & 0&0016 & 131\
25 & 58049.7293 & 0.0013 & 0&0056 & 11\
26 & 58049.8161 & 0.0009 & 0&0036 & 26\
37 & 58050.7943 & 0.0016 & 0&0043 & 17\
48 & 58051.7706 & 0.0012 & 0&0031 & 16\
49 & 58051.8592 & 0.0016 & 0&0029 & 15\
59 & 58052.7441 & 0.0014 & $-$0&0008 & 14\
60 & 58052.8311 & 0.0019 & $-$0&0027 & 21\
70 & 58053.7185 & 0.0026 & $-$0&0039 & 11\
71 & 58053.8096 & 0.0016 & $-$0&0016 & 20\
79 & 58054.5201 & 0.0010 & $-$0&0020 & 180\
82 & 58054.7896 & 0.0059 & 0&0009 & 19\
\
\
\
ASASSN-17nr {#obj:asassn17nr}
-----------
This object was detected as a transient at $V$=14.6 on 2017 October 18 by the ASAS-SN team. The outburst was announced after the observation of $V$=15.4 on 2017 October 21. Subsequent observations detected superhumps (vsnet-alert 21556; e-figure \[fig:asassn17nrshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17nroc2017\]. Although $O-C$ data suggested a positive $P_{\rm dot}$, the value should be treated with caution due to the limited quality of observations. There remains possibilities of one-day aliases, although the $O-C$ analysis favors this selection (e-figure \[fig:asassn17nrshpdm\]).
(85mm,110mm)[asassn17nrshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 58052.8080 & 0.0029 & 0&0026 & 13\
36 & 58054.8301 & 0.0017 & $-$0&0036 & 13\
53 & 58055.7903 & 0.0021 & $-$0&0013 & 11\
54 & 58055.8449 & 0.0020 & $-$0&0030 & 13\
71 & 58056.8071 & 0.0043 & 0&0014 & 16\
72 & 58056.8618 & 0.0015 & $-$0&0003 & 11\
89 & 58057.8210 & 0.0011 & 0&0011 & 18\
106 & 58058.7824 & 0.0051 & 0&0047 & 15\
107 & 58058.8364 & 0.0015 & 0&0023 & 19\
142 & 58060.8021 & 0.0045 & $-$0&0040 & 17\
\
\
\
ASASSN-17of {#obj:asassn17of}
-----------
This object was detected as a transient at $g$=16.2 on 2017 November 3 by the ASAS-SN team. The outburst was announced after observations of $g$=17.0 on 2017 November 6 and $g$=16.8 on 2017 November 7. Subsequent observations detected superhumps (vsnet-alert 21567, 21580; e-figure \[fig:asassn17ofshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17ofoc2017\]. Stages B and C can be recognized.
(85mm,110mm)[asassn17ofshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 58065.2348 & 0.0059 & $-$0&0121 & 18\
1 & 58065.3140 & 0.0022 & 0&0032 & 21\
47 & 58068.2582 & 0.0011 & 0&0021 & 55\
57 & 58068.8988 & 0.0028 & 0&0024 & 53\
58 & 58068.9647 & 0.0017 & 0&0043 & 62\
59 & 58069.0280 & 0.0025 & 0&0036 & 61\
74 & 58069.9920 & 0.0024 & 0&0071 & 60\
79 & 58070.3085 & 0.0023 & 0&0035 & 48\
80 & 58070.3719 & 0.0019 & 0&0028 & 48\
94 & 58071.2616 & 0.0016 & $-$0&0038 & 46\
94 & 58071.2616 & 0.0016 & $-$0&0038 & 45\
109 & 58072.2166 & 0.0016 & $-$0&0093 & 24\
\
\
\
ASASSN-17oo {#obj:asassn17oo}
-----------
This object was detected as a transient at $g$=15.0 on 2017 November 1 by the ASAS-SN team. The outburst was announced after an observation of $g$=15.05 on 2017 November 10. Although subsequent observations detected superhumps (vsnet-alert 21591), individual maxima were difficult to measure. The presence of superhumps was, however, secure as shown in the PDM analysis (e-figure \[fig:asassn17ooshpdm\]). It was likely that we observed superhumps with diminished amplitudes near the end of stage B or in stage C (due to the delay in confirmation of the outburst, our observations started 14 d after the initial outburst detection). The superhump period was measured to be 0.06781(5) d by the PDM method.
(85mm,110mm)[asassn17ooshpdm.eps]{}
ASASSN-17ou {#obj:asassn17ou}
-----------
This object was detected as a transient at $g$=16.6 on 2017 November 10 by the ASAS-SN team. The outburst was announced after observations of $g$=16.9 on 2017 November 11, $g$=17.1 on 2017 November 12 and $g$=17.1 on 2017 November 13. Subsequent observations detected superhumps (vsnet-alert 21588, 21589, 21590; e-figure \[fig:asassn17oushpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17ouoc2017\].
(85mm,110mm)[asassn17oushpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 58072.2353 & 0.0014 & 0&0037 & 21\
1 & 58072.2901 & 0.0002 & 0&0014 & 262\
2 & 58072.3471 & 0.0003 & 0&0012 & 236\
12 & 58072.9174 & 0.0029 & 0&0003 & 43\
13 & 58072.9753 & 0.0017 & 0&0011 & 58\
14 & 58073.0226 & 0.0013 & $-$0&0088 & 49\
23 & 58073.5451 & 0.0017 & $-$0&0004 & 19\
51 & 58075.1502 & 0.0016 & 0&0051 & 20\
52 & 58075.1996 & 0.0006 & $-$0&0027 & 40\
53 & 58075.2584 & 0.0008 & $-$0&0010 & 55\
54 & 58075.3137 & 0.0008 & $-$0&0027 & 51\
70 & 58076.2332 & 0.0004 & 0&0027 & 49\
\
\
\
ASASSN-17pb {#obj:asassn17pb}
-----------
This object was detected as a transient at $V$=15.8 on 2017 November 13 by the ASAS-SN team. The outburst was announced after an observation of $V$=16.1 on 2017 November 14. Subsequent observations detected superhumps (vsnet-alert 21599, 21605, 21642; e-figure \[fig:asassn17pbshpdm\]). The times of superhump maxima are listed in e-table \[tab:asassn17pboc2017\]. The epoch for $E$=0 was a stage A superhump. The epochs for $E$=23 and 24 may be during the transition to stage B. The epoch for $E$=128 probably corresponds to a stage C superhump. On the first two nights (2017 November 15 and 16), the object did not show superhump-like modulations. There was a suggestion of low-amplitude modulations with a period of 0.0169(2) d (e-figure \[fig:asassn17earlyshpdm\]). This period is not related to the superhump one and the origin of this variation is unclear.
(85mm,110mm)[asassn17pbshpdm.eps]{}
(85mm,110mm)[asassn17pbearlypdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 58075.5735 & 0.0009 & $-$0&0129 & 56\
23 & 58077.3332 & 0.0005 & $-$0&0032 & 68\
24 & 58077.4116 & 0.0007 & $-$0&0009 & 53\
47 & 58079.1640 & 0.0034 & 0&0014 & 11\
48 & 58079.2417 & 0.0012 & 0&0030 & 19\
61 & 58080.2283 & 0.0017 & 0&0005 & 47\
62 & 58080.3070 & 0.0010 & 0&0031 & 70\
63 & 58080.3847 & 0.0007 & 0&0047 & 102\
64 & 58080.4646 & 0.0023 & 0&0086 & 13\
74 & 58081.2182 & 0.0022 & 0&0012 & 32\
75 & 58081.2894 & 0.0025 & $-$0&0037 & 46\
84 & 58081.9851 & 0.0020 & 0&0072 & 15\
85 & 58082.0581 & 0.0190 & 0&0042 & 7\
88 & 58082.2825 & 0.0021 & 0&0003 & 28\
89 & 58082.3616 & 0.0007 & 0&0034 & 22\
97 & 58082.9731 & 0.0019 & 0&0062 & 12\
100 & 58083.1950 & 0.0038 & $-$0&0003 & 33\
101 & 58083.2748 & 0.0016 & 0&0035 & 58\
102 & 58083.3417 & 0.0017 & $-$0&0057 & 28\
128 & 58085.3051 & 0.0021 & $-$0&0206 & 20\
\
\
\
CRTS J044027.1$+$023301 {#obj:j0440}
-----------------------
This object (=CSS090219:044027$+$023301, hereafter CRTS J044027) was detected as a transient by CRTS on 2009 February 19. The object has an X-ray counterpart of 1RXS J044027.0$+$023300.
The 2017 outburst was detected by the ASAS-SN team at $V$=14.56 on 2017 August 30. This outburst was originally suspected to be a normal one due to its large fading rate (vsnet-alert 21388). Subsequent observations, however, detected superhumps (vsnet-alert 21397, 21411; figure \[fig:j0440shpdm\]). The detection on 2017 August 30 was probably of a precursor outburst. The times of superhump maxima are listed in e-table \[tab:j0440oc2017\]. Although $E$=1 appears to be a stage A superhump, we could not determine the period. Later observations probably recorded stage B superhumps since the object faded soon after these observations.
There was another superoutburst in the ASAS-SN data peaking on 2016 February 27 at $V$=14.56. Another detection on 2012 September 22 at $V$=15.37 probably corresponds to a normal outburst. CRTS detected additional two faint outbursts in addition to the 2009 one (all of them were apparently normal outbursts).
(85mm,110mm)[j0440shpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57999.6124 & 0.0005 & $-$0&0031 & 47\
49 & 58002.7769 & 0.0012 & 0&0035 & 23\
50 & 58002.8377 & 0.0016 & $-$0&0002 & 22\
64 & 58003.7440 & 0.0024 & 0&0039 & 22\
65 & 58003.8044 & 0.0014 & $-$0&0002 & 22\
66 & 58003.8702 & 0.0020 & 0&0011 & 22\
80 & 58004.7722 & 0.0024 & 0&0008 & 34\
81 & 58004.8343 & 0.0027 & $-$0&0015 & 31\
96 & 58005.8013 & 0.0022 & $-$0&0012 & 23\
97 & 58005.8641 & 0.0021 & $-$0&0029 & 31\
\
\
\
CRTS J080941.3$+$171528 {#obj:j0809}
-----------------------
This object (=CSS120120:080941$+$171528, hereafter CRTS J080941) was detected as a transient by CRTS on 2012 January 20.
The 2017 outburst was detected by the ASAS-SN team at $V$=15.3 on 2017 April 8 and announced after observation of $V$=15.4 on 2017 April 9. The large outburst amplitude and past behavior suggested a superoutburst (vsnet-alert 20888). Subsequent observations detected long-period superhumps (vsnet-alert 20893, 20906, 20912; e-figure \[fig:j0809shpdm\]). The times of superhump maxima are listed in e-table \[tab:j0809oc2017\]. The period for $E \le$12 was substantially longer. The difference of the periods before and after $E$=12 was 1.2%, which is too large to be considered as stage B-C transition. We rather consider that this reflects stage A-B transition, as have been recently recorded in many SU UMa-type dwarf novae with long superhump periods (V1006 Cyg and MN Dra: [@kat16v1006cyg]; CRTS J214738.4$+$244554 and OT J064833.4$+$065624: [@Pdot7], KK Tel, possibly V452 Cas and ASASSN-15cl: [@Pdot8], MASTER OT J021315.37$+$533822.7: [@Pdot9], OT J002656.6$+$284933 = CSS101212:002657$+$284933: [@kat17j0026]).
CRTS J080941 is not only an SU UMa-type dwarf nova in the middle of the period gap but also shows superhump evolution common to many long-period SU UMa-type dwarf novae. Since (supposed) stage A superhumps were detected, determination of the orbital period is highly desired to determine the mass ratio in such a system in the middle of the period gap.
CRTS recorded two past outbursts: 2007 May 11 (15.8 mag) and 2012 January 20 (16.3 mag). No past outbursts were detected in the ASAS-SN data starting from 2012 January.
(85mm,110mm)[j0809shpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57853.3307 & 0.0008 & $-$0&0056 & 84\
1 & 57853.4269 & 0.0004 & $-$0&0101 & 104\
10 & 57854.3463 & 0.0005 & 0&0030 & 97\
11 & 57854.4452 & 0.0009 & 0&0011 & 63\
12 & 57854.5495 & 0.0009 & 0&0047 & 35\
20 & 57855.3607 & 0.0048 & 0&0103 & 18\
22 & 57855.5545 & 0.0011 & 0&0027 & 33\
32 & 57856.5585 & 0.0016 & $-$0&0004 & 27\
42 & 57857.5622 & 0.0014 & $-$0&0037 & 35\
49 & 57858.2770 & 0.0051 & 0&0060 & 54\
52 & 57858.5710 & 0.0016 & $-$0&0021 & 36\
62 & 57859.5741 & 0.0040 & $-$0&0060 & 34\
\
\
\
CRTS J120052.9$-$152620 {#obj:j1200}
-----------------------
This object (=CSS110205:120053$-$152620, hereafter CRTS J120052) was discovered by the CRTS on 2011 February 5. The 2011 and 2016 superoutbursts were reported in @Pdot3 and @Pdot8, respectively. The 2017 superoutburst was detected by the ASAS-SN team at $V$=13.8 on 2017 April 24. Single-night observations were performed and obtained two superhump maxima: BJD 2457867.0363(3) ($N$=159) and 2457867.1236(4) ($N$=167).
This object showed many outbursts. Among them, we listed likely/possible superoutbursts in e-table \[tab:j1200out\] (in the case of ASAS-3 and ASAS-SN, the identifications were based on durations; in the case of CRTS, we selected possible ones based on brightness since they were single-night detections). The maxima after 2006 appear to be well expressed by a supercycle of 200.7(9) d. The period was stable at 192(2) d between 2006 and 2009 and apparently lengthened to 203(1) d after that. The cycle count between 2001 and 2006 was unclear and we couldn’t determine the supercycle uniquely. This object appears to be a relatively ordinary SU UMa-type dwarf nova with frequent outbursts.
[cccccc]{} Year & Month & Day & max& mag& Source\
2001 & 2 & 14 & 51954 & 13.4V & ASAS-3\
2006 & 5 & 21 & 53876 & 14.0V & ASAS-3\
2007 & 6 & 2 & 54254 & 14.1V & ASAS-3\
2007 & 12 & 22 & 54457 & 13.9V & ASAS-3\
2008 & 6 & 7 & 54625 & 14.0C & CRTS, ASAS-3\
2009 & 1 & 8 & 54840 & 13.7V & ASAS-3\
2009 & 7 & 18 & 55031 & 13.9C & CRTS\
2010 & 1 & 18 & 55215 & 13.8C & CRTS\
2015 & 2 & 1 & 57055 & 13.6V & ASAS-SN\
2016 & 3 & 14 & 57462 & 13.8V & ASAS-SN\
2017 & 4 & 24 & 57867 & 13.7V & ASAS-SN\
\
\
CRTS J122221.6$-$311524 {#obj:j1222}
-----------------------
Although we have already reported the 2013 superoutburst and claimed the object to be a best candidate for a period bouncer [@kat13j1222], we treat this object again since @neu17j1222 reported a spectroscopic orbital period of 109.80(7) min \[0.07625(5) d\]. Since we have a period of stage A superhumps \[0.07721(1) d\], we can now directly estimate the mass ratio. The value of $\epsilon^*$ for stage A superhumps is 0.0124(6), which corresponds to $q$=0.032(2). This value supersedes our previous constraint [@kat13j1222] and the $q$ value based on stage B superhumps with a large uncertainty [@neu17j1222]. The value is sufficiently low to give strong credence to the period-bouncer status.
CRTS J162806.2$+$065316 {#obj:j1628}
-----------------------
The object was detected as a transient (=CSS110611:162806$+$065316; hereafter CRTS J162806) by CRTS on 2011 June 11. The 2011 superoutburst was studied in @Pdot4. The 2017 superoutburst was detected by the ASAS-SN team at $V$=14.64 on 2017 March 11. Only one superhump was recorded: BJD 2457825.5805(9) ($N$=76).
Although this field has been monitored by the ASAS-SN team since 2012 April 1, no other secure outburst was recorded. The 2013 October outburst fell in the gap of the ASAS-SN observations. The three outbursts detected by CRTS since 2007 were likely superoutbursts (e-table \[tab:j1628out\]). The intervals of known (likely) superoutbursts were 1522 d, 844 d and 1256 d. Considering the sparse past observations of CRTS, the frequency of outbursts in this system probably is not very low.
[cccccc]{} Year & Month & Day & max& mag& Source\
2005 & 9 & 14 & 53628 & 15.3C & CRTS\
2006 & 6 & 20 & 53907 & 16.0C & CRTS\
2007 & 4 & 10 & 54201 & 14.1C & CRTS\
2011 & 6 & 10 & 55723 & 14.2C & CRTS\
2013 & 10 & 1 & 56567 & 14.1C & CRTS\
2017 & 3 & 11 & 57823 & 14.6V & ASAS-SN\
\
\
CRTS J214934.1$-$121908 {#obj:j2149}
-----------------------
The object was detected as a transient (=CSS120922:214934$-$121908; hereafter CRTS J214934) by CRTS on 2012 September 22. The SU UMa-type nature was confirmed by single-night observations in @Pdot5.
The 2017 superoutburst was detected by the ASAS-SN team at $V$=15.69 on 2017 October 8. Subsequent observations detected superhumps (vsnet-alert 21511; e-figure \[fig:j2149shpdm\]). The times of superhump maxima are listed in e-table \[tab:j2149oc2017\]. Although observations were insufficient, the data suggest stage B-C transition. The period of stage B superhumps given in the table should refer to a lower limit since stage B likely ended before $E$=64.
We listed superoutbursts in the ASAS-SN data in e-table \[tab:j2149out\]. The interval between the 2012 superoutburst and the 2013 one was 284 d. The interval between the 2013 and 2014 superoutbursts was 332 d. Assuming that there were four supercycles between 2014 and 2017, the supercycle was estimated to be 308(3) d.
(85mm,110mm)[j2149shpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 58036.9845 & 0.0019 & $-$0&0020 & 63\
1 & 58037.0564 & 0.0015 & $-$0&0015 & 69\
64 & 58041.5593 & 0.0035 & 0&0040 & 16\
65 & 58041.6313 & 0.0030 & 0&0045 & 27\
92 & 58043.5531 & 0.0033 & $-$0&0011 & 16\
93 & 58043.6280 & 0.0019 & 0&0024 & 28\
106 & 58044.5494 & 0.0033 & $-$0&0043 & 15\
107 & 58044.6231 & 0.0095 & $-$0&0019 & 20\
\
\
\
[ccccc]{} Year & Month & Day & max& $V$ mag\
2013 & 7 & 3 & 56477 & 15.7\
2014 & 5 & 31 & 56809 & 15.4\
2017 & 10 & 7 & 58034 & 15.6\
\
CRTS J223235.4$+$304105 {#obj:j2232}
-----------------------
The object was detected as a transient (=CSS081107:223235$+$304105; hereafter CRTS J223235) by CRTS on 2008 November 7 [@dra14CRTSCVs]. The 2017 outburst was detected by the ASAS-SN team (the name ASASSN-17nw was assigned) at $V$=17.7 on 2017 October 18 and was announced after the object brightened to $V$=15.9 on 2017 October 19. Subsequent observations detected superhumps (vsnet-alert 21551, 21553; e-figure \[fig:j2232shpdm\]). The times of superhump maxima are listed in e-table \[tab:j2232oc2017\].
(85mm,110mm)[j2232shpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 58054.5328 & 0.0017 & 0&0004 & 47\
28 & 58056.2899 & 0.0011 & $-$0&0063 & 44\
29 & 58056.3626 & 0.0017 & 0&0034 & 72\
30 & 58056.4225 & 0.0017 & 0&0003 & 73\
31 & 58056.4874 & 0.0011 & 0&0022 & 107\
32 & 58056.5483 & 0.0015 & 0&0001 & 60\
\
\
\
CTCV J1940$-$4724 {#obj:j1940}
-----------------
This object (hereafter CTCV J1940) was detected as a CV by Calán-Tololo Survey [@aug10CTCVCV2]. @aug10CTCVCV2 obtained an orbital period of 0.0809(30) d using a set of 18 spectroscopic observations taken over a baseline of 0.11 d. Since the period suggested an SU UMa-type dwarf nova, a systematic search for outbursts was conducted (cf. vsnet-alert 21188). Although the outburst detected on 2017 July 1 turned out to be a normal outburst, a precursor outburst (2017 August 19–20) and an apparent superoutburst starting on 2017 August 27 were visually detected by R. Stubbings (vsnet-alert 21370). Subsequent observations detected superhumps (vsnet-alert 21378). The resultant data, however, were not ideally spaced and there remained an ambiguity in choosing the alias. Although the PDM analysis favored a period of 0.07124(8) d, an $O-C$ analysis gave a systematic trend and preferred 0.07667(7) d. We adopted the latter, which is also marginally consistent with the approximate orbital period by @aug10CTCVCV2 in estimating cycle counts in e-table \[tab:j1940oc2017\].
We listed superoutbursts in the ASAS-SN data in e-table \[tab:j1940out\]. The data suggest a supercycle of 168(6) d. This supercycle is also consistent with the ASAS-3 data, which detected several long outbursts though they were not as well sampled as in the ASAS-SN data.
[ccccc]{} Year & Month & Day & max& $V$ mag\
2015 & 11 & 3 & 57330 & 13.0\
2016 & 5 & 28 & 57537 & 13.9\
2016 & 10 & 25 & 57687 & 13.3\
2017 & 4 & 7 & 57851 & 13.5\
2017 & 8 & 29 & 57995 & 13.4\
\
\
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57994.6257 & 0.0016 & $-$0&0029 & 19\
13 & 57995.6269 & 0.0017 & 0&0016 & 18\
26 & 57996.6239 & 0.0016 & 0&0019 & 24\
39 & 57997.6202 & 0.0020 & 0&0016 & 25\
52 & 57998.6136 & 0.0019 & $-$0&0017 & 27\
65 & 57999.6099 & 0.0012 & $-$0&0021 & 18\
66 & 57999.6912 & 0.0018 & 0&0025 & 20\
78 & 58000.6071 & 0.0026 & $-$0&0016 & 25\
79 & 58000.6860 & 0.0029 & 0&0006 & 19\
\
\
\
DDE 51 {#obj:dde51}
------
DDE 51 was discovered as a dwarf nova by D. Denisenko.[^11] Denisenko monitored this object and found an outburst at an unfiltered CCD magnitude of 16.52 on 2017 September 28 (vsnet-alert 21475). The object further brightened and reached an unfiltered CCD magnitude of 15.58 on 2017 September 30. The initial superhump detection was made by Oleg Milantiev (vsnet-alert 21489). Further observations clarified that this object is an SU UMa-type in the period gap by detecting long-period superhumps (vsnet-alert 21490, 21491, 21493; e-figure \[fig:dde51shpdm\]). The times of superhump maxima are listed in e-table \[tab:dde51oc2017\]. It is apparent that there was a stage transition between $E$=12 and $E$=30. Considering the similarity with the SU UMa-type dwarf novae in the period gap V1006 Cyg and MN Dra [@kat16v1006cyg], we consider that superhumps for $E \le$12 were stage A superhumps despite large superhump amplitudes (see also a discussion in subsection \[obj:j0809\]). Determination of the orbital period will be a crucial test for this interpretation.
Among outbursts recorded by the ASAS-SN team (e-table \[tab:dde51out\]), only the 2017 September–October one appears to be a superoutburst. Observations are not yet sufficient to determine the supercycle.
(85mm,110mm)[dde51shpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 58028.4266 & 0.0005 & $-$0&0099 & 112\
1 & 58028.5285 & 0.0021 & $-$0&0085 & 50\
9 & 58029.3409 & 0.0008 & $-$0&0002 & 81\
10 & 58029.4476 & 0.0014 & 0&0059 & 46\
11 & 58029.5420 & 0.0005 & $-$0&0001 & 100\
12 & 58029.6449 & 0.0009 & 0&0022 & 45\
30 & 58031.4572 & 0.0007 & 0&0052 & 87\
49 & 58033.3685 & 0.0013 & 0&0066 & 106\
50 & 58033.4669 & 0.0010 & 0&0044 & 93\
57 & 58034.1696 & 0.0005 & 0&0035 & 180\
70 & 58035.4750 & 0.0004 & 0&0022 & 92\
71 & 58035.5733 & 0.0005 & $-$0&0001 & 96\
76 & 58036.0759 & 0.0005 & $-$0&0001 & 148\
77 & 58036.1743 & 0.0007 & $-$0&0022 & 164\
108 & 58039.2839 & 0.0005 & $-$0&0087 & 85\
\
\
\
[ccccc]{} Year & Month & Day & max& $V$ mag\
2016 & 5 & 12 & 57521 & 16.5\
2016 & 10 & 26 & 57688 & 16.0\
2017 & 6 & 16 & 57921 & 15.8\
2017 & 9 & 29 & 58026 & 15.2\
\
MASTER OT J132501.00$+$431846.1 {#obj:j1325}
-------------------------------
Although this object has not been identified as an SU UMa-type dwarf nova, we include it due to its special characteristics. This object (hereafter MASTER J132501) was detected as a transient at an unfiltered CCD magnitude of 15.4 on 2017 June 5 [@bal17j1325atel10470]. The object was initially proposed to be a large-amplitude WZ Sge-type dwarf nova. There were, however, no early superhumps (vsnet-alert 21107). Based on this negative detection and further brightening, @den17j1325atel10480 suggested this object to be a likely supernova in NGC 5145. Although we had observations on three nights between 2017 June 7 and 11, no periodic signal was detected. The object, however, faded quickly (vsnet-alert 21130), which appears to be inconsistent with the supernova classification. Although the object was again suspected to be a dwarf nova, the behavior is unusual (fading at a rate of $\sim$0.5 mag d$^{-1}$, which suggests an SS Cyg-type dwarf nova, while the amplitude exceeds 8 mag). No spectroscopic observation was reported. This object probably would require future deep observations to clarify the nature.
MASTER OT J174305.70$+$231107.8 {#obj:j1743}
-------------------------------
This object (hereafter MASTER J174305) is a transient detected at an unfiltered CCD magnitude of 15.6 on 2012 April 5 [@bal12j1743atel4022]. Although 2012 observations established the SU UMa-type classification, only two superhumps were recorded [@Pdot4].
The 2017 superoutburst was recorded by the ASAS-SN team (cf. vsnet-alert 21024). Superhumps were detected (vsnet-alert 21034, 21043; e-figure \[fig:j1743shpdm\]). The times of superhump maxima are listed in e-table \[tab:j1743oc2017\]. Although stages B and C were apparently recorded, the determined periods were not precise due to the limited quality of the data. The mean superhump period with the PDM method \[0.06955(3) d\], however, is sufficiently reliable.
The field of this object has been monitored since 2012 March 17 by the ASAS-SN team and the coverage has been particularly good since 2012 September. The object has been regularly caught in outbursts and they are listed in e-table \[tab:j1743out\]. All these outbursts were superoutburst as judged from outburst durations. In additions to them, there were single-night outburst detections at an unfiltered CCD magnitude of 17.3 on 2015 October 6 (E. Muyllaert) and $V$=16.84 (multiple detections on a single night, ASAS-SN). These outbursts were likely normal outbursts. The recorded superoutbursts were well expressed by a supercycle of 208(1) d. These data suggest that MASTER J174305 is a relatively ordinary SU UMa-type dwarf nova with rather frequent superoutbursts.
(85mm,110mm)[j1743shpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57891.5471 & 0.0014 & $-$0&0025 & 58\
7 & 57892.0348 & 0.0006 & $-$0&0022 & 129\
12 & 57892.3855 & 0.0010 & 0&0002 & 75\
13 & 57892.4563 & 0.0009 & 0&0014 & 75\
14 & 57892.5261 & 0.0007 & 0&0015 & 32\
27 & 57893.4328 & 0.0009 & 0&0028 & 113\
28 & 57893.5019 & 0.0007 & 0&0023 & 112\
29 & 57893.5715 & 0.0010 & 0&0022 & 71\
38 & 57894.1926 & 0.0011 & $-$0&0035 & 122\
39 & 57894.2665 & 0.0026 & 0&0007 & 94\
41 & 57894.4039 & 0.0028 & $-$0&0011 & 75\
42 & 57894.4744 & 0.0018 & $-$0&0003 & 77\
43 & 57894.5439 & 0.0015 & $-$0&0004 & 76\
44 & 57894.6130 & 0.0056 & $-$0&0010 & 42\
\
\
\
[ccccc]{} Year & Month & Day & max& $V$-mag\
2012 & 4 & 2 & 56020 & 16.03\
2012 & 10 & 30 & 56231 & 15.66\
2013 & 5 & 27 & 56440 & 15.91\
2014 & 8 & 14 & 56884 & 16.16\
2015 & 2 & 15 & 57069 & 15.98\
2015 & 9 & 5 & 57271 & 16.14\
2016 & 4 & 9 & 57488 & 16.09\
2017 & 5 & 14 & 57888 & 16.08\
\
MASTER OT J192757.03$+$404042.8 {#obj:j1927}
-------------------------------
This object (hereafter MASTER J192757) is a transient detected at an unfiltered CCD magnitude of 15.2 on 2014 February 13 [@bal14j1927atel6024]. The 2017 superoutburst was detected at $V$=14.36 on 2017 April 4 by the ASAS-SN team. The ASAS-SN data recorded $V$=13.83 on 2017 April 1. Subsequent observations detected superhumps (vsnet-alert 20885; figure \[fig:j1927shpdm\]). The times of superhump maxima are listed in e-table \[tab:j1927oc2017\]. Although there remained some ambiguity in the one-day alias, the other aliases appear to be excluded by using the observations on the first night. The best superhump period by the PDM method was 0.08161(5) d.
There were two likely superoutbursts in the past ASAS-SN data: 2014 March 18 ($V$=13.8) and 2015 June 21 ($V$=14.1). Since the object is sufficiently bright, future observations will clarify more detailed development of superhumps and outburst statistics.
(85mm,110mm)[j1927shpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57849.5593 & 0.0003 & $-$0&0014 & 80\
1 & 57849.6437 & 0.0008 & 0&0015 & 30\
12 & 57850.5382 & 0.0004 & $-$0&0001 & 85\
\
\
\
MASTER OT J200904.69$+$825153.6 {#obj:j2009}
-------------------------------
This object (hereafter MASTER J200904) is a transient detected at an unfiltered CCD magnitude of 15.6 on 2014 March 10 [@bal14j2009atel5974]. The 2017 outburst was detected by the ASAS-SN team at $V$=15.70 on 2017 June 2. The large outburst amplitude [@bal14j2009atel5974] suspected an SU UMa-type dwarf nova. Subsequent observations detected large-amplitude superhumps (vsnet-alert 21089, 21093; e-figure \[fig:j2009shpdm\]). The times of superhump maxima are listed in e-table \[tab:j2009oc2017\].
(85mm,110mm)[j2009shpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57909.0440 & 0.0038 & $-$0&0025 & 46\
1 & 57909.1194 & 0.0008 & $-$0&0004 & 81\
2 & 57909.1943 & 0.0007 & 0&0010 & 82\
3 & 57909.2593 & 0.0030 & $-$0&0074 & 32\
4 & 57909.3378 & 0.0015 & $-$0&0022 & 36\
5 & 57909.4139 & 0.0010 & 0&0004 & 78\
6 & 57909.4873 & 0.0007 & 0&0004 & 82\
7 & 57909.5638 & 0.0009 & 0&0035 & 51\
14 & 57910.0791 & 0.0018 & 0&0050 & 73\
15 & 57910.1482 & 0.0017 & 0&0006 & 84\
16 & 57910.2206 & 0.0014 & $-$0&0004 & 75\
19 & 57910.4455 & 0.0012 & 0&0043 & 70\
20 & 57910.5157 & 0.0010 & 0&0010 & 70\
55 & 57913.0799 & 0.0037 & $-$0&0041 & 65\
69 & 57914.1124 & 0.0029 & 0&0007 & 40\
\
\
\
MASTER OT J205110.36$+$044842.2 {#obj:j2051}
-------------------------------
This object (hereafter MASTER J205110) was detected as a quasar or a dwarf nova at an unfiltered CCD magnitude of 15.0 on 2017 September 28 by the MASTER network [@shu17j2051atel10790]. Two past outbursts (2001 June and 2002 August) were detected by D. Denisenko in past images and Denisenko suggested it to be an SU UMa-type dwarf nova based on the past outbursts, ROSAT identification and the SDSS colors (vsnet-alert 21483). Subsequent observations detected superhumps (vsnet-alert 21486, 21487, 21488; e-figure \[fig:j2051shpdm\]). The times of superhump maxima are listed in e-table \[tab:j2051oc2017\].
(85mm,110mm)[j2051shpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 58025.5946 & 0.0005 & 0&0029 & 262\
1 & 58025.6720 & 0.0005 & $-$0&0004 & 265\
18 & 58027.0415 & 0.0014 & $-$0&0030 & 132\
19 & 58027.1245 & 0.0007 & $-$0&0007 & 196\
22 & 58027.3685 & 0.0015 & 0&0011 & 88\
34 & 58028.3343 & 0.0015 & $-$0&0015 & 79\
59 & 58030.3552 & 0.0037 & 0&0016 & 85\
\
\
\
MASTER OT J212624.16$+$253827.2 {#obj:j2126}
-------------------------------
This object (hereafter MASTER J212624) was detected as a transient at an unfiltered CCD magnitude of 14.1 on 2013 June 26 by the MASTER network [@den13j2126atel5111]. The SU UMa-type nature was confirmed during the 2013 superoutburst. The object attracted attention since the superhump period indicates an SU UMa-type dwarf nova in the period gap see @Pdot5 and @Pdot8 for more information.
The 2017 superoutburst was detected by the ASAS-SN team at $V$=14.90 on 2017 August 25. The object further brightened to $V$=14.2 in 0.85 d. The initial observations on 2017 August 26–27 did not show large-amplitude superhumps. They became apparent on the next night and grew further (vsnet-alert 21374, 21377). This was the first time in this object growing superhumps were recorded. The times of superhump maxima are listed in e-table \[tab:j2126oc2017\]. Stage A was impressively long. In determining periods, we neglected 32 $\le E \le$ 35, which were likely a transition phase between stages A and B. A comparison of the $O-C$ diagrams suggests that we observed stage A and early half of stage B (e-figure \[fig:j2126comp\]. We if could have continued observations, we could have detected a strongly positive $P_{\rm dot}$ as in the 2013 superoutburst.
This object adds a new example of long developing time of superhumps in long-$P_{\rm orb}$ systems (see subsection \[obj:j0809\]).
This object showed relatively regular superoutburst (e-table \[tab:j2126out\]). The supercycle was 345(9) d.
(85mm,70mm)[j2126comp.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57992.3487 & 0.0036 & $-$0&0338 & 93\
2 & 57992.5291 & 0.0040 & $-$0&0377 & 62\
21 & 57994.3239 & 0.0003 & 0&0068 & 128\
22 & 57994.4164 & 0.0002 & 0&0072 & 355\
23 & 57994.5108 & 0.0002 & 0&0094 & 323\
24 & 57994.6061 & 0.0004 & 0&0127 & 148\
32 & 57995.3457 & 0.0002 & 0&0152 & 224\
33 & 57995.4367 & 0.0002 & 0&0141 & 332\
34 & 57995.5299 & 0.0002 & 0&0152 & 299\
35 & 57995.6260 & 0.0006 & 0&0191 & 54\
43 & 57996.3555 & 0.0008 & 0&0117 & 74\
45 & 57996.5381 & 0.0002 & 0&0100 & 236\
46 & 57996.6289 & 0.0024 & 0&0087 & 43\
46 & 57996.6289 & 0.0024 & 0&0087 & 43\
55 & 57997.4480 & 0.0002 & $-$0&0013 & 309\
56 & 57997.5387 & 0.0003 & $-$0&0027 & 227\
66 & 57998.4412 & 0.0028 & $-$0&0215 & 65\
67 & 57998.5371 & 0.0008 & $-$0&0176 & 131\
75 & 57999.2675 & 0.0011 & $-$0&0243 & 97\
\
\
\
[ccccc]{} Year & Month & Day & max& $V$ mag\
2013 & 11 & 16 & 56613 & 14.9\
2014 & 9 & 10 & 56911 & 14.5\
2015 & 8 & 26 & 57261 & 14.2\
2016 & 7 & 31 & 57601 & 14.4\
2017 & 8 & 25 & 57991 & 14.2\
\
OT J182142.8$+$212154 {#obj:j1821}
---------------------
This object (hereafter OT J182142) was discovered by K. Itagaki at an unfiltered CCD magnitude of 14.9 on 2010 April 24 (vsnet-alert 11952). Subsequent observations confirmed the SU UMa-type classification [@Pdot2].
The 2017 superoutburst was detected by the ASAS-SN team at $V$=14.82 on 2017 May 31. Subsequent observations confirmed superhumps (vsnet-alert 21083, 21099). The times of superhump maxima are listed in e-table \[tab:j1821oc2017\].
The object faded by 1.4 mag between 2017 June 3 and 5, and it was likely the termination of the superoutburst. It was likely that the superoutburst was not detected sufficiently early and there was a chance that we only observed stage C superhumps (this may have been also the case for the 2010 observations).
This field has been covered by the ASAS-SN team since 2014 March 18. There was only one previous outburst (type unknown) in the record at $V$=15.8 on 2016 February 15.
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57907.1707 & 0.0090 & $-$0&0067 & 30\
1 & 57907.2600 & 0.0010 & 0&0005 & 79\
2 & 57907.3477 & 0.0016 & 0&0060 & 37\
13 & 57908.2475 & 0.0026 & 0&0023 & 45\
15 & 57908.4107 & 0.0007 & 0&0012 & 44\
16 & 57908.4888 & 0.0008 & $-$0&0028 & 40\
39 & 57910.3813 & 0.0016 & 0&0004 & 44\
40 & 57910.4622 & 0.0026 & $-$0&0008 & 41\
\
\
\
OT J204222.3$+$271211 {#obj:j2042}
---------------------
This object (PNV J20422233$+$2712111, hereafter OT J204222) is a transient discovered independently by H. Nishimura, T. Kojima and Kaneko at an unfiltered CCD magnitude of 11.1 on 2017 April 13.[^12] The object was immediately suspected to be a dwarf nova, not a classical nova as originally suspected, by the presence of a faint blue object in SDSS and a GALEX ultraviolet object. The large outburst amplitude (nearly 9 mag) suggested a WZ Sge-type dwarf nova (vsnet-alert 20915, see also $<$https://www.aavso.org/pnv-j204222332712111-new-transient-111-mag-vulpecula$>$).
Despite the large outburst amplitude, observations of this object were rather sparse in the early morning and expected early superhumps were not detected. The object was found to show ordinary superhumps on 2017 April 22 (vsnet-alert 20941, 20959; e-figure \[fig:j2042shpdm\]). A retrospective examination suggested that growing superhumps were probably already present on 2017 April 18 (vsnet-alert 20960), only 5 d after the outburst detection. The times of superhump maxima are listed in e-table \[tab:j2042oc2017\]. Due to the lack of observations, the period of stage A was not determined and $P_{\rm dot}$ for stage B superhumps was not well determined (it may be almost zero considering the large uncertainty).
In the ASAS-SN data, the object was detected at $V$=12.2 on 2017 April 15 but was still in quiescence or still very faint on 2017 April 9. The waiting time of the emergence of superhumps was thus shorter than 9 d.
(85mm,110mm)[j2042shpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57861.8675 & 0.0010 & $-$0&0046 & 417\
65 & 57865.5307 & 0.0015 & 0&0058 & 61\
66 & 57865.5821 & 0.0003 & 0&0010 & 102\
119 & 57868.5587 & 0.0003 & $-$0&0009 & 81\
120 & 57868.6163 & 0.0005 & 0&0006 & 47\
156 & 57870.6483 & 0.0015 & 0&0095 & 25\
165 & 57871.1388 & 0.0029 & $-$0&0058 & 25\
166 & 57871.1956 & 0.0006 & $-$0&0052 & 61\
167 & 57871.2567 & 0.0012 & $-$0&0004 & 16\
\
\
\
PNV J20205397$+$2508145 {#obj:j2020}
-----------------------
This object (hereafter PNV J202053) was detected as a transient by T. Kojima at an unfiltered CCD magnitude of 12.3 on 2017 September 12.[^13] The outburst was recorded in the ASAS-SN data at $V$=12.9 on 2017 September 13. The object was not in outburst on 2017 September 8. A blue quiescent counterpart with $g$=20.41 was identified by Brian Skiff and astrometry by Andrea Mantero (cf. vsnet-alert 21424, 21426). Subsequent photometry detected early superhumps (vsnet-alert 21427, 21431, 21436; e-figure \[fig:j2020eshpdm\]) and the object was identified as a WZ Sge-type dwarf nova. The object started to show ordinary superhumps (e-figure \[fig:j2020shpdm\]) on 2017 September 19 (see vsnet-alert 21455, 21460 and analysis in this subsection). The times of maxima of ordinary superhumps are listed in e-table \[tab:j2020oc2017\]. Stages A and B were very clearly recorded and there was possibly a transition to stage C during the final decline from the superoutburst plateau (e-figure \[fig:j2020humpall\]).
The period of early superhump by the PDM method was 0.056509(5) d (e-figure \[fig:j2020eshpdm\]). The value of $\epsilon^*$ for stage A superhumps was 0.0329(11), which corresponds to $q$=0.090(3). This value appears to be typical for a WZ Sge-type dwarf nova. There was no post-superoutburst rebrightening.
(85mm,110mm)[j2020eshpdm.eps]{}
(85mm,110mm)[j2020shpdm.eps]{}
(160mm,200mm)[j2020humpall.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lrrp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & & & & & & &\
0 & 58016.2541 & 0.0005 & $-$0&0280 & 114 & 111 & 58022.6571 & 0.0002 & 0&0001 & 226\
1 & 58016.3131 & 0.0005 & $-$0&0265 & 119 & 112 & 58022.7139 & 0.0002 & $-$0&0005 & 180\
2 & 58016.3642 & 0.0016 & $-$0&0328 & 72 & 113 & 58022.7706 & 0.0002 & $-$0&0013 & 164\
3 & 58016.4221 & 0.0013 & $-$0&0323 & 74 & 114 & 58022.8302 & 0.0004 & 0&0009 & 50\
18 & 58017.3023 & 0.0008 & $-$0&0136 & 121 & 118 & 58023.0612 & 0.0018 & 0&0022 & 48\
19 & 58017.3617 & 0.0012 & $-$0&0116 & 129 & 119 & 58023.1131 & 0.0017 & $-$0&0034 & 63\
33 & 58018.1731 & 0.0031 & $-$0&0043 & 51 & 127 & 58023.5714 & 0.0026 & $-$0&0046 & 24\
35 & 58018.3057 & 0.0042 & 0&0134 & 25 & 128 & 58023.6304 & 0.0003 & $-$0&0029 & 60\
36 & 58018.3520 & 0.0003 & 0&0023 & 138 & 129 & 58023.6875 & 0.0004 & $-$0&0033 & 58\
37 & 58018.4103 & 0.0003 & 0&0032 & 147 & 130 & 58023.7455 & 0.0004 & $-$0&0027 & 52\
38 & 58018.4711 & 0.0006 & 0&0066 & 88 & 139 & 58024.2650 & 0.0013 & $-$0&0001 & 62\
39 & 58018.5281 & 0.0004 & 0&0062 & 74 & 140 & 58024.3183 & 0.0003 & $-$0&0043 & 117\
41 & 58018.6450 & 0.0002 & 0&0082 & 74 & 141 & 58024.3755 & 0.0004 & $-$0&0044 & 113\
42 & 58018.7022 & 0.0002 & 0&0080 & 156 & 142 & 58024.4342 & 0.0023 & $-$0&0032 & 72\
43 & 58018.7611 & 0.0003 & 0&0095 & 154 & 145 & 58024.6040 & 0.0011 & $-$0&0057 & 34\
44 & 58018.8236 & 0.0007 & 0&0145 & 34 & 146 & 58024.6625 & 0.0005 & $-$0&0046 & 60\
53 & 58019.3379 & 0.0001 & 0&0119 & 290 & 147 & 58024.7193 & 0.0004 & $-$0&0052 & 61\
54 & 58019.3947 & 0.0001 & 0&0113 & 245 & 157 & 58025.2942 & 0.0002 & $-$0&0046 & 116\
55 & 58019.4509 & 0.0002 & 0&0101 & 154 & 158 & 58025.3501 & 0.0004 & $-$0&0062 & 117\
56 & 58019.5055 & 0.0033 & 0&0072 & 31 & 159 & 58025.4096 & 0.0006 & $-$0&0041 & 71\
59 & 58019.6799 & 0.0002 & 0&0093 & 90 & 161 & 58025.5231 & 0.0020 & $-$0&0055 & 52\
60 & 58019.7373 & 0.0002 & 0&0093 & 89 & 162 & 58025.5831 & 0.0006 & $-$0&0029 & 60\
61 & 58019.7945 & 0.0002 & 0&0090 & 77 & 163 & 58025.6371 & 0.0006 & $-$0&0064 & 59\
64 & 58019.9653 & 0.0012 & 0&0076 & 24 & 170 & 58026.0419 & 0.0030 & $-$0&0036 & 51\
69 & 58020.2501 & 0.0063 & 0&0052 & 18 & 171 & 58026.0952 & 0.0031 & $-$0&0077 & 52\
70 & 58020.3080 & 0.0007 & 0&0056 & 78 & 172 & 58026.1582 & 0.0037 & $-$0&0021 & 67\
71 & 58020.3649 & 0.0010 & 0&0051 & 146 & 186 & 58026.9613 & 0.0010 & $-$0&0031 & 51\
72 & 58020.4225 & 0.0005 & 0&0053 & 73 & 187 & 58027.0187 & 0.0015 & $-$0&0031 & 56\
73 & 58020.4794 & 0.0004 & 0&0048 & 73 & 188 & 58027.0754 & 0.0014 & $-$0&0039 & 55\
74 & 58020.5385 & 0.0006 & 0&0064 & 59 & 189 & 58027.1315 & 0.0020 & $-$0&0052 & 56\
76 & 58020.6520 & 0.0002 & 0&0051 & 150 & 196 & 58027.5351 & 0.0010 & $-$0&0036 & 47\
77 & 58020.7096 & 0.0002 & 0&0053 & 158 & 197 & 58027.5944 & 0.0006 & $-$0&0018 & 56\
78 & 58020.7660 & 0.0002 & 0&0043 & 160 & 198 & 58027.6487 & 0.0007 & $-$0&0049 & 60\
79 & 58020.8226 & 0.0004 & 0&0034 & 51 & 199 & 58027.7075 & 0.0008 & $-$0&0035 & 60\
87 & 58021.2844 & 0.0006 & 0&0057 & 128 & 200 & 58027.7661 & 0.0024 & $-$0&0023 & 37\
88 & 58021.3399 & 0.0005 & 0&0038 & 314 & 227 & 58029.3157 & 0.0036 & $-$0&0034 & 32\
89 & 58021.3970 & 0.0007 & 0&0035 & 298 & 229 & 58029.4351 & 0.0017 & 0&0012 & 48\
90 & 58021.4548 & 0.0005 & 0&0039 & 167 & 230 & 58029.5082 & 0.0124 & 0&0168 & 36\
91 & 58021.5106 & 0.0003 & 0&0022 & 111 & 231 & 58029.5570 & 0.0026 & 0&0082 & 58\
93 & 58021.6342 & 0.0007 & 0&0109 & 39 & 232 & 58029.6118 & 0.0016 & 0&0055 & 58\
94 & 58021.6830 & 0.0003 & 0&0023 & 57 & 233 & 58029.6615 & 0.0018 & $-$0&0022 & 36\
95 & 58021.7395 & 0.0003 & 0&0014 & 60 & 246 & 58030.4194 & 0.0031 & 0&0092 & 128\
96 & 58021.7987 & 0.0002 & 0&0032 & 53 & 248 & 58030.5280 & 0.0011 & 0&0028 & 46\
97 & 58021.8557 & 0.0010 & 0&0028 & 37 & 249 & 58030.5850 & 0.0017 & 0&0025 & 60\
105 & 58022.3177 & 0.0012 & 0&0053 & 54 & 250 & 58030.6388 & 0.0011 & $-$0&0012 & 55\
109 & 58022.5451 & 0.0023 & 0&0029 & 29 & 261 & 58031.2644 & 0.0084 & $-$0&0074 & 11\
110 & 58022.5997 & 0.0002 & 0&0001 & 71 & 263 & 58031.3750 & 0.0019 & $-$0&0116 & 39\
\
\
\
SDSS J152857.86$+$034911.7 {#obj:j1528}
--------------------------
This object (hereafter SDSS J152857) was selected as a CV during the course of the SDSS [@szk03SDSSCV2]. Some outbursts were recorded ([@wil10j1924]; [@dra14CRTSCVs]). No secure orbital variations were detected by @wou12SDSSCRTSCVs and it was concluded that the object has a very long orbital period or it is of low inclination. @kat12DNSDSS estimated an orbital period of 0.082(3) d from SDSS colors. The 2017 outburst was detected by the ASAS-SN team at $V$=15.62 on 2017 May 16. Observations on 2017 May 21–22 detected superhumps (vsnet-alert 21056; e-figure \[fig:j1528shpdm\]; e-table \[tab:j1528oc2017\]).
(85mm,110mm)[j1528shpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57895.3966 & 0.0011 & 0&0003 & 41\
1 & 57895.4598 & 0.0011 & 0&0000 & 38\
2 & 57895.5223 & 0.0007 & $-$0&0008 & 64\
3 & 57895.5871 & 0.0028 & 0&0005 & 25\
\
\
\
SDSS J153015.04$+$094946.3 {#obj:j1530}
--------------------------
This object (hereafter SDSS J153015) was originally selected as a CV by the SDSS [@szk09SDSSCV7]. The SU UMa-type nature was confirmed during the 2017 March superoutburst [@Pdot9]. The supercycle was estimated to be very short \[84.7(1.2) d\].
The 2017 June superoutburst was detected by the ASAS-SN team at $V$=15.86 on 2017 June 13. Subsequent observations detected superhumps (vsnet-alert 21121). The times of superhump maxima are listed in e-table \[tab:j1530oc2017b\].
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57919.0487 & 0.0013 & 0&0011 & 86\
1 & 57919.1257 & 0.0020 & 0&0029 & 83\
4 & 57919.3495 & 0.0017 & 0&0008 & 67\
5 & 57919.4249 & 0.0008 & 0&0009 & 86\
13 & 57920.0182 & 0.0031 & $-$0&0083 & 63\
18 & 57920.4043 & 0.0011 & 0&0012 & 88\
26 & 57921.0077 & 0.0013 & 0&0021 & 83\
27 & 57921.0717 & 0.0022 & $-$0&0092 & 83\
31 & 57921.3873 & 0.0016 & 0&0052 & 25\
32 & 57921.4608 & 0.0040 & 0&0034 & 27\
\
\
\
SDSS J204817.85$-$061044.8 {#obj:j2048}
--------------------------
This object (hereafter SDSS J204817) is a CV selected during the course of the SDSS [@szk03SDSSCV2]. The object was identified as an SU UMa-type dwarf nova during the 2009 superoutburst. @wou10CVperiod observed the object in quiescence and obtained an orbital period of 0.060597(2) d. @Pdot2 reported a superhump period of 0.06166(2) d based on the orbital period by @wou10CVperiod.
The 2017 superoutburst was detected by the ASAS-SN team at $V$=14.55 on 2017 May 30. Subsequent observations detected superhumps (vsnet-alert 21087). Although we obtained data on two nights, the data on the first night was of less quality and the apparent hump was not expressed by the 2009 period, we only used the data on the second night. The times of superhump maxima are listed in e-table \[tab:j2048oc2017\]. The quality of the data during the 2017 superoutburst was poorer and we could not give meaningful constraint on the superhump period. We instead provide an updated period analysis of the 2009 data using a modern (LOWESS) detrending method. The updated result favors the alias corresponding to the orbital period (e-figure \[fig:j2048shpdm\]).
In the ASAS-SN data, there was a superoutburst (maximum $V$=14.9 on 2015 October 19) preceded by a precursor outburst on 2015 October 15 ($V$=15.3). The behavior appears to be typical for an SU UMa-type dwarf nova.
(85mm,110mm)[j2048shpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57908.4832 & 0.0100 & 0&0006 & 20\
1 & 57908.5505 & 0.0013 & $-$0&0011 & 30\
2 & 57908.6212 & 0.0031 & 0&0006 & 12\
\
\
\
TCP J00332502$-$3518565 {#obj:j0033}
-----------------------
This object (hereafter TCP J003325) was discovered by Shigehisa Fujikawa at an unfiltered CCD magnitude of 12.8–13.3 on 2017 August 5.[^14] The object was suspected to be a WZ Sge-type dwarf nova (vsnet-alert 21320). There was a 2003 outburst in the ASAS-3 data (D. Denisenko, vsnet-alert 21321). The object was also recorded by the ASAS-SN team at $V$=14.7 (rising) in 2017 August 5 and $V$=12.7 on 2017 August 7. Well-developed superhumps were observed since 2017 August 18 (vsnet-alert 21357; e-figure \[fig:j0033shpdm\]). The times of superhump maxima are listed in e-table \[tab:j0033oc2017\]. We retrospectively identified that ordinary superhumps emerged on 2017 August 13 (0 $\le E \le$ 2) based on the flattening of the outburst light curve and low superhump amplitudes. Although these superhumps were presumably stage A, we could not determine the period due to a 4-d gap in the observation after 2017 August 13. There were possible very low-amplitude early superhumps (e-figure \[fig:j0033eshpdm\]) with a period of 0.05484(16) d.
The relatively early (8 d after the outburst rise) appearance and the presence of the 2003 outburst suggests that this object is not an extreme WZ Sge-type dwarf nova.
(85mm,110mm)[j0033shpdm.eps]{}
(85mm,110mm)[j0033eshpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57979.4129 & 0.0018 & $-$0&0003 & 127\
1 & 57979.4705 & 0.0023 & 0&0022 & 126\
2 & 57979.5338 & 0.0023 & 0&0103 & 127\
91 & 57984.4321 & 0.0002 & $-$0&0016 & 127\
92 & 57984.4871 & 0.0002 & $-$0&0019 & 124\
109 & 57985.4236 & 0.0002 & $-$0&0032 & 127\
110 & 57985.4787 & 0.0003 & $-$0&0033 & 98\
163 & 57988.4006 & 0.0004 & $-$0&0054 & 107\
165 & 57988.5114 & 0.0003 & $-$0&0050 & 127\
166 & 57988.5648 & 0.0002 & $-$0&0068 & 123\
220 & 57991.5505 & 0.0009 & $-$0&0003 & 127\
222 & 57991.6606 & 0.0007 & $-$0&0006 & 109\
255 & 57993.4897 & 0.0016 & 0&0080 & 127\
256 & 57993.5447 & 0.0013 & 0&0078 & 127\
\
\
\
TCP J20100517$+$1303006 {#obj:j2010}
-----------------------
This object (hereafter TCP J201005) was discovered as a transient at an unfiltered CCD magnitude of 12.6 on 2017 June 19 by Tadashi Kojima.[^15] There is an ROSAT X-ray source 1RXS J201006.4$+$130259 (vsnet-alert 21146) and the object was considered to be a dwarf nova. Although initial observations suggested the presence of early superhumps (vsnet-alert 21149, 21153), The object was later found to be an ordinary SU UMa-type dwarf nova (vsnet-alert 21156; the initial claim of early superhumps was due to the relatively strong secondary maximum of superhumps; e-figure \[fig:j2010shpdm\]). The object entered the rapidly fading stage on 2017 June 28 (vsnet-alert 21174).
The times of superhump maxima during the superoutburst and early post-superoutburst phase are listed in e-table \[tab:j2010oc2017\]. Although there were observations on BJD 2457931, the hump profile became complex with strong secondary humps and it was difficult to measure superhump maxima. The $O-C$ values for post-superoutburst maxima are strongly negative, indicating that they were traditional late superhumps ($\sim$0.5 phase jump around the termination of the superoutburst). The superhump stage was unknown. Since the profile was already doubly humped at the time of initial observations (see also e-figure \[fig:j2010shpdm\]), the observations were not early enough to detect typical stage B superhumps The observed superhumps may have been already stage C ones.
The object was recorded to undergo frequent outbursts (cf. vsnet-alert 21192, 21197, 21203) and they were once confused as rebrightenings. They were shown to be normal outbursts occurring every 5–10 d (vsnet-alert 21235). Superoutbursts recorded in the ASAS-SN data are summarized in e-table \[tab:j2010out\]. The coverage by the ASAS-SN observations of this field was not good enough as in other objects and true maxima were not necessarily detected. The data, however, were sufficient to estimate the supercycle of 103.5(1.3) d. The short supercycle suggests a high mass-transfer rate and it is consistent with the appearance of traditional late superhumps.
(85mm,110mm)[j2010shpdm.eps]{}
[rp[55pt]{}p[40pt]{}r@[.]{}lr]{} & & & &\
0 & 57927.3706 & 0.0096 & $-$0&0061 & 42\
1 & 57927.4535 & 0.0010 & $-$0&0039 & 180\
2 & 57927.5285 & 0.0018 & $-$0&0095 & 123\
9 & 57928.0947 & 0.0039 & $-$0&0078 & 48\
10 & 57928.1808 & 0.0008 & $-$0&0023 & 76\
11 & 57928.2610 & 0.0012 & $-$0&0028 & 59\
13 & 57928.4230 & 0.0003 & $-$0&0021 & 276\
14 & 57928.5024 & 0.0004 & $-$0&0034 & 304\
15 & 57928.5929 & 0.0011 & 0&0065 & 41\
17 & 57928.7515 & 0.0008 & 0&0038 & 87\
18 & 57928.8280 & 0.0005 & $-$0&0003 & 110\
19 & 57928.9089 & 0.0007 & $-$0&0001 & 57\
26 & 57929.4755 & 0.0011 & 0&0020 & 42\
30 & 57929.7951 & 0.0015 & $-$0&0010 & 44\
31 & 57929.8812 & 0.0007 & 0&0045 & 142\
38 & 57930.4484 & 0.0007 & 0&0071 & 189\
39 & 57930.5309 & 0.0011 & 0&0090 & 169\
40 & 57930.6122 & 0.0016 & 0&0096 & 77\
42 & 57930.7740 & 0.0011 & 0&0102 & 185\
43 & 57930.8514 & 0.0009 & 0&0069 & 188\
44 & 57930.9385 & 0.0014 & 0&0134 & 80\
67 & 57932.7677 & 0.0020 & $-$0&0123 & 28\
68 & 57932.8509 & 0.0018 & $-$0&0097 & 22\
79 & 57933.7416 & 0.0024 & $-$0&0061 & 25\
80 & 57933.8229 & 0.0023 & $-$0&0055 & 23\
\
\
\
[ccccc]{} Year & Month & Day & max& $V$ mag\
2015 & 6 & 10 & 57184 & 13.4\
2015 & 9 & 24 & 57290 & 13.7\
2016 & 4 & 10 & 57489 & 13.5\
2016 & 7 & 23 & 57593 & 13.7\
2016 & 11 & 14 & 57707 & 13.8\
\
\
\
Alksnis, A., & Zharova, A. V. 2000, IBVS, 4909
Antipin, S. V. 1996, IBVS, 4343
Antipin, S. V., & Pavlenko, E. P. 2002, A&A, 391, 565
, T., [Tappert]{}, C., [Dall]{}, T., & [Maza]{}, J. 2010, MNRAS, 405, 621
, A., [et al.]{} 2010, ApJ, 711, 389
, P., [Denisenko]{}, D., [Gorbovskoy]{}, E., & [Lipunov]{}, V. 2013, Perem. Zvezdy, submitted (arXiv/1307.7396)
, P., [et al.]{} 2014a, Astron. Telegram, 6024
, P., [et al.]{} 2012, Astron. Telegram, 4022
, P., [et al.]{} 2017, Astron. Telegram, 10470
, P., [et al.]{} 2014b, Astron. Telegram, 5974
, K., [et al.]{} 2017, A&A, 603, A72
, K., [Olech]{}, A., [Pospieszy[ń]{}ski]{}, R., [Martinelli]{}, F., & [Marciniak]{}, A. 2014, Acta Astron., 64, 337
, D., [et al.]{} 2010, J. Br. Astron. Assoc., 120, 33
, A. J. 1925, Harvard Coll. Obs. Bull., 825, 1
, R., [et al.]{} 2017, Astron. Telegram, 10334
, W. S. 1979, J. Amer. Statist. Assoc., 74, 829
, J. M. C., [et al.]{} 2019, MNRAS, 488, 4149
, A. B., [Shappee]{}, B. J., [Archer Shappee]{}, B., & [ASAS-SN]{} 2015, American Astron. Soc. Meeting Abstracts, 225, \#344.02
, A. B., [et al.]{} 2014, Astron. Telegram, 6211
, D. 2017, Astron. Telegram, 10480
, D., [et al.]{} 2013, Astron. Telegram, 5111
, M., [et al.]{} 2008, MNRAS, 386, 1568
, A. J., [et al.]{} 2009, ApJ, 696, 870
, A. J., [et al.]{} 2014, MNRAS, 441, 1186
, L. K. 1973, Astron. Tsirk., 774, 5
Fernie, J. D. 1989, PASP, 101, 225
, E. S., [et al.]{} 2013, Astron. Rep., 57, 233
Green, R. F., Ferguson, D. H., Liebert, J., & Schmidt, M. 1982, PASP, 94, 560
Green, R. F., Schmidt, M., & Liebert, J. 1986, ApJS, 61, 305
Grubissich, C., & Rosino, L. 1958, Asiago Contr., 93, 1
Harvey, D., Skillman, D. R., Patterson, J., & Ringwald, F. A. 1995, PASP, 107, 551
Hellier, C. 2001, PASP, 113, 469
Henden, A. A., Munari, U., & Sumner, B. 2001, IBVS, 5140
, C. 1949a, Erg. Astron. Nachr., 12, 12
, C. 1949b, Erg. Astron. Nachr., 12, 1
, C. 1957a, Mitteil. Ver[ä]{}nderl. Sterne, 1, 245
, C. 1957b, Mitteil. Ver[ä]{}nderl. Sterne, 1, 245
, C. 1963, Ver[ö]{}ff. Sternw. Sonneberg, 6, 1
, C. 1964, Astron. Nachr., 288, 49
, C. 1967, Astron. Nachr., 290, 43
, A., [et al.]{} 2017, PASJ, 69, 72
, Keisuke, [Kato]{}, Taichi, [Monard]{}, Berto, [Hambsch]{}, Franz-Josef, [Myers]{}, Gordon, [Starr]{}, Peter, [Cook]{}, Lewis M., & [Nogami]{}, Daisaku 2019, PASJ, 71, 48
, A. B., & [Thorstensen]{}, J. R. 2006, PASP, 118, 1119
, T. 2015, PASJ, 67, 108
, T. 2019, PASJ, submitted
, T., [et al.]{} 2016a, PASJ, 68, 49
, T., [et al.]{} 2014a, PASJ, 66, 90
, T., [et al.]{} 2015, PASJ, 67, 105
, T., [et al.]{} 2013a, PASJ, 65, 23
, T., [et al.]{} 2014b, PASJ, 66, 30
, T., [Hambsch]{}, F.-J., [Monard]{}, B., [Nelson]{}, P., [Stubbings]{}, R., & [Starr]{}, P. 2019a, PASJ, in press (arXiv/1909.00910)
, T., [et al.]{} 2016b, PASJ, 68, 65
, T., [et al.]{} 2009, PASJ, 61, S395
Kato, T., Ishioka, R., & Uemura, M. 2002, PASJ, 54, 1029
, T., [et al.]{} 2017a, PASJ, 69, 75
, T., [et al.]{} 2012a, PASJ, 64, 21
, T., [Maehara]{}, H., & [Uemura]{}, M. 2012b, PASJ, 64, 62
, T., [et al.]{} 2010, PASJ, 62, 1525
, T., [Monard]{}, B., [Hambsch]{}, F.-J., [Kiyota]{}, S., & [Maehara]{}, H. 2013b, PASJ, 65, L11
Kato, T., Nogami, D., Masuda, S., & Baba, H. 1998, PASP, 110, 1400
Kato, T., Nogami, D., Masuda, S., & Hirata, R. 1995, IBVS, 4193
, T., & [Osaki]{}, Y. 2013, PASJ, 65, 115
, T., [et al.]{} 2019b, PASJ, 71, L1
, T., [et al.]{} 2016c, PASJ, 68, L4
, T., [Sekine]{}, Y., & [Hirata]{}, R. 2001, PASJ, 53, 1191
Kato, T., Stubbings, R., Pearce, A., Dubovsky, P. A., Kiyota, S., Itoh, H., & Simonsen, M. 2001, IBVS, 5109
, T., [et al.]{} 2017b, PASJ, 69, L4
Kato, T., Uemura, M., Ishioka, R., Nogami, D., Kunjaya, C., Baba, H., & Yamaoka, H. 2004, PASJ, 56, S1
, N., [Chochol]{}, D., & [Shugarov]{}, S. 2014, Contr. of the Astron. Obs. Skalnaté Pleso, 43, 306
, A. V. 2005, Perem. Zvezdy, Prilozh., 5, 4
, M., [et al.]{} 2016a, PASJ, 68, 55
, M., [et al.]{} 2018, PASJ, 70, 47
, M., [et al.]{} 2016b, PASJ, 68, L2
, T., & [Skiff]{}, B. A. 2000, IBVS, 4896
, C., [et al.]{} 2013, AJ, 145, 145
, C., [Garnavich]{}, P., [Kennedy]{}, M., [Szkody]{}, P., & [Dai]{}, Z. 2018, AJ, 155, 232
Liu, Wu., & Hu, J. Y. 2000, ApJS, 128, 387
Liu, Wu., Hu, J. Y., Zhu, X. H., & Li, Z. Y. 1999, ApJS, 122, 243
, W. J. 1938, Publ. of the Astron. Obs. Univ. of Minnesota, 6, 1
, T., [Parsons]{}, S., & [Dhillon]{}, V. 2017, Astron. Telegram, 10354
Mason, E., & Howell, S. 2003, A&A, 403, 699
, M., [et al.]{} 2019, MNRAS, 486, 5535
Mennickent, R. E., Matsumoto, K., & Arenas, J. 1999, A&A, 348, 466
Motch, C., Haberl, F., Guillout, P., Pakull, M., Reinsch, K., & Krautter, J. 1996, A&A, 307, 459
, P., [et al.]{} 2015, Acta Astron., 65, 313
, C., [et al.]{} 2013, PASJ, 65, 117
, K., [et al.]{} 2017, PASJ, 69, 2
, V. V., [et al.]{} 2017, MNRAS, 467, 597
, V. V., [et al.]{} 2018, A&A, 611, A13
Nogami, D., & Kato, T. 1995, IBVS, 4227, 1
Nogami, D., [et al.]{} 2003, A&A, 404, 1067
Novák, R. 1997, IBVS, 4489
, R., [et al.]{} 2019, PASJ, in press
, T., [et al.]{} 2012, PASJ, 64, L3
, Y., & [Kato]{}, T. 2013a, PASJ, 65, 50
, Y., & [Kato]{}, T. 2013b, PASJ, 65, 95
, E. Y. 1985, Perem. Zvezdy, 22, 261
, A. F., [et al.]{} 2019, MNRAS, 483, 1080
, A. F., [Schmidtobreick]{}, L., [Tappert]{}, C., [G[ä]{}nsicke]{}, B. T., & [Mehner]{}, A. 2018, MNRAS, 481, 2523
, J., [et al.]{} 2005, PASP, 117, 1204
Patterson, J., [et al.]{} 2003, PASP, 115, 1308
, J., [Thorstensen]{}, J. R., & [Knigge]{}, C. 2008, PASP, 120, 510
, E., [et al.]{} 2019, Contr. of the Astron. Obs. Skalnaté Pleso, 49, 204
, E. P., [Samsonov]{}, D. A., [Antonyuk]{}, O. I., [Andreev]{}, M. V., [Baklanov]{}, A. V., & [Sosnovskij]{}, A. A. 2012, Astrophysics, 55, 494
, G. 2002, Acta Astron., 52, 397
, J. L., [et al.]{} 2014, Astron. Telegram, 6688
, G. A. 1969, Mitteil. Ver[ä]{}nderl. Sterne, 5, 88
, F. A. 1993, PhD thesis, Dartmouth Coll., Hanover, NH.
, P., [G[ä]{}nsicke]{}, B. T., [Hagen]{}, H.-J., [Marsh]{}, T. R., [Harlaftis]{}, E. T., [Kitsionas]{}, S., & [Engels]{}, D. 2005, A&A, 431, 269
, G. . 1978, IBVS, 1421
Rosino, L., & Pigatto, L. 1972a, IAU Circ., 2453
Rosino, L., & Pigatto, L. 1972b, IAU Circ., 2464
, A. S. 1991, Astron. Tsirk., 1550, 16
, A. S., & [Alksnis]{}, A. K. 1989, Soviet Astronomy Letters, 15, 382
, A. S., [Goranskij]{}, V. P., & [Samus]{}, N. N. 1992, IBVS, 3756
, J., & [Boyd]{}, D. 2007, J. Br. Astron. Assoc., 117, 25
, J., [Brady]{}, S., [Foote]{}, J., [Starkey]{}, D., & [Vanmunster]{}, T. 2008, J. Br. Astron. Assoc., 118, 288
, H. A., [Thorstensen]{}, J. R., [Peters]{}, C. J., [Kapusta]{}, A. B., & [Taylor]{}, C. J. 2007, PASP, 119, 494
, V., [et al.]{} 2017, Astron. Telegram, 10790
, A. S., [Pavlenko]{}, E. P., [Antonyuk]{}, O. I., [Antonyuk]{}, K. A., [Sosnovsky]{}, A. A., [Galeev]{}, A. I., [Pit’]{}, N. V., & [Babina]{}, Yu. V. 2016, Astrophys. Bull., 71, 293
, A. S., [et al.]{} 2018, Astrophysics, 61, 64
, K. Z., [et al.]{} 2013, Astron. Telegram, 5118
Stellingwerf, R. F. 1978, ApJ, 224, 953
, P., [et al.]{} 2009, AJ, 137, 4011
, P., [et al.]{} 2003, AJ, 126, 1499
, P., [et al.]{} 2006, AJ, 131, 973
Thorstensen, J. R., Fenton, W. H., Patterson, J. O., Kemp, J., Krajci, T., & Baraffe, I. 2002, ApJ, 567, L49
Thorstensen, J. R., Wade, R. A., & Oke, J. B. 1986, ApJ, 309, 721
Uemura, M., [et al.]{} 2002, PASJ, 54, L15
, E. O. 2017, AAVSO Alert Notice, 580
, Y., [et al.]{} 2017, PASJ, 69, 89
Warner, B. 1995, Cataclysmic Variable Stars (Cambridge: Cambridge University Press)
Wenzel, W. 1989, IBVS, 3405
, K. A., [et al.]{} 2010, AJ, 139, 2587
, M. A., [Still]{}, M. D., [Howell]{}, S. B., [Cannizzo]{}, J. K., & [Smale]{}, A. P. 2011, ApJ, 741, 105
, P. A., & [Warner]{}, B. 2010, MNRAS, 403, 398
, P. A., [Warner]{}, B., [de Bud[é]{}]{}, D., [Macfarlane]{}, S., [Schurch]{}, M. P. E., & [Zietsman]{}, E. 2012, MNRAS, 421, 2414
, L., [Mroz]{}, P., [Udalski]{}, A., [Poleski]{}, R., [Kostrzewa-Rutkowska]{}, Z., & [OGLE-IV Team]{} 2014, Astron. Telegram, 6690
, P., [Kato]{}, T., & [Shugarov]{}, S. 2013, PASJ, 65, 54
, S. V., [Tovmassian]{}, G. H., [Napiwotzki]{}, R., [Michel]{}, R., & [Neustroev]{}, V. 2006, A&A, 449, 645
, W., [et al.]{} 2010, Cent. Bur. Electron. Telegrams, 2574
, K. 2004, IBVS, 5599
[^1]: $<$http://www.aavso.org/data-download$>$.
[^2]: $<$http://cv.asassn.astronomy.ohio-state.edu/$>$.
[^3]: $<$http://nesssi.cacr.caltech.edu/catalina/$>$. For the information of the individual Catalina CVs, see $<$http://nesssi.cacr.caltech.edu/catalina/AllCV.html$>$.
[^4]: $<$https://www.aavso.org/$>$.
[^5]: $<$https://groups.yahoo.com/neo/groups/baavss-alert/$>$.
[^6]: $<$https://groups.yahoo.com/neo/groups/cvnet-outburst/$>$.
[^7]: $<$https://www.aavso.org/vsx/$>$.
[^8]: The R Foundation for Statistical Computing:\
$<$http://cran.r-project.org/$>$.
[^9]: These vsnet-alert messages can be seen at $<$http://ooruri.kusastro.kyoto-u.ac.jp/pipermail/vsnet-alert/$>$.
[^10]: CVOM: OGLE Monitoring system of cataclysmic variable stars: $<$http://ogle.astrouw.edu.pl/ogle4/cvom/cvom.html$>$, under OGLE-MC-DN-0016.
[^11]: $<$http://scan.sai.msu.ru/$^\sim$denis/VarDDE.html$>$.
[^12]: $<$http://www.cbat.eps.harvard.edu/unconf/followups/J20422233+2712111.html$>$
[^13]: $<$http://www.cbat.eps.harvard.edu/unconf/followups/J20205397$+$2508145.html$>$.
[^14]: $<$http://www.cbat.eps.harvard.edu/unconf/followups/J00332502-3518565.html$>$.
[^15]: $<$http://www.cbat.eps.harvard.edu/unconf/followups/J20100517+1303006.html$>$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Cosmic rays that escape their acceleration site interact with the ambient medium and produce gamma rays as the result of inelastic proton–proton collisions. The detection of such diffuse emission may reveal the presence of an accelerator of cosmic rays, and also constrain the cosmic ray diffusion coefficient in its vicinity. Preliminary results in this direction have been obtained in the last years from studies of the gamma–ray emission from molecular clouds located in the vicinity of supernova remnants, which are the prime candidate for cosmic ray production. Hints have been found for a significant suppression of the diffusion coefficient with respect to the average one in the Galaxy. However, most of these studies rely on the assumption of isotropic diffusion, which may not be very well justified. Here, we extend this study to the case in which cosmic rays that escape an accelerator diffuse preferentially along the magnetic field lines. As a first approximation, we further assume that particles are strongly magnetized and that their transport perpendicular to the magnetic field is mainly due to the wandering of the field lines. The resulting spatial distribution of runaway cosmic rays around the accelerator is, in this case, strongly anisotropic. An application of the model to the case of the supernova remnant W28 demonstrates how the estimates of the diffusion coefficient from gamma–ray observations strongly depend on the assumptions made on the isotropy (or anisotropy) of diffusion. For higher levels of anisotropy of the diffusion, larger values of the diffusion coefficient are found to provide a good fit to data. Thus, detailed models for the propagation of cosmic rays are needed in order to interpret in a correct way the gamma–ray observations.'
author:
- |
L. Nava$^{1}$[^1] and S. Gabici$^{1}$\
$^{1}$APC, AstroParticule et Cosmologie, Université Paris Diderot, CNRS/IN2P3, CEA/Irfu, Observatoire de Paris, Sorbonne Paris Cité,\
10, rue Alice Domon et Léonie Duquet, 75205 Paris Cedex 13, France
title: 'Anisotropic CR diffusion and $\gamma$-ray production close to supernova remnants, with an application to W28'
---
\[firstpage\]
cosmic rays – gamma rays – ISM: supernova remnants.
Introduction
============
Galactic Cosmic Rays (CRs) are mainly constituted by relativistic protons and are believed to be accelerated at SuperNova Remnant (SNR) shocks via first order Fermi mechanism [@hillas]. Though very popular, this scenario still needs to be conclusively proven by observations.
If CRs are indeed accelerated at SNRs, these objects must be gamma–ray sources. This is because the CRs accelerated at the shock undergo inelastic proton–proton interactions with the ambient medium and produce neutral pions which in turn decay into gamma rays [@dav; @nt]. Several SNRs have been detected in gamma rays at both TeV [e.g. @jim] and GeV [e.g. @giordano] energies, in agreement with such expectations. However, it is often difficult to determine whether the origin of the gamma–ray emission is hadronic, and thus related to the acceleration of CRs, or due to leptonic mechanisms such as inverse Compton scattering. For this reason, multi-wavelength studies of SNRs have been extensively carried out in an attempt to solve this degeneracy. Though for some individual SNRs it has been possible to ascribe the gamma–ray emission exclusively and quite confidently to hadronic [e.g. @fermitycho; @giovanni] or leptonic [e.g. @fermiRXJ; @donRXJ] processes, in other cases this ambiguity remains a problem.
An alternative way to reveal the presence of a CR source is by searching for the radiation produced by CRs that escape the acceleration site [@atoyan; @gabici07; @rodriguez; @gabici09]. At some stage of the dynamical evolution of the SNR, CRs are expected to leave the shock region and escape into the interstellar medium. The details of the escape mechanism are still not very well understood [see e.g. @gabiciescape and references therein], but it is generally believed that the ability of a SNR in confining particles decreases gradually with the shock speed, with higher energy particles leaving the shock earlier than low energy ones. Once escaped, CRs diffuse away from the SNR and produce gamma–rays in interactions with the ambient gas. To date, some possible evidence for particle escape from SNRs has been pushed forward by the observation of diffuse gamma–ray emission from the vicinity of the shell of the SNRs W28 [@W28hess; @W28agile; @W28fermi] and W44 [@W44fermi]. In both cases, the emission is clearly located outside of the shell and it is spatially coincident with the location of massive Molecular Clouds (MCs). This would favor a scenario in which the MCs are illuminated by the runaway CRs and, being very massive, become prominent gamma–ray sources [for the case of W28 see e.g. @gabiciW28 and discussion in Section \[sect:W28\]].
Besides revealing the presence of a CR source, the gamma–ray emission from runaway CRs can also be used to constrain the particles’ diffusion coefficient in the region surrounding the accelerator. This is very important for several reasons: first of all, a theoretical determination of the diffusion coefficient is a very complex task [see e.g. @lazarian; @lazarian08] and observational constraints are needed in order to guide and constrain models. In addition to that, the diffusion of CRs is believed to be a non–linear process in which CRs themselves generate via streaming instability the turbulence they scatter off [e.g. @kulsrud]. This is particularly relevant close to CR sources, where the CR density is expected to be very high, and possibly sufficient to suppress significantly the diffusion coefficient through streaming instability [@plesser; @malkov]. Thus, an empirical determination of the diffusion coefficient can reveal precious information on the ways in which particles and waves interact in astrophysical plasmas.
Most of the studies aimed at the determination of the CR diffusion coefficient from gamma–ray observations rely on the assumption of [*isotropic diffusion*]{} [@diego; @fujita; @li; @gabiciW28; @ohira; @lazarianW28]. The common rationale of these approaches can be summarized as follows: if SNRs are the sources of CRs, they have to convert a fraction $\eta \approx 10\%$ of their explosion energy $E_{SN} = 10^{51} E_{51} {\rm erg}$ into accelerated particles. If the diffusion of CRs proceeds isotropically, after a time $t$ from escape CRs of a given energy $E$ are distributed roughly homogeneously within a distance $R_d(E) \approx \sqrt{6\,D(E) \times t}$ from the SNR. Here, $D(E)$ is the energy dependent diffusion coefficient of CRs. Gamma–ray observations of MCs located in the vicinity of W28 or W44 tells us which is the CR density $n_{CR}(E)$ needed to explain the observed emission. According to what said above, such density has to be of the order of $n_{CR}(E) \approx f_{sp}(E) ~ \eta E_{SN}/R_d^3$, where the factor $f_{sp}(E)$ contains the information on the shape of the spectral energy distribution of escaping CRs. For aged SNRs such as W28 and W44, the time $t$ after the escape can be identified with the SNR age $t_{age}$, and thus an expression for the diffusion coefficient can be obtained, and reads: $D \approx (f_{sp} \eta E_{SN}/n_{CR})^{2/3} t_{age}^{-1}$. Since the values of all the physical quantities present on the right side of the equation can be inferred from observations, the expression provides a direct estimate of the diffusion coefficient.
As an example, we summarize here the results obtained by @gabiciW28 in interpreting the gamma–ray emission observed from the MCs located close to the SNR W28. They obtained a good fit to the gamma–ray spetrum measured by H.E.S.S. at photon energies $\gtrsim 300$ GeV by assuming a diffusion coefficient for $\approx 3$ TeV CRs of the order of $D(3~{\rm TeV}) \approx 5 \times 10^{27} (\eta/0.1)^{2/3}$ cm$^2$/s. The corresponding diffusion length $R_d$ of these particles is of the order of 100 pc. Also the broad band gamma–ray spectrum from GeV to TeV energies can be fitted by adjusting the energy dependence of the diffusion coefficient and the distances between the SNR and the clouds. The important point here is the fact that the estimated diffusion coefficient is more than one order of magnitude smaller than the one normally adopted to describe the propagation of CRs of energy $\gtrsim$ TeV in the galactic disk, which is $\approx 10^{29}\,$cm$^2$/s [@andyreview; @fiorenzareview]. These results are very similar to the ones obtained by other authors by means of similar modeling [@fujita; @li; @ohira; @lazarianW28] and seem to point toward a drop of the diffusion close to the SNR W28.
However, the validity of the assumption of isotropic diffusion of CRs needs to be discussed. In fact, if the intensity of the turbulent field $\delta B$ on scales resonant with the Larmor radius of particles is significantly smaller than the mean large scale field $B_0$ (i.e. if $\delta B/B_0 \ll 1$), then [*CR diffusion becomes anisotropic*]{}, with particles diffusing preferentially along the magnetic field lines [e.g. @fabien]. In the limiting (but still reasonable) case in which the perpendicular diffusion coefficient can be set equal to zero, the transport of CRs across the mean field is mainly due to the wandering of magnetic field lines [@parker]. This is the situation that we investigate in this paper.
To give a qualitative idea of the role that anisotropic diffusion can play in these study, let us consider an idealized case in which particles that escape a SNR diffuse along a magnetic flux tube characterized by a very long coherence length (i.e. the magnetic flux tube is preserved for a long distance). In this case, after a time $t$ particle will diffuse up to a distance $R_d \approx \sqrt{2\,D_\parallel \times t}$ along the tube (here $D_\parallel$ is the [*parallel*]{} diffusion coefficient), while their transverse distribution will be equal to the radius of the SNR shock at the time of their escape, $R_{sh}$, which is of the order of $\approx$ 1–10 pc. Thus, the enhanced CR density in the flux tube will be proportional to $n_{CR} \propto (R_d R_{sh}^2)^{-1}$ instead of $\propto R_d^{-3}$ as in the isotropic case. It is easy to see that the estimates of the diffusion coefficient based on the two opposite assumptions of isotropic and one-dimensional diffusion will differ by a factor of $\approx (R_d/R_{sh})^{4/3}$, which can be much larger than an order of magnitude! Thus, it is of paramount importance to investigate how the interpretation of gamma–ray observations depends on the assumptions made concerning CR diffusion.
In Section \[sect:model\] we develop a model for CR propagation in which CRs are strongly magnetized and diffuse uniquely along the magnetic field lines. The wandering of the field lines is also taken into account, and a diffusion coefficient $D_m$ for the magnetic field lines that depends on the properties of the turbulent field is defined [see e.g. @duffy]. In Section \[sect:results\] the model is used to predict the spatial distribution of runaway CRs and their spectrum. Finally, we apply the model to the case of the SNR W28 in Section \[sect:W28\]. A good fit to data is obtained, and the estimate of the parallel diffusion coefficient is found to depend on the level of anisotropy of the diffusion. For higher level of anisotropy, i.e. smaller values of $D_m$, larger values of the particles’ diffusion coefficient are needed in order to fit data. A discussion of the results and of future perspectives in this line of research can be found in Section \[sect:conclusions\].
Cosmic–ray transport in the presence of magnetic field line wandering {#sect:model}
=====================================================================
Consider a magnetic flux tube whose mean magnetic field $B_0$ is assumed to lie along the $z$-axis, perpendicular to the $(x,y)$ plane. The wandering of magnetic field lines is due to long wavelength perturbations, i.e. perturbations on scales much larger than the particles’ gyroradii, with root mean square amplitude $\delta B$. The condition for the validity of quasi–linear theory is $\delta B/B_0 \ll L_{\perp}/L_{\parallel}$, where $L_{\perp}$ and $L_{\parallel}$ are the field coherence lengths perpendicular and parallel to $B_0$, respectively [@kadomtsev]. According to quasi–linear theory, the field lines passing in the vicinity of $(x_0,y_0)$ at $z = 0$ are spread over a larger region as they reach a given $z$. The probability distribution describing this spreading of field lines is gaussian and characterized by $\langle (x-x_0)^2 \rangle = \langle (y-y_0)^2 \rangle = 2 D_m z$, where brackets indicate an ensemble average and $D_m$ is a diffusion coefficient for field lines [@parker]. For a broad band Fourier spectrum of the perturbation, the coherence lengths can be expressed as $L_{\perp,\parallel} = 2 \pi/\Delta k_{\perp,\parallel} \approx 2 \pi / k_{\perp,\parallel}$, with $k_{\perp}$ and $k_{\parallel}$ the characteristic wave–vectors of the perturbation [@achterberg]. Under these circumstances, the diffusion coefficient is $D_m = (\delta B/B_0)^2 L_{\parallel}/4$ [@kadomtsev; @duffy]. It is also possible to define the Lyapunov length $\lambda_L = L_{\perp}^2 (\delta B/B_0)^{-2}/L_{\parallel}$ which describes the exponential separation of field lines whose initial separation is smaller than $L_{\perp}$ [@isichenko]. This can be interpreted as the length above which the flux tube is disrupted by field line divergence. For fiducial values of the parameters the length of the flux tube is of the order of a few hundred parsecs [see e.g. @plesser].
We are interested here in studying the propagation of CRs in the presence of magnetic field line wandering. Energetic particles diffuse along and across the magnetic field line as the result of resonant interactions with magnetic field perturbations. Such perturbations are characterized by length scales of the order of the particles’ Larmor radii. Such scales are much smaller than the ones responsible for field line wandering. According to quasi–linear theory, the ratio between the parallel to perpendicular diffusion coefficient is $D_{\parallel}/D_{\perp}= 1 + (\lambda_{\parallel}/r_g)^2$, where $\lambda_{\parallel}$ is the particle’s mean free path along the field line and $r_g$ is its gyroradius. In the interstellar medium it is believed that $\lambda_{\parallel} \gg r_g$ which implies $D_{\perp} \ll D_{\parallel}$ [e.g. @fabien]. Thus, in the following we will neglect the diffusion of particles perpendicular to the field lines. In other words, a given particle remains attached to the same field line. Under this conditions, in a time interval $\Delta t$ a particle diffuses along a given field line a distance $\langle (\Delta z)^2 \rangle = 2 D_{\parallel} \Delta t$, but over such a distance $\Delta z$ along the $z$–axis the field line is displaced by an amount $\langle (\Delta x)^2 \rangle = 2 D_m \Delta z$. This leads to [@getmantsev; @rosenbluth; @cptuskin]: $$\langle (\Delta x)^2 \rangle \propto D_m \sqrt{D_{\parallel} \Delta t}$$ which describes a sub–diffusive transport of particles perpendicular to the mean magnetic field $B_0$.
This behavior of energetic particles due to the combination of the particle diffusion along the field lines and the random walk of the field lines themselves has been often referred to as [*compound diffusion*]{}, or [*anomalous diffusion*]{}. Models of compound diffusion have been developed and used in a great variety of contexts, to study phenomena like the heat transport in Tokamak [@rosenbluth; @isichenko], the propagation of energetic particles in the solar wind [@parker; @zimbardo], the confinement of CRs in the Galaxy [@getmantsev; @lingenfelter; @cptuskin] and their acceleration at astrophysical shocks [@achterberg; @duffy; @john]. In this paper, we apply the formalism of compound diffusion to another context, which is the propagation of CRs in the vicinity of their sources, i.e. SNRs, after they escape the acceleration region. We will show that in this situation an accurate modeling is needed in order to interpret in a correct way the gamma–ray observations of MCs located in the vicinity of SNRs.
In order to describe the compound diffusion of CRs we adopt the mathematical formalism developed by @webb and we define $P_{\rm FRW} (x|z)$ as the probability to find a field line displaced by an amount $\Delta x$ after a step of length $z$ along the direction of the umperturbed field $B_0$. From what said above, it follows that: $$\label{eq:frw}
P_{\rm FRW} (\Delta x|z)=\frac{1}{\sqrt{4\pi D_m z}}\exp\left[-\frac{(\Delta x)^2}{4D_m z}\right]$$ which corresponds to a diffusive behavior of field lines (FLW stands for Field line Random Walk). A similar equation holds for the displacement $\Delta y$. This has to be combined with the probability $P_{\rm \parallel} (z|\Delta t)$ that a particle moves a distance $z$ along the field line in a time $\Delta t=t-t_0$. For diffusive transport of particles along the field we have: $$\label{eq:parallel}
P_{\rm \parallel} (z|\Delta t)=\frac{1}{\sqrt{4\pi D_\parallel \Delta t}}\exp\left[-\frac{z^2}{4 D_\parallel\Delta t}\right]$$ The probability for a particle to reach the position $(x,y,z)$ at the time $t$, when its position at the time $t_0$ was ($x_0,y_0,z_0=0$), is then the product of $P_{\rm FRW}$ with $P_{\rm \parallel}$: $$\label{eq:prob}
P(\Delta x, \Delta y, z; \Delta t)=P_{\rm \parallel} (z|\Delta t)\, P_{\rm FRW} (\Delta x|z)\,P_{\rm FRW} (\Delta y|z)$$
In order to model the escape of CRs from a SNRs we assume, following @plesser, that particles are injected in the flux tube in the $xy$-plane at $z = 0$, within a circular region whose radius is equal to the SNR shock radius $R_{sh}(E)$. Since CRs of different energy are expected to escape the remnant at different times, the radius of the injection region is an energy dependent quantity. Following @gabici09 we assume a power law scaling to connect the particle energy of the runaway CRs with the time after the supernova explosion: $$\label{eq:escapetime}
E_{esc} = E_{MAX} \left( \frac{t_{esc}}{t_{Sed}} \right)^{-\delta}$$ where the implicit assumption has been made that the maximum energy of CRs accelerated in a SNR $E_{MAX}$ is reached at the time $t_{Sed}$ which marks the transition between the free expansion and the Sedov phases of the SNR evolution, and that CRs are gradually released in the interstellar medium from that time on. The Sedov phase is characterized by a scaling $R_{sh} \propto t^{2/5}$ which gives: $$R_{sh}(E_{esc}) \propto \left( \frac{E_{esc}}{E_{MAX}} \right)^{-\frac{2}{5 \delta}}$$ which is what is assumed in the following. Other parameterizations of the escape time of CRs can be easily implemented.
The spatial distribution of CRs can now be obtained by integrating the probability function given by Eq. \[eq:prob\] within the range $R_0 = \sqrt{(x_0^2+y_0^2)} \le R_{sh}(t_{esc}(E))$. To do so, it is convenient to adopt a cylindrical coordinate system and express the Field line Random Walk part of Eq. \[eq:prob\] as a function of the quantities $R=\sqrt{x^2+y^2}$, $R_0 = \sqrt{(x_0^2+y_0^2)}$ and $cos(\Delta\varphi)=(xx_0+yy_0)/(RR_0)$ which leads to: $$\begin{aligned}
f_{CR}(R,z,t,E) = A ~ \frac{E^{-\Gamma}}{\pi R_{sh}^2} ~ P_{\parallel}(z|t-t_{esc}) \times \nonumber
\\
\times \int_0^{2 \pi} d\varphi \int_0^{R_{sh}(t_{esc})} {\rm d}R_0\,R_0\,P_{FRW}(R,R_0,\Delta\varphi|z) \end{aligned}$$ where it has been assumed that the total spectrum of CRs released in the interstellar medium during the whole life of the SNR is a power law $A E^{-\Gamma}$ with normalization: $$\begin{cases}
A=\frac{\eta E_{\rm SN} (\Gamma-2)}{E_{\rm MAX}^{2-\Gamma}}\left[ \left( \frac{E_{\rm MIN}}{E_{\rm MAX}}\right)^{2-\Gamma}-1\right]^{-1} ~~~~~~~\rm for~~\Gamma\neq2 \\
A=\frac{\eta E_{\rm SN}}{\ln(E_{\rm MAX}/E_{\rm MIN})} ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\,\rm for~~\Gamma=2
\end{cases}$$ Here, $E_{SN}$ it the supernova explosion energy, $\eta$ is the fraction of this energy converted into CRs which are released in the interstellar medium, and $E_{MIN}$ and $E_{MAX}$ represent the extension in energy of the CR spectrum.
To describe the diffusion of CRs along field lines we adopt a diffusion coefficient which is a power-law in energy: $$\label{eq:diffusion}
D_\parallel(E)=\widetilde{D}_{\parallel}~\left(\frac{E}{10\rm\,GeV}\right)^{s}$$ where $\widetilde{D}_\parallel$ and $s$ are considered here as free parameters.
Before proceeding in computing the spatial distribution of CRs around a SNR we notice that, for $z \ll L_{\parallel}$, Eq. \[eq:frw\] does not provide a good description of the field line wandering. The reason is that in this regime the lateral displacement of a field line after a step $z$ along $B_0$ is of the order of $\approx (\delta B/B_0) z = bz$, since $b = \delta B/B_0$ represents the angle between the unperturbed ($B_0$) and total ($B_0+\delta B$) magnetic field [@isichenko]. An accurate and quantitative analysis of the behavior of a magnetic flux tube in this regime goes beyond the scope of this paper [see @isichenko for a more detailed discussion]. However, in order to describe this regime in a qualitative way, for $z \ll L_{\parallel}$ we substitute Eq. \[eq:frw\] with: $$\label{eq:conical}
P(R | z) = \frac{\vartheta\left[(b z)^2 - (R-R_0)^2\right]}{\pi (b z)^2}$$ where $\vartheta[s]$ is the Heaviside function, equal to 1 for $s>0$ and 0 for $s<0$. Eq. \[eq:conical\] roughly mimics the behavior of a flux tube characterized by a opening angle $b$. In the intermediate region $z \approx L_{\parallel}$ we use an interpolating function to bridge the behaviors described by Eqns. \[eq:frw\] and \[eq:conical\].
Results {#sect:results}
=======
In this Section we compute the spatial distribution of CRs expected in the vicinity of a SNR at a given time after the explosion. We consider a typical supernova, characterized by the following fiducial values of parameters: an explosion energy of $E_{SN} = 10^{51}\,$erg, a mass of the ejecta equal to $M_{ej} = 1.4 \, M_{\odot}$, and a density of the circumstellar medium $n_0 = 1\,$cm$^{-3}$. We further assume that a fraction $\eta = 0.1$ of the supernova explosion energy is converted into CRs, which are injected in the interstellar medium with a power law differential energy spectrum $dN/dE \propto E^{-\alpha}$ which extends from $E_{\rm MIN}=1\,$GeV to $E_{\rm MAX}=5\,$PeV (approximately the position of the knee in the CR spectrum). It is known from CR data that $\alpha$ should be in the range $\approx2.1-2.4$ [@fiorenzareview; @andyreview]. We adopt here $\alpha=2.2$ as a representative value. As described in Sec. \[sect:model\], CRs are gradually released from the SNR during the Sedov phase that goes from $t \approx 280$ yr to $t \approx 3.6 \times 10^4\,$yr [@cioffi]. For the parallel diffusion coefficient of CRs (Eq. \[eq:diffusion\]) we assume $\widetilde{D}_\parallel = 10^{28}\,$cm$^2$/s and $s = 0.5$.
-0.38cm ![Cosmic ray over-density around a typical supernova remnant (see text for details) for a particle energy of $E=1\,$TeV at a time $t=10\,$kyr after the explosion. The left panel refers to an isotropic diffusion coefficient of cosmic rays equal to $D = 5 \times 10^{26} (E/10~{\rm GeV})^{0.5}\,$cm$^2$/s, while the right panel refers to an anisotropic diffusion scenario with $D_\parallel = 10^{28}(E/10~{\rm GeV})^{0.5}\,$cm$^2$/s, $D_m = 1\,$pc, and $b^2 = (\delta B/B_0)^2 = 0.2$. The black cross marks the a position at which the CR over-density is equal in the two panels.[]{data-label="fig:iso"}](cr_map_ani_iso.eps "fig:")
As a first step, we compare in Fig. \[fig:iso\] the results that are obtained if an isotropic diffusion coefficient is assumed [as, e.g., in @atoyan; @gabici09], with the ones obtained for the anisotropic diffusion model that we consider in this paper. In both panels of Fig. \[fig:iso\], the SNR is located at the centre of the field and the color code refers to the excess of CRs with respect to the average density of CRs in the Galaxy, which is [e.g. @pdg]: $$N_{CR}^{gal}(E) \approx 1.8 ~ \left( \frac{E}{\rm GeV} \right)^{-2.7} {\rm GeV^{-1}cm^{-2}s^{-1}sr^{-1}}$$ Over-densities are plotted for a particle energy of $1\,$TeV and for a time $t = 10\,$kyr after the supernova explosion. Here the diffusion coefficient of the magnetic field lines is set equal to $D_m = 1\,$pc, with $b^2 = (\delta B/B_0)^2 = 0.2$ (different values of $D_m$ will be explored in the following). This corresponds to a parallel coherence length of the perturbation of $L_{\parallel} = 20\,$pc.
{width="1.\textwidth"}
{width="1.\textwidth"}
The spatial distribution of CRs is strikingly different in the two scenarios: spherically symmetric in the left panel, and strongly elongated in the direction of the magnetic field flux tube in the right panel. A filamentary diffusion of CRs was also found in the numerical simulations by @giacinti. The same parameters have been used to compute the over–densities in the two scenarios, with the exception of the CR diffusion coefficient, which in the left panel has been assumed to be isotropic and equal to $D = \widetilde{D}\,(E/10\,{\rm GeV})^{0.5}\,$cm$^2$/s with $\widetilde{D} = 5 \times 10^{26}\,$cm$^2$/s. The choice of two significantly different values for $\widetilde{D}$ and $\widetilde{D}_\parallel$, with $\widetilde{D} \ll \widetilde{D}_\parallel$ has been made in order to obtain the same level of CR over–density in the vicinity of the SNR. As an example, the black cross in Fig. \[fig:iso\] shows a position, located $30\,$pc away from the centre of the explosion, where the CR over-density is identical in the two panels. To get comparable values for the CR over–density, a much smaller (isotropic) diffusion coefficient $\widetilde{D}$ is needed in order to compensate for the larger solid angle over which CRs can propagate. As already stressed in the introduction, this fact must be taken into account when interpreting the gamma–ray observations of molecular clouds illuminated by CRs escaping from SNRs. This will be discussed in Sec. \[sect:W28\], when the model developed here will be applied to fit the gamma–ray observations of the SNR W28.
In Fig. \[fig:cr\_map\] we show the CR over–density around the SNR for different values of the particle energy and of the time after the supernova explosion. The upper, middle, and lower panels refer to a time of 6, 19, and 60 kyr after the explosion, respectively. Plots on the first, second, and third column refer to particle energies of $30\,$GeV, $3\,$TeV, and $300\,$TeV, respectively. The escape of CRs is described by Eq. \[eq:escapetime\], which states that higher energy CRs are released first, and lower energy CRs escape at later times. This is the reason why there is no CR excess in the top–left panel of Fig. \[fig:cr\_map\]: for the choice of parameters made here, for a SNR age of $6\,$kyr particles with an energy of $30\,$GeV are still confined within the SNR shock. As the age of the SNR increases, CRs diffuse further away along the flux tube and fill a broader and broader region. As a consequence of that, the CR over–density decreases accordingly. It is evident from these maps that a molecular cloud located in the vicinity of the SNR can be illuminated by the escaping CRs and become a bright gamma–ray source only if it is located within the flux tube. A nearby cloud which is not magnetically connected with the SNR will not be illuminated by CRs, despite its proximity to the SNR.
All the plots in Fig. \[fig:cr\_map\] refer to a region of size $\approx 200\,$pc around the SNR. As said in Sec. \[sect:model\], this roughly represents the expected length of a magnetic flux tube in the Galaxy [e.g. @plesser]. For distances larger than a few hundred parsecs from the SNR, the flux tube loses its identity and it is disrupted due to the exponential divergence of field lines. Thus, the results presented in this paper are accurate and reliable for distances up to a few hundred parsecs and less.
Since we assumed here that CRs are strongly magnetized, i.e. their remain attached to field lines, their transport across the mean magnetic field is solely governed by the diffusion of field lines, which is described by the diffusion coefficient $D_m$. This quantity determines how quickly the field lines diverge as a function of the displacement $z$ along the mean field. This is demonstrated in Fig. \[fig:Dm\], where the CR over–density for particles of energy $3\,$TeV is shown. The age of the SNR is $19\,$kyr. Three different values for $D_m$ are considered: 0.5, 1, and $2\,$pc for the left, middle, and right panel, respectively. Unfortunately, our knowledge of the properties of the interstellar magnetic field is not good enough to allow a determination or an estimate of this parameter. For a SNR located in a diffuse interstellar medium (i.e. with no massive molecular clouds) the morphology of the resulting gamma–ray emission due to CR proton–proton interaction with the ambient gas would closely follow the spatial distribution of CRs. Thus, observing the diffuse gamma–ray emission generated by runaway CRs around SNRs might serve as a tool to explore the structure of the interstellar magnetic field. The detection of such diffuse emission is within the capabilities of future gamma–ray instruments such as the Cherenkov Telescope Array [@acero; @sabrinona].
Finally, we show in Fig. \[fig:spectra\] the spectra of escaping CRs at different distances from the SNR and at different times after the supernova explosion. Each panel refers to a different epoch: 6, 19, and $60\,$kyr for the top, middle, and bottom panel, respectively. Solid curves show the spectra at three different positions on the $z$-axis: $40\,$pc, $100\,$pc, and $200\,$pc. For each of these positions we also show the spectra at different distances from the $z$-axis: $25\,$pc (dotted lines) and $50\,$pc (dashed lines).
At high energies, in almost all cases the energy spectra are power laws with slope $\approx \alpha+s/2$, where $\alpha$ is the slope of the injection spectrum of runaway CRs and $s$ is the slope of the energy dependent diffusion coefficient (Eq. \[eq:diffusion\]). Such a behavior can be inferred from Eq. \[eq:parallel\]. If one moves to larger values of $z$, the on–axis (i.e. R = 0) high energy spectrum preserves the same slope, but its normalization decreases as $\approx 1/z$. This is due to the fact that the transverse section of the magnetic flux tube is increasing proportionally to $z$, while the CR intensity along a field line is independent of $z$ for $z^2 \ll 4 D_{\parallel} t$ (see Eqns. \[eq:frw\] and \[eq:parallel\]). A feature common to all the spectra plotted in Fig. \[fig:spectra\] is the presence of a low energy cutoff. The cutoff is due to the fact that at a given time, only particles of sufficiently large energy had enough time to propagate over the distance $z$. The cutoff is moving towards lower energies if a longer time is considered, because particles with lower energies have then the time to reach a given position $z$ along the axis. Finally, some curves (for example the ones with $z = 40\,$pc and $R = 25\,$pc) exhibit a quite sharp high energy cutoff. This cutoff is due to the fact that particles of different energy are injected within different transverse sections of the flux tube. Higher energy particles are released earlier from the SNR, when the shock radius is smaller, lower energy ones are injected later, when the shock radius is larger. While diffusing along the field lines, CRs are displaced in the transverse direction due to field line wandering. Higher energy particles, which have been injected in a smaller region around $z = 0$, need on average a larger transverse displacement in order to reach a given distance $\hat{R}$ from the $z$–axis. Thus, for small enough $z$, the opening of the magnetic flux tube might not be enough to allow high energy particles to reach $\hat{R}$, and this explains the presence of the cutoff.
![Spectra of runaway cosmic rays at different positions and times after the explosion. The age of the supernova remnant is $t=6\,$kyr (upper panel), $t=19\,$kyr (middle panel) and $t=60\,$kyr (lower panel). Solid lines refer to spectra along the $z$-axis, oriented as the mean magnetic field, at three different positions ($z=40\,$pc, $z=100\,$pc and $z=200\,$pc). For each distance $z$, three different values of $R$ (the perpendicular distance from the $z$-axis) are also considered: $R=0\,$pc, $R=25\,$pc, and $R=50\,$pc. The black lines show the CR background.[]{data-label="fig:spectra"}](CR_spectrum_3times_idl.eps){width="47.00000%"}
[![Gamma–ray emission from the three molecular clouds surrounding the supernova remnant W28. Fermi data are shown as open (blue) circles. Filled (black) circles refer to HESS data. The dashed lines show the contribution to the gamma–ray emission from the cosmic ray galactic background, the long-dashed lines show the contribution from cosmic rays that escaped from W28 and the solid (red) line is the total.[]{data-label="fig:w28"}](w28_mc1.ps "fig:"){width=".46\textwidth"} ![Gamma–ray emission from the three molecular clouds surrounding the supernova remnant W28. Fermi data are shown as open (blue) circles. Filled (black) circles refer to HESS data. The dashed lines show the contribution to the gamma–ray emission from the cosmic ray galactic background, the long-dashed lines show the contribution from cosmic rays that escaped from W28 and the solid (red) line is the total.[]{data-label="fig:w28"}](w28_mc2.ps "fig:"){width=".46\textwidth"} ![Gamma–ray emission from the three molecular clouds surrounding the supernova remnant W28. Fermi data are shown as open (blue) circles. Filled (black) circles refer to HESS data. The dashed lines show the contribution to the gamma–ray emission from the cosmic ray galactic background, the long-dashed lines show the contribution from cosmic rays that escaped from W28 and the solid (red) line is the total.[]{data-label="fig:w28"}](w28_mc3.ps "fig:"){width=".46\textwidth"}]{}
In the next Section we apply the model developed above to a specific object, i.e. the SNR W28 and the molecular clouds located in its proximity. Such clouds have been detected in gamma rays and this has been interpreted by many authors as the result of their being illuminated by CRs that escaped the SNR. We will demonstrate that a good agreement can be reached between the predictions of the model and observations and we will discuss the impact of this on the attempts to derive the particle diffusion coefficient close to SNRs by means of gamma–ray observations.
Application to the supernova remnant W28 {#sect:W28}
========================================
W28 is an old SNR in its radiative phase of evolution, located in a region rich of dense molecular gas with average density $\gtrsim 5\,{\rm cm^{-3}}$. At an estimated distance of $\sim 2\,{\rm kpc}$ the SNR shock radius is $\sim 12\,{\rm pc}$ and its velocity $\sim 80\,{\rm km/s}$ [e.g. @rho]. By using the dynamical model by @cioffi and assuming that the mass of the supernova ejecta is $\sim 1.4\,M_{\odot}$, it is possible to infer the supernova explosion energy ($E_{SN} \sim 0.4 \times 10^{51}{\rm erg}$), initial velocity ($\sim 5500\,{\rm km/s}$), and age ($t_{age} \sim 4.4 \times 10^4 {\rm yr}$).
Gamma ray emission has been detected from the region surrounding W28 both at TeV [@W28hess] and GeV energies [@W28fermi; @W28agile], by HESS, FERMI, and AGILE, respectively. The TeV emission correlates quite well with the position of three massive molecular clouds, one of which is interacting with the north-eastern part of the shell (and corresponds to the TeV source HESS J1801-233), and the other two being located to the south of the SNR (TeV sources HESS J1800-240 A and B). The masses of these clouds can be estimated from CO measurements and result in $\approx 5$, $6$, and $4 \times 10^4 M_{\odot}$, respectively, and their projected distances from the centre of the SNR are $\approx$ 12, 20, and $20\,$pc, respectively [@W28hess]. The GeV emission roughly mimics the TeV one, except for the fact that no significant emission is detected at the position of HESS J1800-240 A.
The gamma–ray emission from the clouds in the W28 region has been interpreted by many authors as the result of the interaction of CRs that escaped W28 with the dense gas in the cloud [@fujita; @li; @gabiciW28; @ohira; @lazarianW28]. All these approaches started from the assumption of isotropic diffusion of CRs, and a general consensus was found on the fact that, in order to fit observations, the diffusion coefficient had to be suppressed by a factor of $\approx 10 ... 100$ with respect to the average value in the Galaxy, which is $D_{gal} \approx D_0\,(E/10\,{\rm GeV})^{\delta}$ with $D_0 \approx 10^{28}...10^{29}\,$cm$^2$/s and $s \approx 0.3 ... 0.7$ [@fiorenzareview].
In this section we take a different approach and we apply the model developed in Sec. \[sect:model\] to estimate the spectrum of CRs and derive the $\gamma$-ray emission expected from the clouds in the W28 region. This approach is radically different from the ones mentioned above because it relies on the more physical assumption that the diffusion of CRs is not isotropic, but proceeds mainly along the magnetic field lines.
The results of our modeling are shown in Fig. \[fig:w28\], where the gamma–ray data from the three molecular clouds are plotted as blue open symbols (data from FERMI) and black filled dots (data from HESS). The emission from the sources HESS J1801-233, and HESS J1800-240 A and B is plotted in the top, middle, and bottom panel, respectively. The black dashed lines represent the contribution to the gamma–ray emission from the proton–proton interactions of the CRs in the galactic background with the inter–cloud gas. The blue long–dashed lines represents the contribution to the emission from the runaway CRs that escaped from W28. The solid red line is the total emission. The gamma–ray fluxes have been computed following [@kamae] with an additional multiplicative factor 1.5 to take into account elements heavier than hydrogen in both cosmic rays and ambient gas [@mori].
A good agreement with observations is obtained is a parallel diffusion coefficient $\widetilde{D}_\parallel = 10^{28}\,$cm$^2$/s with $s = 0.5$ is adopted, together with a diffusion coefficient for field lines $D_m = 1\,$pc with $b^2 = (\delta B/B_0)^2 = 0.2$. Moreover, we assumed that $\approx 20\%$ of the total explosion energy has been converted into CRs with a spectrum proportional to $E^{-2.2}$ and extending from $1\,$GeV to $5\,$PeV. In order to be illuminated by the escaping CRs, the three molecular clouds have to be located in the proximity of the axis of the magnetic flux tube (i.e. the direction of the local mean field). The spectra reported in the figure refers to the positions $z = 10$, $165$, and $35\,$pc and $R = 6.5$, $0$, and $14\,$pc (top to bottom panel, respectively).
It has to be noted that, due to the number of parameters involved in the model, other sets of parameter values might be found that provide an equally satisfactory fit to data. This is not surprising, given that several previous modelings of this source provided an equally good fit to data by using a radically different picture (i.e. isotropic diffusion of CRs) for the transport of particles. Moreover, while a quite small normalization of the (isotropic) diffusion coefficient, roughly of the order of $\widetilde{D} \approx 5 \times 10^{26}$ cm$^2$/s had to be adopted in order to fit data satisfactorily, in the anisotropic case we obtain a good agreement with data for a significantly larger value of the (parallel) diffusion coefficient of $\widetilde{D}_\parallel\approx 10^{28}$ cm$^2$/s. It might be noticed that this number is close to the standard values inferred for the diffusion of CRs in the Galaxy. Thus, any attempt to constrain the CR diffusion coefficient from the observations of gamma–rays from the vicinity of SNRs needs to take into account that an intrinsic uncertainty exists, and it is related to the unknown nature of the CR transport in the interstellar medium, and in particular to the unknown relative relevance of the transport parallel and perpendicular to the magnetic field lines.
Conclusions {#sect:conclusions}
===========
The details of the transport of CRs in the Galaxy are still little understood. Studies of the composition of CRs provide us with an estimate of the average confinement time of CRs within the Galaxy, which can be translated into a spatially averaged diffusion coefficient for CRs [e.g. @andyreview; @fiorenzareview]. Whether the CR diffusion coefficient has large spatial variations or it is rather uniform throughout the Galaxy is not known, thought a suppression of diffusion close to CR sources might be expected due to CR streaming instability [@plesser; @malkov]. To this purpose, the detection of gamma–ray emission from the vicinity of CR accelerators might be used to constrain the CR diffusion coefficient, and thus assess the importance of such suppression [e.g. @atoyan; @gabici09]. This is because CRs escaping the accelerators would produce gamma rays via proton–proton interactions with the ambient medium. Both the morphology of the resulting emission and its spectrum would depend on the functional form (i.e. energy dependence, level of anisotropy) of the diffusion coefficient.
An object that has been extensively investigated in this context is the SNR W28. Three massive molecular clouds, with total mass in the $\approx 10^5 M_{\odot}$ range, are located in the vicinity of the SNR shell and emit gamma rays. This has been interpreted as the result of the illumination of the clouds by the CRs that escaped the SNR. Several models have been proposed to fit these observations, and all of them are based on the assumption that the diffusion of CRs proceeds isotropically [e.g. @W28agile; @gabiciW28 and see Sec. \[sect:W28\] for a complete list of references]. There is a general consensus on the fact that the (isotropic) diffusion coefficient has to be suppressed by a factor of $\approx 10...100$ with respect to the average Galactic one in order to explain the observations. This implies coefficients in the range $\widetilde{D} \approx 10^{26} ... 10^{27}\,$cm$^2$/s.
In this paper, the assumption of isotropy of diffusion has been relaxed, and a more physically motivated situation have been investigated, in which CRs propagate mainly along the magnetic field lines. We considered here the limiting scenario in which the diffusion of CRs across field lines is very small and thus can be neglected. In such a situation, the transverse displacement of CRs is uniquely due to the wandering of the field lines [@parker]. Spectra and morphology of the spatial distribution of CRs around SNRs have been computed and described. The main feature is the elongated, filamentary distribution of CRs, as opposed to the spherical distribution found in the case of isotropic diffusion.
In order to fit the gamma–ray data from the W28 region within this scenario, one has to assume that the molecular clouds in its vicinity are magnetically connected to the SNR through a magnetic field flux tube. If this is the case, an accurate fit to data can be obtained. Under this assumption, particles are bound to the flux tube and thus forced to propagate along a specific direction. For plausible values of the diffusion coefficient of magnetic field lines, in order to obtain the correct CR over–density at the location of the molecular clouds a large (parallel) diffusion coefficient of the order of $\widetilde{D}_\parallel \approx 10^{28}\,$cm$^2$/s has to be adopted.
The fact that a very good agreement has been found with data in the two radically different scenarios characterized by isotropic and anisotropic diffusion tells us that more data needs to be collected from more SNRs in order to infer with reasonable confidence the properties of the diffusion of particles escaping their accelerators. The diffuse emission that these runaway particles would produce in their interaction with the ambient gas is, even in the absence of very massive clouds, within the capabilities of the Cherenkov Telescope Array [@acero; @sabrinona]. These observations will provide us with precious informations about the properties of the transport of CRs in the Galaxy, but also with a direct evidence for the fact that SNRs are indeed the accelerators of galactic CRs.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank F. Casse, A. Marcowith, F. Piazza, V. Ptuskin, R. Schlickeiser, and L. Sironi for helpful discussions. The work of LN and SG has been supported by ANR under the JCJC Programme.
[99]{}
Abdo, A. A., et al., 2010, ApJ, 718, 348
Abdo, A. A., et al., 2011, ApJ, 734, 28
Acciari, V., A., et al., 2011, ApJ Lett., 730, L20
Acero, F., et al., 2012, accepted to Astropartcle Physics, arXiv:1209.0582
Achterberg, A., Ball, L., 1994, A&A, 284, 687
Aharonian, F. A., Atoyan, A. M., 1996, A&A, 309, 917
Aharonian, F. A., et al., 2008, A&A, 481, 401
Casanova, S., Jones, D. I., Aharonian, F. A., Fukui, Y., Gabici, S., et al., 2010, PASJ, 62, 1127
Casse, F., Lemoine, M., Pelletier, G., 2002, Phys. Rev. D, 65, 023002
Castellina, A., Donato, F., 2011, arXiv:1110.2981
Chuvilgin, L. G., Ptuskin, V. S., 1993, A&A, 279, 278
Cioffi, D. F., McKee, C. F., Bertschinger, E., 1988, ApJ, 334, 252
Drury, L. O’C., Aharonian, F. A., Völk, H. J., 1994, A&A, 287, 959
Duffy, P., Kirk, J. G., Gallant, Y. A., Dendy, R. O., 1995, A&A Lett., 302, L21
Ellison, D. C., Patnaude, D. J., Slane, P., Raymond, J., 2010, ApJ, 712, 287
Fujita, Y., Ohira, Y., Tanaka, S. J., Takahara, F., 2009, ApJ, 707, L179
Gabici, S., Aharonian, F. A., 2007, ApJ Lett., 665, L131
Gabici, S., Aharonian, F. A., Casanova, S., 2009, MNRAS, 396, 1629
Gabici, S., Casanova, S., Aharonan, F. A., Rowell, G., SF2A–2010: Proceedings of the Annual meeting of the French Society for Astronomy and Astrophysics. Eds.: S. Boissier, M. Heydari-Malayeri, R. Samadi and D. Valls-Gabaud, p.313 – arXiv:1009.5291
Gabici, S., 2011, MmSAI, 82, 760
Getmantsev, G. G., 1963, Soviet Astr., 6, 477
Giacinti, G., Kachelriess, M., Semikoz, D. V., 2012, PRL, 108, 261101
Giordano, F., 2011, MmSAI, 82, 743
Giuliani, A., et al., 2010, A&A, 516, L11
Hillas, A. M., 2005, J. Phys. G: Nucl. Part. Phys., 31, R95
Hinton, J. A., Hofmann, W., 2009, ARA&A, 47, 523
Isichenko, M. B., 1991, Plasma Physics and Controlled Fusion, 33, 795
Jokipii, J. R., Parker, E. N., 1969, ApJ, 155, 777
Kadomtsev, B. B., Pogutse, O. P., 1979, Nucl. Fusion Suppl., 1, 649
Kamae, T., Karlsson, N., Mizuno, T., Abe, T., Koi, T., 2006, ApJ, 647, 692
Kirk, J. G., Duffy, P., Gallant, Y. A., 1996, A&A, 314, 1010
Kulsrud, R., Pearce, W. P., 1969, ApJ, 156, 445
Li, H., Chen, Y., 2010, MNRAS, 409, L35
Lingenfelter, R. E., Ramaty, R., Fisk, L. A., 1971, Astrophys. Lett., 8, 93
Malkov, M. A., Diamond, P. H., Sagdeev, R. Z., Aharonian, F. A., Moskalenko, I. V., arXiv:1207.4728
Mori, M., 1997, ApJ, 478, 225
Morlino, G., Caprioli, D., 2012, A&A, 538, A81
Naito, T., Takahara, F., 1994, J. Phys. G: Nucl. Part. Phys., 20, 477
Ohira, Y., Murase, K., Yamazaki, R., 2011, MNRAS, 410, 1577
Particle Data Group 2008, Phys. Lett. B, 667, 212
Ptuskin, V. S., Zirakashvili, V. N., Plesser, A. A., 2008, Adv. Space Res., 42,486
Rechester, A. B., Rosenbluth, M. N., 1978, Phys. Rev. Lett., 40, 38
Rho, J., Borkowski, K. J., 2002, ApJ, 575, 201
Rodriguez Marrero, A. Y., Torres, D. F., de Cea del Pozo, E., Reimer, O., Cillis, A. N., 2008, ApJ, 689, 213
Strong, A. W., Moskalenko, I. V., Ptuskin, V. S., 2007, ARNPS, 57, 285
Torres, D. F., Rodriguez Marrero, A. Y., de Cea del Pozo, E., 2008, MNRAS, 387, L59
Uchiyama, Y., Funk, S., Katagiri, H., Katsuta, J., Lemoine-Goumard, M., Tajima, H., Tanaka, T., Torres, D. F., 2012, ApJ Lett., 749, L35
Yan, H., Lazarian, A., 2004, ApJ, 614, 757
Yan, H., Lazarian, A., 2008, ApJ, 673, 942
Yan, H., Lazarian, A., Schlickeiser, R., 2012, ApJ, 745, 140
Webb, G. M., Zank, G. P., Kaghashvili, E. Kh., Le Roux, J. A., 2006, ApJ, 651, 211
Zimbardo, G., Pommois, P., Veltri, P., 2006, ApJ, 639, L91
\[lastpage\]
[^1]: E-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider generalised Kreĭn-Feller operators $\Delta_{\nu, \mu} $ with respect to compactly supported Borel probability measures $\mu$ and $\nu$ under the natural restrictions ${\operatorname{supp}}(\nu)\subset{\operatorname{supp}}(\mu)$ and $\mu$ atomless. We show that the solutions of the eigenvalue problem for $\Delta_{\nu, \mu} $ can be transferred to the corresponding problem for the classical Kreĭn-Feller operator $\Delta_{\nu, \Lambda}=\partial_{\mu}\partial_{x}$ with respect to the Lebesgue measure $\Lambda$ via an isometric isomorphism of the underlying Banach spaces. In this way we reprove the spectral asymptotic on the eigenvalue counting function obtained by Freiberg. Additionally, we investigate infinitesimal generators of generalised Liouville Brownian motions associated to generalised Kreĭn-Feller operator $\Delta_{\nu, \mu}$ under Neumann boundary condition. Extending the measure $\mu$ and $\nu$ to the real line allows us to determine its walk dimension.'
address:
- 'FB 3 – Mathematik, Universität Bremen, Bibliothekstr. 1, 28359 Bremen, Germany'
- 'School of Mathematics, University of Birmingham, Edgbaston, Birmingham, B15 2TT, UK'
author:
- Marc Kesseböhmer
- Aljoscha Niemann
- Tony Samuel
- Hendrik Weyer
title: 'Generalised Kreĭn-Feller operators and Liouville Brownian motion via transformations of measure spaces'
---
[^1]
Introduction and motivation
===========================
The classical Kreĭn-Feller differential operator $\Delta_{\nu, \Lambda}$, introduced in [@Fe57; @KK68], where $\nu$ denotes a non-atomic compactly supported Borel probability measure on $\mathbb{R}$ and where $\Lambda$ denotes the one-dimensional Lebesgue measure, has been investigated with respect to its spectral properties first by Fujita [@Fu87], Küchler [@MR574035], Langer [@MR0314125] and Kotani and Watanabe [@MR661628] and more recently by Arzt [@A15b], Ehnes [@Ehnes2019] and Freiberg [@MR2017701; @MR2030736; @Fr05]. The case when $\nu$ is purely atomic has also been studied in [@MR2513598]; where it was shown that the eigenvalues of $\Delta_{\nu, \Lambda}$ have a dependence not only on the positions of the atoms of $\nu$ but also on the weights of the atoms. Returning to the case when $\nu$ is a non-atomic, it has been established that $\Delta_{\nu, \Lambda}$ is the infinitesimal generator of a Liouville Brownian motion (also known as gap diffusion, skip-free diffusion, quasi-diffusion or generalised diffusion), see [@Burkhardt1983; @MR2817342; @MR3034785; @X_Jin_2017; @MR574035; @MR0314125; @MR3005002; @MR3272822].
Here, we investigate generalised Kreĭn-Feller operators $\Delta_{\nu, \mu}$ for Borel measures $\nu$ and $\mu $ on the real line under the natural assumptions ${\operatorname{supp}}(\nu) \subseteq {\operatorname{supp}}( \mu)$ and $\mu$ atomless. In the case that $\nu=\mu =\Lambda$ the operator coincides with the classical second order weak derivative. For arbitrary $\mu=\nu$, atomless and compactly supported, a harmonic calculus for $\Delta_{\mu, \mu}$ was developed in [@FZ02] and, when $\mu$ is a self-similar measure supported on a Cantor set, it is now well established that the eigenvalue counting function of $\Delta_{\mu, \mu}$ is comparable to the square-root function. In [@KSW16] the exact eigenvalues and eigenfunctions of $\Delta_{\mu, \mu}$ were obtained and it was shown that the eigenvalues do not depend on the given measure. Moreover, the eigenfunctions are given by a composition of the appropriated classical trigonometric functions with a phase space transformation induced by the distribution function of $\mu$. The case when the measure $\mu$ has a continuous as well as an atomic part, was the subject of [@KSW17; @KSW2019b]. Here, it has been shown that, if $\mu$ has a continuous part, then the eigenvalues may depend on the position of the atoms, and otherwise not.
In the present article, we elaborate on the connections between the generalised and the classical Kreĭn-Feller operators by establishing a suitable phase space transformation determined by the distribution function of $\mu$. As a first application of this observation we show that the spectral properties of the generalised Kreĭn-Feller operators can be reduced to those of the to the classical ones; and as a second application, we connect properties of the associated Liouville Brownian motions for generalised Kreĭn-Feller operators to that of the classical Kreĭn-Feller operators with a special focus on the concept of walk dimension. This complements and partially resembles the general framework established by Dynkin [@DynkinI_II Vol.I,6].
Setup and statement of main results
===================================
Our setting
-----------
Let $\mu$ and $\nu$ denote two Borel probability measures on $[0,1]$ with ${\operatorname{supp}}(\nu) \subseteq {\operatorname{supp}}(\mu)$, $\nu(\{0,1\})=0$ and $\mu$ atomless. Denote the distribution function of $\mu$ and $\nu$ by $F_{\mu}$ and $F_{\nu}$, respectively. Let $(C_{\nu, \mu}, \lVert \cdot \rVert_{\infty} )$ denote the Banach space of continuous function on $[0,1]$ which are linear in scale $F_{\mu}$ on intervals where $F_{\nu}$ is constant. Namely, on each connected component $J$ of $[0,1] \setminus {\operatorname{supp}}(\nu)$ the function $f$ is linear in $F_{\mu}$, that is $f(x)=a_J F_{\mu}(x)+b_J$ for all $x \in J$ and some $a_J, b_J \in \mathbb{R}$. As indicated above, we let $\Lambda$ denote the one-dimensional Lebesgue measure restricted to $[0,1]$.
Set $\S^w \coloneqq L^2(\nu)$ and $\S^s \coloneqq C_{\nu,\mu}$, where $w$ stands for [*weak*]{} and $s$ stands for [*strong*]{}; we sometimes write $\S^*(\mu,\nu)$ instead of $\S^*$ to stress the dependence of the underlying measure spaces for $* \in \{s,w\}$. In what follows we will mainly be concerned with the Banach spaces $(\S^{w},\Vert\cdot\Vert_{2})$ and $(\S^{s},\Vert\cdot\Vert_{\infty})$. Now fix $* \in \{s,w\}$; a function $f$ belonging to the set $C([0,1] )$ of continuous function on $[0,1] $ is said to lie in $\mathcal{D}^*(\Delta_{\nu,\mu})$ if there exist $a,b \in \R$ and $g \in \S^*$ with $$\begin{aligned}
\label{KreinFeller}
f(x)=a+bF_{\mu}(x)+\int_{[0,x]}( F_{\mu}(x)-F_{\mu}(y) ) g(y)\;{\mathrm{d}}\nu(y)
\end{aligned}$$ for all $x \in [0,1]$. Alternatively, by Fubini we can write $ f(x)=a+bF_{\mu}(x)+\int_{[0,x]}\int_{[0,y]} g(s)\;{\mathrm{d}}\nu(s)\;{\mathrm{d}}\mu(y)$. By the uniqueness of densities we observe that $f$ determines $a$, $b$ and $g$ uniquely and by setting $ \Delta_{\nu,\mu}f \coloneqq g$ we define an injective linear operator $\Delta_{\nu,\mu}:\mathcal{D}^*(\Delta_{\nu,\mu})\to \S^*$. The first derivative of $ f \in \mathcal{D}^{\text{*}}(\Delta_{\nu,\mu})$ is defined by $$\begin{aligned}
\label{eq:first_derivative}
\nabla_{\mu}f(x) \coloneqq \nabla_{\mu}f(0)+\int_{[0,x]} \Delta_{\nu,\mu}f(y)\;{\mathrm{d}}\nu(y),
\;\; \text{where} \;\;
\nabla_{\mu}f(0) \coloneqq \lim_{x \downarrow 0^+} \dfrac{f(x)-f(0)}{F_{\mu}(x)-F_{\mu}(0)}
\;\;
\text{and}
\;\;
0^+ \coloneqq \inf ({\operatorname{supp}}(\mu)).\end{aligned}$$ The existence of the above limit follows from and the assumption that ${\operatorname{supp}}(\nu) \subseteq {\operatorname{supp}}( \mu)$. Additionally, from together with Lebesgue’s dominated convergence theorem, we have that $f$ is constant on every interval of constancy of $F_{\mu}$, $a=f(0)$, $b= \nabla f_{\mu}(0)$ and $\mathcal{D}^s ( \Delta_{\nu,\mu} ) \subset \mathcal{D}^w (\Delta_{\nu,\mu}) \subset C_{\nu,\mu}$. Moreover, for all $f \in\mathcal{D}^*(\Delta_{\nu,\mu})$ and all $x \in (0,1)$, $$\begin{aligned}
\nabla_{\mu}(f)(x)
= \lim_{ \substack{ y \downarrow x^+}} \dfrac{f(y)-f(x)}{F_{\mu}(y)-F_{\mu}(x)} + \Delta_{\nu,\mu} f (x)\nu(\{ x^+ \} ) (1 - \mathds{1}_{\{x^+\}}(x) )
= \lim_{ \substack{ y \uparrow x^-}} \dfrac{f(x)-f(y)}{F_{\mu}(x)-F_{\mu}(y)} + \Delta_{\nu,\mu} f (x)\nu(\{ x^- \} ).
\end{aligned}$$ Here, $x^+ \coloneqq \inf\{ y \in [0,1] \colon F_{\mu}(x) < F_{\mu}(y) \}$ and $x^- \coloneqq \sup\{ y \in [0,1] \colon F_{\mu}(x)>F_{\mu}(y) \}$. Observe, if $\nu$ is atomless, then $\nabla_{\mu}(f)$ is independent of $\nu$ and given by an ‘ordinary’ differential quotient.
For $ \gamma =(\alpha,\beta) \in [0,{\pi/2}]^{2}$ we consider the following eigenvalue problem for $\Delta_{\nu,\mu}$, see [@Fu87], $$\begin{aligned}
\label{EWC}
\Delta_{\nu,\mu}f= \lambda f
\end{aligned}$$ with [*Robin boundary conditions*]{} $$\begin{aligned}
\label{BC}
f(0)\cos(\alpha)-\nabla_{\mu}f(0)\sin(\alpha)=0
\quad \text{and} \quad
f(1)\cos(\beta)+\nabla_{\mu}f(1)\sin(\beta)=0.
\end{aligned}$$ We refer to the particular case $\gamma=(\pi/2,\pi/2)$ as the [*Neumann case*]{} and the case $\gamma=(0,0)$ as the [*Dirichlet case*]{}. We denote by $\mathcal{D}^{*}_{\gamma}( \Delta_{\nu,\mu} )$ the set of all $f \in \mathcal{D}^{*}( \Delta_{\nu,\mu} )$ which satisfy and . Combining and with Fubini’s theorem and the assumptions on $\nu,\mu$ one obtains a [*Gauss-Green formula*]{}; namely, for $f,g \in \mathcal{D}_{\gamma}^s( \Delta_{\nu,\mu} )$, $$\begin{aligned}
\int (\Delta_{\nu,\mu}f )g\mathrm{d} \nu
&= \left(
\nabla_{\mu}f (1)-\nabla_{\mu}f (0)\right)g(0)+ \int\nabla_{\mu}g(y)\left(
\nabla_{\mu}f (1)+\Delta_{\nu,\mu}f (y)\nu \left(\{y\} \right)-\nabla_{\mu}f (y)
\right)
{\mathrm{d}}\mu(y)
\\
&= \nabla_{\mu}f (1) g(1) - \nabla_{\mu}f (0) g(0) - \int \nabla_{\mu}f \nabla_{\mu} g \;\mathrm{d} \mu\end{aligned}$$ where $\int \nabla_{\nu,\mu}g(y)\Delta_{\nu,\mu}f (y)\nu \left(\{y\} \right) \,{\mathrm{d}}\mu(y)=0$ since $\mu$ is atomless. Considering $g = f$ we obtain that $\Delta_{\mu,\nu}$ with domain $\mathcal{D}^{*}_{\gamma}( \Delta_{\nu,\mu} )$, for $\gamma\coloneqq (\alpha,\beta) \in [0,\pi/2]^{2}$, is non-positive and since $$\begin{aligned}
&\int (\Delta_{\nu,\mu}f )gd \nu-\int (\Delta_{\nu,\mu}g )f \mathrm{d} \nu\\
&=-\tan(\beta) \nabla_{\mu}f(1) \nabla_{\mu}g(1)-\tan(\alpha) \nabla_{\mu}f(0) \nabla_{\mu}g(0)+\tan(\beta) \nabla_{\mu}f(1) \nabla_{\mu}g(1)+\tan(\alpha) \nabla_{\mu}f(0) \nabla_{\mu}g(0)
=0,\end{aligned}$$ the boundary conditions force $\Delta_{\mu,\nu}$ also to be symmetric.
Main results
------------
In we establishes a strong connection between $\Delta_{\nu,\mu}$ and $\Delta_{\nu \circ F_{\mu}^{-1}, \Lambda}$. Indeed, by utilising the *pseudo-inverse* $$\begin{aligned}
\check{F}^{-1}_{\mu}(x) \colon x \mapsto \inf\{ y \in [0,1] \colon F_{\mu}(y) \geq x \}
\end{aligned}$$ of $F_{\mu}$, we prove, for $* \in \{s,w\}$ and $\gamma \in [0,\pi/2]^{2}$, that $\varphi: f \mapsto f \circ \check{F}_{\mu}^{-1}$ is an isometric isomorphism on $\S^{*}$ with $$\begin{aligned}
\Delta_{\nu \circ F_{\mu}^{-1}, \Lambda }\circ\varphi=\varphi\circ \Delta_{\nu,\mu }\quad \text{and}\quad\varphi ( \mathcal{D}^*_{\gamma}(\Delta_{\nu,\mu}))=\mathcal{D}^*_{\gamma}(\Delta_{\nu \circ F_{\mu}^{-1},\Lambda}).\end{aligned}$$ With this at hand, one may conclude that the spectral properties of $\Delta_{\nu \circ F_{\mu}^{-1},\Lambda}$ will be inherit from $\Delta_{\nu ,\mu}$ and vice versa. For instance, using our correspondence theorem (), we obtain that $\Delta_{\nu ,\mu}$ is a densely defined, self adjoint operator with compact resolvent. Further, we are able the prove the following.
1. We obtain concerning the asymtotic growth rate of the eigenvalue counting function of $\Delta_{\nu,\mu}$, first provided in [@Fr05] for a certain class of self-similar measures. This is achieved via an application of and the identification of $\eta = \nu \circ F_{\mu}^{-1}$ as a certain self-similar measure, in tandem with the corresponding result for the classical Kreĭn-Feller operator $\Delta_{\eta,\Lambda}$, see [@Fu87]. This generalises several of the results of [@KSW16], where the case $\nu=\mu$ was considered.
2. Letting $(X_{t})_{t \geq 0}$ denote the a Liouville Brownian motion with speed measure $\nu \circ F_{\mu}^{-1}$ (see ), utilising our correspondence theorem (), we show that the infinitesimal generator of $(\check{F}^{-1}_{\mu}( X_t ))_{ t\geq 0}$ is the generalised Kreĭn-Feller operator $\Delta_{\nu,\mu}$ with Neumann boundary condition. Additionally, we compute the walk dimension of $(X_t)_{t \geq 0}$ and $(\check{F}^{-1}_{\mu}( X_t ))_{ t\geq 0}$.
Kreĭn-Feller operators
======================
Properties of classical Kreĭn-Feller operators
----------------------------------------------
In this section we collect important properties of classical Kreĭn-Feller operators, that is we consider the case $\mu= \Lambda$, with respect to weak and strong solutions. Most of these results are nowadays folklore and can be found for instance in [@MR0314125 Behauptung 2.2, Satz 2.1, Satz 3.1] with more or less detailed proofs. Since we could not find references where all the facts are proven in detail we decided to give here a quick overview and reduce all properties essentially on two key observations, namely the symmetry as a consequence of the Gauss-Green formula and surjectivity as stated in the following lemma.
\[Surjec\] For $\gamma=(\alpha, \beta) \in [0,{\pi/2}] ^{2}$ and $* \in \{s,w\}$ we have that the map $\Delta_{\nu,\Lambda}: \mathcal{D}^{*}_{\gamma}( \Delta_{\nu,\Lambda} ) \rightarrow \S_{\gamma}^* $ is surjective, where $\S_{(\pi/2,\pi/2)}^*\coloneqq\left\{ g \in \S^* : \int g {\mathrm{d}}\nu=0 \right\}$ and $\S_{\gamma}^* \coloneqq \S^
*$ for $\gamma\not=(\pi/2,\pi/2)$.
For $\gamma=(\alpha, \beta) \in [0,{\pi/2}]^{2}\setminus\{(\pi/2,\pi/2)\}$ we have that $\Delta_{\nu,\Lambda}$ is also injective and for its inverse $\Delta_{\nu,\Lambda}^{-1}: \S_{\gamma}^* \rightarrow \mathcal{D}^{*}_{\gamma}( \Delta_{\nu,\Lambda} )$ we have the following kernel representation $$\begin{aligned}
\Delta_{\nu,\Lambda}^{-1}g:x\mapsto \int K_{\alpha,\beta}(x,y) g(y) \,{\mathrm{d}}\nu(y)\end{aligned}$$ with continuous kernel given by $$\begin{aligned}
K_{\alpha,\beta}(x,y)\coloneqq
\begin{cases}
A_{\alpha,\beta}(1+\tan(\beta)-y)( \tan(\alpha)+x)+{\mathbbm{1}}_{[0,x]}(y)(x-y) &\mbox{for $\alpha,\beta\in [0,\pi/2)$},\\
(y-1)(x+\tan(\alpha) )+{\mathbbm{1}}_{[0,x]}(y)(x-y) &\mbox{for $\beta=\pi/2$ and $\alpha \in [0,\pi/2)$},\\
(y-1-\tan(\beta))+{\mathbbm{1}}_{[0,x]}(y)(x-y) &\mbox{for $\alpha= \pi/2$ and $\beta \in [0,\pi/2)$}.
\end{cases}\end{aligned}$$ with $A_{\alpha,\beta}:= {-1}/(1+\tan(\alpha)+\tan(\beta))$, for $\alpha,\beta\in (0,\pi/2)$. For the Neumann case $\alpha=\beta=\pi/2$ the operator $\Delta_{\nu,\Lambda}$ is not injective with kernel $ \Delta_{\nu,\Lambda}^{-1}(\{0\})=\R {\mathbbm{1}}$.
We only consider the case $\alpha,\beta \in [0,\pi/2)$ the other cases can be proved along the same lines. For fixed $ g \in \S^{*}$ and $x \in [0,1]$ we set $$f(x):=b\tan(\alpha)+bx+\int_{[0,x]} (x-y)g(y) {\mathrm{d}}\nu(y),$$ with $$b\coloneqq\dfrac{-1}{(1+\tan(\alpha)+\tan(\beta))} \left(\tan(\beta)\int_{[0,1]}g(y) {\mathrm{d}}\nu(y)+
\int_{[0,1]} (1-y)g(y) {\mathrm{d}}\nu(y)\right).$$ This imposes the right boundary conditions on $f$ and consequently we have $f \in \mathcal{D}^{w}_{\gamma}( \Delta_{\nu,\Lambda} ) $ with $\Delta_{\nu, \Lambda}f=g$.
As a direct consequence it follows that only in the Neumann case one has an eigenfunction with eigenvalue equal to zero.
Recall the following abstract facts on linear operators: Let us assume that $A: \operatorname{dom}(A) \subset H \rightarrow
H$ is a linear, symmetric and surjective operator on a Hilbert space $H$. Then one easily verifies that the annihilator of $\operatorname{dom}(A)$ is trivial. i.e. $\operatorname{dom}(A)^\perp=\{0\}$, and equivalently, $\operatorname{dom}(A)$ is dense in $H$. We can deduce further that $A$ is also self-adjoint: The inclusion $\operatorname{dom}(A) \subset \operatorname{dom}(A^*)$ holds by symmetry of $A$, where $A^*$ denotes the adjoint of $A$. For the reverse inclusion note that for fixed $f \in \operatorname{dom}(A^*)$ there exists, by surjectivity of $A$, an element $g \in \operatorname{dom}(A)$ such that $A^*f=Ag$. Then, using symmetry again, for each $h \in \operatorname{dom}(A)$, we have $\langle f ,A h \rangle = \langle A^* f,h\rangle =
\langle A g,h \rangle =
\langle g , A h\rangle $ and by surjectivity of $A$ we deduce $f=g \in \operatorname{dom}(A)$.
Now, we can apply this observation to our special setting, namely, for $\gamma \in [0,\pi/2]^2$ we set $H=\S_{\gamma}^w$, $A= \Delta_{\nu, \mu}$ and $\operatorname{dom}(A)={\mathcal{D}}^w_{\gamma}(\Delta_{\nu, \mu} )\cap \S_{\gamma}^w$. In the case of $\gamma \in [0, \pi/2] \setminus \{ (\pi/2,\pi/2)\}$ it follows immediately that $ \Delta_{\nu, \mu}$ with domain ${\mathcal{D}}^w_{\gamma}(\Delta_{\nu, \mu} )$ is a densely defined, self-adjoint operator on $L^{2}(\nu)$. In the case of Neumann boundary condition, i.e. $\gamma=(\pi/2,\pi/2)$, note that for the kernel we have $\Delta_{\nu,\ \mu}^{-1}(\{0\})= \R\mathds{1}$ and hence $L^2(\nu) = \S_{\gamma}^w \oplus \R\mathds{1}$. Therefore we have $ {\mathcal{D}}^w_{\gamma}(\Delta_{\nu, \mu} )=\{f +a:f\in {\mathcal{D}}^w_{\gamma}(\Delta_{\nu, \mu} )\cap \S_{\gamma}^w, a \in \R\}$ is dense in $L^2(\nu) $. Using this oberservation it follows again that $\Delta_{\nu, \mu} $ with domain ${\mathcal{D}}^w_{\gamma}(\Delta_{\nu, \mu} )$ is a densely defined and self-adjoint operator on $L^2(\nu) $. The following proposition summarizes this observation.
\[adjoint\] The partially defined operator $\Delta_{\nu,\Lambda}:L^{2}(\nu)\to L^{2}(\nu) $ with domain $ \mathcal{D}^{w}_{\gamma}( \Delta_{\nu,\Lambda}) $ for $\gamma \in [0,{\pi/2}]^{2}$ is self-adjoint, non-positive and, in particular, closed.
Fix $\gamma=(\alpha,\beta) \in [0,\pi/2]^2\setminus\{ (\pi/2, \pi/2)\}$ then the operator $R_0\coloneqq -\Delta_{\nu,\Lambda}^{-1}: \S_{\gamma}^w \to \S_{\gamma}^w$ is compact and self-adjoint.
Lemma \[Surjec\] shows that $R_0$ is a Hilbert-Schmidt operator with continuous (bounded) kernel function, therefore the compactness follows. Furher, the operator $R_0$ is self-adjoint this follows from symmetry of $\Delta_{\nu, \Lambda}$ in tandem with the fact that $R_0$ is also bounded.
\[Spectral\] Let $(\alpha,\beta) \in [0,\pi/2]^2$ , then the operator $\Delta_{\nu,\mu}$ with domain ${\mathcal{D}}_{\gamma}^{w}(\Delta_{\nu,\mu} )$ gives rise to an orthonormal (possibly finite) basis of eigenfunctions with eigenvalues $\lambda_{n}\leq 0$. If $L^{2}(\nu)$ is not finite dimensional then we have a countable number of eigenvalues with $\lim_{n \rightarrow \infty} -\lambda_n=\infty$, in particular $\Delta_{\nu,\Lambda}$ is an unbounded operator, otherwise there are only finitely many eigenfunctions and $\Delta_{\nu,\Lambda}$ is bounded.
At first we consider the case $\gamma=(\alpha,\beta) \in [0,\pi/2]^2\setminus\{ (\pi/2, \pi/2)\}$. Note that if $f \in {\mathcal{D}}^w_{\gamma}\left(\Delta_{\nu, \Lambda} \right)$ is an eigenfunction of $\Delta_{\nu, \Lambda}$ with eigenvalue $\lambda<0$, then applying Lemma \[Surjec\] gives $$\Delta_{\nu, \Lambda}f=\lambda f \Longleftrightarrow \lambda^{-1} f = \Delta_{\nu, \Lambda}^{-1}f.$$ Then the statement follows directly from the spectral theorem for linear, compact and self-adjoint operators applied to $ \Delta_{\nu, \Lambda}^{-1}$.
For the case $\gamma=(\alpha,\beta)=(\pi/2,\pi/2)$ we have to consider the resolvent operator $R^{\lambda}_{\nu,\Lambda}\coloneqq(\lambda I -\Delta_{\nu,\Lambda})^{-1}$ with $\lambda>0$ , from the integral representation of the resolvent operator $R^{\lambda}_{\nu,\Lambda}\coloneqq(\lambda I -\Delta_{\nu,\Lambda})^{-1}$ of $\Delta_{\nu,\Lambda}$ with domain ${\mathcal{D}}^w_{\gamma}\left(\Delta_{\nu, \Lambda} \right)$ given in [@MR0314125 1.2], one may conclude that $R^{\lambda}_{\nu,\Lambda}$ is compact and self-adjoint, see also [@MR574035 Theorem 1, p. 251]. Again applying the spectral theorem proves the statement.
Clearly, $L^{2}(\nu)$ is finite dimensional if and only if the support of $\nu$ is a finite set.
\[densecm\] For $\gamma=(\alpha,\beta )\in [0,{\pi/2}]^{2} $, set $$\begin{aligned}
C_{\nu,\Lambda}^{\gamma}
\coloneqq \begin{cases}
\left\{ f \in C_{\nu,\Lambda} \colon f(0)=0, \ f(1)=0\right\}& \text{ if } \alpha=\beta=0, \\
\left\{ f \in C_{\nu,\Lambda} \colon f(0)=0\right\}& \text{ if } \alpha=0, \ \beta \in (0,\pi/2],\\
\left\{ f \in C_{\nu,\Lambda} \colon f(1)=0\right\} &\text{ if } \alpha \in (0,\pi/2], \ \beta=0, \\
C_{\nu,\Lambda
}& \text{ if } \alpha,\beta \in (0,\pi/2].
\end{cases}
\end{aligned}$$ The set $\mathcal{D}^s_{\gamma}(\Delta_{\nu, \, \Lambda})$ is dense in $\left(C_{\nu,\Lambda}^{\gamma}, \lVert \cdot \rVert_{\infty}\right)$.
The result can be found in [@MR0314125 Behauptung 2.4] without a detailed proof. We will sketch a proof of this result for the case $\gamma=(\alpha,\beta) \in (0,\pi/2) \times [0,\pi/2)$; the other cases follow in a similar fashion. Fix $\Phi \in C_{\mu,\Lambda}' $ such that $$\forall f \in \mathcal{D}^s_{\gamma}(\Delta_{\nu, \, \Lambda}): \
\Phi(f)=\int f(x) \,{\mathrm{d}}\phi(x)=0,$$ where $\phi$ is the signed distribution function representing $\Phi$. By definition $\phi$ is of bounded variation and local constant on the complement of ${\operatorname{supp}}(\nu)$. With $E:=-(1+\tan(\beta))A_{\alpha,\beta} $ it follows from the proof of Lemma \[Surjec\] for all $g \in C_{\nu,\Lambda}^{\gamma}$ $$\begin{aligned}
\int (E+s\cdot A_{\alpha,\beta})g(s)\,{\mathrm{d}}\nu(s) \cdot \int (\tan(\alpha)+x) \,{\mathrm{d}}\phi(x)&= \int \int (x-s)g(s) \,{\mathrm{d}}\nu (s)\,{\mathrm{d}}\phi(x) \label{eq:dual1}\\
&= \int \left(\phi(1)(1-s)- \int_{[s,1]}\phi(s) \,{\mathrm{d}}s \right)g(s)\,{\mathrm{d}}\nu(s).\nonumber\end{aligned}$$ Furher, we have $$\int (\tan(\alpha)+x) \,{\mathrm{d}}\phi(x)=\tan(\alpha)\left( \phi(1)-\phi(0)\right)+\phi(1)-\int \phi(s) \,{\mathrm{d}}s\eqqcolon B_{\phi}.$$ Combining these identities gives $$\begin{aligned}
\int \left(E \cdot B_{\phi} +s \cdot A_{\alpha,\beta} \cdot B_{\phi} -\phi(1)(1-s)+ \int_{[s,1]}\phi(s) \,{\mathrm{d}}s\right)g(s) \,{\mathrm{d}}\nu(s)=0\end{aligned}$$ If we consider $g(s):=E \cdot B_{\phi} +s \cdot A_{\alpha,\beta} \cdot B_{\phi} -\phi(1)(1-s)+ \int_{[s,1]}\phi(x) \,{\mathrm{d}}x \in \mathcal{C}_{\nu,\Lambda}^{\gamma}$ it follows, that for all $s \in [0,1]$ $$\int_{[s,1]}\phi(x) \,{\mathrm{d}}x=\phi(1)-E \cdot B_{\phi} -s( A_{\alpha,\beta} \cdot B_{\phi}+\phi(1)),$$
which is only possible if $\phi(s)=A_{\alpha,\beta} B_{\phi}+\phi(1)$, for all $s \in [0,1]$. Therefore $\Phi$ is a Dirac-measure in $\{0\}$ with weight $\phi(0)$ and we have by , or all $f \in \mathcal{C}_{\nu,\Lambda}^{\gamma}$, $$\phi(0) \ \tan(\alpha) \ \int\left(E+A_{\alpha,\beta} \cdot s\right)f(s) \,{\mathrm{d}}\nu(s)=0$$ For the particular choice $f \in \mathcal{C}_{\nu,\Lambda}^{\gamma}$ given by $f:s\mapsto E+A_{\alpha,\beta} \cdot s$ the above integral is positive and hence $
\phi(0)=0$. Consequently, $\Phi=0$ and we have shown that the annihilator of $ \mathcal{D}^s_{\gamma}(\Delta_{\nu, \Lambda}
)$ is trivial and therefore $ \mathcal{D}^s_{\gamma}(\Delta_{\nu, \, \Lambda})$ is dense in $ \mathcal{C}_{\nu,\Lambda}^{\gamma}$.
Generalized Kreĭn-Feller operators and transformations of measure spaces
------------------------------------------------------------------------
Let us first state two basic key observations. three key observation.
\[identity\] The function $\check{F}^{-1}_{\mu} \circ F_{\mu}$ equals the identity $\nu$-almost everywhere.
Note, $\check{F}^{-1}_{\mu}( F_{\mu}(x) )\neq x$ if and only if there exists $\varepsilon>0$ with $F_{\mu}(x-\varepsilon)=F_{\mu}(x)$. This means, if $\check{F}^{-1}_{\mu}( F_{\mu}(x) )\neq x$, then $x$ belongs to an interval of constancy for $F_\mu$. This in tandem with our hypothesis ${\operatorname{supp}}(\nu) \subseteq {\operatorname{supp}}(\mu)$, implies that the countable union of these intervals have $\nu$-measure zero.
\[lem: isomorphism\] The mapping $\varphi: \S^{*}(\nu,\mu) \to \S^{*}(\nu\circ {F}^{-1}_{\mu} ,\Lambda))$ defined by $\varphi(f) \coloneqq f \circ \check{F}^{-1}_{\mu}$, is an isometric isomorphism with inverse $\varphi^{-1}(f) \coloneqq f \circ F_{\mu}$, $*\in \{s,w\}$.
This is a consequence of together with the push-forward formula for measures in the strong case and the definition of $C_{\nu,\mu}$ in the weak case.
The following theorem is the main observation in this section needed for all subsequent corollaries.
\[prop:laplace\_backward\] For $\gamma \in [0,{\pi/2}]^{2}$, we have that $\Delta_{\nu \circ F_{\mu}^{-1}, \Lambda }\circ\varphi=\varphi\circ \Delta_{\nu,\mu }$ and $\varphi ( \mathcal{D}^*_{\gamma}(\Delta_{\nu,\mu}))=\mathcal{D}^*_{\gamma}(\Delta_{\nu \circ F_{\mu}^{-1},\Lambda})$.
If $f \in \mathcal{D}^*_{\gamma}(\Delta_{\nu,\mu})$, then, for all $x \in {\operatorname{supp}}(\mu)$, $$\begin{aligned}
f(x)=a+bF_{\mu}(x)+
\int \mathds{1}_{[0, x]}(y)(F_{\mu}(x)-F_{\mu}(y) ) \Delta_{\nu, \mu } f(y) \;{\mathrm{d}}\nu(y),\end{aligned}$$ where $a=f(0)$ and $b= \nabla_{\mu} f(0)$. Using and replacing $x$ with $\check{F}_{\mu}^{-1}(x)$ gives $$\begin{aligned}
f(\check{F}_{\mu}^{-1}(x))
&=a+bF_{\mu}(\check{F}_{\mu}^{-1}(x))+
\int \mathds{1}_{[0, \check{F}_{\mu}^{-1}(x)]}(y)(F_{\mu}(\check{F}_{\mu}^{-1}(x))-F_{\mu}(y) ) \Delta_{\nu, \mu }(f )(y)\;{\mathrm{d}}\nu(y) \\
&=a+b x+
\int \mathds{1}_{[0, \check{F}_{\mu}^{-1}(x)]}(y)(x-F_{\mu}(y) ) \Delta_{\nu, \mu }(f )(\check{F}_{\mu}^{-1}(F_{\mu}(y)))\;{\mathrm{d}}\nu(y) \\
&=a+bx+
\int \mathds{1}_{[0, x]}(F_{\mu}(y))(x-F_{\mu}(y) ) \Delta_{\nu, \mu }(f )(\check{F}_{\mu}^{-1}(F_{\mu}(y)))\;{\mathrm{d}}\nu(y) \\
&=a+bx+
\int \mathds{1}_{[0, x]}(y)(x-y ) \Delta_{\nu, \mu }(f ) \circ \check{F}_{\mu}^{-1}(y) \;{\mathrm{d}}(\nu \circ F_{\mu}^{-1})(y).\end{aligned}$$ Therefore, $a=f ( \check{F}^{-1}_{\mu}(0) )$ and $b=\nabla_{\mu}f ( \check{F}^{-1}_{\mu}(0) )$. This shows $$\begin{aligned}
f \circ \check{F}_{\mu}^{-1} \in \mathcal{D}^*(\Delta_{\nu \circ F_{\mu}^{-1}, \, \Lambda})
\quad \text{and} \quad
\Delta_{\nu \circ {F}_{\mu}^{-1}, \, \Lambda} (f \circ \check{F}_{\mu}^{-1}) = \Delta_{\nu, \, \mu}f\circ \check{F}^{-1}_{\mu}.
\end{aligned}$$ If $f \circ \check{F}_{\mu}^{-1} \in \mathcal{D}^*_{\alpha,\beta}(\Delta_{\nu \circ F_{\mu}^{-1}, \, \Lambda})$, then, as above, for $x \in {\operatorname{supp}}(\mu)$ with $F_{\mu}(x-\varepsilon)<F_{\mu}(x)$ for all $\varepsilon>0$, $$\begin{aligned}
f( x )
=f(\check{F}_{\mu}^{-1}( F_{\mu}( x)) )
&= c+dF_{\mu}(x) + \int \mathds{1}_{[0, F_{\mu}(x)]}(y)(F_{\mu}(x)-y ) \Delta_{\nu \circ F_{\mu}^{-1}, \Lambda}(f \circ \check{F}_{\mu}^{-1})(y) \;{\mathrm{d}}(\nu \circ F_{\mu}^{-1})(y)\\
&=
c+dF_{\mu}(x)+
\int \mathds{1}_{[0, F_{\mu}(x)]}(F_{\mu}(y))(F_{\mu}(x)-F_{\mu}(y) ) \Delta_{\nu \circ F_{\mu}^{-1}, \Lambda}
(f \circ \check{F}_{\mu}^{-1})
\circ F_{\mu}(y)\;{\mathrm{d}}\nu (y) \\ &=
c+dF_{\mu}(x)+
\int \mathds{1}_{[0, x]}(y)(F_{\mu}(x)-F_{\mu}(y) ) \Delta_{\nu \circ F_{\mu}^{-1}, \Lambda}
(f \circ \check{F}_{\mu}^{-1})
\circ F_{\mu}(y)\;{\mathrm{d}}\nu (y),\end{aligned}$$ with $c=f (\check{F}^{-1}_{\mu}(0 ))$ and $d =\nabla f (\check{F}^{-1}_{\mu} (0 ) )$. The case $x \in {\operatorname{supp}}(\mu)$ and $F_{\mu}(x-\varepsilon)=F_{\mu}(x)$ for some $\varepsilon>0$ implies that $x$ lies in an interval of constancy. Notice, $$\begin{aligned}
\check{F}_{\mu}^{-1}( [0,1])=\{0\} \cup {\operatorname{supp}}(\mu) \cap \{ x \in [0,1] \text{ is not a right endpoint of an interval of constancy}\}, \end{aligned}$$ thus we can consider a modification of $f$ such that $f$ is constant on each interval of constancy of $F_{\mu}$. Thus, we have $c=f(0), \ d=\nabla_{\mu}f(0)$. Therefore, we have $$\begin{aligned}
\Delta_{\nu \circ F_{\mu}^{-1}, \Lambda}(f \circ \check{F}_{\mu}^{-1} ) \circ F_{\mu}
= \Delta_{\nu,\mu} ( f ) \quad \text{and} \quad f\in \mathcal{D}^{*}(\Delta_{\nu, \mu}).
\end{aligned}$$ It remains to verify that the boundary conditions are preserved. Combining the above with gives $$\begin{aligned}
\nabla_{\mu}f(1)= \nabla_{\mu}f(0)+\int_{[0,1]} \Delta_{\nu,\mu}f(y)\;{\mathrm{d}}\nu(y)
&= \nabla_{\Lambda}f(0)+\int_{[0,1]} \Delta_{\nu,\mu}f ( \check{F}_{\mu}^{-1}(y) )\;{\mathrm{d}}\nu \circ F_{\mu}^{-1}(y) \\
&= \nabla_{\Lambda}f(0)+\int_{[0,1]}
\Delta_{\nu \circ {F}_{\mu}^{-1}, \, \Lambda} (f \circ \check{F}_{\mu}^{-1})
\;{\mathrm{d}}\nu \circ F_{\mu}^{-1}(y)
= \nabla_{\Lambda} (f \circ \check{F}_{\mu}^{-1} )(1). \qedhere\end{aligned}$$
For $\gamma=(\alpha,\beta )\in [0,{\pi/2}]^{2} $, set $$\begin{aligned}
C_{\nu,\mu}^{\gamma}
\coloneqq \begin{cases}
\left\{ f \in C_{\nu,\mu} \colon f(0)=0, \ f(1)=0\right\}& \text{ if } \alpha=\beta=0, \\
\left\{ f \in C_{\nu,\mu} \colon f(0)=0\right\}& \text{ if } \alpha=0, \ \beta \in (0,\pi/2],\\
\left\{ f \in C_{\nu,\mu} \colon f(1)=0\right\} &\text{ if } \alpha \in (0,\pi/2], \ \beta=0, \\
C_{\nu,\mu
}& \text{ if } \alpha,\beta \in (0,\pi/2].
\end{cases}
\end{aligned}$$ The set $\mathcal{D}^s_{\gamma}(\Delta_{\nu, \, \mu})$ is dense in $(C_{\nu,\mu}^{\gamma}, \lVert \cdot \rVert_{\infty})$.
This follows from , Lemma \[densecm\] and Theorem \[prop:laplace\_backward\].
For each $\gamma \in [0,{\pi/2}]^{2}$, the operator $\Delta_{\nu, \, \mu}$ with domain ${\mathcal{D}}_{\gamma}^w(\Delta_{\nu,\mu})$ is densely defined and self-adjoint.
The denseness follows from , Lemma \[adjoint\] and Theorem \[prop:laplace\_backward\]. Furhermore, for $f \in \operatorname{dom}(\Delta_{\nu,\mu}^*)$ by Theorem \[prop:laplace\_backward\] and Lemma \[identity\], $$\begin{aligned}
g\mapsto
\left\langle f ,\Delta_{\nu,\mu} g \right\rangle_{L^2(\nu)}
=\left\langle f\circ \check{F}^{-1}_{\mu} , \Delta_{\nu \circ \check{F}^{-1}_{\mu},\Lambda} (g \circ \check{F}^{-1}_{\mu}) \right\rangle_{L^2(\nu \circ F_{\mu}^{-1})}
\end{aligned}$$ defines a continuous linear functional on $ \operatorname{dom}(\Delta_{\nu,\mu})$. Combining Proposition \[adjoint\] and Theorem \[prop:laplace\_backward\] we deduce $ f \circ \check{F}^{-1}_{\mu} \in \operatorname{dom}\big( \Delta_{\nu \circ \check{F}^{-1}_{\mu},\Lambda}^{*}\big)= \operatorname{dom}\big( \Delta_{\nu \circ \check{F}^{-1}_{\mu},\Lambda}\big)={\mathcal{D}}_{\gamma}^w( \Delta_{\nu \circ \check{F}^{-1}_{\mu},\Lambda})$ and consequently $f \in {\mathcal{D}}_{\gamma}^w(\Delta_{\nu,\mu})=\operatorname{dom}(\Delta_{\nu,\mu})$.
There is an analogous theorem in general theory of Markov processes \[[@DynkinI_II], p. 325, Theorem 10.13\]. Taking into account that $\Delta_{\nu,\mu}$ with domain ${\mathcal{D}}_{\gamma}^{*}(\Delta_{\nu,\mu})$ has a probabilistic interpretation as infinitesimal generator of a Markov process, thus the result above is not surprising.
Corollary \[Spectral\] implies that $\Delta_{\nu \circ {F}_{\mu}^{-1}, \, \Lambda}$ with domain ${\mathcal{D}}_{\gamma}^w \left( \Delta_{\nu, \Lambda} \right)$ gives rise to an orthonormal basis of eigenfunctions with non-positive eigenvalues. In the case of ${\operatorname{supp}}(\nu)$ is infinite we have $(\lambda_n)_{n \in \mathbb{N}}$ with $\lim_{n \to \infty}-\lambda_n = \infty$ otherwise there are only finitely many eigenvalues. Using the one-to-one correspondence established in to relate the spectral properties of $\Delta_{\nu,\mu}$ with those of $\Delta_{\nu \circ {F}_{\mu}^{-1}, \, \Lambda}$, we obtain the following.
\[thm Spec\] For fixed $ \gamma \in [0,{\pi/2}]^{2}$, the operators $\Delta_{\nu \circ {F}_{\mu}^{-1}, \, \Lambda} $ with domain $\mathcal{D}^w_{\gamma}( \Delta_{\nu \circ F_{\mu}^{-1},\Lambda} )$ and $ \Delta_{\nu,\mu} $ with domain $\mathcal{D}^w_{\gamma}( \Delta_{\nu,\mu})$ have the same eigenvalues $(\lambda_n)_{n \in \mathbb{N}}$. Further, if $f$ is an eigenfunction of $ \Delta_{\nu,\mu} $, then $f \circ \check{F}^{-1}_{\mu}$ is an eigenfunction of $ \Delta_{\nu \circ {F}_{\mu}^{-1}, \, \Lambda} $, and if $f$ is an eigenfunction of $\Delta_{\nu \circ {F}_{\mu}^{-1}, \, \Lambda}$, then $f \circ F_{\mu}$ is an eigenfunction of $ \Delta_{\nu,\mu}$. In particular, if $(f_n)_{n \in \mathbb{N}}$ denotes the orthonormal basis consisting of eigenfunctions of $ \Delta_{\nu \circ {F}_{\mu}^{-1}, \, \Lambda}$ then $(f_n \circ F_{\mu})_{n \in \mathbb{N}}$ forms an orthonormal basis consisting of $ \Delta_{\nu,\mu}$-eigenfunctions.
can be seen as a generalisation of [@KSW16] where the $\Delta_{\mu,\mu}$-Laplacian has been considered.
\[Reso\] Assume that ${\operatorname{supp}}(\nu) \subseteq {\operatorname{supp}}(\mu)$ and $\gamma \in[0,{\pi/2}]^{2}$. For $\lambda>0$, letting $R_{\nu , \mu}^{\lambda}$ denote the resolvent operator of $ \Delta_{\nu, \mu}$ with domain ${ \mathcal{D}^*_{\gamma}(\Delta_{\nu,\mu})}$, for all $f \in \S^*$ $$\begin{aligned}
R_{\nu,\mu}^{\lambda}( f ) \circ \check{F}_{\mu}^{-1}=R_{\nu \circ F_{\mu}^{-1},\Lambda}^{\lambda}(f \circ \check{F}_{\mu}^{-1}).\end{aligned}$$ In particular, we have $$\begin{aligned}
\big \Vert R_{\nu \circ F_{\mu}^{-1},\Lambda}^{\lambda} \big\Vert_{L^2(\nu \circ F_{\mu}^{-1} ) }=\big \Vert R_{\nu,\mu}^{\lambda} \big\Vert_{L^2 (\nu) },\end{aligned}$$ and $$\begin{aligned}
\big \Vert R^{\lambda}_{\nu \circ F_{\mu}^{-1},\Lambda} \big\Vert_{C_{\nu \circ F_{\mu}^{-1},\Lambda}}
=
\big\Vert R^{\lambda}_{\nu,\mu}\big\Vert_{C_{\nu,\mu}}\end{aligned}$$
Note that the resolvent $R^{\lambda}_{\nu,\mu}$ is well defined for all $\lambda>0$ (see [@MR2017701]) and for $f \in \S^*$ we have $$\begin{aligned}
\Delta_{\nu,\mu} R_{\nu ,\mu}^{\lambda}f
=\lambda R_{\nu, \mu}^{\lambda}f-f, \ R_{\nu ,\mu}^{\lambda}f \in \mathcal{D}^*_{ \gamma}( \Delta_{\nu,\mu} ) .\end{aligned}$$ gives $$\begin{aligned}
\Delta_{\nu\circ F_{\mu}^{-1},\Lambda} (R_{\nu ,\mu}^{\lambda}f\circ \check{F}^{-1}_{\mu}) \circ {F}_{\mu}
=\lambda R_{\nu,\mu}^{\lambda}f-f\end{aligned}$$ and hence $$\begin{aligned}
\Delta_{\nu\circ F_{\mu}^{-1},\Lambda} (R_{\nu ,\mu}^{\lambda}f\circ \check{F}^{-1}_{\mu})
=\lambda R_{\nu,\mu}^{\lambda}f\circ \check{F}_{\mu}^{-1} -f \circ \check{F}_{\mu}^{-1}\end{aligned}$$ proving the first part. The statement on the norms now follows from .
\[Feller\] Let $\gamma \in[0,{\pi/2}]^{2}$. Then the operator $\Delta_{\mu,\nu}$ with domain $ {\mathcal{D}}_{\gamma}^s(\Delta_{\nu,\mu})$ is an infinitesimal generator of a strongly continuous semi-group of contraction on $C_{\nu,\mu}^{\gamma}$.
It is well known that this holds true for the classical Krein-Feller operator [@MR0314125 Behauptung 4.1]. This in tandem with the Yosida-Hille Theorem [@ma1992introduction p. 11, Theorem 1.12] and , yields $$\begin{aligned}
\big\Vert R^{\lambda}_{\nu \circ F_{\mu}^{-1},\Lambda} \big\Vert_{C^{\gamma}_{\nu \circ F_{\mu}^{-1},\Lambda}}
=
\big\Vert R^{\lambda}_{\nu,\mu} \big\Vert_{C_{\nu,\mu}^{\gamma}} \leq 1/\lambda ,\end{aligned}$$ for all $\lambda>0$. Further $\Delta_{\nu, \mu}$ is densely defined and closed operator in $C_{\nu,\mu}^{\gamma}$. Applying again the theorem of Yosida-Hille gives the statement.
Applications
============
Spectral asymptotics
--------------------
In this section we review the asymptotic spectral properties of $\Delta_{\nu,\mu}$ for certain classes of self-similar measures. We will show how the results in [@Fr05] can be deduced from [@Fu87] with the help of the above established isomorphism. Let us give general assumption on the self-similar measures which are clearly fulfilled under the assumptions (A.1) – (A.4) of [@Fr05].
\[StrongAssumption\] Let $S_i:[0,1]\to [0,1]$, $i=1,\dots,M$, denote a family of affine contractions fulfilling the open set condition (OSC), that is, $S_i((0,1)) \cap S_j((0,1))= {\varnothing}$, for all $j \neq i$, and let $\nu$ denote the associated self-similar measure with probability weight vector $(p_{1},\ldots ,p_{M}) \in (0,1)^M$ uniquely determined by $$\begin{aligned}
\nu(A)=\sum_{i=1}^{M} p_i \nu(S_i^{-1}(A))\end{aligned}$$ for $A \in \mathfrak{B}([0,1])$. For fixed $(\sigma_{1},\ldots ,\sigma_{M})\in (0,1)^{M}$ with $\sum_{i=1}^M \sigma_i \leq 1$ let $\mu$ denote an atomless probability measure with ${\operatorname{supp}}(\nu) \subseteq {\operatorname{supp}}(\mu)$ such that for all $A \in \mathfrak{B}([0,1])$, $i=1,\dots ,M$, we have $\mu(S_i(A))=\sigma_i \mu(A)$.
\[Freiberg\] For the decreasing sequence of eigenvalues $( \lambda_n )_{n \in \mathbb{N}}$ of $\Delta_{\nu,\mu}$ with domain $ \mathcal{D}^{w}_{\gamma}(\Delta_{\nu ,\mu})$, $\gamma\in [0,\pi/2]^{2}$, we have $$\begin{aligned}
-\lambda_n\asymp n^{1/ u} \quad \text{and} \quad N_{\Delta_{\nu,\mu}}(x)\asymp x^{ u},\end{aligned}$$ where $ u\in (0,1)$ denotes the unique number with $\sum_{i=1}^M ( \sigma_i p_i )^{ u}=1.$
We have $\sigma_i F_{\mu}(x)=\sigma_i \mu([0,x])= \mu(S_i( [0,x] ) )=F_{\mu}(S_i(x))- F_{\mu}(S_i(0))$, for all $x \in [0,1]$, $i \in \{1,\dots,M\}$. Setting ${\widetilde}{S}_i(x)\coloneqq\sigma_i x+F_{\mu}(S_i(0))$ we obtain $F_{\mu} \circ S_i= {\widetilde}{S}_i \circ F_{\mu}$, and hence, for all $A \in \mathfrak{B}([0,1])$, $$\begin{aligned}
\sum_{i=1}^M p_i\nu( F_{\mu}^{-1} \circ {\widetilde}{S}_i^{-1}(A))
&=\sum_{i=1}^M p_i\nu( S_i^{-1} \circ F_{\mu}^{-1}(A))
=\nu ( F_{\mu}^{-1}(A) ). \end{aligned}$$ This shows that $ \nu \circ F_{\mu}^{-1}$ is the unique self similar measure with contractions $( {\widetilde}{S}_i )_{i=1,\dots N}$ and weights $(p_i)_{i=1,...,M}$. Further, $$\begin{aligned}
{\widetilde}{S}_i( [0,1] )=[F_{\mu}(S_i(0)),\sigma_i+F_{\mu}(S_i(0))]=[F_{\mu}(S_i(0)),F_{\mu}(S_i(1))],\end{aligned}$$ which shows that the OSC is satisfied for $( {\widetilde}{S}_i \colon i=1,\dots ,M )$ given that $( {S}_i \colon i=1,\dots ,M )$ satisfies the OSC. Therefore, we can apply the classical result from Fujita [@Fu87] for spectral dimension of $\Delta_{\nu \circ {F}_{\mu}^{-1}, \, \Lambda} $ with domain ${\mathcal{D}^{w}_{\gamma} (\Delta_{\nu \circ {F}_{\mu}^{-1}, \, \Lambda} )}$. Combining this with completes the proof.
Let $S_i(x) \coloneqq s_ix+b_i$ for $x \in [0,1]$ with $s_i,b_i \geq 0$ and $s_i+b_i \leq 1$ for $i=1,\dots,M$ and assume the OSC. For $N\leq M$ let $\nu$ denote the self-similar measure with respect to $(S_{i}\colon i=1,\ldots ,N)$ and probability weight vector $(p_{1},\ldots , p_{N})\in (0,1)^{N}$ and support $L=\bigcup_{i=1}^N S_i(L)$, and $\mu$ the self-similar measure with respect to $(S_{i}\colon i=1,\ldots ,M)$ and probability weight vector $(\sigma_{1},\ldots , \sigma_{M})\in (0,1)^{M}$ and support $K=\bigcup_{i=1}^M S_i(K)$. Then we have $L \subset K$ and $$\begin{aligned}
\nu \circ F_{\mu}^{-1}(A)= \sum_{i=1}^N p_i \nu(S_i^{-1}\circ F_{\mu}^{-1}(A)).\end{aligned}$$ If for $\delta_{K}\coloneqq \dim_{H}(K)$ and $\delta_{L}\coloneqq \dim_{H}(L)$ we chose $\sigma_{m}\coloneqq s_{m}^{\delta_{K}}$, $m =1,\ldots, M$, and $p_{\ell}\coloneqq s_{\ell}^{\delta_{L}}$, $\ell =1,\ldots, L$, then we find the well-known relation $ u=\delta_{K}/(\delta_{K}+\delta_{L})$.
Liouville Brownian motion and walk dimension {#sec:LBM}
--------------------------------------------
In this section we will construct Liouville Brownian motion via a time change of a fixed Brownian motion. We will give a short overview about basic properties of this stochastic process. After which, we will compute the walk dimension for special classes of gap diffusions. Additionally, we show that the infinitesimal generator coincides with a generalised Kreĭn-Feller operator.
For the tuple $\left(\Omega, \mathcal{F}, ( \mathcal{F}_t)_{t \geq 0}, ( \theta_t)_{t \geq 0}, \mathbb{P}_x\right)$ where $\mathbb{P}_x$ denotes the probability measure such that, $( \mathcal{F}_t)_{t \geq 0}$ is the right-continuous completed natural filtration and $(\theta_t )_{t \geq 0}$ is the shift-operator. The expectation with respect to $\mathbb{P}_x$ is denoted by $\mathbb{E}_x$. We call the stochastic process $(B_t )_{t \geq 0}$ a Brownian motion if the following are satisfied.
1. $\mathbb{P}_x(B_0=x)=1$
2. For $0 \leq s_0< \dots <s_n$ with $n \in \N$ the increments $B_{s_1}-B_{s_0}, \dots ,B_{s_n}-B_{s_{n-1}}$ are stochastically independent.
3. The distribution of the increment $B_{s}-B_{t}$ follows a Gaussian distribution with variance $s-t$ where $s>t \geq 0$ and mean $0$.
Let $(L_{x}^t)_{t \geq 0, x \in \mathbb{R}}$ the jointly continuous version of the local time of the Brownian motion, see [@Revuz2013 Chapter VI].
Let $m$ denote a non-zero Borel measure on $(\R,\mathcal{B})$.
\[LBM\] We define for $ t \geq 0$ the [*(inverse) time-change function*]{} $$\begin{aligned}
\Phi_t \coloneqq \int_{\R} L_x^t \;{\mathrm{d}m}(x),\;\; \ \hat{\Phi}^{-1}_t \coloneqq \inf\left\{s \geq 0 \colon \Phi_s>t\right \}\end{aligned}$$ where $\hat{\Phi}^{-1}_t$ is called right-continuous inverse of $\Phi_t$. Now, we define the new process for $x \in {\operatorname{supp}}(m)$ $$\begin{aligned}
\left((X_t)_{ t \geq 0} \coloneqq (B_{\hat{\Phi}^{-1}_t})_{t \geq 0}, (\mathcal{F}_{\hat{\Phi}^{-1}_t})_{t \geq 0}, \mathbb{P}_x\right )\end{aligned}$$ which will be called $m$-Liouville Brownian motion with speed measure $m$ starting in $x$.
\[DefWalkD\] Let $(X_t)_{t \geq 0}$ a Liouville Brownian motion with speed measure $m$. Then the *walk dimension* $d_w(x)$ in $x \in {\operatorname{supp}}(m)$ is defined by $$\begin{aligned}
\label{WalkD}
d_W(x) \coloneqq
\lim_{R \to \infty} \dfrac{\log(\mathbb{E}_x[\tau_{(-R+x,R+x)}] )}{\log(R)}\end{aligned}$$ assuming that the limit exists, where $\tau_{(-R+x,R+x)} \coloneqq \inf \{ t \geq 0 \colon X_t \notin (-R+x,R+x)\}$
Assuming the support of $m$ is bounded, then the walk dimension does not exist. In this case it can be useful to consider the so-called *local walk dimension* $$\begin{aligned}
d_{LW}(x) \coloneqq \lim_{R \downarrow 0} \dfrac{\log(\mathbb{E}_x[\tau_{(-R+x,R+x)}] )}{\log(R)}.\end{aligned}$$ There are other definitions of the walk dimension; for example in [@KGOLMANKHANEH2018960] the walk dimension $d_{w}(x)$ is defined by $$\begin{aligned}
\mathbb{E}_x [ (X_t-x )^2 ] \asymp t^{2/d_w(x)}.\end{aligned}$$
It remains open in which cases this definition coincides with the Definition \[DefWalkD\].
\[SkipfreeP\] For all $t \geq 0$, we have $\mathbb{P}_x$-almost everywhere, that $X_t \in {\operatorname{supp}}(m)$.
We have $\left((X_t )_{t \geq 0}, (\mathcal{F}_{\hat{\Phi}^{-1}_t})_{t \geq 0}, (\theta_{\hat{\Phi}^{-1}_t} )_{t \geq 0}, ( \mathbb{P}_x )_{x \in {\operatorname{supp}}(m)}\right)$ defines a Feller process. Namely, the tuple defines is a strong Markov process such that, for all $f \in C_b ({\operatorname{supp}}(m) )$, the function $x \mapsto \mathbb{E}_x[f(X_t) ]$ belongs to $C_b({\operatorname{supp}}(m) )$ and $\lim_{ t \downarrow 0} \mathbb{E}_x[f(X_t)] =f(x)$. Here, $C_b ({\operatorname{supp}}(m) )$ denotes the set of bounded continuous function with domain ${\operatorname{supp}}(m)$.
To compute the walk dimension and infinitesimal generator, we need the following lemma.
\[PotentialGapD\] Fix $x_0,x_1 \in {\operatorname{supp}}(m)$ with $x_0<x_1$ and set $\tau_{(x_0,x_1)} \coloneqq \inf\{ t \geq 0 \colon X_t \notin (x_0,x_1)\}$. For $x \in (x_0,x_1) \cap {\operatorname{supp}}(m)$, the following holds.
1. For all $t \geq 0$ and all bounded measurable functions $f$ on ${\operatorname{supp}}(m)$ we have $$\begin{aligned}
\qquad\quad\mathbb{E}_x\left[ \int_0^{\tau_{[x_0,x_1]}}f(X_s)\mathrm{d}s \right]=\int_{[x_0,x_1]} G_{x_0,x_1}(x,y)f(y)\;{\mathrm{d}m}(y),
\quad \text{where} \quad
G_{x_0,x_1}(x,y) \coloneqq 2\dfrac{(x \wedge y-x_0)(x_1-x \vee y)}{x_1-x_0}
$$ for $x,y \in (x_0,x_1)$. In particular, we obtain $\mathbb{E}_x [ \tau_{(x_0,x_1)} ] < \infty$.
2. We have $$\begin{aligned}
\mathbb{P}_x\left(X_{\tau_{(x_0,x_1)}} =x_0\right)=\dfrac{x_1-x}{x_1-x_0}
\quad \text{and} \quad
\mathbb{P}_x\left(X_{\tau_{(x_0,x_1)}} =x_1\right)=\dfrac{x-x_0}{x_1-x_0}.\end{aligned}$$
3. We have for $x<x_1$ with $x,x_1 \in {\operatorname{supp}}(m)$ $$\begin{aligned}
\mathbb{E}_x\left[ \int_0^{\tau_{(-\infty,x_1]}}f(X_s)ds\right]=\int_{(-\infty,x_1]}2(x_1-x \vee y)f(y)\;{\mathrm{d}m}(y)\end{aligned}$$ and for $x_2 \in supp(m)$ such that $x_2<x$ $$\begin{aligned}
\mathbb{E}_x\left[ \int_0^{\tau_{(x_2,\infty)}}f(X_s)ds\right]=\int_{[x_2,\infty)}2(x \wedge y-x_2)f(y)\;{\mathrm{d}m}(y).\end{aligned}$$
The proof for (i) and (ii) one can find in [@Burkhardt1983], p. 42, Lemma 2.4.5. The proof of (iii) follows in a similar way as for (ii) taking into account that the following hold for $a<x,y<b$ $$\mathbb{E}_x[ L_y^{\tau_{(-\infty,a)}} ]=2(y \wedge x-a)
\quad \text{and} \quad
\mathbb{E}_x[ L_y^{\tau_{(b,\infty)}} ]=2(b-y \vee x).
\qedhere$$
Now, we will give the connection between generalised Kreĭn-operators $D_{\nu}D_{\mu}$ and the infinitesimal generator of transformed gap diffusions. We consider, Borel measures $\mu,\nu$ with compact support and $
\nu( \{0,1\})=0$ and ${\operatorname{supp}}(\nu) \subseteq {\operatorname{supp}}(\mu) \subset [0,1]$. In the case if $F_{\mu}$ is strictly increasing see [@Burkhardt1983] p. 64-65. The only difference to the calculation in [@Burkhardt1983] is the replacement of the inverse with the pseudo-inverse and ${\operatorname{supp}}(\mu) \neq [0,1]$ if $F_{\mu}$ is not strictly increasing. Now, let $(X_t)_{ t \geq 0}$ a Liouville Brownian motion with speed measure $\nu \circ F_{\mu}^{-1}$. We call $Y_t \coloneqq \check{F}_{\mu}^{-1}(X_t)$ (generalised) $\nu$-$\mu$-Liouville Brownian motion with speed measure $\nu \circ F_{\mu}^{-1}$ which is again strong Markov process thus it is a deterministic transformation of $(X_t)_{t \geq 0}$.
Let $(Z_t)_{t \geq 0}$ denote a Markov process with compact state space $E$ and let $C(E)$ denote the continuous functions on $E$. A function $f \in C(E)$ is said to belong the domain $\mathcal{D}(A)$ of the infinitesimal generator of $(Z_t)_{t \geq 0}$ if the limit $$\begin{aligned}
Af(x)=\lim_{t \downarrow 0} ( \mathbb{E}_x[f(Z_t) ]-f(x))/t
\end{aligned}$$ exists in $C(E)$ with respect to $\lVert \cdot \rVert_{\infty}$.
Now, we compute the infinitesimal generator of $(Y_t)_{t \geq 0}$.
\[GapDSec\] Let $(X_t)_{ t \geq 0}$ a $\nu \circ F_{\mu}^{-1}$-Liouville Brownian motion and let $A$ denote the infinitesimal generator of $(Y_t)_{t \geq 0}=(\check{F}^{-1}_{\mu}(X_t))_{t \geq 0}$. For $f \in \mathcal{D}(A)$ there exists a continuous continuation of $f$ in $C_{\nu,\mu}$ (also denoted by $f$) such that $$\begin{aligned}
\label{SecondOrder}
f(x)=f(0)
+\int_{0}^x (F_{\mu}(x)-F_{\mu}(y) )2Af(y)d\nu(y), \; \; x \in \R, \end{aligned}$$ and the Neumann boundary condition $\nabla_{\mu} f(0)=\nabla_{\mu} f(1)=0$ are satisfied.\
We will present two proofs. The first proof uses and a result of [@Burkhardt1983]. The second proof will make use of Lemma \[PotentialGapD\].
Let $A_Y$ the infintesimale generator of $(Y_t)_{t \geq 0}$ on $C_{\nu,\mu}$ and $A_X$ the infinitesimal generator of $(X_t)_{t \geq 0}$ on $C_{\nu \circ F_{\mu}^{-1},\Lambda}$. At first note that we have for $f \circ \check{F}_{\mu}^{-1} \in \mathcal{D}(A_X)$ and $x \in F_{\mu}( {\operatorname{supp}}(\nu))$ $$\begin{aligned}
\lim_{t \downarrow 0} ( \mathbb{E}_x[ f( \check{F}_{\mu}^{-1}(X_t))-f(\check{F}^{-1}_{\mu}(x)) ] )/t
=\lim_{t \downarrow 0} ( \mathbb{E}_{\check{F}_{\mu}^{-1}(x)}[ f (Y_t) -f(\check{F}^{-1}_{\mu}(x))] )/t
=\lim_{t \downarrow 0} ( \mathbb{E}_{\check{F}_{\mu}^{-1}(x)}[ f (Y_t) -f(Y_0)] )/t\end{aligned}$$ From this, it follows that $$\begin{aligned}
f \circ \check{F}_{\mu}^{-1} \in \mathcal{D}(A_X) \rightarrow
f \in \mathcal{D}(A_Y)
\quad \text{and} \quad
A_X(f \circ \check{F}_{\mu}^{-1})=A_Y ( f) \circ \check{F}_{\mu}^{-1}.\end{aligned}$$ From [@Burkhardt1983 p.49], it is known that $2A_X(f)=\Delta_{\nu \circ F_{\mu}^{-1}, \Lambda}(f), \ f \in \mathcal{D}(A_X)$ with $\nabla_{\mu} f(0)=\nabla_{\mu} f(1)=0$. This in tandem with and the fact that $A_y(f) \in C_{\nu,\mu}$ yields $$\Delta_{\nu,\mu}(f)=
\Delta_{\nu \circ F_{\mu}^{-1}, \Lambda}(f \circ \check{F}_{\mu}^{-1})\circ F_{\mu}=2A_Y(f)
\quad \text{and} \quad
\nabla_{\mu} f \circ \check{F}_{\mu}^{-1}(0)=\nabla_{\mu} f \circ \check{F}_{\mu}^{-1}(1)=0.
\qedhere$$
First, note that ${\operatorname{supp}}(\nu \circ F_{\mu}^{-1})=F_{\mu}({\operatorname{supp}}(\nu))$. Therefore, we have for all $t \geq 0$ $ \check{F}^{-1}_{\mu}(X_t) \in \check{F}^{-1}_{\mu}(F_{\mu}({\operatorname{supp}}(\nu)) ) $ almost everywhere by . We will show that holds on $ \check{F}^{-1}_{\mu}(F_{\mu}({\operatorname{supp}}(\nu)) )$ which determines $f$ uniquely. To see this note that by we have $\nu( \check{F}^{-1}_{\mu}(F_{\mu}({\operatorname{supp}}(\nu)) )\cap {\operatorname{supp}}(\nu))=1$ proving that the set $\check{F}^{-1}_{\mu}(F_{\mu}({\operatorname{supp}}(\nu)) )$ intersects ${\operatorname{supp}}(\nu)$ densely. Let $x_1 \in \check{F}^{-1}_{\mu}(F_{\mu}({\operatorname{supp}}(\nu)) )\cap {\operatorname{supp}}(\nu)$. Clearly, $\check{F}^{-1}_{\mu}(F_{\mu}(x_{1}))=x_{1}$. Applying (iii) and yields $$\begin{aligned}
\mathbb{E}_{F_{\mu}(0^+)}\left[ \int_{0}^{\tau_{(-\infty,F_{\mu}(x_1))}} 2Af\circ \check{F}_{\mu}^{-1}(X_s)\;{\mathrm{d}}s \right]
&=\int_{[F_{\mu}(0^+),F_{\mu}(x_1)]}(F_{\mu}({x_1})-y ) 2Af \circ \check{F}_{\mu}^{-1}(y)\;{\mathrm{d}}\nu \circ F_{\mu}^{-1}(y) \\
&=\int_{[0,x_1]}(F_{\mu}(x_1)-F_{\mu}(y)) 2Af(y)\;{\mathrm{d}}\nu (y).\end{aligned}$$ On other site applying Dynkin’s formula ([@Revuz2013], p. 284 Proposition 1.5) and using the fact that $Af$ is bounded and $\tau_{(-\infty,F_{\mu}(x_1))}$ is integrable yields $$\begin{aligned}
\mathbb{E}_{F_{\mu}(0^+)}\left[ \int_{0}^{\tau_{(-\infty,F_{\mu}(x_2))}} 2Af\circ \check{F}_{\mu}^{-1}(X_s) \;{\mathrm{d}}s \right]
=\mathbb{E}_{F_{\mu}(x_0)}[ f (\check{F}_{\mu}^{-1}(X_{\tau_{(-\infty,F_{\mu}(x_1))}}) )]-f( \check{F}_{\mu}^{-1}(F_{\mu}(0^+)))
=f(x_1)-f( \check{F}_{\mu}^{-1}(F_{\mu}(0^+))).\end{aligned}$$ Therefore, we have $$\begin{aligned}
\label{Neumann}
f(x_1)=f( \check{F}_{\mu}^{-1}(F_{\mu}(0^+)))+\int_{[0,x_1]}(F_{\mu}(x_1)-F_{\mu}(y)) 2Af(y)\;{\mathrm{d}}\nu (y).\end{aligned}$$ This combined with $f(0) = f (\check{F}^{-1}_{\mu} (F_{\mu}(0^+ )))$ shows for all $x\in {\operatorname{supp}}(\nu)$ and implies $\nabla_{\mu}f(0)=0$. In the same way applying the second part of Lemma \[PotentialGapD\] (iii) for $x=1^-$ and $x_2<1$ gives $$f(x_1)-f(1)=\int_{[x_2,1]} 2 ( F_{\mu}(1)-F_{\mu}(y))Af(y) d \nu(y).$$ Therefore, it follows $\nabla_{\mu}f(1)=0$.
In following we consider the case $\mu=\nu$. The occupation formula of the local time yields for $ t \geq 0$ $$\begin{aligned}
\Phi_t=\int L^{t}_x\;{\mathrm{d}}\nu \circ F_{\mu}^{-1}(x)=
\int L^{t}_x \;{\mathrm{d}}\Lambda(x)
=\int_{0}^t \textbf{1}_{[0,1]}(B_s) \;{\mathrm{d}}\Lambda(s).\end{aligned}$$ Now, for the case that $\mu$ is the $1/3$-Cantor, we demonstrate in Fig. \[fig\] a simulation of a $\mu$-$\mu$-Liouville Brownian path. First we need a simulation of a standard Brownian path $B_{t}$ as depicted in Fig. \[fig\](), the corresponding time-change function $\Phi_t$ with respect to $\Lambda$, see Fig. \[fig\](), and the associated $\Lambda$-Liouville Brownian path $B_{\check{\Phi}^{-1}_{t}}$ as shown in Fig. \[fig\](). Finally, its image under $\check{F}_{\mu}^{-1}$ then shows a realisation of a $\mu$-$\mu$-Liouville Brownian path as shown in Fig. \[fig\]().
[0.49]{}
[0.49]{}
\
[0.49]{}
[0.49]{}
The above theorem allows us to consider the set of (continuous) elements $f$ from $\mathcal{D}(2A)$. With this at hand we observe the following.
$\mathcal{D}(A)=\mathcal{D}^s_{\pi/2, \pi/2}( \Delta_{\nu,\mu})$
We have $\mathcal{D}(A) \subseteq \mathcal{D}^s_{\pi/2, \pi/2 }( \Delta_{\nu,\mu} )$ and $2Af= \Delta_{\nu,\mu}f$ for all $f \in \mathcal{D}(A)$ by . The operator $\Delta_{\nu,\mu}$ restricted to $ \mathcal{D}^s_{\pi/2, \pi/2 }( \Delta_{\nu,\mu} )$ is an infinitesimal generator of a Feller process (). Therefore, [@Revuz2013 Exercise 1.18] shows that the transition functions of both processes are equal.
With we are able to give easy condition to calculate the walk dimension (under some additional assumptions).
\[ThmWalkDim\] Let $(X_s)_{s \geq 0}$ a $m$-Liouville Brownian motion starting in $x \in {\operatorname{supp}}(m)$. If, for $R(r)\coloneqq \inf\{s\geq r\colon s\in {\operatorname{supp}}(m)\}$ and $L(r)\coloneqq \inf\{s\geq r\colon -s\in {\operatorname{supp}}(m)\}$, we have $$\begin{aligned}
\lim_{r \to \infty}\frac{\log R(r)}{\log r} =\lim_{r \to \infty}\frac{\log L(r)}{\log r}=1,\end{aligned}$$ then, for $t\geq 0$, $$\begin{aligned}
\label{Condition1WD}
d_W(x)=1+t\iff \lim_{r \to \infty} \frac{\log{m((-r,r))} }{\log r}=t.\end{aligned}$$ In particular, if the walk dimension exists, then it is independent of the starting point in ${\operatorname{supp}}(m)$.
Set $S(r)\coloneqq R(r+x)\wedge L(r-x) $ and $T(r)\coloneqq R(r+x)\vee L(r-x)$. Using and estimating the Green kernel gives $$\begin{aligned}
(S(r)/2) m((-S(r)/2,S(r)/2)) \leq
\mathbb{E}_x [\tau_{(x-r,x+r)} ]
\leq {T(r)} m((-T(r),T(r))).\end{aligned}$$ Taking logarithms, using our hypothesis, and dividing by $\log(r)$ then proves the equivalence.
These ideas would also carry over to calculate the local walk dimension.
Assume $m$ is a finite and non-zero measure on $[0,1)$, let ${\widetilde}{m} = m \star \delta_{\mathbb{Z}}$ and let $(X_t)_{t \geq 0}$ denote a Liouville Brownian motion with speed measure ${\widetilde}{m}$. For all $x \in {\operatorname{supp}}( {\widetilde}{m})$, we have $d_W(x)=2$.
Since $ {\widetilde}{m}((-r+x,r+x)) =2r\cdot m([0,1))+O(1)$ and $L(r-x)-r,R(r+x)-r\in [0, 1]$, the claim is a consequence of with $t=1$.
Let $\nu$ be a Borel measure on $\R$ and $ F_{\mu}$ is a non-decreasing and continuous function on $\R$ with associated Lebesgue-Stieltjes measure $\mu$ such that ${\operatorname{supp}}(\nu) \subset {\operatorname{supp}}(\mu)$. Denote by $(X_t)_{t \geq 0}$ the $m$-Liouville Brownian motion with $m \coloneqq \nu \circ F_{\mu}^{-1}$ . Let $d_W$ denote the walk dimension of $(X_t)_{t \geq 0}$ and $\check{d}_W$ denote the walk dimension of $(\check{F}_{\mu}^{-1}(X_t))_{t \geq 0}$. For $x \in {\operatorname{supp}}(m)$, if $$\begin{aligned}
\forall r \in \mathbb{R} :
F_{\mu}(-r)=-F_{\mu}(r), \
a \coloneqq \lim_{r \rightarrow \infty} \dfrac{\log F_{\mu}(r) }{\log r}>0 \quad \mbox{ and }\quad
\lim_{r \rightarrow \infty} \dfrac{\log R(r) }{\log r}=\lim_{r \rightarrow \infty} \dfrac{\log L(r) }{\log r}=1,\end{aligned}$$ then $$\begin{aligned}
\lim_{r \rightarrow \infty} \frac{\log{m(\left(-r,r\right))} }{\log r}=t
\iff
1+t=d_W\left(x\right)=
a^{-1} \cdot \check{d}_W(\check{F}_{\mu}^{-1}(x)).\end{aligned}$$
Let $x \in \check{F}_{\mu}^{-1}\left({\operatorname{supp}}(m)\right)$ and, for $r>0$, define $$R^{\check{F}_{\mu}^{-1}}(r)
\coloneqq \inf\{ s \geq r \colon s \in \check{F}_{\mu}^{-1}({\operatorname{supp}}(m))\}
\quad \text{and} \quad
L^{\check{F}_{\mu}^{-1}}(r)\coloneqq \inf\{ s \geq r \colon -s \in \check{F}_{\mu}^{-1}({\operatorname{supp}}(m)) \}.$$ By assumption $(F_{\mu}(-r)=-F_{\mu}(r))$ and so $$\begin{aligned}
F_{\mu}(R^{\check{F}_{\mu}^{-1}}(r+x))
= R(F_{\mu}(x+r)), \ F_{\mu}(L^{\check{F}_{\mu}^{-1}}(r-x))
= L(F_{\mu}(r-x)),
\end{aligned}$$ therefore we have $$\begin{aligned}
\inf\{ t \geq 0 \colon \check{F}_{\mu}^{-1}\left(X_t \right) \in \{ R^{\check{F}^{-1}}(r+x),L^{\check{F}^{-1}}(r-x)\}
&= \inf\{ t \geq 0 \colon X_t \in \{ R(F_{\mu}(r+x)),L(F_{\mu}(r-x))\}\end{aligned}$$ Define for $r>0$ large $$m(r,x):=\min \{F_{\mu}(r+x)-F_{\mu}(x),F_{\mu}(r-x)-F_{\mu}(x)\}, \ M(r,x):=\max \{F_{\mu}(r+x)-F_{\mu}(x),F_{\mu}(r-x)-F_{\mu}(x)\},$$ then $$\begin{aligned}
\mathbb{E}_{F(x)}[
\tau_{(F_{\mu}(x)-m(r,x),F_{\mu}(x)+m(r,x) )} ] \leq \mathbb{E}_{F(x)}[ \tau_{(F_{\mu}(-r+x),F_{\mu}(r+x))}] \leq \mathbb{E}_{F_{\mu}(x)}[ \tau_{(F_{\mu}(x)-M(r,x),F_{\mu}(x)+M(r,x) )} ].\end{aligned}$$ The statement follows now with help of Theorem \[ThmWalkDim\] and the assumption on $F_{\mu}$.
[10]{}
P. Arzt. Measure theoretic trigonometric functions. , 2(2):115–169, 2015.
R. Beals and P. C. Greiner. Strings, waves, drums: spectra and inverse problems. , 7(2):131–183, 2009.
G Burkhardt. . PhD thesis, TU Dresden, 1983.
Alexander M. G. Cox, David Hobson, and Jan Obłój. Time-homogeneous diffusions with a given marginal at a random time. , 15:S11–S24, 2011.
E. B. Dynkin. , volume 122 of [ *Translated with the authorization and assistance of the author by J. Fabius, V. Greenberg, A. Maitra, G. Majone. Die Grundlehren der Mathematischen Wissenschaften, Bände 121*]{}. Academic Press Inc., Publishers, New York; Springer-Verlag, Berlin-Göttingen-Heidelberg, 1965.
T. Ehnes. . , 2019.
E. Ekström, D. Hobson, S. Janson, and J. Tysk. Can time-homogeneous diffusions produce any distribution? , 155(3-4):493–520, 2013.
W. Feller. Generalized second order differential operators and their lateral conditions. , 1:459–504, 1957.
U. Freiberg. Analytical properties of measure geometric [K]{}rein-[F]{}eller-operators on the real line. , 260:34–47, 2003.
U. Freiberg. A survey on measure geometric [L]{}aplacians on [C]{}antor like sets. , 28(1):189–198, 2003. Wavelet and fractal methods in science and engineering, Part I.
U. Freiberg. Spectral asymptotics of generalized measure geometric [L]{}aplacians on [C]{}antor like sets. , 17(1):87–104, 2005.
U. Freiberg and M. Zähle. Harmonic calculus on fractals—a measure geometric approach. [I]{}. , 16(3):265–277, 2002.
T. Fujita. A fractional dimension, self-similarity and a generalized diffusion operator. In [*Probabilistic methods in mathematical physics ([K]{}atata/[K]{}yoto, 1985)*]{}, pages 83–90. Academic Press, Boston, MA, 1987.
Alireza K. Golmankhaneh and Alexander S. Balankin. Sub- and super-diffusion on cantor sets: Beyond the paradox. , 382(14):960 – 967, 2018.
X. Jin. Spectral representation of one-dimensional [L]{}iouville brownian motion and [L]{}iouville brownian excursion, 2017.
I. S. Kac and M. G. Kre[ĭ]{}n. Criteria for the discreteness of the spectrum of a singular string. , 1958(2 (3)):136–153, 1958.
M. Kesseböhmer, T. Samuel, and H. Weyer. A note on measure-geometric [L]{}aplacians. , 181(3):643–655, 2016.
M. Kesseböhmer, T. Samuel, and H. Weyer. Measure-geometric [L]{}aplacians for discrete distributions. In Niemeyer, Pears, Rock, and Samuel, editors, [*Horizons of Fractal Geometry and Complex Dimensions*]{}, volume 731 of [*Contemp. Math.*]{}, chapter 7, pages 133–142. Amer. Math. Soc., Providence, R.I., 2019.
M. Kesseböhmer, T. Samuel, and H. Weyer. Measure-geometric [L]{}aplacians for partially atomic measures. , 2019.
S. Kotani and S. Watanabe. Kreĭn’s spectral theory of strings and generalized diffusion processes. In [*Functional analysis in [M]{}arkov processes ([K]{}atata/[K]{}yoto, 1981)*]{}, volume 923 of [*Lecture Notes in Math.*]{}, pages 235–259. Springer, Berlin-New York, 1982.
U. K[ü]{}chler. Some asymptotic properties of the transition densities of one-dimensional quasidiffusions. , 16(1):245–268, 1980.
H. Langer, L. Partzsch, and D. Schütze. Über verallgemeinerte gewöhnliche [D]{}ifferentialoperatoren mit nichtlokalen [R]{}andbedingungen und die von ihnen erzeugten [M]{}arkov-[P]{}rozesse. , 7:659–702, 1971/72.
Z.M. Ma, Z.M. Ma, and M. R[ö]{}ckner. . Universitext (1979). Springer-Verlag, 1992.
John M. Noble. Time homogeneous diffusions with a given marginal at a deterministic time. , 123(3):675–718, 2013.
D. Revuz and M. Yor. . Springer Science and Business Media, Berlin Heidelberg, third edition, 2013.
R. Rhodes and V. Vargas. Spectral dimension of [L]{}iouville quantum gravity. , 15(12):2281–2298, 2014.
[^1]:
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We have fabricated PbTiO$_{3}$/SrRuO$_{3}$ superlattices with ultra-thin SrRuO$_{3}$ layers. Due to the superlattice geometry, the samples show a large anisotropy in their electrical resistivity, which can be controlled by changing the thickness of the PbTiO$_{3}$ layers. Therefore, along the ferroelectric direction, SrRuO$_{3}$ layers can act as dielectric, rather than metallic, elements. We show that, by reducing the concentration of PbTiO$_{3}$, an increasingly important effect of polarization asymmetry due to compositional inversion symmetry breaking occurs. The results are significant as they represent a new class of ferroelectric superlattices, with a rich and complex phase diagram. By expanding our set of materials we are able to introduce new behaviors that can only occur when one of the materials is not a perovskite titanate. Here, compositional inversion symmetry breaking in bi-color superlattices, due to the combined variation of A and B site ions within the superlattice, is demonstrated using a combination of experimental measurements and first principles density functional theory.'
author:
- 'S.J. Callori'
- 'J. Gabel'
- Dong Su
- 'J. Sinsheimer'
- 'M.V. Fernandez-Serra'
- 'M. Dawber'
title: 'Ferroelectric PbTiO$_{3}$/SrRuO$_{3}$ superlattices with broken inversion symmetry'
---
Artificially layered perovskite oxide superlattices provide many opportunities to develop systems with novel and tunable properties[@Dawber08; @Zubko11]. As far as ferroelectric superlattices are concerned, the insulating titanium perovskite oxides (e.g. PbTiO$_{3}$, BaTiO$_{3}$, CaTiO$_{3}$ and SrTiO$_{3}$) have to date been the most popular “building blocks”, but the need for new functionalities, particularly related to magnetism, requires the use of a wider set of materials and a deep understanding of the new physical phenomenon related to interfaces. In this letter an unconventional approach is demonstrated: we use a material that is normally metallic to play the role of a dielectric, in a ferroelectric-dielectric superlattice.
The much-studied compound SrRuO$_{3}$ provides the proof of concept that metallic magnetic oxides can transform into thin-film dielectric components in certain heterostructures. In bulk, SrRuO$_{3}$ has the distorted perovskite orthorhombic Pnma structure, is metallic and is ferromagnetic below Tc = 160 K [@Allen96; @Singh96; @Mazin97; @Dodge00; @Capogna02]. It is also a commonly used electrode material for oxides, and the interface with ferroelectric oxides has been much studied [@Junquera03; @Gerra06; @Stengel06; @AguadoPuente08; @Shin10]. However, SrRuO$_{3}$ becomes insulating in layers of thickness less than 4 unit cells; this behavior has been observed in thin films [@Toyota05; @Schultz09; @Xia09] and in SrTiO$_{3}$/SrRuO$_{3}$ superlattices[@Izumi98; @Kumigashira08]. First-principles investigations [@Zayak06; @Rondinelli08; @Mahadevan10] and experiment [@Herranz03; @Kim05; @Moore07; @Grutter10; @Ziese10; @Choi10] indicate that epitaxial strain, size effects, chemical pressure, surface reconstruction and interaction with the substrate may all play an important role in the observed behavior. In a recent contribution Verissimo-Alves et al. [@VerissimoAlves12] showed from first principles calculations that a highly confined 2DEG is formed at the interface in SrTiO$_{3}$/SrRuO$_{3}$ superlattices. We will show in this letter that a similar effect occurs in PbTiO$_{3}$/SrRuO$_{3}$ superlattices, but that in the direction perpendicular to the interfaces PbTiO$_{3}$/SrRuO$_{3}$ superlattices containing single unit cell layers of SrRuO$_{3}$ are insulating and can be ferroelectric.
A second motivation for creating PbTiO$_{3}$/SrRuO$_{3}$ superlattices is that, as they have both A and B site variation, inversion symmetry can be compositionally broken [@Sai00; @Sai01]. As a result, an asymmetry is introduced in the ferroelectric double-well potential which can lead to “self-poling" materials. Self-poling ferroelectric materials are useful in piezoelectric applications where the desired mode of operation is to apply an electric field either with or against a fixed polarization direction to achieve, respectively, an expansion or contraction of the material. In a tri-color superlattice compositional breaking of inversion symmetry can occur with only A or B site variation. The effect has been seen with A site variation in tri-layer superlattices containing BaTiO$_{3}$, CaTiO$_{3}$ and SrTiO$_{3}$ [@Warusawithana03; @Lee05], and shown to be an appealing route towards magnetoelectric materials with tri-color variation on the B site[@Hatt07; @Yamada08], but until now this behavior has not been shown experimentally in bi-color superlattices.
Epitaxial growth of PbTiO$_{3}$/SrRuO$_{3}$ superlattices can be achieved on SrTiO$_3$ substrates as both PbTiO$_{3}$ and SrRuO$_{3}$ have in-plane lattice parameters close to that of SrTiO$_3$, which at room temperature is cubic with a=3.905 Å. At room temperature bulk PbTiO$_{3}$ is tetragonal (a=3.904 Å, c=4.15 Å) and orthorhombic SrRuO$_{3}$ can be considered as pseudo-cubic with a=3.93 Å. For this study, the n/1 PbTiO$_{3}$/SrRuO$_{3}$ (n unit cells (u.c.) PbTiO$_{3}$/1 u.c. SrRuO$_{3}$) superlattices were deposited using off-axis RF magnetron sputtering on (001) SrTiO$_3$ substrates, which had been treated with buffered HF and annealed to ensure TiO$_{2}$ termination. The SrRuO$_{3}$ layer thickness was grown to 1 u.c. for all the samples considered here, with the aim being that the SrRuO$_{3}$ layers should act as dielectrics, rather than metals. By contrast the thickness of the PbTiO$_{3}$ layer was changed from sample to sample, so that the relative effect of bulk ferroelectricity vs. interfacially driven compositional inversion symmetry breaking could be assessed. The total number of bilayers in the superlattice was varied from one sample to another so that the total thickness of each sample was in each case as close as possible to 100nm. Growth rates for the two materials within the superlattice were obtained from x-ray diffraction measurements performed on many preliminary samples. Bottom SrRuO$_{3}$ electrodes (20 nm in thickness) were deposited *in situ* for the samples used for electrical measurements and gold top electrodes were added to the samples post-deposition. The superlattices were grown at a temperature of 550$^{o}$ C and the SrRuO$_{3}$ electrodes were grown at 620$^{o}$ C.
Experimentally, it has been demonstrated by Rjinders et al [@Rijnders04] that when grown by pulsed laser deposition the termination of a SrRuO$_{3}$ film is affected by both deposition conditions and layer thickness and that as a film grows on Ti terminated SrTiO$_{3}$ there is a conversion from a RuO$_{2}$ to a SrO termination layer. At a growth temperature of 700$^{o}$C this transition already occurs for a single unit cell SrRuO$_{3}$ layer, but at lower deposition temperatures, this transition occurs later in the growth, as the RuO$_{2}$ layer is comparably more stable at these conditions. As a consequence of the small thickness of our SrRuO$_{3}$ layers, the low deposition temperatures used in our process, and the different kinetic regime of sputtering as compared to pulsed laser deposition, a RuO$_{2}$ termination of our SrRuO$_{3}$ layers, may still be possible.
The epitaxial growth of our samples was confirmed by x-ray diffraction and high resolution scanning transmission electron microscopy (HR STEM). Fig. \[TEM\](b) shows a HR STEM cross section of an 8/1 PbTiO$_{3}$/SrRuO$_{3}$ superlattice. The PbTiO$_{3}$ layers are the brightest because of the high atomic number of Pb, while the SrRuO$_{3}$ layers are less bright than PbTiO$_{3}$, but have enhanced brightness compared to SrTiO$_{3}$ (not shown) because the Ru ion has a higher atomic number than Ti. In addition to the STEM image shown, we also carried out STEM-EELS line scans [@SI], which support the ideality of our grown structures.
![*(a) A HR-STEM image of an 8/1 PbTiO$_3$/SrRuO$_3$ superlattice. (b) The three types of interfaces considered theoretically in this paper, illustrated for the case of a 3/1 PbTiO$_3$/SrRuO$_3$ superlattice.*[]{data-label="TEM"}](fig1.pdf){width="8.5cm"}
Although the interface most likely to form in PbTiO$_{3}$/SrRuO$_{3}$ superlattices if the atoms in the material are deposited in the same ratio as the parent targets breaks inversion symmetry it is possible to consider interfaces that might form which do not. In Fig. \[TEM\] (b) we show three kinds of interfaces which we have studied using first principles calculations, which we illustrate for the case of a 3/1 superlattice. The ideal structure which breaks inversion symmetry is the first from the left in Fig. \[TEM\] (b). In order to study the significance of the symmetry breaking effect, we also simulated two different kinds of interface which conserve symmetry, one in which one Pb-O plane has been replaced by a Sr-O plane, which is shown in the center of Fig. \[TEM\] (b), and another, (less likely due to the high volatility of Ru), unit cell in which a Ti-O plane has been replaced by a Ru-O plane is shown on the right.
We investigated the 3 kinds of interface shown above using first principles calculations. These were performed using density functional theory, using a basis of numerical atomic orbitals as implemented in the [siesta]{} code. We used the same basis set and pseudopotentials as Verissimo-Alves et al. [@VerissimoAlves12]. We studied the influence of spin polarization, the use of the generalized gradients approximation within the commonly used Wu-Cohen parametrization[@Wu06], and the effect of correlations within the LDA+U and LSDA+U approximations[@Sanvito04]. The different approximations used can affect the electronic properties of the metallic layer along the parallel direction. However the electrical anisotropy is mostly dependent on the superlattice periodicity, and therefore results are computed using the local density approximation (LDA), which is the best method to characterize both the structural and electronic properties of the superlattices. Full details regarding the calculations can be found in the supplemental materials to this letter[@SI]. We examined the electrical conductivity of the superlattices both in-plane ( $\sigma_{xx}$) and out-of-plane ($\sigma_{zz}$) by calculating the diagonal elements of the conductivity tensor within the relaxation time approximation to the Boltzman transport equation[@Allen88]. The anisotropy in $\sigma$ is fully determined by the anisotropy of the Fermi surface geometry, as determined by $\sigma_{\alpha\alpha}=-e^{2}\tau\sum_{k}v_{k\alpha}^{2}\delta\left(\epsilon_{F}-\epsilon_{k}\right)$. The relaxation time, $\tau$ is the only variational parameter in the expression, we choose $\tau=1.3\times10^{-14}$ s after experimental results in bulk SrRuO$_{3}$[@Santi97]. This approximation ignores the anisotropy of the electron-phonon scattering, although this is known to be a much smaller effect[@Allen88]. The results shown in Fig. \[DoubleWell\] (a) are for the ideal interfaces, but these quantities were also calculated for the two other cases and are qualitatively similar, again an indication that the anisotropy is almost fully determined by the inter- SrRuO$_{3}$ layer distance. We find, in agreement with ref. [@VerissimoAlves12], that the electrons in the single unit cell layers of SrRuO$_{3}$ are confined to that layer, so while the in-plane conductivity, $\sigma_{xx}$, of the structures does not change dramatically as the spacing between the layers is varied the out-of-plane conductivity, $\sigma_{zz}$, decreases exponentially with a characteristic length of 1.3$\mathrm{\AA}$ as the thickness of the PbTiO$_{3}$ layers is increased. In the two-component PbTiO$_{3}$/SrRuO$_{3}$ superlattices in which the inversion symmetry is broken by the ideal interface structure, our calculations predict a self-poling behavior. In Fig. \[DoubleWell\] (b) we show the energy of the superlattice as a function of the polarization. We computed the total energy of the system for atomic displacements along the line $\vec{r}=\vec{r}_{P_+} +u(\vec{r}_{P_-} -\vec{r}_{P_+})$, interpolating linearly between the two minima of the energy. These two minima are characterized by two different polarization states, $P_{up}$ (higher energy minimum) and $P_{down}$ (lower energy minimum). As shown in Table \[secondtable\], the simulations for the ideal interface show that, as the PbTiO$_{3}$ layer thickness is reduced, there is an increasingly large difference in the values of the stable up and down polarizations, until for the 5/1 superlattice, when the potential well has just one minimum, only the down polarization is stable.
![*(a) Calculated conductivity both in-plane and out of plane plotted a function of the number of unit cells of PbTiO$_{3}$ included in each bilayer (with zero corresponding with SrRuO$_{3}$ strained in-plane to the SrTiO$_{3}$ lattice parameter). (b) Calculated total energy per unit cell as a function of polarization.*[]{data-label="DoubleWell"}](fig2.pdf){width="8.5cm"}
Superlattice
---------------------------------------- ------------- ----------- ----------- ------
Sr Excess Ru excess
$P_ {down}$ $P_{up}$ $P$ $P$
(PbTiO$_{3}$)$_{5}$(SrRuO$_{3}$)$_{1}$ 35.8 unstable 2.0 23.3
(PbTiO$_{3}$)$_{6}$(SrRuO$_{3}$)$_{1}$ 39.2 16.8 16.4 32.5
(PbTiO$_{3}$)$_{7}$(SrRuO$_{3}$)$_{1}$ 45.0 45.1 27.2 44.2
(PbTiO$_{3}$)$_{9}$(SrRuO$_{3}$)$_{1}$ 50.9 50.2 42.9 52.7
: *Calculated stable polarization magnitudes (DFT LDA) in the up and down directions for a selection of ideal superlattices.*[]{data-label="secondtable"}
The preference of one polarization state over another for superlattices with broken compositional inversion symmetry was seen in all of the calculation schemes used. In the spin polarized calculations for samples with single unit cell layers of SrRuO$_{3}$ the spin polarization is not affected by the direction of the polarization. However, in simulations of superlattices with symmetry breaking interfaces that have SrRuO$_{3}$ layers thicker than one unit cell the magnetization is different for the two polarization directions. Although it is not the focus of the present paper, this finding demonstrates the potential for the compositional breaking of inversion symmetry at the PbTiO$_{3}$/SrRuO$_{3}$ interface to enable a form of coupling between magnetism and ferroelectricity.
Experimental values for the switched ferroelectric polarization of the samples were obtained from polarization-electric field hysteresis loops performed on a number of samples. Polarization hysteresis was observed in samples with a PbTiO$_{3}$ layer thickness of 5 u.c or greater. The experimentally measured polarization as a function of the total number of unit cells in each bilayer is shown in Fig. \[polfig\], along with characteristic loops measured at 3 different frequencies on the 7/1 shown as an inset. Successful hysteresis loops confirm that the thin layers of SrRuO$_{3}$ in the material are acting as dielectric layers, and allow a continuous polarization in the structure. An independent confirmation of ferroelectricity comes from x-ray diffraction reciprocal space maps around superlattice Bragg peaks, shown in the supplemental information[@SI]. These show diffuse scattering from the in-plane periodicity of stripe domains with polarization oriented up and down with respect to the substrate, and are similar to those seen in PbTiO$_{3}$/SrTiO$_{3}$ superlattices.[@Zubko10; @Jo11]. These domains features were observed in the 7/1, 9/1, and 13/1 superlattices, but not in the 5/1 superlattice. The lack of domain features in the 5/1 PbTiO$_{3}$/SrRuO$_{3}$ superlattice may be due to the relative instability of one polarization state with respect to each other, ie., while the polarization can be switched under field from one direction to the other, its equilibrium configuration is dominated by a single polarization direction. Direct comparison between the calculated stable polarizations in the layer and the experimentally measured switched polarizations are difficult to make, as the two quantities, while related, are not identical. However, it can be seen that the polarization values and their dependence on the number of unit cells in each bilayer are qualitatively similar for experiment and the 3 theoretical cases considered in Table \[secondtable\]. For an equivalent composition the Ru excess superlattices have higher polarizations than the Sr excess superlattices. In the ideal case marked differences in the polarization value for the up and down state become noticeable for superlattices whose bilayers contain less than 7 unit cells of PbTiO$_{3}$. When the PbTiO$_{3}$ layer thickness is 3 layers or less the samples become fairly conductive in the out of plane direction, both in theory and experiment.
![*Switched polarization plotted as a function of the total number of unit cells in each bilayer. Inset: Polarization-field hysteresis loops measured at 3 different frequencies on a 7/1 sample.* []{data-label="polfig"}](fig3.pdf){width="8.5cm"}
An indirect probe of the stable polarization is the average tetragonality (c/a) of ferroelectric superlattices, which can be measured using x-ray diffraction [@DawberPRL05; @DawberAM07; @Bousquet08]. Our experimental measurements and first principles results from DFT LDA calculations are shown in Fig. \[cafig\]. As with the polarization values in Fig. \[polfig\], we have plotted the results in terms of the total number of unit cells per bilayer, and included for each set of values the nominal composition. However, the Sr and Ru excess structures deviate from these ideal structures as described earlier; for the former one Pb-O plane has been replaced by a Sr-O plane, and for the latter a Ti-O plane has been replaced by a Ru-O plane. It is difficult to make a definite conclusion which interface is present in our experimental samples solely from comparing the experimental results with the theoretical predictions shown. However, we suggest that a comparison of the data over the whole range of the plot would tend to exclude the Sr excess interface, and although the Ru excess interface matches the experimental data relatively well, this interface is unlikely to occur in experiment due to the high volatility of Ru. An interesting point in this figure is that the average tetragonality begins to increase again as the number of unit cells in each bilayer is decreased. This is because, in contrast to PbTiO$_{3}$/SrTiO$_{3}$ where SrTiO$_{3}$ grown on SrTiO$_{3}$ has a tetragonality of 1 and the c/a montonically decreases, as the amount of PbTiO$_{3}$ decreases, SrRuO$_{3}$ grown on SrTiO3 has a tetragonality of 1.03. It appears that the c axis lattice parameter of paraelectric PbTiO$_{3}$ in these superlattices is quite low, both in experiment and theory and is certainly well below the value of approximately 1.03 that is predicted by Landau theory[@DawberAM07]. The upturn in the c/a value occurs for the Ru excess samples first as these are, for each sample, essentially half a unit cell closer to being SrRuO$_{3}$ than the corresponding ideal superlattice.
![*Average tetragonality, or the ratio of the average out of plane lattice parameter $c$ to the lattice parameter of the SrTiO$_{3}$ substrate $a$, (i.e. $c/a$), plotted as a function of the total number of unit cells in each bilayer. As well as the experimental results results from DFT LDA calculations for three different kinds of interfaces are shown.*[]{data-label="cafig"}](fig4.pdf){width="8.5cm"}
The effect that the compositional breaking of inversion symmetry has on functional properties is most evident in the dielectric response of the samples. In Fig. \[ptosrodielectricexpt\], we show dielectric constant measured as a function of electric field (at a frequency of 10kHz). The measurements show an evolution from a typical butterfly loop for PbTiO$_{3}$ rich samples (13/1 and 9/1 samples) to a highly asymmetric curve with greatly enhanced peak dielectric constant for the 3/1 sample. An unusual characteristic where two peaks are seen on each voltage trace displayed in the 5/1 sample. The composition at which the transition from conventional ferroelectric behavior occurs (approx. 7/1) matches the composition highlighted by theory at which compositional inversion symmetry breaking becomes a dominant factor. To summarize our picture of this system: The 3/1 sample can be characterized as a spontaneously polarized non-switchable insulator (or, in other words, an interfacially driven pyroelectric). Samples between 7/1 and 3/1 are best described as asymmetric ferroelectrics. Samples with PbTiO$_{3}$ layers thicker than 7 unit cells, while still containing interfaces which compositionally break inversion symmetry are not greatly affected by them.
![*Dielectric constant-field loops for 5 of our samples.*[]{data-label="ptosrodielectricexpt"}](fig5.pdf){width="8.5cm"}
Besides the direct results presented here, our study offers a general demonstration of the possibilities unlocked by expanding the material set used in ferroelectric superlattices. Our findings should motivate a broader exploration of candidate materials for the development of new artificially layered ferroelectrics. In particular, besides the self-poling behavior that compositional broken inversion symmetry produces, the use of thin metallic materials as dielectric layers has intriguing potential for the development of highly coupled multiferroics.
We acknowledge very useful discussions with Phil Allen. MD acknowledges support from the National Science Foundation under DMR1055413. MD and DS acknowledge support from a SBU/BNL seed grant. MVFS acknowledges support from DOE award DEFG02-09ER16052. Use of the Center for Functional Nanomaterials, at Brookhaven National Laboratory, was supported by U.S. Department of Energy, Office of Basic Energy Sciences, under Contract No. DE-AC02-98CH10886.
[99]{}
M. Dawber, N. Stucki, C. Lichtensteiger, S. Gariglio and J.-M. Triscone, *J. Phys. Condens. Matter* **20** 264015 (2008).
P. Zubko, S. Gariglio, M. Gabay, P. Ghosez and J.-M. Triscone, *Annu. Rev. Condens. Matter. Phys.*, **2** 141 (2011).
P.B. Allen, H. Berger, O. Chauvet, L. Forro, T. Jarlborg, A. Junod, B. Revaz and G. Santi, *Phys. Rev. B*, **53** 4393 (1996).
D. J. Singh, *J. Appl. Phys.* **79**, 4818 (1996).
I. I. Mazin and D.J. Singh, *Phys. Rev. B* **56**, 2556 (1997).
J.S. Dodge, C.P. Weber, J. Corson, J. Orenstein, Z. Schlesinger, J.W. Reiner and M.R. Beasley, *Phys. Rev. Lett.* **85**, 4932 (2000).
L. Capogna, A. P. Mackenzie, R.S. Perry, S.A. Grigera, L.M. Galvin, P. Raychaudhuri, A.J. Schofield, C.S. Alexander, G. Cao, S.R. Julian and Y. Maeno, *Phys. Rev. Lett.* **88**, 076602 (2002).
J. Junquera and P. Ghosez, *Nature* **422**, 506 (2003).
G. Gerra, A. K. Tagantsev, N. Setter, and K. Parlinski, *Phys. Rev. Lett.* **96** 107603 (2006).
M. Stengel and N. A. Spaldin, *Nature* **443**, 679 (2006).
P. Aguado-Puente, J. Junquera, *Phys. Rev. Lett.* **100** 177601 (2008).
J. Shin, A. Borisevich, V. Meunier, J. Zhou, E. W. Plummer, S.V. Kalinin and A.P. Baddorf, *ACS Nano* **4**, 4190 (2010).
D. Toyota, I. Ohkubo, H. Kumigashira, M. Oshima, T. Ohnishi, M. Lippmaa, M. Kawasaki, and H. Koinuma, *Appl. Phys. Lett.* **87**, 162508 (2005), *J. Appl. Phys* **99** 08N505 (2006).
M. Schultz, S. Levy, J.W. Reiner and L. Klein, *Phys. Rev. B* **79** 125444 (2009).
J. Xia, W. Siemons, G. Koster, M.R. Beasley and A. Kapitulnik, *Phys. Rev. B* **79**, 140407(R) (2009).
M. Izumi, K. Nakazawa and Y. Bando, *J. Phys. Soc. Japan* **67** 651 (1998).
H. Kumigashira, M. Minohara, M. Takizawa, A. Fujimori, D. Toyota, I. Ohkubo, M. Oshima, M. Lippmaa, and M. Kawasaki, *Appl. Phys. Lett.* **92** 122105 (2008).
A. T. Zayak, X. Huang, J. B. Neaton, and Karin M. Rabe, *Phys. Rev. B* **74** 094104 (2006), *Phys. Rev. B* **77**, 214410 (2008) .
J. M. Rondinelli, N. M. Caffrey, S. Sanvito and N. A. Spaldin, *Phys. Rev. B* **78** 155107 (2008).
P. Mahadevan, F. Aryasetiawan, A. Janotti, T. Sasaki , *Phys. Rev. B* **80**, 035106 (2009).
G. Herranz, B. Martinez, J. Fontcuberta, F. Sanchez, C. Ferrater, M.V. Garcia-Cuenca, M Varela *Phys. Rev. B* **67** 174423 (2003).
K.W. Kim, J.S. Lee, T. W. Noh, S.R. Lee and K. Char, *Phys. Rev. B* **71** 125104 (2005).
R. G. Moore, Jiandi Zhang, V. B. Nascimento, R. Jin, Jiandong Guo, G.T. Wang, Z. Fang, D. Mandrus, E. W. Plummer, *Science* **26** 318, 615 (2007).
A. Grutter, F. Wong, E. Arenholz, M. Liberati, and Y. Sukuki, *J. Appl. Phys.* **107** 09E138 (2010).
M. Ziese, I. Vrejoiu and D. Hesse, *Phys. Rev. B* **81** 184418 (2010).
K. J. Choi, S. H. Baek, H. W. Jang, L. J. Belenky, M. Lyubchenko, and C.-B. Eom, *Adv. Mater.* **22** 759762 (2010).
M.Verissimo-Alves, P. Garcia-Fernandez, D.I. Bilc, Ph. Ghosez, and J. Junquera, *Phys. Rev. Lett.* **108** 107003 (2012).
N. Sai, B. Meyer, and D. Vanderbilt, *Phys. Rev. Lett.*, **84**, 5636 (2000).
N. Sai, B. Meyer, and D. Vanderbilt, *Fundamental Physics of Ferroelectrics*, ****, 218 (2001).
M.P. Warusawithana, E.V. Colla, J.N. Eckstein, and M. B. Weissman *Phys. Rev. Lett.* **90** 036802 (2003).
H.N. Lee, H.M. Christen, M.F. Chisholm, C.M. Rouleau and D.H. Lowndes, *Nature* **433**, 395 (2005).
A.J. Hatt and N.A. Spaldin *Appl. Phys. Lett.* **90** 242916 (2007).
H. Yamada, H. Sato, H. Akoh, N. Kida, T. Arima, M. Kawasaki and Y. Tokura *Appl. Phys. Lett.*
G. Rijnders, D.H.A. Blank, J. Choi, and C.B. Eom, *Appl. Phys. Lett.* **92** 062508 (2008). **84**, 505 (2004).
Details in Supplementary Materials.
J.M. Soler, E. Artacho, J.D. Gale, A. García, J. Junquera, P. Ordejón, and D. Sánchez-Portal, *J. Phys. Condens. Matter* **14** 2745 (2002).
J. Junquera, M. Zimmer, P. Ordejón, Ph. Ghosez, *Phys. Rev. B* **67** 155327 (2003).
Z. Wu and R.E. Cohen, *Phys. Rev. B.* **73** 235116 (2006).
M. Wierzbowska, D. Sanchez-Portal, and S. Sanvito. *Phys. Rev. B* **70** 235209 (2004).
P. B. Allen, W. E. Pickett and H. Krakauer *Phys. Rev. B* **37** 7482 (1988).
G. Santi and T. Jarlborg. *J . Phys. Condens. Mat.* **9** 956 (1997).
P. Zubko, N. Stucki, C. Lichtensteiger, and J.-M. Triscone, *Phys. Rev. Lett*, **104**, 187601 (2010).
J.Y. Jo, P.Chen, R.J. Sichel, S.J. Callori, J. Sinsheimer, E.M. Dufresne, M. Dawber and P.G. Evans, *Phys. Rev. Lett*, **107** 055501 (2011).
M. Dawber, C. Lichtensteiger, M. Cantoni, M. Veithen, P. Ghosez, K. Johnston, K.M. Rabe, and J.-M. Triscone, *Phys. Rev. Lett.*, **95**, 177601 (2005).
M. Dawber, N. Stucki, C. Lichtensteiger, S. Gariglio, P. Ghosez and J.-M. Triscone, *Advanced Materials*, **19**, 4153 (2007).
E. Bousquet, M. Dawber, N. Stucki, C. Lichtensteiger, P. Hermet, S.Gariglio, J.-M. Triscone, and P. Ghosez, *Nature*, **452**, 732 (2008).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study probabilistic single-item second-price auctions where the item is characterized by a set of attributes. The auctioneer knows the actual instantiation of all the attributes, but he may choose to reveal only a subset of these attributes to the bidders. Our model is an abstraction of the following Ad auction scenario. The website (auctioneer) knows the demographic information of its impressions, and this information is in terms of a list of attributes (e.g., age, gender, country of location). The website may hide certain attributes from its advertisers (bidders) in order to create thicker market, which may lead to higher revenue. We study how to hide attributes in an optimal way. We show that it is NP-hard to solve for the optimal attribute hiding scheme. We then derive a polynomial-time solvable upper bound on the optimal revenue. Finally, we propose two heuristic-based attribute hiding schemes. Experiments show that revenue achieved by these schemes is close to the upper bound.'
author:
- |
Mingyu Guo\
Department of Computer Science\
University of Liverpool, UK\
[email protected] Argyrios Deligkas\
Department of Computer Science\
University of Liverpool, UK\
[email protected]
bibliography:
- '../mg.bib'
title: Revenue Maximization via Hiding Item Attributes
---
Introduction
============
One advantage of Internet advertising is that it offers advertisers the ability to target customers based on various traits such as demographics. [@Even-Dar07:Sponsored] showed that, for sponsored search of a given keyword, instead of running a single auction for the keyword, we can split the whole auction into many separate auctions based on visitors/impressions’ [*contexts*]{} (e.g., demographics). For example, if we know and only know the visitors’ locations, then each location defines a context. In this example scenario, splitting based on context means separate auction for each location. Splitting based on context increases the advertisers’ welfare. The explanation is simple: after splitting, advertisers can tailor their bids to the context. As a result, advertisers generally only win (impressions from) visitors that they aim to target, and the payments are also lower, since advertisers only face competition from those targeting similar visitors. On the other hand, splitting may reduce the revenue received by the auctioneer (publisher, e.g., website) due to the [*thin market problem*]{}: there may be few competitors for some contexts. Actually, if for every context, there is only one advertiser interested in it, then the total revenue is $0$ under the standard second-price auction.
[@Ghosh07:Computing] observed that having a single auction for all contexts and having separate auction for each context are not the only two options. There are other ways to split based on context, and it may lead to much higher revenue. The idea explored in [@Ghosh07:Computing] is to [*cluster*]{} the contexts into bundles, and run separate auction for each bundle. For example, suppose there are three different contexts: Beijing, Chicago, and London (assuming the only contextual information is the location and visitors are only from these three cities). We can have one auction for the bundle Beijing and Chicago (and a second auction for London only). The interpretation (due to [@Emek12:Signaling]) is that if a visitor is from Beijing or Chicago, then the auctioneer informs the advertisers that the impression is from one of these two cities, [*but not exactly which*]{}. When this happens, both advertisers targeting Beijing and advertisers targeting Chicago will compete in the auction. Their bids depend on how much they value impressions from Beijing and Chicago, respectively. Their bids also depend on the conditional probability that the impression is from Beijing (or Chicago) given that the impression is from one of these two cities.
To put it more formally, [@Ghosh07:Computing] studied probabilistic single-item second-price auctions (again, interpretation due to [@Emek12:Signaling]). In such an auction, there is only one item for sale under a second-price auction, but the item has different possible [*instantiations*]{}. The auctioneer knows the actual instantiation but the bidders do not. The auctioneer may choose to hide certain information from the bidders if this increases the revenue. The probabilistic single-item second-price auction model is an abstraction of the following Ad auction scenario. We have a website that sells one advertisement slot. That is, there is only one item – the only advertisement slot, but the item takes many possible instantiations, due to the fact that visitors/impressions have different demographic profiles. The auctioneer knows every visitor’s demographic profile, and he may hide certain information from the advertisers. As mentioned above, [@Ghosh07:Computing] considered hiding information by [*clustering*]{}: the auctioneer tells the bidders that the actual instantiation is among several instantiations. [@Emek12:Signaling; @Miltersen12:Send] studied the exact same model and went one step further. These two papers studied hiding information by [*signaling*]{}: the auctioneer sends out different signals, and the bidders infer the probability distribution of the actual instantiation, based on the signal received. It is easy to see that signaling is more general than clustering. Interestingly, for full information settings (settings where the auctioneer knows the bidders’ exact valuations), [@Ghosh07:Computing] showed that it is NP-hard to solve for the optimal clustering scheme (optimal in terms of revenue). On the other hand, [@Emek12:Signaling; @Miltersen12:Send] both independently showed that, under the same full information assumption, it takes only polynomial time to solve for the optimal signaling scheme. This is mostly due to the fact that instantiations are treated as divisible goods under signaling schemes,
In this paper, we continue the study of revenue-maximizing probabilistic single-item second-price auctions. We observe that in practice, [*Ad impressions are categorized based on multiple attributes*]{}. Given this, we argue that the most natural way to hide information is by [*hiding attributes*]{}. For example, let there be three attributes, each with two possible values:
- Age: Teenager, Adult
- Gender: Male, Female
- Location: US, Non-US
Together there are $2^3$ possible instantiations. Under the clustering scheme studied in [@Ghosh07:Computing], the website is allowed to hide information by bundling any subset of instantiations. However, not all bundles are natural. For example, consider the bundle $\{$(Teenager, Male, US), (Adult, Female, Non-US)$\}$. By creating this bundle, the website basically may tell the advertisers that a visitor is either a teenage US male or an adult Non-US female. This does not appear natural. The signaling scheme studied in [@Emek12:Signaling; @Miltersen12:Send] is even more general than clustering, so it may also lead to unnatural bundles.
On the other hand, attribute hiding always leads to natural bundles. For example, the website may hide the location attribute. That is, if the actual instantiation is (Teenager, Male, US), then the website may inform the advertisers that the visitor is a teenage male. By hiding the location attribute, we essentially created a bundle (Teenager, Male, ?), which consists of both (Teenager, Male, US) and (Teenager, Male, Non-US).
Based on the above example, it is easy to see that attribute hiding is clustering with a particular structure. It should be noted that this relationship between attribute hiding and clustering does not mean previous results on clustering apply to our model. For example, one of the two main results from [@Ghosh07:Computing] is a constructed clustering scheme that guarantees one half of the optimal revenue (and one half of the optimal social welfare). The construction does not apply to our model since it generally leads to unnatural bundles.
In this paper, we first show that it is NP-hard to solve for the optimal attribute hiding scheme.[^1] We then derive a polynomial-time solvable upper bound on the optimal revenue. Finally, we propose two heuristic-based attribute hiding schemes. Experiments show that revenue achieved by these schemes is close to the upper bound.
Besides the aforementioned related work in the computer science literature, bundling has also been well-studied in the economics literature. [@Palfrey83:Bundling] observed that for small numbers of bidders, a revenue-maximizing auctioneer may choose to bundle the items, and this makes bidders universally worse-off. On the other hand, for large numbers of bidders, the auctioneer may choose to unbundle the items, and this hurts the high-demand bidders while benefiting the low-demand bidders. [@Chakraborty99:Bundling] quantitatively analyzed the bundling behavior of the auctioneer. The result is that under a Vickrey auction, for each pair of objects, there is a unique critical number. If there are fewer bidders than this number, the seller chooses to bundle the items, and vice versa. [@Avery00:Bundling] studied more sophisticated bundling policy, including bundling with discounts and probabilistic bundling (the probability of bundling occurring depends on the bids).
Model Description
=================
There is a single item for sale characterized by $k$ attributes (attribute $1$ to $k$). Attribute $i$ has $C_i$ possible values, ranging from $0$ to $C_i-1$. $m$ is the total number of possible instantiations. $m=\prod_{i}C_i$. In this paper, when we mention polynomial time or NP-hardness, we mean in terms of $m$.
An instantiation whose $i$-th attribute equals $a_i$ is written as $$(a_1,a_2,a_3,\ldots,a_k)$$
The space of all possible instantiations $\Omega$ is $$\{0,\ldots,C_1-1\}\times\{0,\ldots,C_2-1\}\times\ldots\times\{0,\ldots,C_k-1\}$$
A [*natural bundle*]{} $b$ is an element from the following set of all natural bundles (denoted by $\mathcal{B}$): $$\{0,\ldots,C_1-1,?\}\times\{0,\ldots,C_2-1,?\}\times\ldots\times\{0,\ldots,C_k-1,?\}$$
Natural bundles are bundles of instantiations resulting from hiding attributes. An attribute of a natural bundle either takes a specific value, or is represented by a question mark, which means that this attribute is hidden. For example, let $k=5$, given the instantiation $(a_1,a_2,a_3,a_4,a_5)$, if we hide attributes $1$ and $3$, then it results in the natural bundle $(?,a_2,?,a_4,a_5)$. This bundle has size $C_1C_3$. As another example, every instantiation itself corresponds to a natural bundle of size $1$ (no attribute hidden). An instantiation $\omega$ belongs to a natural bundle $b$ if and only if for every attribute, either $\omega$ and $b$ share the same attribute value, or the attribute is hidden for $b$. Unlike the total number of arbitrary bundles, which equals $2^m$, the total number of natural bundles is polynomial in $m$, as shown below:
$$|\mathcal{B}|=\prod_{1\le i\le k}(C_i+1)\le
\prod_{1\le i\le k}C_i^2=m^2$$
The probabilities’ of different instantiations are based on a [*publicly known*]{} distribution $\Delta(\Omega)$. To simplify the presentation, when discussing bidders’ valuations, we factor in the probabilities. For example, if bidder $i$ values $\omega$ at $5$ when $\omega$ is the actual instantiation, and $\omega$ happens with probability $0.1$, then we say bidder $i$’s valuation for $\omega$ is $0.5$.
Let $n$ be the number of bidders. Let $v_i(\omega)$ be bidder $i$’s (expected) valuation for instantiation $\omega$. Following [@Ghosh07:Computing; @Emek12:Signaling; @Miltersen12:Send][^2], we assume full information: the auctioneer knows the bidders’ true valuations. Again, following previous models, we only consider bidders with additive valuations. That is, bidder $i$’s valuation for bundle $b$, denoted by $v_i(b)$, equals $\sum_{\omega\in b}v_i(\omega)$.
Following previous models, the auction is the Vickrey auction. We use $2(b)$ to denote the revenue for selling $b$ as a bundle. $2(b)$ is the second highest value in $\{v_i(b)|1\le i\le n\}$.
An [*attribute hiding scheme*]{} is a way to cluster the instantiations into natural bundles. An attribute hiding scheme is characterized by a set of bundles $\{b_1,b_2,\ldots,b_t\}$, satisfying
- All bundles are natural: $b_i\in \mathcal{B}$ for $1\le i\le t$
- The bundles are disjoint[^3]: for every pair of $b_i$ and $b_j$, there exists an attribute, so that for this attribute, $b_i$ and $b_j$ take different values (neither is $?$).
Under the attribute hiding scheme $\{b_1,b_2,\ldots,b_t\}$, instantiations covered by $b_i$ will have their attributes hidden to match $b_i$. Essentially, instantiations in $b_i$ are sold in a bundle. Instantiations not covered by any $b_i$ are sold without hiding attributes (sold separately as natural bundles of size $1$).
Under attribute hiding scheme $\{b_1,b_2,\ldots,b_t\}$, the revenue of the auctioneer equals $$\sum_{1\le i\le t}2(b_i)+\sum_{\omega\in \Omega-\cup_{1\le i\le t}b_i}2(\omega)$$
We introduce another function $r$. For $b\in \mathcal{B}$, $r(b)$ represents the extra revenue obtained by selling $b$ as a bundle, rather than selling instantiations in $b$ separately. We have $$r(b)=2(b)-\sum_{\omega\in b}2(\omega)$$
The revenue of the auctioneer can then be rewritten as $$\sum_{1\le i\le t}r(b_i)+\sum_{\omega\in \Omega}2(\omega)$$
The second term of the above expression does not depend on the attribute hiding scheme. Therefore, the problem of designing optimal attribute hiding scheme is equivalent to the problem of searching for a set of disjoint natural bundles $\{b_1,b_2,\ldots,b_t\}$, so that $\sum_{1\le i\le t}r(b_i)$ is maximized.
Hardness Result
===============
Previously, [@Ghosh07:Computing] showed that it is NP-hard to solve for the optimal clustering scheme. The proof was by reduction from [*3-partition*]{}: given $3z$ integers, determine whether it is possible to partition them into $z$ groups with equal sums. In this section, we prove a similar result. We show that it is also NP-hard to solve for the optimal attribute hiding scheme. Our proof is by reduction from [*monotone one-in-three 3SAT*]{} [@Schaefer78:The]. Monotone one-in-three 3SAT is a variant of 3SAT. Monotone means that the literals are just variables, never negations. One-in-three means that the determination problem is to see whether there is an assignment so that for each clause, exactly one literal is true. We emphasize again that our result is not implied by the hardness result from [@Ghosh07:Computing].
It is NP-hard to solve for the optimal attribute hiding scheme.
Let us consider the following monotone one-in-three 3SAT instance with $D$ clauses:
$$(x_{f(1)}\vee x_{f(2)}\vee x_{f(3)})\land
(x_{f(4)}\vee x_{f(5)}\vee x_{f(6)})\land\ldots$$ $$\ldots\land(x_{f(3D-2)}\vee x_{f(3D-1)}\vee x_{f(3D)})$$
There are $3D$ literals, and they are from a list of $E$ variables ($x_1$ to $x_E$, $f$’s range is between $1$ and $E$). According to [@Schaefer78:The], it is NP-complete to determine whether there exists an assignment of the $x_i$, so that the 3SAT instance is true, and for each clause, there is exactly one true literal.
We will construct a probabilistic single-item auction scenario with $m$ possible instantiations and $n$ bidders. Both $m$ and $n$ are polynomial in $E$. We will show that for the constructed scenario, if we are able to solve for the optimal attribute hiding scheme in polynomial time (in $m$), then we are able to determine the above 3SAT instance in polynomial time (in $E$). This implies that it is NP-hard to solve for the optimal attribute hiding scheme.
Our construction is as follows. Let the number of attributes $k$ be $\lceil\log_2(D)\rceil+\lceil\log_2(E)\rceil+11$. All attributes are binary. The total number of instantiations $m$ is polynomial in $E$ as shown below. $$m=2^{\lceil\log_2(D)\rceil+\lceil\log_2(E)\rceil+11}
\le 2^{\log_2(D)+\log_2(E)+13}$$ $$=8192DE\le 8192E^4$$
Our proof relies on the following seven families of natural bundles (Family \[eq:f1\] to \[eq:f7\]): $$\label{eq:f1}
(\underline{e},\underline{d},0,?,?,0,1,0,1,0,1,0,1)$$ $$\label{eq:f2}
(\underline{e},\underline{d},?,0,?,0,1,0,1,0,1,0,1)$$ $$\label{eq:f3}
(\underline{e},\underline{d},?,?,0,0,1,0,1,0,1,0,1)$$ $$\label{eq:f4}
(\underline{e},\underline{?},0,0,0,?,?,0,1,0,1,0,1)$$ $$\label{eq:f5}
(\underline{?},\underline{d},1,?,?,0,1,?,?,0,1,0,1)$$ $$\label{eq:f6}
(\underline{?},\underline{d},?,1,?,0,1,0,1,?,?,0,1)$$ $$\label{eq:f7}
(\underline{?},\underline{d},?,?,1,0,1,0,1,0,1,?,?)$$
In the above, $\underline{e}$ is the binary representation of integer $e$ ($1\le e\le E$). The representation’s width is $\lceil\log_2(E)\rceil$. Similarly, $\underline{d}$ is the binary representation of integer $d$ ($1\le d\le D$). The representation’s width is $\lceil\log_2(D)\rceil$. Finally, $\underline{?}$ is $?$ repeated $\lceil\log_2(E)\rceil$ times (Family \[eq:f5\], \[eq:f6\], and \[eq:f7\]) or $\lceil\log_2(D)\rceil$ times (Family \[eq:f4\]).
We recall that the problem of designing optimal attribute hiding scheme is equivalent to the search of disjoint natural bundles $\{b_1,b_2,\ldots,b_t\}$, so that $\sum_{1\le i\le t}r(b_i)$ is maximized. Given a natural bundle $b$, $r(b)$ depends on the bidders’ valuations. We will construct a set of bidders, so that for any natural bundle $b$, $r(b)=0$ by default. The exceptions are:
- For $i=1,2,3$, we use $b^i(e,d)$ to represent the natural bundle characterized by $e$ and $d$ in Family $i$. $r(b^i(e,d))=1$ if and only if, in the 3SAT instance, variable $e$ appears in the $i$-th position of clause $d$.
- We use $b^4(e)$ to represent the natural bundle characterized by $e$ in Family \[eq:f4\]. Let $\#e$ be the number of times variable $e$ appears in the 3SAT instance. It is without loss of generality to assume $\#e\le D$ (no literal appears twice in a clause). Let $r(b^4(e))=\#e(1-\epsilon)$. Here, $\epsilon$ is a constant that is less than $\frac{1}{D}$. The idea is to make sure that $\#e(1-\epsilon)>\#e-1$.
- We use $b^5(d)$ to represent the natural bundle characterized by $d$ in Family \[eq:f5\]. $r(b^5(d))=3$.
- We use $b^6(d)$ to represent the natural bundle characterized by $d$ in Family \[eq:f6\]. $r(b^6(d))=3$.
- We use $b^7(d)$ to represent the natural bundle characterized by $d$ in Family \[eq:f7\]. $r(b^7(d))=3$.
For now, we simply assume that it is possible to construct a polynomial number of bidders, so that the values of $r(b)$ for different $b$ are indeed as described above. We will provide the specific construction toward the end. Let $O$ be an optimal attribute hiding scheme corresponding to the above construction. If $r(b)=0$, then it is without loss of generality to assume $b\notin O$. Therefore, we can ignore bundles not in the above seven families. Some bundles from Family \[eq:f1\] to \[eq:f3\] can also be ignored for the same reason. For presentation purposes, we call the remaining bundles [*helpful*]{} bundles. A bundle $b$ is helpful if and only if $r(b)>0$.
Let us consider a fixed variable $e$ ($1\le e\le E$). $e$ appears $\#e$ times in the 3SAT instance, so there are exactly $\#e$ pairs of $d$ ($1\le d\le D$) and $i$ ($1\le i\le 3$), so that $b^i(e,d)$ is helpful. We use $b_{e,1},b_{e,2},\ldots,b_{e,\#e}$ to denote these $\#e$ helpful bundles. They are the only helpful bundles that intersect $b^4(e)$. If some of these bundles are not in $O$, then none of them is in $O$. The reason is that $r(b^4(e))=\#e(1-\epsilon)>\#e-1$, so it is better off to add $b^4(e)$ into $O$ (and push out $b_{e,1}$ to $b_{e,\#e}$ if they are in $O$). In summary, for $e$ from $1$ to $E$, we must have one of the following two:
- $b_{e,1},b_{e,2},\ldots,b_{e,\#e}$ are all in $O$. $b^4(e)$ is not in $O$.
- None of $b_{e,1},b_{e,2},\ldots,b_{e,\#e}$ is in $O$. $b^4(e)$ is in $O$.
Let $T$ be the set of $e$ values where $b_{e,1},b_{e,2},\ldots,b_{e,\#e}$ are all in $O$. Let $F$ be the set of $e$ values where none of $b_{e,1},b_{e,2},\ldots,b_{e,\#e}$ is in $O$. We use $O_{1234}$ to denote the set of helpful bundles in $O$ that belong to Family \[eq:f1\] to \[eq:f4\]. We have $$\sum_{b\in O_{1234}}r(b)=\sum_{e\in T}\#e + \sum_{e\in F}\#e(1-\epsilon)$$ $$=\epsilon\sum_{e\in T}\#e + \sum_{e\in T}\#e(1-\epsilon)+\sum_{e\in F}\#e(1-\epsilon)$$ $$=\epsilon\sum_{e\in T}\#e + (1-\epsilon)3D$$
Let us then consider a fixed variable $d$ ($1\le d\le D$), $b^5(d)$, $b^6(d)$, and $b^7(d)$ pair-wise intersect. Therefore, in $O$, at most one of them can appear. Actually, exact one of them appears. If none of them appears in $O$, then we can add $b^5(d)$ into $O$, which results in higher revenue. Let $e_2$ and $e_3$ be the second and third variables in clause $d$ of the 3SAT instance. The only helpful bundles $b^5(d)$ intersects with are $b^2(e_2,d)$ and $b^3(e_3,d)$. By removing these two from $O$ (if they are in $O$ to start with) and adding $b^5(d)$ into $O$, the revenue increases. Therefore, for any $d$ from $1$ to $D$, $O$ contains exactly one of $\{b^5(d),b^6(d),b^7(d)\}$. We use $O_{567}$ to denote the set of helpful bundles in $O$ that belong to Family \[eq:f5\] to \[eq:f7\]. We have $$\sum_{b\in O_{567}}r(b)=3D$$ Hence, $$\sum_{b\in O}r(b)=
\sum_{b\in O_{1234}}r(b)+
\sum_{b\in O_{567}}r(b)=
\epsilon\sum_{e\in T}\#e + (2-\epsilon)3D$$
Let $d$ be a specific value between $1$ and $D$. If $b^5(d)$ belongs to $O$, then among helpful bundles characterized by $d$ from Family \[eq:f1\] to \[eq:f3\], the only helpful bundle that can coexist with $b^5(d)$ is $b^1(e_1,d)$, where $e_1$ is the first variable in clause $d$ of the 3SAT instance. In general, no matter which among $\{b^5(d),b^6(d),b^7(d)\}$ appears in $O$, among helpful bundles characterized by $d$ from Family \[eq:f1\] to \[eq:f3\], there is at most one that can be in $O$. Therefore, the total number of helpful bundles from Family \[eq:f1\] to \[eq:f3\] in $O$ is at most $D$. we have $$\sum_{e\in T}\#e\le D$$ $$\sum_{b\in O}r(b)=\epsilon\sum_{e\in T}\#e + (2-\epsilon)3D
\le \epsilon D + (2-\epsilon)3D=6D-2D\epsilon$$
If we are able to solve for the optimal attribute hiding scheme in polynomial time, then we are also able to determine in polynomial time whether $\sum_{b\in
O}r(b)$ is equal to the upper bound $6D-2D\epsilon$. If they are equal, then we have a satisfactory assignment of the 3SAT instance. For variable $e$, $b_{e,1}$ to $b_{e,\#e}$ determine whether $e$ is true or not. If they are all in $O$, then $e$ is set to be true. Otherwise (if none of them is in $O$), $e$ is set to be false. When the upper bound is reached, $\sum_{e\in T}\#e=D$, which implies that under the above assignment, there are exactly $D$ true literals. Next, we show that two true literals cannot appear in the same clause. That is, there is exactly one true literal for each clause under the assignment, and all clauses are satisfied (there are $D$ true clauses). Given $d$, let the variables in clause $d$ be $e_1,e_2,e_3$. $b^1(e_1,d)$, $b^2(e_2,d)$, and $b^3(e_3,d)$ are all helpful bundles. We proved that among helpful bundles characterized by $d$ from Family \[eq:f1\] to \[eq:f3\], there is at most one that can be in $O$. Therefore, only one of $b^1(e_1,d),b^2(e_2,d),b^3(e_3,d)$ can be in $O$. That is, only one of $e_1,e_2,e_3$ is set to be true.
The other direction can be shown similarly. If there is a satisfactory assignment of the 3SAT instance, then $\sum_{b\in O}r(b)$ should match the upper bound $6D-2D\epsilon$.
In conclusion, for the constructed auction setting, it is NP-hard to determine whether the optimal revenue $\sum_{b\in O}r(b)+\sum_{\omega\in \Omega}2(\omega)$ reaches $6D-2D\epsilon+\sum_{\omega\in\Omega}2(\omega)$.
Finally, we still need to show that it is possible to construct a polynomial number of bidders, so that the values of $r(b)$ are exactly as described above. Due to space constraint, we present the construction and omit the proof.
- We construct two bidders who both value every instantiation equally, and the valuation for every instantiation is $L$ ($L>D$).
- For every helpful bundle $b$, we construct two new bidders. By default, both bidders value all instantiations in $b$ at $L$ and value all instantiations outside of $b$ at $0$. The exceptions are that one bidder values instantiation $b|_?^0$ at $r(b)+L$ and the other bidder values instantiation $b|_?^1$ at $r(b)+L$. Here, $b|_?^y$ is the instantiation resulting from replacing all $?$ in $b$ by $y$.
Tree-Structured Attribute Hiding Schemes
========================================
In this section, we study a special family of attribute hiding schemes, which we call the [*tree-structured*]{} schemes. Let $b$ be a [*non-unit*]{} natural bundle (bundle of size greater than $1$). For $b$, at least one attribute is hidden. Let $x$ be one of the hidden attributes of $b$. We can split $b$ into $C_x$ disjoint natural bundles by revealing attribute $x$. The resulting bundles are $b|_x^0, b|_x^1, \ldots, b|_x^{C_x-1}$. $b|_x^i$ represents the natural bundle obtained by replacing the $x$-th attribute of $b$ by $i$. If $b$ belongs to an attribute hiding scheme $O$, then after splitting $b$, the new scheme becomes $$(O-\{b\})\cup\{b|_x^0,b|_x^1,\ldots,b|_x^{C_x-1}\}$$ It is easy to see that the new scheme is still feasible (the bundles remain disjoint).
Tree-structured attribute hiding schemes are results of [*recursive splitting*]{} (revealing attribute) starting from $\{(?,?,\ldots,?)\}$. At every step, we either terminate and keep the current scheme, or pick a non-unit bundle from the current scheme, and split (reveal) one of its attributes.
An attribute hiding scheme $O$ is tree-structured if and only if it satisfies one of the following:
- $O=\{(?,?,\ldots,?)\}$: the scheme is simply hiding all attributes and selling all instantiations in a single bundle.
- There exists a tree-structured attribute hiding scheme $O'$. There exists a bundle $b\in O'$ whose $x$-th attribute is hidden. After splitting $b$ by revealing attribute $x$, the resulting scheme is equivalent to $O$.[^4]
Let us consider an example with three binary attributes. $\{(?,?,?)\}$ is, by definition, a tree-structured attribute hiding scheme. Starting from $\{(?,?,?)\}$, if we pick $(?,?,?)$ and reveal its second attribute, then we get
The leaves $\{(?,0,?),(?,1,?)\}$ characterize a new tree-structured attribute hiding scheme. If we further split the first bundle $(?,0,?)$ based on its third attribute, then we get
Again, the leaves $\{(?,0,0),(?,0,1),(?,1,?)\}$ characterize a new tree-structured attribute hiding scheme.
If there are at most two attributes, then all attribute hiding schemes are tree-structured.[^5]
If there are at least three attributes, then there exist attribute hiding schemes that are not tree-structured.
We construct the following natural bundles. For $i$ from $1$ to $k$, let $b_i$’s $i$-th attribute be hidden, let $b_i$’s $((i\bmod k)+1)$-th attribute be $1$, and let $b_i$’s all other attributes be $0$. $$\begin{aligned}
b_1&=&(?,1,0,0,\ldots,0,0)
\nonumber\\
b_2&=&(0,?,1,0,\ldots,0,0)
\nonumber\\
b_3&=&(0,0,?,1,\ldots,0,0)
\nonumber\\
&\ldots&
\nonumber\\
b_{k-1}&=&(0,0,0,0,\ldots,?,1)
\nonumber\\
b_k&=&(1,0,0,0,\ldots,0,?)
\nonumber\end{aligned}$$ The $b_i$ are disjoint. $\{b_1,b_2,\ldots,b_k\}$ is not tree-structured because starting from $(?,?,\ldots,?)$, if we ever reveal an attribute (e.g., attribute $x$), then $b_x$ cannot be in the final scheme.
As we mentioned earlier, tree-structured attribute hiding schemes are results of recursive splitting starting from the bundle of all instantiations. At every step, we either terminate or split a non-unit bundle in some way. For every natural bundle $b$, let $t(b)$ be the optimal revenue for selling instantiations in $b$, as a result of making optimal recursive splitting decisions on $b$. $t((?,?,\ldots,?))$ is then the optimal revenue of tree-structured attribute hiding schemes. Given a bundle, we either sell it as a whole, or split it in some way as a first step. Let $h(b)$ be the set of hidden attributes of $b$. We have
$$t(b)=\max\{2(b),\max_{x\in h(b)}\sum_{0\le i\le C_x-1}t(b|_x^i)\}$$
If $b$ has size $1$, then $h(b)=\emptyset$. That is, for unit bundles, $t(b)=2(b)$. Given the values of $t(b)$ for all $b$ with $|h(b)|=y$, we can then easily compute the values of $t(b)$ for all $b$ with $|h(b)|=y+1$. The total number of natural bundles $|\mathcal{B}|$ is polynomial in $m$. For every $b$, $t(b)$ is the maximum of at most $k+1$ values, which is at most $\log_2m+1$. Therefore, the optimal revenue $t((?,?,\ldots,?))$ can be computed in polynomial time. The corresponding optimal scheme can be obtained along the way.
Upper Bound and Weighted Matching
=================================
Our objective is to find a set of disjoint natural bundles, denoted by $O$, which maximizes $\sum_{b\in O}r(b)$. We can model it as an integer program. We introduce $|\mathcal{B}|$ binary variables. For $b\in \mathcal{B}$, let $z_b$ be a binary variable. If $z_b=1$, then it means $b\in O$. The number of binary variables $|\mathcal{B}|$ is polynomial in $m$. The objective is to maximize $\sum_{b\in
\mathcal{B}}z_br(b)$. The constraints are that bundles in $O$ are disjoint. That is, for $b_1,b_2\in \mathcal{B}$, if $b_1$ and $b_2$ intersect, $z_{b_1}+z_{b_2}\le 1$. The number of constraints is at most $|\mathcal{B}|^2$, which is polynomial in $m$. In summary, the optimal revenue can be solved for based on an integer program with polynomial numbers of variables and constraints. One upper bound can then be solved for in polynomial time if we consider the linear relaxation (replacing binary variables by non-integer variables).
Some preprocessing can vastly reduce the number of variables in the above program. We first observe that, by definition, $r(b)=0$ for all $b$ with size $1$. That is, we can safely set $z_b=0$ for all $b$ with size $1$. We then observe that, for any natural bundle $b$ with size greater than $1$, if the following expression is true, then it means that instead of selling $b$ as a single bundle, we can achieve higher revenue by recursively splitting it, in which case we can safely set $z_b=0$.
$$2(b)<\max_{x\in h(b)}\sum_{0\le i\le C_x-1}t(b|_x^i)$$
In Section \[sec:experiment\], our simulation shows that when computing the upper bound, the above observations indeed vastly reduce the number of variables in the linear program. For example, for settings with $10$ binary attributes and $10$ bidders, originally, there are as many as $(2+1)^{10}=59049$ variables. After preprocessing, there are only $220.28$ variables on average over repeated simulations.
We then discuss another heuristic for generating attribute hiding schemes with high revenue. This heuristic only applies to settings where all attributes are binary. If all attributes are binary, then a natural bundle with only one attribute hidden contains exactly two instantiations. The heuristic is based on [*maximum weighted matching*]{}. We view all instantiations as vertices. If two instantiations can be merged into a natural bundle $b$, and $r(b)>0$, then we create an edge with weight $r(b)$ between them. Maximum weighted matching can be solved in polynomial time. The matching result characterizes the optimal attribute hiding scheme under the additional constraint that at most one attribute is hidden.[^6]
Experiments {#sec:experiment}
===========
In this section, we evaluate the performances of the proposed heuristic-based attribute hiding schemes. For different values of $k$, $\bar{C}$, and $n$, we construct problem instances with $k$ attributes, each attribute taking $\bar{C}$ possible values, and $n$ bidders. The total number of possible instantiations is then $\bar{C}^k$. For each instantiation, bidders’ valuations are drawn independently from $U(0,1)$. For every setup, we repeat $100$ times and report the averages.
Setup Tree Match UB \#Opt \#Var HM
------------- ------- ------- ------- ------- -------- ------
$k=n=3$
$\bar{C}=2$ 13.33 11.58 15.42 47 5.82 1.08
$k=n=5$
$\bar{C}=2$ 3.953 3.810 4.354 35 15.8 1.54
$k=n=10$
$\bar{C}=2$ 0.836 0.927 0.950 0 220.28 4.76
$k=n=3$
$\bar{C}=3$ 9.251 NA 10.58 25 13.28 0.96
$k=n=5$
$\bar{C}=3$ 1.767 NA 1.976 0 45.39 0.3
$k=n=8$
$\bar{C}=3$ 0.296 NA 0.361 0 326.18 0.01
The table fields are described below:
- Tree, Match, UB: Comparing to selling all instantiations separately, the extra revenue in terms of percentage. Tree is short for optimal tree-structured scheme. Match is short for optimal scheme based on maximum weighted matching (only applies to $\bar{C}=2$). UB is short for upper bound on the optimal revenue.
- \#Opt: Among $100$ repeated simulations, how many times one of the heuristic-based schemes reaches the upper bound (therefore guarantees optimality[^7]).
- \#Var: How many variables are in the linear program for computing upper bound.
- HM: How many natural bundles with at least two hidden attributes cannot be recursively split to achieve higher revenue.
Future Research
===============
Given the fact that it is NP-hard to solve for the optimal attribute hiding scheme, one direction of future research is to study whether there are heuristic-based attribute hiding schemes that guarantee a constant fraction of the optimal revenue. A similar direction is to see how much revenue we lose by not allowing unnatural bundles. A preliminary result shows that the optimal revenue by clustering (allowing unnatural bundles) can be as high as twice the optimal revenue by hiding attributes. The construction is as follows. There are $m$ instantiations and $m$ bidders. Bidder $i$ only values instantiation $i$ positively. Let instantiation $1$ be $(0,0,\ldots,0)$ and bidder $1$’s valuation for it be $\frac{m}{2}$. Let instantiation $m$ be $(1,1,\ldots,1)$ and bidder $m$’s valuation for it be $\frac{m}{2}$. For $1<i<m$, let bidder $i$’s valuation for instantiation $i$ be $1$. With this setup, the optimal revenue by clustering is $\frac{2m-2}{2}$. The optimal revenue by hiding attributes is $\frac{m}{2}$. The ratio $\frac{2m-2}{m}$ approaches $2$ for large $m$.
[^1]: We mentioned earlier that [@Ghosh07:Computing] proved a similar result. The authors showed that it is NP-hard to solve for the optimal clustering scheme. It should be noted that our NP-hardness result is not implied by this earlier result, which relied on reduction involving unnatural bundles. Actually, our requirement on bundles being natural greatly adds to the difficulty of the reduction, and our proof is based on completely new techniques.
[^2]: Besides the full information setting, [@Emek12:Signaling] also discussed the more general Bayesian setting.
[^3]: If under an attribute hiding scheme, two different natural bundles share one common instantiation, then for this instantiation, it is not clear which attributes we should hide.
[^4]: Two schemes are equivalent if they share the same set of non-unit bundles.
[^5]: This proposition implies that if there are at most two attributes ($m$ can still be large), then we can solve for the optimal attribute hiding scheme in polynomial time, because it must be tree-structured.
[^6]: In Section \[sec:experiment\], our simulation shows that there are generally very few natural bundles with at least two hidden attributes cannot be recursively split to achieve higher revenue. This somewhat justifies the heuristic requirement that at most one attribute is hidden.
[^7]: Even if the heuristic-based schemes do not reach the upper bound, they may still possibly be optimal.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider radially twisted nanotubes in the low-energy approximation where the dynamics is governed by a one-dimensional Dirac equation. The mechanical deformation of the nanotubes is reflected by the presence of an effective vector potential. We discuss twisted carbon and boron-nitride nanotubes, where deformations give rise to periodic and nonperiodic finite-gap Hamiltonians. The intimate relation of these systems with the integrable Ablowitz-Kaup-Newell-Segur hierarchy is exploited in the study of their spectral properties as well as in the computation of the (local) density of states. We show that a nonlinear hidden supersymmetry generated by local supercharges arises naturally in the finite-gap configurations of twisted nanotubes with time-reversal symmetry. The properties of the van Hove singularities are encoded in its structure.'
author:
- |
Francisco Correa$^{1}$, Vít Jakubský$^{2}$\
[*$^{1}$Centro de Estudios Científicos (CECs), Avenida Arturo Prat 514, Valdivia, Chile*]{}\
[*${}^2$Nuclear Physics Institute, Řež near Prague, 25068, Czech Republic*]{}
title: 'Finite-gap twists of carbon nanotubes and an emergent hidden supersymmetry'
---
Introduction
============
Since their discovery in 1991 [@NanotubesObserved], carbon nanotubes have attracted massive attention from both experimental and theoretical physicists. Their remarkable mechanical and electronic properties, extreme mechanical strength [@UltimateStrength], and elasticity [@Elasticity] as well as conductivity [@experimentnanotuba] make their use in future electronic devices very promising [@NanotubeElectronics]. Carbon nanotubes are also relevant as a low-dimensional test field of fundamental physical theories. For instance, the Klein tunneling, a well known phenomenon in relativistic quantum theory [@Klein], has not been observed for elementary particles so far. However, it was predicted and observed in carbon nanotubes where it stays behind the absence of backscattering on impurities [@Ando].
The single-wall carbon nanotubes are small cylinders with atom-thick shells, that can be created by rolling up graphene nanoribbons [@RollTheRibbon]. Despite their structural simplicity (as quasi one-dimensional objects), the nanotubes exhibit a remarkable variety of electronic properties. They can be either metallic or semiconducting, depending on the orientation of the hexagonal lattice in the nanotube. The gap between the valence and conduction band can be regulated by either external fields [@ABoscillations] or by mechanical deformations [@KaneMele]. This paves the way to strain engineering [@StrainEngineering] where graphene-based devices would be fine-tuned by deformations.
In the present paper, we will consider a class of exactly solvable models of single-wall carbon nanotubes subject to radial twist (axial torsion). The systems will be studied in the low-energy regime where the dynamics of charge carriers is well approximated by a $(1+1)$-dimensional Dirac equation [@Semenoff]. In particular, the twist configurations under investigation will be described by a finite-gap Hamiltonian.
Besides the analysis of the spectral properties, the computation of the local density of states (LDOS) of the systems will be addressed. This quantity is measurable by means of scanning tunneling microscopy experiments [@STM] and is very important for the specification of the electronic properties of nanostructures. When integrated spatially, it provides the density of states (DOS) that reflects the probability of inserting an electron at given energy into the system.
In general, it is a rather complicated task to analytically compute the local density of states. However, as it was suggested in [@Dunne] and demonstrated later in [@BdG], the computation can be considerably simplified for a broad class of Dirac and Bogoliubov - de Gennes Hamiltonians that are characterized by a finite number of gaps in the spectrum. This approach is based on the intrinsic properties of the finite-gap systems that arise from the stationary AKNS (Ablowitz-Kaup-Newell-Segur) hierarchies of integrable systems [@Gesztesy].
In the next section, we will set up the theoretical framework in which the twisted nanotubes will be analyzed. It will be explained how the Dirac Hamiltonian emerges in the description of twisted carbon (and boron-nitride) nanotubes. In Sec. \[AKNSLDOS\], the relevant points of the construction of the AKNS hierarchies are summarized and the closed formula for the LDOS is discussed. Section \[examples\] is devoted to the presentation of explicit examples [of carbon and boron-nitride nanotubes]{} where the LDOS and DOS will be computed explicitly. We will show in Sec. \[susy\] that the singular points of DOS, so-called van Hove singularities, are closely related to the nonlinear supersymmetry that arises naturally for the nanotubes with finite-gap twists and conserved time-reversal symmetry. The last section is left for the discussion and outlook.
The model {#model}
=========
Carbon atoms have four valence electrons; three of them are tightly bound in the interatomic bonds while the fourth one is free and can contribute to the electronic properties of the crystal. The properties of the collective excitations of these electrons in graphene are well described by the tight-binding Hamiltonian [@CastroNeto]. The interactions between the nearest neighbors are only assumed, being specified by a constant hopping parameter. The spin degree of freedom of electrons can be neglected; it is irrelevant in the considered interactions.
The analysis of the band structure displays the specific feature of graphene: the Fermi surface is formed by six discrete points where valence and conduction bands meet. They are located at the corners of the hexagonal first Brillouin zone and are called Dirac points. Only two of them are inequivalent[^1] and correspond to different electronic states. Let us denote them as $\mathbf{K}$ and $\mathbf{K'}\equiv-\mathbf{K}$. In the close vicinity of the Dirac points, the energy surface acquires conelike shape. It suggests that the dispersion relation is linear in this region. Indeed, taking $\mathbf{k}=\pm\mathbf{K}+\mathbf{\delta k}$ with $\mathbf{\delta k}\sim 0$ and expanding the tight-binding Hamiltonian up to the terms linear in $\mathbf{\delta k}$, we get the stationary equation for the two-dimensional massless Dirac particle [@Semenoff]. The Hamiltonian acquires the same form in both valleys of $\mathbf{K}$ and $\mathbf{K'}$. In the coordinate representation, we have [^2] $$\label{singlevalleyhd}
h(\pm \mathbf{K})\psi_{\pm\mathbf{K}}=(-i\sigma_2\partial_{x}+ i\sigma_1\partial_{y})\psi_{\pm\mathbf{K}}=\lambda\psi_{\pm\mathbf{K}} \, .$$ The spinorial degree of freedom in (\[singlevalleyhd\]), the pseudospin, arises due to the two carbon atoms in the elementary cell; the hexagonal lattice can be thought of as assembled from two triangular lattices. The spin-up or down components of the wave functions are non-vanishing only on one of the two triangular sublattices. The operator $h(\mathbf{K})$ acts on the spinors $(\psi_{\mathbf{K}A},\psi_{\mathbf{K}B})^t$, while $h(\mathbf{K'})$ on the spinors $(\psi_{\mathbf{K'}B},\psi_{\mathbf{K'}A})^t$. Here, the first index denotes the valley, the second distinguishes between the sublattices $A$ and $B$ and $t$ denotes transposition.
The formula (\[singlevalleyhd\]) was introduced already in 1984 by Semenoff [@Semenoff] and makes the basis for considering the condensed matter system as a convenient test field for a low-dimensional quantum field theory. Indeed, it makes it possible to observe phenomena in this condensed matter system that are native in high-energy quantum physics, see e.g. [@Shytov], [@Gusynin].
A single-wall carbon nanotube is rolled up from a straight graphene strip. The actual orientation of the hexagonal lattice in the strip is uniquely determined by the chiral (circumference) vector $\mathbf{C_h}$, which is a linear combination of the translation vectors of the lattice [@Blase]. Its length corresponds to the diameter of the nanotube. We can fix the coordinates such that $y$ goes in the circumference direction. Then the chiral vector gets the simple form, $\mathbf{C_h}=(0,C_h)$.
The effect of rolling up the strip is reflected by the periodic boundary condition imposed on the wave functions, $\psi_{\mathbf{K}}(x,y+C_h)=\psi_{\mathbf{K}}(x,y)$. It leads to the quantization of the momentum in the circumference direction which acquires discrete values $k_y$. In the low-energy approximation, only the value of $k_y$ that minimizes the energy is relevant. The system is then governed by a truly one-dimensional Hamiltonian $-i\sigma_2\partial_{x}+\sigma_1k_y$. The actual value of this fixed $k_y$ depends on the character of the nanotube. Instead of going into more details that can be found, for instance in Ref. [@Blase], let us notice that $k_y=0$ corresponds to metallic nanotubes as there is no gap in the spectrum. When $k_y\neq 0$, there is a small gap in the spectrum and the nanotube is semiconducting. For purposes of our current analysis, we can suppose that the nanotubes are metallic (i.e. the angular momentum is vanishing, $k_y=0$) and are infinitely long. The latter approximation is rather reasonable due to the recent experiments where ultralong single-wall nanotubes were created [@ultralong].
Up to now, we considered systems where neither external fields nor any strains were present. By deforming the crystal mechanically, the interatomic distances in the lattice are modified. Thus, the hopping parameter ceases to be constant and becomes position dependent. This leads to the appearance of gauge fields in the tight-binding Hamiltonian. It can be approximated in the low-energy limit by Dirac operator with nonvanishing vector potential [@KaneMele; @Kolesnikov; @GaugeFields].
We shall consider radial twist of the nanotubes. Let us mention that both single-wall and multi-wall carbon nanotubes with radial twist were used in construction of nanoelectromechanical devices, e.g. single-molecule torsional pendulum [@pendulum], abacus-type resonators [@resonators] or even rotors [@rotors]. In these systems, the nanotube served as the torsional string that was twisted by deflection of small paddles attached to it; see also [@TwistedNanotubes] for a brief review.
The radial twist shifts the atoms in the lattice perpendicularly to the axis, preserving the tubular shape of the nanotube. The displacement is reflected by a deformation vector, which measures the difference between actual and equilibrium position of atoms. It can be written in our specific case as $\mathbf{d}=(0,d_y(x))^t$. We consider the situation where the displacement is smooth and small on the scale of the interatomic distance. Then the interaction does not mix the valleys of $\mathbf{K}$ and $\mathbf{K'}$, and the system can be studied in the vicinity of one Dirac point only. The stationary equation for low energy Dirac fermions in the $\mathbf{K}$-valley acquires the following simple form [@KaneMele], [@twisting], $$\label{eq1}
h(\mathbf{K})\phi=(-i\sigma_2\partial_x+\Sigma (x)\sigma_1)\phi=\lambda\phi \, ,$$ where the vector potential $\Sigma (x)$ reflects the twist. It is related to the displacement vector by $\mathbf{d}=\zeta\,(0,\int \Sigma (x) dx)$, where $\zeta$ is a constant dependent on the crystal characteristics.[^3] In this framework, the constant vector potential $\Sigma (x)=\gamma>0$ would reflect a linear displacement ${\mathbf d}=\zeta\,(0,\gamma x)$.
Finally, let us consider the following generalization of (\[eq1\]), where a mass term is included, $$\label{eq1m}
h(\mathbf{K})\phi=( -i\sigma_2\partial_x+\Sigma(x)\sigma_1+M\sigma_3)\phi=\lambda\phi \, .$$ The analogue of (\[singlevalleyhd\]) with the mass term $M$ was proposed by Semenoff for description of the quasiparticles in the boron-nitride crystal in the low-energy approximation [@Semenoff]. The boron-nitride crystal has the same hexagonal structure as graphene. However, the atoms in the elementary cell of the crystal cease to be equivalent. It gives rise to the potential term with $\sigma_3$ that distinguishes between the two triangular sublattices $A$ and $B$.
We will consider (\[eq1m\]) as the Hamiltonian of radially twisted carbon ($M=0$) or boron-nitride ($ M\neq 0$), depending on the value of $M$, nanotubes. Boron-nitride nanotubes were studied theoretically and observed experimentally, see e.g. [@PreparationBNT], [@TheoryBNT]. Contrary to the carbon nanotubes, they are always semiconducting.
We shall consider the scenario where the term $\Sigma(x)\sigma_1+M\sigma_3$ in (\[eq1m\]) belongs to the broad class of the finite-gap potentials. In the next section, we will show how the peculiar properties of finite-gap systems can be utilized for computation of the local density of states.
Finite-gap twists and the LDOS via AKNS hierarchy {#AKNSLDOS}
=================================================
We review here some properties of the integrable ANKS hierarchies associated with the Dirac Hamiltonian (\[eq1m\]). To explain more easily the main features, we use the following unitarily transformed Hamiltonian, $$\begin{aligned}
\nonumber
\tilde{h}&=&\exp\left(-i\frac{\sigma_1\pi}{4}\right)h(\mathbf{K})\,\exp\left(i\frac{\sigma_1\pi}{4}\right)\\&=&\left(\begin{array}{cc}-i\partial_x&\Delta(x)\\\Delta(x)^*&i\partial_x\end{array}\right)\, ,
\label{h1}\end{aligned}$$ where $\Delta=\Sigma(x)+i M$. This form with diagonal derivative term, the Bogoliubov-de Gennes type Hamiltonian, is used frequently in the analysis of Gross-Neveu and Nambu-Jona-Lasinio models [@Dunne], [@Thies] and will make the presentation more coherent with the specialized literature [@Gesztesy].
The vector potential $\Delta(x)=\Delta$ in (\[h1\]) is called finite gap (or algebro-geometric in the mathematical literature) when it solves one of the equations from the stationary AKNS hierarchy of the nonlinear differential equations, namely AKNS$_N$. One of the most intriguing properties of the Hamiltonian (\[h1\]) with a finite gap potential is manifested in its spectrum; it consists of a finite number of bands [@Gesztesy], [@concini]. The actual number of bands (or gaps) is fixed by the AKNS$_N$ equation solved by $\Delta$. The values of band-edge energies of a finite-gap system can be obtained in purely algebraic manner; see [@Gesztesy]. These features are intimately related with the existence of an integral of motion of the Hamiltonian (\[h1\]). Nonperiodic finite-gap systems can be obtained as the infinite-period limit of the periodic ones. In this context, the nonperiodic systems are known as kink or kink antikink models in analogy with the soliton solutions in the Korteweg-de Vries (KdV) hierarchy.
Another relevant feature of this class of models is that they can approximate very well any condensed matter systems described, in the low-energy approximation, by the Hamiltonian (\[eq1\]) with a generic periodic potential. The Hamiltonian with a generic periodic potential has an infinite number of spectral gaps, the width of which [decreases]{} rapidly with the increasing absolute value of energy. Hence, the spectrum of such system can be fitted well by a finite- gap one.
The stationary AKNS hierarchy of nonlinear differential equations can be constructed in terms of a Lax pair which consists of the Hamiltonian $\tilde{h}$ and a matrix differential operator $\tilde{S}_{N+1}$, defined as $$\label{S}
\tilde{S}_{N+1}=i\sum_{l=0}^{N+1}\left(\begin{array}{cc}g_{N+1-l}&f_{N-l}\\f^*_{N-l}&g_{N+1-l}\end{array}\right)\sigma_3h^l \, ,\quad N\in\mathbb{N} \, .$$ The functions ${f}_{n}(x)$ and ${g}_n(x) $ are defined, recursively, in the following manner, $$\begin{aligned}
{f}_n&=&-\frac{i}{2}{f}_{n-1}^{\,\prime}+\Delta \,
{g}_n \, ,\\
{g}_n^\prime &=& i\left(\Delta^*{f}_{n-1}-\Delta\, {f}_{n-1}^*\label{int}
\right)\, ,\\
{g}_0&=&1,\quad
{f}_{-1}=0 \, .
\label{recursion}\end{aligned}$$ The functions $f_{n}$ and $g_n$ depend on $\Delta(x)$ and its derivatives and also contain integration constants that appear due to the integration of (\[int\]); see Ref. [@BdG] for details.
The operator (\[S\]) satisfies the following commutation relation for any positive integer $N$, $$\begin{aligned}
\label{commutator}
[\tilde{S}_{N+1},\tilde{h}]&=&2i\left(\begin{array}{cc}0&f_{N+1}\\-f^*_{N+1}&0
\end{array}
\right) \, ,\end{aligned}$$ The stationary AKNS hierarchy of nonlinear differential equations is then defined in terms of the vanishing commutator (\[commutator\]), $$\label{AKNS1}
\text{AKNS}_N=f_{N+1}=0 \, .$$ The Hamiltonian $\tilde{h}$ and the operator $\tilde{S}_{N+1}$ are called the Lax pair of the stationary AKNS hierarchy. When a function $\Delta$ satisfies the $(N+1)^{\text{th}}$ order differential equation AKNS$_{N}$, all the next equations of the hierarchy with greater values than $N$ (and with the integration constants fixed appropriately[^4]) are immediately solved.
The operators $\tilde{h}$ and $\tilde{S}_{N+1}$ satisfy the remarkable algebraic relation, $$\label{BuchanalChaudy}
\tilde{S}_{N+1}^2=\prod_{n=0}^{2N+1}(\tilde{h}-\lambda_{n}) \, ,$$ where $\lambda_{n}$ are band-edge energies. The operator valued function on the righthand side is known as the spectral polynomial. The integral $\tilde{S}_{N+1}$ annihilates all the singlet eigenstates $\tilde{\phi}_n$ of $\tilde{h}$, corresponding to the band-edge energies, $(\tilde{h}-\lambda_n)\tilde{\phi}_n=0$, $$\label{Skernel}
\tilde{S}_{N+1} \tilde{\phi}_n=0, \quad n=0,1,...,2N+1 \, .$$
The local density of states $\rho(x,\lambda)$ is defined in terms of the trace of the Green’s function, $\tilde{R}(x,\lambda)\equiv \tilde{G}(x,x,\lambda)$, $$\begin{aligned}
\label{LDOS}
\rho(x,\lambda)&=&-\frac{1}{\pi}\lim_{\mbox{Im}\lambda\rightarrow 0_+}\mbox{Im}\,Tr \,\tilde{R}(x,\lambda)\,,\end{aligned}$$ where the trace is computed over matrix degrees of freedom. The function $\tilde{R}(x,\lambda)$ is also called diagonal resolvent or Gorkov resolvent.
The spatial integration of LDOS leads to the formula for DOS. In case of periodic quantum systems, the integration can be performed over one period $L$ [^5], $$\label{dos}
\text{DOS}=\frac{1}{L}\int_{L} \rho(x,\lambda)\, dx\,.$$
Explicit calculation of the Green’s function can be quite difficult. Nevertheless, the definition (\[LDOS\]) suggests that the need of its explicit knowledge might be avoided; only the diagonal resolvent is required to find LDOS. This fact was utilized in [@Dunne] and further developed in [@BdG]. Indeed, an exact form of the diagonal resolvent was found for a wide class of Hamiltonians (\[h1\]). The approach was based on the fact that $\tilde{R}(x,\lambda)$ has to satisfy the Dikii-Eilenberger equation [@eilenberger], $$\begin{aligned}
\frac{\partial}{\partial x}\tilde{R}(x; \lambda)\, \sigma_3-i\, \left[
\begin{pmatrix}
\lambda &-\Delta(x) \cr
\Delta^*(x) & -\lambda
\end{pmatrix}, \tilde{R}(x; \lambda)\,\sigma_3
\right]
&=&0\, ,\nonumber\\
\label{dikii}\end{aligned}$$ where $\lambda$ belongs to the spectrum of $\tilde{h}$. Additionally, $\tilde{R}(x,\lambda)$ has to satisfy the following requirements, $$\label{constrains}
\tilde{R}=\tilde{R}^{\dagger},\quad \det \tilde{R}=-\frac{1}{4}\, ,$$ where the latter one fixes the normalization of $\tilde{R}$. For more details on the properties of $\tilde{R}$ and derivation of (\[constrains\]), see e.g. the Appendix in [@kos].
Making the following ansatz for the diagonal resolvent [@BdG], $$\begin{aligned}
\tilde{R}(x; \lambda)=\sum_{n=0}^{N+1} \beta_n(\lambda)\, \begin{pmatrix}
{g}_n(x) & {f}_{n-1}(x)\cr
{f}_{n-1}^*(x) & {g}_n(x)
\end{pmatrix}\, ,
\label{poly}
\end{aligned}$$ and substituting (\[poly\]) into (\[dikii\]), the Dikii-Eilenberger equation transforms into the two (mutually conjugated) nonlinear differentials equations of the form of the AKNS hierarchy. The diagonal entries in (\[dikii\]) vanish identically due to the recurrence relations (\[recursion\]). The resulting equation can be written as $$\begin{aligned}
\label{AKNS}
&&\sum_{n=0}^{N+1}\beta_n(\lambda){f}_n-\lambda\sum_{n=0}^{N+1}\beta_n(\lambda){f}_{n-1}=0 \, ,\end{aligned}$$ which can be solved by fixing properly the constants $\beta_n(\lambda)$, see footnote $4$ and Ref. [@BdG]. It can be shown that the ansatz (\[poly\]) fulfills the requirements (\[constrains\]). [^6]
Making the inverse transformation (\[h1\]), we can find the Lax operator associated with the finite-gap Hamiltonian (\[eq1m\]) as $$\begin{aligned}
\label{Ss}
&{S}_{N+1}=\exp\left(i\frac{\sigma_1\pi}{4}\right)\tilde{S}_{N+1}\exp\left(-i\frac{\sigma_1\pi}{4}\right)& \\
&=-i\sum_{l=0}^{N+1}\left(g_{N+1-l}\mathbf{1}+\sigma_3\mbox{Im}f_{N-l}+\sigma_1\mbox{Re}f_{N-l}\right)\sigma_2h^l& \, .\notag\end{aligned}$$ The diagonal resolvent for the Hamiltonian $h$ can be obtained directly from (\[eq1m\]), since the trace of an operator is invariant with respect to similarity transformations.
Exactly solvable models of the twisted nanotubes {#examples}
================================================
The periodic systems described by (\[eq1\]) can be classified in terms of a quantity which we call average twist. It is defined as $$\label{constant}
\Sigma_c =\frac{\mbox{max}(\Sigma)+\mbox{min}(\Sigma)}{2},$$ and corresponds to the value around which the potential is oscillating. We will present two- and four-gap systems, denoted as $\Sigma(x)=\Delta_2$ and $\Sigma(x)=\Delta_4$, respectively, where the average twist is vanishing. Then we will consider two simple cases where it acquires nonzero values. They correspond to the one-, $\Sigma(x)=\Delta_1$, and three-gap $\Sigma(x)=\Delta_3$ systems. The mass term will be identically zero in all these models, $M=0$. We will see that the actual value of the average twist is in correlation with the qualitative spectral properties of these models.
As the last example, we will consider a nonperiodic system with a constant mass, $M\neq0$. It will serve for illustration of a twisted boron-nitride nanotube.
Configurations with zero average twist
--------------------------------------
### Two-gap system
First, let us consider the system governed by (\[eq1\]) with the vector potential $$\label{crystal1}
\Delta_{2} = m k^2 \frac{{\rm sn }\,m x \, {\rm cn }\, mx}{{\rm dn }\,m x} \, ,$$ where $m$ is a real parameter and $k\in(0,1)$. This vector potential is induced by the deformation specified by the following displacement vector (see Fig. \[twist\]), $${\bf d}=(0,-\zeta\, \ln {\rm dn }\,m x) \, .$$
[cc]{} ![Illustration of the two-gap (upper), three-gap (middle) and four-gap (lower) configurations of the twisted carbon nanotubes. In the system without twist, the black line would be straight. []{data-label="twist"}](dib1.eps "fig:")\
\
![Illustration of the two-gap (upper), three-gap (middle) and four-gap (lower) configurations of the twisted carbon nanotubes. In the system without twist, the black line would be straight. []{data-label="twist"}](dib2.eps "fig:")\
\
![Illustration of the two-gap (upper), three-gap (middle) and four-gap (lower) configurations of the twisted carbon nanotubes. In the system without twist, the black line would be straight. []{data-label="twist"}](dib3.eps "fig:")
The crystal kink two-gap potential (\[crystal1\]) is given in terms of doubly periodic Jacobi elliptic functions depending on the modular parameter $k$. It has a real period $L=2 K(k)$, where $K(k)$ is the complete elliptic integral of the first kind. For the definitions and properties of the elliptic functions, we recommend Refs. [@Abramovitz; @WW]. The infinite-period limit ($k \to 1$) of (\[crystal1\]) is called the single kink vector potential $\Delta_2= \tanh x$. Let us notice that the properties of the Dirac electron in graphene in the presence of a single-kink-type vector potential were analyzed in [@milpas]. The potential $\Delta_{2}$ vanishes in the limit when the modular parameter goes to zero.
The spectrum of the one-dimensional Dirac Hamiltonian $h(\mathbf{K})$ has two gaps located symmetrically with respect to zero. The band-edge energies are $\lambda_0=-\lambda_3=-m$ and $\lambda_1=-\lambda_2=-m \sqrt{1-k^2}$. The corresponding eigenstates ($(h(\mathbf{K})-\lambda_n)\phi_n=0$, $n=0,1,2,3$) are $$\begin{aligned}
\label{2gf1}
\phi_{0}=&\left( -{\rm sn }\,m x ,\, \frac{{\rm cn }\,m x}{{\rm dn }\,m x}\right)^t\, ,& \phi_{3}= \sigma_3 \phi_{0} \, ,\\
\phi_{1}=&\left(\frac{1}{\sqrt{1-k^2}} {\rm cn }\,m x ,\, \frac{{\rm sn }\,m x}{{\rm dn }\,m x} \right)^t \, ,& \phi_{2}= \sigma_3 \phi_{1} \, .
\label{2gf2}\end{aligned}$$ The band-edge energies are nondegenerate, while the energies from the interior of the bands are doubly degenerated.
Using directly the formula (\[poly\]) for $N=2$, we can find the explicit form of the diagonal resolvent. Its trace then reads $$\label{2gaptrace}
{\rm Tr}\, R_2(x; \lambda)= \frac{\lambda^2+\frac{1}{2}\Delta_2^2}{\sqrt{(m^2-\lambda^2)(\lambda^2+m^2(k^2-1))}} \, ,$$ and the associated density of states acquires the following form $$\label{dos2}
\text{DOS}_{2}=-\frac{1}{\pi}\lim_{\mbox{Im}\lambda\rightarrow 0_+}\mbox{Im}\,\frac{\lambda^2-m^2\frac{E(k)}{K(k)}}{\sqrt{(m^2-\lambda^2)(\lambda^2+m^2(k^2-1))}} \, ,$$ where we have used Eqs. (\[LDOS\]) and (\[dos\]). Notice that $\text{DOS}_{2}$ is identically zero when $\lambda$ belongs to the prohibited gaps. The function in the argument is purely real for these values of $\lambda$ and, thus, the imaginary part is vanishing identically, see Fig. \[dostwogap\].
![Illustration of DOS (\[dos2\]) of the Hamiltonian (\[eq1m\]) with $\Sigma(x)=\Delta_{2}$ and $M=0$ with $m=1.5$ and $k=0.7$.[]{data-label="dostwogap"}](dos2gap.eps)
### Four-gap system
As the next example, we shall consider the $2 K(k)$-periodic system described by the Hamiltonian (\[eq1\]) with the vector potential $$\label{crystal4}
\Delta_4 = 6m k^2 \frac{ {\rm sn }\,m x \, {\rm cn }\, mx \, {\rm dn }\, mx}{1+k^2+\delta-3k^2{\rm sn }^2\,m x }\, ,$$ where $\delta=\sqrt{1-k^2+k^4}$. The crystal kink four-gap potential (\[crystal4\]) is an isospectral deformation of the crystal kink potential $\Delta_{4'} = 2 m k^2 \frac{{\rm sn }\,m x \, {\rm cn }\, mx}{{\rm dn }\,m x}$. Both potentials reduce to the single kink $\Delta_{4} =\Delta_{4'} =2 m \tanh mx$ when $k\to 1$. The associated displacement vector in this case takes the form $${\bf d}=(0,- \zeta\,\ln (1+k^2+\delta-3k^2{\rm sn }^2\,m x))$$ and is illustrated in Fig. \[twist\]. The spectrum of $h(\mathbf{K})$ has five bands and eight band-edge states $\phi_n$, $n=0,...,7$, which can be defined with help of and an operator ${\cal D}=\frac{d}{dx}+\Delta_4$ as follows, [^7] $$\label{phi}
\phi_n= \left(\psi_n,\frac{1}{\lambda_n}{\cal D}\psi_n\right)^t,\quad (h(\mathbf{K})-\lambda_n)\phi_n=0.$$ Keeping in the mind the spectral symmetry $\lambda\leftrightarrow -\lambda$ (which is valid for any model (\[eq1m\]) with $M=0$), it is sufficient to find just the first four eigenstates $\phi_0,\ \phi_1,\ \phi_2,\ \phi_3$, since the remaining four can be obtained as $\phi_{n+4}=\sigma_3\phi_n$, where $n=0,1,2,3$. They are given in terms of the following functions
&\_0=m(1+k\^2--3k\^2[sn]{}\^2 mx ), \_0=-2m ,&\
&\_1=[cn]{} x [sn]{} x , \_1=-m ,&\
&\_2 =[dn]{} x [sn]{} x , \_2=-m , &\
&\_3= [cn]{} x [dn]{} x , \_3=-m ,&
where $\lambda_n$ are the corresponding eigenvalues.
The local density of states can be computed using the method described in the preceding section, $$\label{4gaptrace}
{\rm Tr}\, R_4(x; \lambda)=\frac{\alpha_1+\alpha_2\Delta_4^2 +3\Delta_4^4+\Delta_4'^2-2\Delta_4 \Delta_4''}{8\sqrt{(\lambda_0-\lambda^2)(\lambda^2 -\lambda_1^2)(\lambda^2 -\lambda_2^2)(\lambda^2 -\lambda_3^2)}},$$ where the constants are $$\begin{aligned}
\alpha_1 &=& 8(\lambda^4-5m^2\delta\lambda^2+4m^4\delta), \\
\alpha_2 &=& 4(\delta^2-5m^2\delta^2) \, .\end{aligned}$$ The explicit (analytical) form of the density of states is rather cumbersome. In Fig. \[dosfourgap\], we present the numerical computed DOS of the current four-gap system.
![Illustration of DOS of the Hamiltonian (\[eq1m\]) with $\Sigma(x)=\Delta_{4}$ and $M=0$ with $m=1$ and $k=0.6$.[]{data-label="dosfourgap"}](dos4gap.eps)
Let us note that for the system described by the isospectral potential $\Delta_{4'}$, the resolvent trace $$\label{4gaptraceb}
{\rm Tr}\, R_{4'}(x; \lambda)=\frac{\eta_1+\eta_2\Delta_{4'}^2+\eta_3 \Delta_{4'}^4}{8\sqrt{(\lambda_0-\lambda^2)(\lambda^2 -\lambda_1^2)(\lambda^2 -\lambda_2^2)(\lambda^2 -\lambda_3^2)}}$$ can be written just in terms of a polynomial in $\Delta_{4'}$, where $\eta_1$, $\eta_2$ and $\eta_1$ are constant depending on $\lambda$.
Semiconducting carbon nanotubes via nonzero average twist
---------------------------------------------------------
### One-gap system
The simplest example of a finite-gap system with the nonzero average twist is given by the Hamiltonian (\[eq1\]) with the constant vector potential $\Sigma(x)=\gamma$. The two band-edge energies correspond to $\lambda_0=-\lambda_1=-\gamma$. The local density of states can be found in the following form $$\label{0gaptrace}
{\rm Tr}\, R_4=\frac{\lambda}{\sqrt{\gamma^2-\lambda^2}}\, .$$ The constant potential can be regarded as periodic with the period being equal to any real number $L$. We can compute the average twist as $\Sigma_c=\gamma$.
The spectrum of the system has two bands separated by a gap of width $2\Sigma_c$. This suggests that the central gap is twice the average twist.
### Three-gap system
Let us test the suggestion in the case of a more complicated system. Its Hamiltonian (\[eq1\]) has the $2K(k)$-periodic vector potential $$\label{crystal2}
\Delta_3=\frac{{\rm cn}\,b\, {\rm dn}\,b}{{\rm sn}\,b}+k^2\, {\rm sn}\,b\,{\rm sn}
\,(x) \,{\rm sn}\,\left(x+b\right),$$ which is called the crystal kink-antikink, three-gap potential, [@BdG]. The real parameter $b\in(0, K(k))$ represents the distance between the kink and the antikink.
The vector potential is induced by the displacement ${\bf d}=(0,\zeta\,F(x))$, where $F(x)$ is as follows, $$\notag
F(x)\!=\!\frac{{\rm cn}\,b\, {\rm dn}\,b}{{\rm sn}\,b} \, \Pi(k^2 \, {\rm sn}^2 b,{\rm am}\; x |k) -\frac{1}{2}\ln \left(\!1\!-\! k^2 \, {\rm sn }^2 b \, {\rm sn }^2 x\right).$$ The function $\Pi(a;x | \phi)$ is the incomplete elliptic integral of the third kind and ${\rm am}\; x $ is the Jacobi amplitude. See Fig. \[twist\] for illustration.
When $b= K(k)$, (\[crystal2\]) is reduced to the two-gap vector potential (\[crystal1\]). In the infinite period limit, the single kink-antikink solution is recovered [@Jackiw], $\lim_{k\rightarrow1}\Delta_1(x)=\coth b +\tanh x -\tanh (x+b)$.
The spectrum of (\[eq1\]) with (\[crystal2\]) contains three gaps positioned symmetrically with respect to zero. The three band-edge states with negative energies are $$\begin{aligned}
\notag
\phi_{0}&=&\left( -{\rm sn }\,(x),\,\,{\rm sn }\,(x+b) \right)^t \, , \quad \lambda_{0}=-\frac{1}{{\rm sn }\,b} \, , \\
\phi_{1}&=&\left(-{\rm cn }\,(x) ,\,\, {\rm cn }\,(x+b)\right)^t \, , \quad \lambda_{1}=-\frac{{\rm dn }\,b}{{\rm sn }\,b} \, , \notag \\
\phi_{2}&=&\left( -{\rm dn }\,(x),\,\, {\rm dn }\,(x+b) \right)^t \, , \quad \lambda_{2}=-\frac{{\rm cn }\,b}{{\rm sn }\,b}\, . \label{states3}\end{aligned}$$ The positive energy states are obtained as $\phi_{n+3}=\sigma_3\phi_n$ and correspond to the energies $\lambda_{n+3}=-\lambda_n$, where $n=0,1,2$.
The trace of diagonal resolvent can be computed from (\[poly\]) for $N=2$ in the following form $$\label{3gaptrace}
{\rm Tr}\, R_3(x; \lambda)=\frac{\frac{\lambda}{2}\left(\alpha+\Delta^2_3 \right)}{\sqrt{( \lambda_0^2-\lambda^2)(\lambda_1^2-\lambda^2 )(\lambda_2^2-\lambda^2 )}}\, ,$$ where $\alpha=1+k^2+2\lambda^2-\frac{3}{{\rm sn}^2\,b}$. The actual integration of the formula above, needed for analytical form of DOS, is rather complicated. We present Fig. \[dosthreegap\] of DOS for the three-gap case that was obtained numerically.
The average twist associated with the potential (\[crystal2\]) can be found as $$\Sigma_c=\frac{{\rm cn}\, b}{{\rm sn}\, b}.$$ Checking the corresponding band-edge energies $\lambda_2$ and $\lambda_3$ in (\[states3\]), we can see that that the gap between the positive and negative energies is exactly of width $2\Sigma_c$.
Comparing the spectra of the presented systems, we can see that the nonvanishing average twist (\[constant\]) is proportional to the magnitude of the central spectral gap in the system. In the two- and four-gap systems, the average twist is vanishing and there is no gap between positive and negative energies. These nanotubes are conducting in the sense that infinitesimal excitation is sufficient to kick the electrons from valence band to conduction band. The systems with the nonvanishing average twist are different. They have a gap between positive and negative energies that are equal to $2\Sigma_c$ and, hence, are semiconducting.
![Illustration of DOS of the Hamiltonian (\[eq1m\]) with $\Sigma(x)=\Delta_{3}$ and $M=0$ with $k=0.2$ and $b=1.5$.[]{data-label="dosthreegap"}](dos3gap.eps)
Boron-nitride nanotubes
-----------------------
As the last example, we shall consider a *nonperiodic* system with the nontrivial mass term. We take the potential term of (\[eq1m\]) in the following form, $$\label{complexcrystal}
\Sigma(x)= N \sin \frac{\theta}{2} \tanh(\sin \frac{\theta}{2}x ), \quad M=-N \cos \frac{\theta}{2},$$ where $N$ is a positive integer and $\theta$ is a real parameter. The potential is classified as $(N+1)$-gap as it solves the corresponding equation of the AKNS. It has $N+2$ singlet states in the spectrum. Two of them correspond to the energies that form the threshold of the continuum spectrum, the rest is associated with bound states of the system. In the case of $N=1$, the eigenstates are then given as $$\begin{aligned}
&\psi_0=\left( \tanh \left(\sin \frac{\theta}{2}x \right),-\cot \frac{\theta}{4} \right)^t,& \, &\lambda_0=-1,& \\
&\psi_1=\left({\rm sech} \, \left(\sin \frac{\theta}{2}x\right),\,0\right)^t,& \, &\lambda_1=\cos \frac{\theta}{2},& \\
&\psi_2=\left(\tanh \left(\sin \frac{\theta}{2}x \right),\tan \frac{\theta}{4} \right)^t,& \, &\lambda_2=1.&\end{aligned}$$ The trace of the diagonal resolvent can be computed in the following manner, $$\label{complextrace}
{\rm Tr}\, R(x; \lambda)=\frac{\lambda^2-\lambda \cos \frac{\theta}{2}-\frac{1}{2}\sin^2 \frac{\theta}{2} \, {\rm sech} ^2\, \left(\sin \frac{\theta}{2}x \right) }{\sqrt{1-\lambda^2}(\lambda-\cos \frac{\theta}{2})} \, .$$
It is worth noticing that in the examples of the carbon nanotubes, the trace of diagonal resolvent (\[2gaptrace\]), (\[4gaptrace\]), (\[4gaptraceb\]), and (\[3gaptrace\]) could be written in terms of the finite-gap potential and its derivatives. A similar result is known for the Schrödinger systems with Lamé potential. The trace of the diagonal resolvent corresponded in that case to a polynomial of the finite-gap potential [@belokolos]. This is related to the fact that the square of the Dirac operator with the four-gap potential $\Delta_{4'}$ corresponds to an extended Schrödinger operator with two-gap Lamé potential.
In the current case with the nonvanishing mass term, the trace of diagonal resolvent can be written as a function of the amplitude of the corresponding complex potential $\Delta$ \[see (\[h1\])\], where $|\Delta|^2=\Sigma(x)^2+M^2$. For (\[complexcrystal\]), we can write $$\nonumber
{\rm Tr}\, R(x; \lambda)=\frac{P_{2N}(x,\theta)}{\sqrt{N^2-\lambda^2}\prod_{n=1}^{2N}(\lambda-\lambda_n)^2}.$$ Here, $$\label{complexgeneral}
P_{2N}(x,\theta)=\sum_{n=0}^N c_n(\lambda) \left( \sin^2 \frac{\theta}{2} \, {\rm sech} ^2\, (\sin \frac{\theta}{2}x) \right)^n$$ with $c_n(\lambda) $ being specific constants [^8].
van Hove singularities and the nonlinear supersymmetry {#susy}
======================================================
The densities of states have a set of singular points that are called van Hove singularities. A closer inspection of the corresponding formulas (\[2gaptrace\]), (\[4gaptrace\]), (\[0gaptrace\]), (\[3gaptrace\]), and (\[complextrace\]) shows that the number as well as the position of the van Hove singularities coincide precisely with the singlet band-edge energies of the finite-gap systems. In this section, we will show that this coincidence is reflected by a nonlinear supersymmetry that underlies the finite-gap configurations of the twisted nanotubes.
Quantum systems in presence of a magnetic field cease to be time-reversal invariant. The time-reversal operator changes the sign of momentum while it preserves the coordinate. It changes the sign of the magnetic field. It can be represented by an anti-unitary operator $\mathcal{T}$ that satisfies $\mathcal{T}^{\dagger}\mathcal{T}=1$, $\mathcal{T}i\mathcal{T}=-i$ and $\mathcal{T}^2=-1$. The latter equality arises due to the half-integer spin of the considered particles.
One can check that the anti-unitary operator $\sigma_2 T$ ($T$ denotes complex conjugation, $T^2=1$) does not commute with the Hamiltonian (\[eq1m\]) due to the symmetry breaking term $\Sigma(x)$ (and $M$). However, we have to keep in mind that these terms arise from the tight-binding model, which, despite the deformations of the crystal, is time-reversal invariant [@CastroNeto].
The time-reversal symmetry of the system in the low-energy regime emerges when dynamics in both valleys corresponding to $\mathbf{K}$ and $\mathbf{K'}$ is taken into account. The total Hamiltonian reads $$\label{totalh}
\mathcal{H}=\left(\begin{array}{cc}h(\mathbf{K})&0\\0&h(-\mathbf{K})\end{array}\right)\, ,$$ where the energy operators $h(\pm\mathbf{K})$ of the subsystems are given as $$\label{htwovalley}
h(\pm\mathbf{K})=-i\sigma_2\partial_{x}\pm \Sigma(x)\sigma_1 \pm M\sigma_3\, .$$ The operator (\[totalh\]) acts on the bispinors $\Psi=(\psi_{\mathbf{K}A},\psi_{\mathbf{K}B},\psi_{\mathbf{K}'B},\psi_{\mathbf{K}'A})$, where we use the notation introduced in the second section below (\[singlevalleyhd\]).
The Hamiltonian $\mathcal{H}$ commutes with the time-reversal operator $\mathcal{T}$ which is defined in the following manner[^9], $$\label{tr}
[\mathcal{H},\mathcal{T}]=0,\quad \mathcal{T}=\sigma_1\otimes\sigma_2\, T.$$ As the considered system consists of a single fermion, the Kramer’s theorem applies; all the energy levels of (\[totalh\]) have to be at least doubly degenerate. In case of a periodic system, the band structure of $h(\mathbf{K})$ is determined by $2N+2$ nondegenerate band-edge energies $\lambda_n$. In the infinite period limit, the operator has $N+2$ singlet states. As we can see from (\[htwovalley\]), the operators $h(\pm\mathbf{K})$ are unitarily equivalent, $h(\mathbf{K})=\sigma_2h(-\mathbf{K})\sigma_2$. Hence, $\mathcal{H}$ has the same band structure as $h(\mathbf{K})$, but the degeneracy is doubled as is required by the Kramer’s degeneracy theorem.
Degeneracy of energy levels is reflected by a set of integrals of motion that are based on the Lax integral $S_{N+1}$, see (\[Ss\]). In the individual subsystems governed by $h(\pm\mathbf{K})$, the degeneracy is associated with two diagonal operators, $\mathcal{Q}_0$, and $\mathcal{Q}_3$, $$\label{q03}
\mathcal{Q}_0=\left(\begin{array}{cc}S_{N+1}&0\\0&\sigma_2S_{N+1}\sigma_2\end{array}\right),\quad \mathcal{Q}_3=\tau_{30}\mathcal{Q}_0,$$ where $\tau_{30}=\sigma_3\otimes \mathbf{1}$. The intervalley (Kramer’s) degeneracy is naturally reflected by the operators $\mathcal{Q}_{1}$ and $\mathcal{Q}_{2}$, $$\label{set3}
\mathcal{Q}_1=\tau_{12}\mathcal{Q}_0,\quad \mathcal{Q}_2=\tau_{22}\mathcal{Q}_0,$$ where $\tau_{ab}=\sigma_a\otimes\sigma_b$, $a,b=1,2$. All these operators commute with the total Hamiltonian, $$\label{set4}
[\mathcal{Q}_a, \mathcal{H}]=0\, .$$ By construction, these operators close Lie algebra $so(3)\oplus u(1)$, $$\label{set2}
[\mathcal{Q}_0,\mathcal{Q}_a]=0,\quad[\mathcal{Q}_a,\mathcal{Q}_b]=2i\varepsilon_{abc}\mathcal{Q}_c, \quad a,b,c=1,2,3.$$ The existence of the operators (\[set3\]) is a direct consequence of the time-reversal symmetry of (\[totalh\]). Indeed, (\[tr\]) implies the unitary equivalence of the valley Hamiltonians $h(\pm\mathbf{K})$ and enables the construction of antidiagonal operators (\[set3\]).
The action of the integrals is quite nontrivial and determined by the properties of the Lax operator $S_{N+1}$. It can be inferred from (\[Skernel\]) that all doublet states $\Psi_{2-deg}$ of $\mathcal{H}$, corresponding to the band-edge energies $\lambda_n$, are annihilated by the integrals of motion $ \mathcal{Q}_a $, $$\mathcal{Q}_a \Psi_{2-deg}=0, \quad a=0,1,2,3 \, .$$ Let us denote by the subscript $_{\mathbf{K}}$ and $_{\mathbf{K}'(=-\mathbf{K})}$ the states that are nonvanishing in one valley only, i.e. $\frac{1}{2}\left(1\pm\tau_{30}\right)\Psi_{\pm \mathbf{K}}=\Psi_{\pm \mathbf{K}}$. We can find mutual eigenstates $\Psi^{\pm}_{\mathbf{K}}$ and $\Psi^{\pm}_{\mathbf{K'}}$ of the Hamiltonian $\mathcal{H}$, the valley-index operator $\tau_{30}$, and the integrals $\mathcal{Q}_0$ and $\mathcal{Q}_3$. They satisfy the following relations, $$(\mathcal{H}-\lambda)\Psi^{\pm}_{\mathbf{K}(\mathbf{K}')}=0,\quad (\tau_{30}-1)\Psi^{\pm}_{\mathbf{K}}=(\tau_{30}+1)\Psi^{\pm}_{\mathbf{K'}}=0$$ and $$\label{r1}
\mathcal{Q}_i \Psi^{\pm}_{\mathbf{K}} = \pm\gamma_\lambda {\Psi}^{\pm}_{\mathbf{K}}, \quad \mathcal{Q}_i \Psi^{\pm}_{\mathbf{K'}} = \pm\gamma_\lambda {\Psi}^{\pm}_{\mathbf{K'}} \quad i=0,3 \, .$$ The eigenvalues $\gamma_{\lambda}$ can be determined from the spectral polynomial (\[BuchanalChaudy\]) as $$\label{gamma}
\gamma_\lambda = \sqrt{P(\lambda)}= \prod_{n=0}^{2N+1}(\lambda-\lambda_{n})^{1/2}\,.$$ Hence, the operators $\mathcal{Q}_0$ and $\mathcal{Q}_3$ act on the basis of $\Psi^{\pm}_{\pm\mathbf{K}}$ as the multiplication by $\sqrt{P(\lambda)}$, i.e. as the square root of the operator $P(\mathcal{H})$. As mentioned above, the roots of the spectral polynomial (\[gamma\]) coincide with the van Hove singularities of the analyzed finite-gap systems. The two antidiagonal operators $\mathcal{Q}_1$ and $\mathcal{Q}_2$ switch the valley index, $$\label{r2}
\mathcal{Q}_1 \Psi^{\pm}_{\mathbf{K}(\mathbf{K}')} = \pm\gamma_\lambda {\Psi}^{\pm}_{\mathbf{K}'(\mathbf{K})}, \quad \mathcal{Q}_2 \Psi^{\epsilon}_{\pm\mathbf{K}} = \pm i\epsilon\gamma_\lambda {\Psi}^{\epsilon}_{\mp\mathbf{K}} \, ,$$ where $\epsilon=\pm$.
The action of the operators $\mathcal{Q}_a$ on the valley index is not indicated by the algebra (\[set2\]). To reflect better the properties of the system, we can define a superalgebra graded by the valley index operator $\tau_{30}$. We denote $\mathcal{F}_{1(2)}\equiv \mathcal{Q}_{1(2)}$ as fermionic operators that change the valley index of the wave functions ($\{\mathcal{F}_{1(2)},\tau_{30}\}=0$) and $\mathcal{B}_{1(2)}\equiv\mathcal{Q}_{0(3)}$ as bosonic operators that preserve the valley index ($[\mathcal{B}_{1(2)},\tau_{30}]=0$). The superalgebra is nonlinear and contains other fermionic operators $\tau_{12}$ and $\tau_{22}$, $$\begin{aligned}
\label{susy1}
&&[\mathcal{H},\mathcal{B}_a]=[\mathcal{H},\mathcal{F}_a]=0,\ \{\mathcal{F}_a,\mathcal{F}_b\}=2\delta_{ab}P(\mathcal{H}),\\
&& [\mathcal{B}_a,\mathcal{F}_{b}]=2i\,\delta_{a2}\,\epsilon_{3bc}\,\tau_{c2}\,P(\mathcal{H}),\\
&& [\mathcal{B}_a,\tau_{22}]=-2i\delta_{2a}\,\mathcal{F}_{1},\quad [\mathcal{B}_a,\tau_{12}]=2i\,\delta_{2a}\,\mathcal{F}_{1},\\
&& \{\mathcal{F}_a,\tau_{b2}\}=2\delta_{ab}\,\mathcal{B}_1.\label{susy3}\end{aligned}$$ The fact that we deal with finite-gap systems is manifested in the anticommutator of the fermionic operators where the spectral polynomial $P(\mathcal{H})$ emerges naturally. It underlies nonlinearity of the superalgebra and manifests the intimate relationship of between the algebraic structure and the van Hove singularities of the considered models.
Let us stress that the superalgebra (\[susy1\])-(\[susy3\]) exists for any finite-gap configuration of the twisted nanotubes described by $\mathcal{H}$ as long as the Hamiltonian possesses the time-reversal symmetry.
The choice of the grading operator was not unique. We could use either $\tau_{12}$ or $\tau_{22}$ equally well; both of them either commute or anticommute with the considered operators. Notice that $\tau_{12}$ corresponds to the unitary component of the time-reversal (\[tr\]). Choosing any of them as the new grading operator, qualitatively the same superalgebra would be obtained. The operators (\[q03\])-(\[set3\]) would be just permuted in the role of bosonic and fermionic generators.
Let us notice that in examples presented in the previous section, the single valley Hamiltonians with the vector potentials (\[crystal1\]) and (\[crystal4\]) commute with the operator $\sigma_3\mathcal{R}$ where $\mathcal{R}$ is the parity[^10], $\mathcal{R}x\mathcal{R}=-x$. Hence, the corresponding Hamiltonian (\[totalh\]) is commuting with $\tau_{33}\mathcal{R}$. The latter operator also commutes with $\tau_{30}$ and $\tau_{22}$, whereas it anticommutes with all the operators $\mathcal{Q}_a$, $a=0,..,3$. It means that $\tau_{33}\mathcal{R}$ could be regarded as a grading operator of a distinct, $N=4$ superalgebra that would be generated by four fermionic operators (\[q03\]) and (\[set3\]). The nonlinear superalgebra of Bogoliubov-de Gennes Hamiltonians generated by nonlocal supercharges was discussed in the literature. We refer to [@BdG] for more details, see also [@varios; @papers].
The formulas for LDOS and DOS computed in the third section with the use of the formula (\[poly\]) have to be multiplied by four to get the correct form for the corresponding twisted nanotubes. Indeed, we have to take into account the valley degeneracy that we discussed in this section, as well the double degeneracy of energy levels due to (real) spin$-\frac{1}{2}$ of the particle that was neglected up this moment.
Finally, let us discuss briefly the settings where an external magnetic field is present in addition to the twists. The magnetic field breaks the time-reversal symmetry. When the vector potential $\Delta_{mg}$ is included into the Hamiltonian, we have $$\label{totalhmg}
h(\pm\mathbf{K})=-i\partial_{x}\sigma_2\pm \Sigma(x) \sigma_1\pm M\sigma_3+\Delta_{mg}\sigma_1.$$ We can see that as long as mass term $M$ is vanishing and either magnetic field *or* twists are switched on (i.e. $\Sigma(x)\Delta_{mg}=0$), all the energy levels have even degeneracy due to the unitary equivalence of $h(\mathbf{K})$ and $h(\mathbf{K'})$. The situation changes when both $\Sigma(x)$ and $\Delta_{mg}$ are nonzero. In that case, we can still have a finite-gap configuration in one valley described by $h(\mathbf{K})$. However, in the second valley the finite-gap potential is violated in general by the changed sign of $\Delta_{mg}$.
Curiously enough, we can still get a finite-gap configuration for each valley by the fine-tuning of the external field. As an example, let us consider the situation when the low-energy dynamics in the $\mathbf{K}$ valley is described by $$\label{twistmg1}
h(\mathbf{K})=-i\partial_x\sigma_2+(\coth b+\tanh x-\tanh (x+b))\sigma_1,$$ which is an infinite-period limit of the three-gap system (\[crystal2\]). Let us suppose that the vector potential in (\[twistmg1\]) is induced both by radial twist and by external magnetic field, where $\Sigma(x)=\frac{1}{2}\left(2\coth b+\tanh x-\tanh (x+b)\right)$ and $\Delta_{mg}=\frac{1}{2}\left(\tanh x-\tanh (x+b)\right)$. Then the subsystem in the $\mathbf{K}'$-valley is described by $$\label{twistmg2}
h(\mathbf{K'})=-i\partial_x\sigma_2-\coth b\, \sigma_1,$$ which is just the trivial one-gap system. In the current setting, deformation associated with $\Sigma(x)$ is asymptotically uniform but gets changed in the localized region where the (asymptotically vanishing) external magnetic field is nonzero. The spectrum of the corresponding total Hamiltonian $\mathcal{H}$ has two singlet discrete energy levels corresponding to the bound states and two doubly degenerate levels $\pm \coth\, b$ corresponding to the threshold of the positive and negative continuum. The other energy levels are four-fold degenerate.
It is worth noticing that the discussed framework can be understood in the context of (planar) graphene crystal in the presence of the external magnetic field and strain, both of which depend on $x$ only. Due to separability of the stationary equation, the one-dimensional Hamiltonian can be written as $$\label{plane}
h(\mathbf{K})=-i\sigma_2\partial_{x}+(k_y+A_y(x))\sigma_1,$$ where $k_y$ corresponds to the momentum that is parallel with the (pseudo-) magnetic barrier. The operator (\[plane\]) describes a massless Dirac particle that moves with fixed direction in the presence of vector potential $A_y$, associated with the strain and the external magnetic field. In this context, the setting with the single-valley Hamiltonians (\[twistmg1\]) and (\[twistmg2\]) with the inhomogeneous external magnetic field perpendicular to the surface and given by $\Delta_{mg}$ is rather realistic.
Discussion and Outlook {#conclusion}
======================
The one-dimensional Dirac operator with finite-gap potential appears in a variety of physically interesting models [@Dunne], [@BdG], [@Feinberg], [@PlyushchayBdG]. In the present paper, we illustrated how the machinery of the AKNS hierarchy can be used in the analysis of the twisted nanotubes in the low-energy regime, particularly, for the computation of the local density of states.
We showed that the finite-gap, time-reversal invariant configurations possess a hidden nonlinear supersymmetry that is associated with the Kramer’s degeneracy of energy levels. Physics of these systems, namely the presence of the two valleys at $\mathbf{K}$ and $\mathbf{K'}$ and the preserved time-reversal symmetry, is responsible for the form of the Hamiltonian (\[totalh\]) which consists of two copies of the (unitarily) equivalent single-valley energy operators.
The current situation differs from the quantum models with bosonized supersymmetry [@bosonized], where a nonlocal integral of motion was identified as the grading operator. Both the Hamiltonian (\[totalh\]) and its integrals of motion (\[q03\]) and (\[set3\]), forming $so(3)\oplus u(1)$ Lie algebra, can be graded by a local operator, e.g. by the valley index operator $\tau_{30}$. This framework represents a nontrivial example of the hidden supersymmetry in the sense that it naturally emerges within the unextended, physical Hamiltonian (\[totalh\]).
The explicit results for the presented finite-gap systems can be extended with the use of Darboux transformation [@DiracDarboux]. Within this framework, one can construct new finite-gap Hamiltonians $h_2$ from a known one, namely $h_1$. The transformation is given in terms of a matrix differential operator $D$, which intertwines two one-dimensional Dirac Hamiltonians, $Dh_1=h_2D$. It maps the eigenstates of $h_1$ into those of $h_2$, keeping the operators (almost) isospectral. Moreover, the diagonal resolvent of $h_2$ can be computed directly from the diagonal resolvent of $h_1$ with the use of $D$, see [@twisting] for details.
In the paper, the operator (\[eq1m\]) was almost exclusively interpreted as the effective Hamiltonian of the twisted carbon (or boron-nitride) nanotube. As we discussed in the end of the preceding section, the results can be also used in the analysis of the Dirac electrons in graphene in the presence of (pseudo-)magnetic barriers. Such systems with Kronig-Penney or a piece-wise constant (pseudo-) magnetic fields induced by either external field or strains were considered in the literature, see [@KP] or [@steplikeB]. In this context, the DOS computed for the twisted nanotubes can be interpreted as the partial density of states in graphene for the $k_y=0$ channel. It could facilitate the computation of the transition coefficient in the normal direction to the magnetic barrier. The known results [@partialLDOS] on the relation between one-dimensional DOS and the phase of the transmission amplitude could be particularly helpful in this context.
Considering Dirac electrons in graphene, it is desirable to extend the analysis for $k_y\neq 0$ as well. Keeping in mind [@milpas] or [@Miserev], the infinite-period limit of the finite-gap models could be a feasible starting point in this respect. The analysis of periodic systems could make it possible to observe the phenomena that appear in graphene superlattices, the new generation of Dirac points in particular [@newgeneration], [@steplikeB2].
Study of the finite-gap configurations of electrostatic potential represents another possible direction for future research. Spectral properties of Dirac electron in graphene in presence of both periodic electrostatic and magnetic fields were discussed in various works, see e.g. [@Novikov], [@Dugaev], [@Wang]. In this context, the mapping between the systems with magnetic and electrostatic field [@tan] could provide an interesting way to extend our results.
The finite-gap systems are an approximation of more realistic settings. They can serve as a test field for numerical or perturbative methods and can also provide qualitative insight into the experimental data. Although, to our best knowledge, the experiments with the single-wall carbon nanotubes with the periodically modulated twist have not been prepared yet, the building blocks of such settings seem to be available, see e.g. [@pendulum]-[@rotors].
As an experimental implementation of the proposed models, we can imagine a long suspended nanotube anchored to a substrate at the ends and a periodic array of small paddles attached to it. By deflection of the paddles, the twist of the nanotube could be altered. Let us mention that the latter configuration (with one paddle) was employed in [@torsionalprops] for measurement of the torsional properties of the nanotubes.
In the presented finite-gap models, the possible interaction of the nanotube with the anchors and the paddles is not taken into account. Still, we think that the results (e.g. the suggested dependence of the central gap on the average twist) might provide interesting qualitative insight into the spectral properties of the settings realized in the experiments.
The authors would like to thank to Gerald Dunne and Dmitry V. Kolesnikov for discussions. F.C. was supported by the Fondecyt Grant No. 11121651 and by the Conicyt grant 79112034 and ACT-91. F.C. wishes to thank the kind hospitality of the Nuclear Physics Institute of the ASCR and the Physics Department of the University of Connecticut. The Centro de Estudios Científicos (CECs) is funded by the Chilean Government through the Centers of Excellence Base Financing Program of Conicyt. V.J. was supported by GAČR Grant No. P203/11/P038 of the Czech Republic.
[99]{}
S. Iijima, Nature [**354**]{}, 56 (1991).
Q. Zhao, M. Buongiorno Nardelli, and J. Bernholc, Phys. Rev. B [**65**]{}, 144105 (2002).
B. I. Yakobson, C. J. Brabec, and J. Bernholc, Phys. Rev. Lett. [**76**]{}, 2511 (1996).
S. J. Tans *et al.*, Nature (London) [**386**]{}, 474 (1997). P. G. Collins and P. Avouris, Sci. Amer. [**283**]{}, 62 (2000);
P. Avouris, J. Appenzeller, R. Martel, and S. J. Wind, Proc. IEEE [**91**]{}, 1772 (2003);
M. P. Anantram and F. Léonard, Rep. Prog. Phys. [**69**]{}, 507 (2006).
O. Klein, Z. Phys. [**53**]{}, 157 (1929). T. Ando, T. Nakanishi, and R. Saito, J. Phys. Soc. Jpn. [**67**]{}, 2857 (1998);
P. L. McEuen, M. Bockrath, D. H. Cobden, Y-G. Yoon, and S. G. Louie, Phys. Rev. Lett. [**83**]{}, 5098 (1999);
T. Ando, H. Suzuura, Physica E [**18**]{}, 202 (2003);
V. Jakubský, L. -M. Nieto and M. S. Plyushchay, Phys. Rev. D [**83**]{}, 047702 (2011). O. O. Kit, T. Tallinen, L. Mahadevan, J. Timonen, and P. Koskinen, Phys. Rev. B [**85**]{}, 085428 (2012).
see chapter E in [@Blase] and references therein.
C. L. Kane and E. J. Mele, Phys. Rev. Lett. [**78**]{}, 1932 (1997).
V. M. Pereira, A. H. Castro Neto, Phys. Rev. Lett. [**103**]{}, 046801 (2009).
G. W. Semenoff, Phys. Rev. Lett. [**53**]{}, 2449 (1984).
J. Tersoff, D. R. Hamann, Phys. Rev. B [**31**]{}, 805 (1985);
Z. F. Wang, Ruoxi Xiang, Q. W. Shi, Jinlong Yang, Xiaoping Wang, J. G. Hou, and Jie Chen, Phys. Rev. B [**74**]{}, 125417 (2006).
G. Basar, G. V. Dunne, Phys. Rev. Lett. [**100**]{}, 200404 (2008);
G. Basar, G. V. Dunne, Phys. Rev. D [**78**]{}, 065022 (2008). F. Correa, G. V. Dunne and M. S. Plyushchay, Annals Phys. [**324**]{}, 2522 (2009) F. Gesztesy, H. Holden, *Soliton Equations and Their Algebro-Geometric Solutions,* Cambridge University Press, New York (2003).
A. H. Castro Neto, F. Guinea, N. M. R. Peres, K. S. Novoselov, A. K. Geim, Rev. Mod. Phys. [**81**]{}, 109 (2009).
A. Shytov, M. Rudner, Nan Gu, M. Katsnelson, L. Levitov, Solid State Commun. [**149**]{}, 1087 (2009).
V. P. Gusynin, S. G. Sharapov, J. P. Carbotte, Int. J. Mod. Phys. B [**21**]{}, 4611 (2007).
J. -C. Charlier, X. Blase and S. Roche, Rev. Mod. Phys. [**79**]{}, 677 (2007). Xueshen Wang et al., Nano. Lett. [**9**]{}, 3137 (2009).
D. V. Kolesnikov, V. A. Osipov, Phys. Part. Nucl. [**40**]{}, 502 (2009).
M. A. H. Vozmediano, M. I. Katsnelson and F. Guinea, Phys. Rept. [**496**]{}, 109 (2010) and references therein. J. C. Meyer, M. Paillet, S. Roth, Science [**309**]{}, 1539 (2005).
H. B. Peng, C. W. Chang, S. Aloni, T. D. Yuzvinsky, and A. Zettl, Phys. Rev. B [**76**]{} 035405 (2007).
A. M. Fennimore, T. D. Yuzvinsky, Wei-Qiang Han, M. S. Fuhrer, J. Cumings and A. Zettl, Nature [**424**]{}, 409 (2003).
E. Joselevich, ChemPhysChem [**7**]{}, 1405 (2006).
V. Jakubský and M. S. Plyushchay, Phys. Rev. D [**85**]{}, 045035 (2012).
E. J. M. Hamilton et.al., Science [**260**]{}, 659 (1993).
A. Rubio, J. L. Corkill, and M. L. Cohen, Phys. Rev. B [**49**]{}, 5081 (1994).
M. Thies, J. Phys. A [**39**]{}, 12707 (2006). C. De Concini, R. A. Johnson, Ergodic Theory Dynamical Systems [**7**]{}, 1 (1987).
G. Eilenberger, Z. Phys. [**214**]{}, 195 (1968).
I. Kosztin, Š. Kos, M. Stone, and A. J. Leggett, Phys. Rev. B [**58**]{}, 9365 (1998).
M. Abramowitz and I. A. Stegun, “Jacobian elliptic functions and theta functions", Ch. 16 in *Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables,* (Dover, New York, Dover, 1972), pp. 567
E. T. Whittaker and G. N. Watson, *A course of modern analysis* (Cambridge University Press, Cambridge, England 1980).
E. Milpas, M. Torres and G. Murguía, J. Phys.: Condens. Matter [**23**]{}, 245304 (2011).
D. J. Fernandez, B. Mielnik, O. Rosas-Ortiz and B. F. Samsonov, Phys. Lett. A [**294**]{}, 168 (2002) \[quant-ph/0302204\]. F. Correa, V. Jakubský and M. S. Plyushchay, J. Phys. A [**41**]{}, 485303 (2008) \[arXiv:0806.1614 \[hep-th\]\]. R. Jackiw and C. Rebbi, Phys. Rev. D [**13**]{}, 3398 (1976). E. D. Belokolos et al, *Algebro-geometric approach to nonlinear integrable equations,* (Springer, Berlin, 1994).
F. Correa, V. Jakubský, L. -M. Nieto and M. S. Plyushchay, Phys. Rev. Lett. [**101**]{}, 030403 (2008) \[arXiv:0801.1671 \[hep-th\]\]. J. Feinberg, Annals Phys. [**309**]{}, 166 (2004). A. Arancibia, J. M. Guilarte and M. S. Plyushchay, arXiv:1210.3666 \[math-ph\]; A. Arancibia and M. S. Plyushchay, Phys. Rev. D [**85**]{}, 045018 (2012)
M. S. Plyushchay, Annals Phys. [**245**]{}, 339 (1996) \[hep-th/9601116\]; Int. J. Mod. Phys. A [**15**]{}, 3679 (2000) \[hep-th/9903130\]. L.M. Nieto, A.A. Pecheritsin, B.F. Samsonov, Annals Phys. [**305**]{}, 151 (2003).
M. Ramezani Masir, P. Vasilopoulos, F. M. Peeters, New J. Phys. 11 095009 (2009); S. Gattenlohner, W. Belzig, M. Titov, Phys. Rev. B [**82**]{}, 155417 (2010).
L. Dell’Anna and A. De Martino, Phys. Rev. B [**79**]{}, 045420 (2009)͒.
̈Y. Avishai, Y. B. Band, Phys. Rev. B [**32**]{}, 2674 (1985).
D. S. Miserev, M. V. Entin, J. Experiment. and Theor. Phys. [**115**]{}, 694 (2012).
Ch.-H. Park, Li Yang, Y-W. Son, M. L. Cohen, and S. G. Louie, Phys. Rev. Lett. [**101**]{}, 126804 (2008);
L. Brey, H. A. Fertig, Phys. Rev. Lett. [**103**]{}, 046809 (2009).
L. Dell’Anna and A. De Martino, Phys. Rev. B [**83**]{}, 155449 (2011).
D. S. Novikov, Phys. Rev. B [**72**]{}, 235428 (2005).
V. K. Dugaev, M. I. Katsnelson, arXiv:1206.4526.
Li-Gang Wang, Shi-Yao Zhu, Phys. Rev. B [**81**]{}, 205444 ͑ (2010͒);
M. Ramezani Masir, P. Vasilopoulos, F. M. Peeters, J. Phys.: Condens. Matter [**22**]{}, 465302 (2010).
Liang Zheng Tan, Ch.-H. Park, and S. G. Louie, Phys. Rev. B [**81**]{}, 195426 (2010).
A. R. Hall, L. An, J. Liu, L. Vicci, M. R. Falvo, R. Superfine, and S. Washburn, Phys. Rev. Lett. [**96**]{}, 256102 (2006).
[^1]: The remaining four Dirac points can be reached by translational vectors of the reciprocal lattice, and, hence, do not represent distinct electronic states.
[^2]: We set $\hbar=e=1$. $\lambda=\frac{E}{v_F}$ where $E$ is energy and $v_F$ the Fermi velocity of the quasi-particle.
[^3]: $\zeta=\left(-\frac{t}{a}\frac{\partial \ln t}{\partial \ln a}\right)^{-1}$ where $t$ is the hopping parameter and $a$ is the lattice constant. See [@GaugeFields] for more details.
[^4]: The AKNS$_{N}$ can be written as a linear combination $\sum_{l=0}^{N+1}c_{l}\hat{{f}}_{l}$ where the functions $\hat{{f}}_{l}$ are defined like in (\[recursion\]) but with all the integration constants that emerge in (\[int\]) fixed to zero.
[^5]: For the nonperiodic settings, the spatial integration can be divergent.
[^6]: The ansatz (\[poly\]) is manifestly hermitian. Additionally, it also satisfies the second condition in (\[constrains\]). Indeed, one can check directly that $(\mbox{det}\tilde{R}(x,\lambda))'=0$ with the use of (\[AKNS\]).
[^7]: This way to express the eigenfunctions is just the essence of usual supersymmetric quantum mechanics applied for finite-gap potentials. To avoid the details here, we refer to [@samsonovetal; @trisusy].
[^8]: For $N=2$ the band edge energies are $\lambda_0=-\lambda_4=-2, \, \lambda_1=-\lambda_3= -\sqrt{\frac{7+\cos \theta}{2}} $ and $\lambda_2=2\cos \frac{\theta}{2} $. The constants in (\[complexgeneral\]) are then given in the following form $$\begin{aligned}
c_0(\lambda)=\lambda ^4-2 \lambda ^3 \cos \left(\frac{\theta }{2}\right)-\frac{1}{2}
\lambda ^2 (\cos (\theta )+7)+ \\
\frac{1}{2} \lambda \left(15 \cos
\left(\frac{\theta }{2}\right)+\cos \left(\frac{3 \theta
}{2}\right)\right)\, \\
c_1(\lambda)=2 \left(-\lambda ^2+\lambda \cos \left(\frac{\theta}{2}\right)+\cos (\theta
)+1\right), \, c_2=\frac{9}{4}\end{aligned}$$
[^9]: In [@Gusynin], the real spin of electrons in taken into account. There, the time-reversal operator is defined as $(\sigma_{1}\otimes\sigma_1)T\sigma_2$, where the last Pauli matrix acts on the spin degree of freedom of the electrons. As we do not consider real spin of electrons in our model, we have to define $\mathcal{T}$ as in (\[tr\]) to keep $\mathcal{T}^2=-1$.
[^10]: The three-gap setting with (\[crystal2\]) has the nonlocal integral of slightly modified form, see [@BdG].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We propose simultaneous confidence bands of the hyperbolic-type for the contrasts between several nonlinear (curvilinear) regression curves. The critical value of a confidence band is determined from the distribution of the maximum of a chi-square random process defined on the domain of explanatory variables. We use the volume-of-tube method to derive an upper tail probability formula of the maximum of a chi-square random process, which is asymptotically exact and sufficiently accurate in commonly used tail regions. Moreover, we prove that the formula obtained is equivalent to the expectation of the Euler-Poincaré characteristic of the excursion set of the chi-square random process, and hence conservative. This result is therefore a generalization of Naiman’s inequality for Gaussian random processes. As an illustrative example, growth curves of consomic mice are analyzed.
[*Keywords and phrases*]{}: Chi-square random process, expected Euler-characteristic heuristic, Gaussian random field, growth curve, Naiman’s inequality, volume-of-tube method.
author:
- 'Xiaolei Lu[^1], Satoshi Kuriki[^2]'
title: Simultaneous confidence bands for contrasts between several nonlinear regression curves
---
Introduction {#sec:introduction}
============
This paper concerns multiple comparisons of $k$ ($\ge 3$) nonlinear (curvilinear) regression curves estimated from independent $k$ groups. Suppose that for each group $i=1,\ldots,k$, and for each explanatory variable $x_j\in\mathcal{X}$, $j=1,\ldots,n$, we have observations $y_{ij1},\ldots,y_{ij r_i}$ as objective variables with $r_i$ replications, which are assumed to follow the model $$\begin{aligned}
\label{model}
y_{ijh} = g_i(x_j) + \varepsilon_{ijh}, \quad i=1,\ldots,k,\ \ j=1,\ldots,n,\ \ h=1,\ldots,r_i.\end{aligned}$$ Here, $\mathcal{X}\subseteq{\mathbb{R}}$ is the domain of explanatory variables, and random errors $\varepsilon_{ijh}$ are assumed to be independently distributed as the normal distribution ${{\mathcal{N}}}(0,\sigma(x_j)^2)$. The variance function $\sigma(x)^2$ is supposed to be known, or at least known up to a constant $\sigma(x)^2 =\sigma^2\sigma_0(x)^2$. In the case of the latter, we suppose that an independent estimator $\widehat\sigma^2$ of $\sigma^2$ is available. In addition, we assume that the true regression curve has the form $$\label{beta_f}
g_i(x)=\beta_i^\top f(x),\ \ x\in\mathcal{X},$$ where $f(x)=(f_1(x),\ldots,f_p(x))^\top$ is a known regression basis vector function, $\beta_i=(\beta_{i1},\ldots,\beta_{ip})^\top$ is an unknown parameter vector. Then, the least squares estimator $\widehat\beta_i$ of $\beta_i$ has the multivariate normal distribution ${{\mathcal{N}}}_p(\beta_i,r_i^{-1}\Sigma)$, where $$\Sigma = \left(\sum_{j=1}^n \frac{1}{\sigma(x_j)^2}f(x_j)f(x_j)^\top\right)^{-1}$$ is the inverse of the $p\times p$ information matrix. When $\sigma(x)^2=\sigma^2\sigma_0(x)^2$, we have $\Sigma=\sigma^2\Sigma_0$, where $\Sigma_0$ is $\Sigma$ with $\sigma(x_j)$ replaced by $\sigma_0(x_j)$.
Let $\mathcal{C}$ denote the set of vectors $c=(c_1,\ldots,c_k)^\top$ such that $\sum_{i=1}^k c_i=0$. The focus of this paper is the construction of $1-\alpha$ simultaneous confidence bands for all the contrasts $\sum_{i=1}^k c_i g_i(x) = \sum_{i=1}^k c_i \beta_i^\top f(x)$ between the $k$ regression curves for all $x \in \mathcal{X}$ and $c\in\mathcal{C}$, where $\mathcal{X}\subset{\mathbb{R}}$ is a finite or half-infinite interval, a finite union of such intervals, or an infinite interval $(-\infty,\infty)$. Specifically, according to the traditional form of the point estimate plus or minus a probability point times the estimated standard error, we construct a $1-\alpha$ simultaneous confidence band of the form $$\begin{aligned}
\label{bands}
\sum_{i=1}^k c_i \beta_i^\top f(x) \in \sum_{i=1}^k c_i\widehat\beta_i^\top f(x) \pm b_{1-\alpha}
\sqrt{\left(\sum_{i=1}^k \frac{c_i^2}{r_i}\right) f(x)^\top\Sigma f(x)},\end{aligned}$$ where $\widehat\beta_i^\top f(x)$ is the estimator of $\beta_i^\top f(x)$ in (\[beta\_f\]). This form is referred to as a hyperbolic-type ([@Liu10]). The critical value $b_{1-\alpha}$ is determined such that the event in (\[bands\]) for all $x \in \mathcal{X}$ and $c \in \mathcal{C}$ holds with a probability of at least $1-\alpha$. Our problem typically arises from growth curve analysis and longitudinal data analysis.
Throughout this paper, we assume that the regression curve $g_i(x)$ is a linear combination of a finite number of known basis functions in (\[beta\_f\]). Although it is a conventional regression model, we must always be careful regarding the approximation bias caused by model misspecification. This issue is examined in Section \[sec:simulation\].
The problem concerning the construction of simultaneous confidence bands in a regression model originates with Working and Hotelling [@Working-Hotelling29]. They formalized this problem as the construction of confidence intervals for an estimated regression line, and provided a critical value by making use of the Cauchy-Schwarz inequality. Specifically, Working and Hotelling [@Working-Hotelling29] treated the case of
- one regression model (equivalent to the case $k=2$ in our problem),
- the simple regression $f(x)=(1,x)^\top$, and
- the unrestricted domain of the explanatory variables $\mathcal{X}=(-\infty,\infty)$.
Subsequently, many reports concerning the relaxation of these conditions have appeared in literature.
In the case of one regression model, Wynn and Bloomfield [@Wynn-Bloomfield71] pointed out that the use of the Cauchy-Schwarz inequality leads to conservative bands unless both (ii) and (iii) hold. They illustrated improved confidence bands for the quadratic regression $f(x)=(1,x,x^2)^\top$. Uusipaikka [@Uusipaikka83] constructed exact confidence bands for linear regression when $\mathcal{X}$ is a finite interval. See Liu et al. [@Liu-Lin-Piegorsch08] and Liu [@Liu10] for historical reviews. The problem of $k$ ($\ge 3$) regression curve comparisons was considered by Spurrier [@Spurrier99; @Spurrier02] and Lu and Chen [@Lu-Chen09], who proposed procedures based on simple linear regression. However, it is difficult to extend these methods to nonlinear regression.
One exception is Naiman’s [@Naiman86] integral-geometric approach. In the unit sphere ${\mathbb{S}}^{p-1}$ of the $p$-dimensional Euclidean space, he defined a trajectory $$\label{Gamma}
\Gamma = \overline{\{ \psi(x) \mid x\in \mathcal{X} \}} \subset {\mathbb{S}}^{p-1}$$ of a normalized basis vector function $$\label{psi}
\psi(x) = \frac{\Sigma^{1/2} f(x)}{\Vert\Sigma^{1/2} f(x)\Vert},$$ and evaluated the volume of the tubular neighborhood of $\Gamma$. In the case of one regression model, he constructed a simultaneous confidence band with the critical value obtained from this volume. The volume formula for such tubes originated from Hotelling [@Hotelling39] and Weyl [@Weyl39]. Currently, this idea is understood in the volume-of-tube method framework ([@Adler-Taylor07], [@Kuriki-Takemura01], [@Kuriki-Takemura09], [@Sun93], [@Takemura-Kuriki02]). As shown in Section \[sec:pivot\], we require the tail probability of the maximum of a Gaussian random field or chi-square random process as a pivotal quantity. The volume-of-tube method is a methodology to evaluate such tail probabilities.
In this paper, we adopt this integral-geometric approach. In the case of $k\ge 3$, we define a subset $M$ in (\[M\]) of a unit sphere, and by evaluating the volume of its tubular neighborhood, we obtain the critical value $b_{1-\alpha}$ in (\[bands\]) by means of the volume-of-tube method. Moreover, we prove that the proposed confidence band is conservative. It is known that Naiman’s [@Naiman86] confidence band is conservative (Naiman’s inequality, see also [@Johnstone-Siegmund89]), and our result is regarded as its generalization.
Note that, in the setting of this paper, the covariance matrices of the estimators $\widehat\beta_i$ are identical up to a multiplicative constant. This property arises from the condition that the explanatory variables $x_j$ are common between $k$ groups in the model (\[model\]). This represents the so-called the balanced case. For the unbalanced case, the problem of constructing simultaneous confidence bands is quite tedious and only simulation-based approaches are available ([@Jamshidian-Liu-Bretz10], [@Liu10], [@Liu-Jamshidian-Zhang04], [@Liu-Wynn-Hayter08]). In this paper, we address only the balanced case.
Moreover, note that in the one-group case ($k=1$), various simultaneous confidence bands by means of the volume-of-tube method have been proposed. Johansen and Johnstone [@Johansen-Johnstone90] demonstrated the usefulness of Hotelling’s volume formula for the construction of simultaneous bands. The application to the B-spline regression is found in Zhou et al. [@Shen-Wolfe-Zhou98]. Sun and Loader [@Sun-Loader94] proposed a modification to the volume-of-tube formula when a small approximation bias caused by model misspecification exists. In succeeding papers, Sun and her coauthors developed this idea further in various model settings ([@Faraway-Sun95], [@Sun-Loader-McCormick00], [@Sun-Raz-Faraway99]). See also Krivobokova et al. [@Krivobokova-Kneib-Claeskenset10]. The crucial difference between this paper and existing work is that in this paper, we need to treat a Gaussian random field with a general dimensional ($k-1$ dimensional) index set, and need the volume formula up to an arbitrary order.
The layout of this paper is as follows. In Section \[sec:pivot\], we define a Gaussian random field and a chi-square random process as pivotal quantities. We show that the critical value $b_{1-\alpha}$ is determined from the upper tail probability of the maximum of a Gaussian random field or chi-square random process. In Section \[sec:tube\], the volume-of-tube method and its related method known as the expected Euler-characteristic heuristics are briefly summarized. The primary results are provided in Section \[sec:main\]. Some simulation study under model misspecification is conducted in Section \[sec:simulation\]. Section \[sec:growth\] is devoted to the growth curve data analysis. Proof details are located in the Appendix.
Random fields as pivotal quantities {#sec:pivot}
===================================
Our problem is to determine the critical value $b_{1-\alpha}$ in (\[bands\]). First, assume that $\Sigma$ is fully known. Define a pivotal quantity $$\label{Txc}
T(x,c)
= \frac{\sum_{i=1}^k c_i(\widehat\beta_i-\beta_i)^\top f(x)}{\sqrt{\left(\sum_{i=1}^k\frac{c_i^2}{r_i}\right) f(x)^\top\Sigma f(x)}}.$$ Then, the critical value $b_{1-\alpha}$ is the solution $b$ of the equation $$\begin{aligned}
\Pr\bigl(T(x,c)\le b,\,\forall x \in \mathcal{X},\,\forall c \in \mathcal{C}\bigr)
= \Pr\biggl(\max_{x\in\mathcal{X},c\in\mathcal{C}} T(x,c)\le b\biggr)
= 1-\alpha.\end{aligned}$$ In this expression, we use $T(x,c)$ instead of $|T(x,c)|$, because $c\in\mathcal{C}$ implies $-c\in\mathcal{C}$ and $|T(x,c)|$ is equal to $T(x,c)$ or $T(x,-c)$. Inverting $|T(c,x)|\le b_{1-\alpha}$ yields the $1-\alpha$ simultaneous confidence band in (\[bands\]).
In the following, we show that $b_{1-\alpha}^2$ is the upper $\alpha$ point of the maximum of a chi-square random process. We can assume that $\sum_{i=1}^k c_i^2/r_i=1$ without loss of generality, because $T(x,c)$ is a homogeneous function in $c$. Let $\rho=(\sqrt{r_1},\ldots,\sqrt{r_k})^\top$, and define a $k\times (k-1)$ matrix $H$ such that $\rho^\top H=0$, $H^\top H=I_{k-1}$, and $H H^\top = I_k - \rho \rho^\top/(\rho^\top \rho)$. (An example of $H$ is given in Remark \[rem:H\] below.) Then the $c=(c_1,\ldots,c_k)^\top$ such that $\sum_{i=1}^k c_i^2/r_i=1$ and $\sum_{i=1}^k c_i=0$ are represented as $$c = {\mathrm{diag}}(\sqrt{r_1},\ldots,\sqrt{r_k}) H h,\ \ h\in{\mathbb{S}}^{k-2},$$ where ${\mathbb{S}}^{k-2}$ is the set of ($k-1$)-dimensional unit column vectors.
Let $\Sigma^{1/2}$ be a matrix such that $(\Sigma^{1/2})^\top \Sigma^{1/2} = \Sigma$, and let $\Sigma^{-1/2}$ be its inverse. Then, $\eta_i=\sqrt{r_i}(\Sigma^{-1/2})^\top(\widehat\beta_i-\beta_i)$ is distributed normally as ${{\mathcal{N}}}_p(0,I)$, independently for $i=1,\ldots,k$. Let $\psi:\mathcal{X}\to{\mathbb{S}}^{p-1}$ as defined in (\[psi\]). Then, $T(x,c)$ is rewritten as $$\begin{aligned}
T(x,c)
=& \sum_{i=1}^k \frac{c_i}{\sqrt{r_i}} \sqrt{r_i} \{(\Sigma^{-1/2})^\top(\widehat\beta_i-\beta_i)\}^\top \frac{\Sigma^{1/2} f(x)}{\Vert\Sigma^{1/2} f(x)\Vert} \nonumber \\
=& c^\top {\mathrm{diag}}(\sqrt{r_1},\ldots,\sqrt{r_k})^{-1}
\begin{pmatrix} \eta_1^\top \\ \vdots \\ \eta_k^\top \end{pmatrix}_{k\times p} \psi(x) \nonumber \\
=& h^\top \begin{pmatrix} \xi_1^\top \\ \vdots \\ \xi_{k-1}^\top \end{pmatrix}_{(k-1)\times p} \psi(x) \nonumber \\
=& \xi^\top \{h\otimes\psi(x)\},
\label{Txc2}\end{aligned}$$ where $\xi_i$ are $p\times 1$ vectors defined by $(\xi_1,\ldots,\xi_{k-1})_{p\times(k-1)}=(\eta_1,\ldots,\eta_k)_{p\times k}H$, $\xi = (\xi_1^\top,\ldots,\xi_{k-1}^\top)^\top$ is a $p(k-1)\times 1$ vector, and ‘$\otimes$’ is the Kronecker product. Vectors $\eta_i$ consist of independent standard Gaussian random variables ${{\mathcal{N}}}(0,1)$, therefore, so does vector $\xi$. When $x$ and $h$ are fixed, because $\Vert\psi(x)\Vert=\Vert h\otimes\psi(x)\Vert=1$, $\xi_i^\top \psi(x)$ is distributed as ${{\mathcal{N}}}(0,1)$ independently for $i=1,\ldots,k$, and $\xi^\top \{h\otimes\psi(x)\}$ is distributed as ${{\mathcal{N}}}(0,1)$.
From (\[Txc2\]), we can see that $$\label{maxT}
\max_{c\in\mathcal{C}} T(x,c) = \sqrt{\sum_{i=1}^{k-1} \bigl\{\xi_i^\top \psi(x)\bigr\}^2}.$$ For each fixed $x$, this is distributed as the square root of the chi-square distribution $\chi^2_{k-1}$ with $k-1$ degrees of freedom.
When $\Sigma=\sigma^2\Sigma_0$ with $\Sigma_0$ known, and an independent estimator $\widehat\sigma^2\sim\sigma^2\chi^2_\nu/\nu$ of unknown $\sigma^2$ is available, we redefine $T(x,c)$ in (\[Txc\]) by replacing $\Sigma$ in the denominator with $\widehat\sigma^2\Sigma_0$. Thus, instead of (\[Txc2\]) and (\[maxT\]), we have $$T(x,c) = \frac{1}{\tau}\xi^\top \{h\otimes\psi(x)\}, \quad
\max_{c\in\mathcal{C}} T(x,c) = \sqrt{\frac{1}{\tau^2} \sum_{i=1}^{k-1} \bigl\{\xi_i^\top \psi(x)\bigr\}^2}, \quad \tau^2=\frac{\widehat\sigma^2}{\sigma^2}.$$
Now, we consider the object in (\[Txc2\]) as a random function of $(x,h)$: $$\label{Z}
Z(x,h) = \xi^\top \{h\otimes\psi(x)\},\ \ (x,h)\in\mathcal{X}\times{\mathbb{S}}^{k-2},$$ where $\xi\sim {{\mathcal{N}}}_{p(k-1)}(0,I)$. Then, $Z(x,h)$ is the Gaussian random field with mean 0, variance 1, and covariance function $${\mathrm{Cov}}\bigl(Z(x,h),Z(\tilde x,\tilde h)\bigr) = \psi(x)^\top\psi(\tilde x) \cdot h^\top \tilde h.$$ Similarly, we define the chi-square random process with $k-1$ degrees of freedom: $$\label{Y}
Y(x) = \sum_{i=1}^{k-1} \bigl\{\xi_i^\top\psi(x)\bigr\}^2,\ \ x\in\mathcal{X}.$$
We summarize the results of this section below.
When $\Sigma$ is known, the critical value $b_{1-\alpha}$ is determined as the solution $b=b_{1-\alpha}$ of $$\Pr\biggl(\max_{x\in\mathcal{X},h\in{\mathbb{S}}^{k-2}} Z(x,h)\ge b \biggr)
= \Pr\biggl(\max_{x\in\mathcal{X}} Y(x)\ge b^2\biggr) = \alpha,$$ where $Z(x,h)$ is the Gaussian random field defined in (\[Z\]), and $Y(x)$ is the chi-square random process defined in (\[Y\]).
When $\Sigma=\sigma^2\Sigma_0$ with $\Sigma_0$ known, the critical value $b_{1-\alpha}$ is determined as the solution $b=b_{1-\alpha}$ of $${{\mathrm{E}}}\biggl[\Pr\biggl(\max_{x\in\mathcal{X},h\in{\mathbb{S}}^{k-2}} Z(x,h)\ge b\tau \,\big|\, \tau^2 \biggr)\biggr]
= {{\mathrm{E}}}\biggl[\Pr\biggl(\max_{x\in\mathcal{X}} Y(x)\ge b^2\tau^2 \,\big|\, \tau^2 \biggr)\biggr] = \alpha,$$ where the expectation is taken over $\tau^2\sim\chi^2_\nu/\nu$, with $\nu$ being the degrees of freedom of the estimator of $\sigma^2$.
\[rem:H\] An example of $k\times (k-1)$ matrix $H$ such that $\rho^\top H=0$, $H^\top H=I_{k-1}$, $H H^\top = I_k - \rho \rho^\top/(\rho^\top \rho)$ with $\rho=(\sqrt{r_1},\ldots,\sqrt{r_k})^\top$ is given as $$H = \begin{pmatrix}
\frac{\sqrt{r_1 r_2}}{\sqrt{R_1 R_2}} & \frac{\sqrt{r_1 r_3}}{\sqrt{R_2 R_3}} & \dots & \frac{\sqrt{r_1 r_k}}{\sqrt {R_{k-1}R_k}} \\
-\frac{R_1}{\sqrt{R_1 R_2}} & \frac{\sqrt{r_2 r_3}}{\sqrt{R_2 R_3}} & \dots & \frac{\sqrt{r_2 r_k}}{\sqrt{R_{k-1}R_k}} \\
& -\frac{R_2}{\sqrt{R_2 R_3}} & \dots & \frac{\sqrt{r_3 r_k}}{\sqrt{R_{k-1}R_k}} \\
& & \ddots & \vdots \\
0 & & & -\frac{R_{k-1}}{\sqrt{R_{k-1}R_k}}
\end{pmatrix}_{k\times (k-1)},$$ where $R_i=\sum_{j=1}^i r_j$.
Preliminaries on the volume-of-tube method {#sec:tube}
==========================================
In this section, we summarize the volume-of-tube method for evaluating the upper tail probability of the maximum of a Gaussian random field.
Let $\xi$ be a Gaussian random vector distributed as ${{\mathcal{N}}}_n(0,I)$. Let $M$ be a closed subset of ${\mathbb{S}}^{n-1}$, which is the unit sphere (the set of unit column vectors) of ${\mathbb{R}}^n$. Then, the random map $u\mapsto \xi^\top u$, $u\in M$, is a Gaussian random field with mean 0, variance 1, and covariance function ${\mathrm{Cov}}(\xi^\top u,\xi^\top v)=u^\top v$. The volume-of-tube method approximates the distribution of the maximum $\max_{u\in M} \xi^\top u$. The maximum of $Z(x,h)$ in (\[Z\]) can be treated in this framework by setting $$\label{M}
M = \overline{\{ h\otimes\psi(x) \mid (x,h)\in\mathcal{X}\times{\mathbb{S}}^{k-2} \}}\quad\mbox{and}\quad n=p(k-1).$$ The dimension of $M$ is $d=\dim M=k-1$. To apply the volume-of-tube method, we require the following assumption on $M$. Let the symbol ‘$\sqcup$’ denote disjoint union.
\[as:M\] $M$ is a $d$-dimensional closed piecewise $C^2$-manifold, or $M$ is a $d$-dimensional $C^2$-manifold with piecewise $C^2$-boundary. We write $M={\mathrm{Int}}M \sqcup \partial M$, where ${\mathrm{Int}}M$ and $\partial M$ denote the interior and boundary of $M$, respectively. In the former case, $\partial M=\emptyset$.
When $M$ is defined by (\[M\]), we can provide a sufficient condition for Assumption \[as:M\].
\[as:psi\] $\psi:\mathcal{X}\to{\mathbb{S}}^{p-1}$ is a one-to-one map of class piecewise $C^2$. There does not exist $x,\tilde x\in\mathcal{X}$ such that $\psi(x)=-\psi(\tilde x)$.
Under Assumption \[as:psi\], the map $(x,h)\mapsto h\otimes\psi(x)$ is a piecewise $C^2$ one-to-one map.
Consider the polynomial regression with a basis function vector $f(x)=(1,x,\ldots,x^{p-1})^\top$. When the domain of $x$ is a finite interval $\mathcal{X}=[a,b]$, we have $$\begin{aligned}
{\mathrm{Int}}M =& \{ h\otimes \psi(x) \mid x\in (a,b),\,h\in{\mathbb{S}}^{k-2} \}, \\
\partial M =& \{ h\otimes \psi(a) \mid h\in{\mathbb{S}}^{k-2} \} \sqcup \{ h\otimes \psi(b) \mid h\in{\mathbb{S}}^{k-2} \}.\end{aligned}$$ When $\mathcal{X}=(-\infty,\infty)$, $\psi(\pm\infty)=(\pm 1)^{p-1}\Sigma^{1/2} e_p/\sqrt{e_p^\top\Sigma e_p}$ with $e_p=(0,\ldots,0,1)^\top$, and hence $h\otimes \psi(\infty)=(-1)^{p-1}h\otimes\psi(-\infty)$. This denotes that $M$ is a closed manifold without boundary.
Consider the trigonometric regression with a basis function vector $$f(x)=\left(1,\sqrt{2}\cos x,\sqrt{2}\sin x,\ldots,\sqrt{2}\cos mx,\sqrt{2}\sin mx\right)^\top.$$ When $\mathcal{X}=[0,2\pi)$, $M$ is a closed manifold without boundary.
We now define “tube”, the key concept of the volume-of-tube method. The set of ${\mathbb{S}}^{n-1}$ points whose great circle distance from $M$ is less than or equal to $\theta$ is the tube about $M$ with radius $\theta$, and has the expression $$M_\theta = \Bigl\{ v\in {\mathbb{S}}^{n-1} \,\big|\, \min_{u\in M} \cos^{-1} \bigl(u^\top v\bigr)\le\theta \Bigr\}.$$
If the radius $\theta$ is sufficiently small, the tube $M_\theta$ does not have self-overlap. Whereas, when the radius $\theta$ is large, the tube does have self-overlap. The threshold radius between the two cases is known as the critical radius $\theta_{\mathrm{c}}$. We let $\theta_{\mathrm{c}}=\pi/2$ when the threshold radius is more than $\pi/2$. Under Assumption \[as:M\], we can prove that $\theta_{\mathrm{c}}>0$. Figure 1 in Kuriki and Takemura [@Kuriki-Takemura09] depicts an example of a tube and its critical radius.
The support cone (or tangent cone) of $M$ at $u\in M$ is denoted by $S_u M$. (See Section 1.2 of Takemura and Kuriki [@Takemura-Kuriki02] for the definition.) The cone with base set $M$ is denoted by ${\mathrm{co}}(M) = \bigsqcup_{\lambda\ge 0} \lambda M$. Then, the support cone of ${\mathrm{co}}(M)$ at $u\in M$ is decomposed as $S_u({\mathrm{co}}(M)) = S_u M\oplus{\mathrm{span}}\{u\}$, where ${\mathrm{span}}\{u\}$ is the linear space spanned by $u$. The normal cone of ${\mathrm{co}}(M)$ at $u\in M$ is defined by the dual of the support cone: $N_u({\mathrm{co}}(M)) = S_u({\mathrm{co}}(M))^*$.
Note that the ($m-1$)-dimensional volume of ${\mathbb{S}}^{m-1}$ is $\Omega_m=2\pi^{m/2}/\Gamma(m/2)$. For $m\times m$ matrix $A=(a_{ij})$, let ${\mathrm{tr}}_0 A=1$ and $${\mathrm{tr}}_e A = \sum_{1\le k_1<\ldots<k_e\le m} \det(a_{k_i k_j})_{1\le i,j\le e},\ \ 1\le e\le m$$ ([@Muirhead05], Appendix A.7). Note that ${\mathrm{tr}}_1 A={\mathrm{tr}}A$, ${\mathrm{tr}}_m A=\det A$. The upper probability of the chi-square distribution with $m$ degrees of freedom is denoted by ${{\overline G}}_m(\cdot)$. Now we can provide the upper tail probability formula for the Gaussian field $\xi^\top u$, $u\in M$. The theorem below is a special case of Proposition 2.2 of Takemura and Kuriki [@Takemura-Kuriki02].
As $b\to\infty$, $$\label{p}
\Pr\biggl(\max_{u\in M} \xi^\top u\ge b\biggr) =
{{\overline P_{\mathrm{tube}}}}(b) + O({{\overline G}}_n(b^{2}(1+\tan^2\theta_{\mathrm{c}}))),$$ where $$\label{p-hat}
{{\overline P_{\mathrm{tube}}}}(b)
= \sum_{0\le e\le d,\,e:\rm even } w_{d+1-e} {{\overline G}}_{d+1-e}(b^2)
+ \sum_{0\le e\le d-1} w'_{d-e} {{\overline G}}_{d-e}(b^2),$$ with $$\begin{aligned}
w_{d+1-e} =& \frac{1}{\Omega_{d+1-e}\Omega_{n-d-1+e}}\int_{{\mathrm{Int}}M}\biggl\{\int_{N_u({\mathrm{co}}(M))\cap{\mathbb{S}}^{n-1}} {\mathrm{tr}}_e H(u,v) \,{{\mathrm{d}}}v\biggr\} \,{{\mathrm{d}}}u,
\label{w} \\
w'_{d-e} =& \frac{1}{\Omega_{d-e}\Omega_{n-d+e}}\int_{\partial M}\biggl\{\int_{N_u({\mathrm{co}}(M))\cap{\mathbb{S}}^{n-1}} {\mathrm{tr}}_e H'(u,v) \,{{\mathrm{d}}}v\biggr\} \,{{\mathrm{d}}}u.
\label{w'}\end{aligned}$$ Here, $H(u,v)$ is the second fundamental form of ${\mathrm{Int}}M$ at $u$ in the direction of $v$, and $H'(u,v)$ is the second fundamental form of $\partial M$ at $u$ in the direction of $v$. ${{\mathrm{d}}}u$ is the volume element of ${\mathrm{Int}}M$ or $\partial M$, and ${{\mathrm{d}}}v$ is the volume element of $N_u({\mathrm{co}}(M))\cap{\mathbb{S}}^{n-1}$.
In (\[p\]), because $\theta_{\mathrm{c}}>0$, the error term $O\bigl({{\overline G}}_n(b^{2}(1+\tan^2\theta_{\mathrm{c}}))\bigr)=O(b^{n-2}e^{-b^2(1+\tan^2\theta_{\mathrm{c}})/2})$ is exponentially smaller than each term ${{\overline G}}_j(b^2)=O(b^{j-2}e^{-b^2/2})$. Hence, (\[p-hat\]) can be used as an approximation formula when $b$ is large. The method in which ${{\overline P_{\mathrm{tube}}}}(b)$ is used as an approximate value is referred to as the volume-of-tube method, or simply the tube method. This name comes from the volume formula for $M_\theta$ below.
\[prop:tube\] For the radius $\theta\in [0,\theta_{\mathrm{c}}]$, the ($n-1$)-dimensional spherical volume of the tube $M_\theta$ is given by $$\begin{aligned}
{\mathrm{Vol}}_{n-1}(M_\theta)
= \Omega_n \Biggl\{ & \sum_{0\le e\le d,\,e:\rm even } w_{d+1-e} {{\overline B}}_{\frac{1}{2}(d+1-e),\frac{1}{2}(n-d-1+e)}(\cos^2\theta) \\
& +\sum_{0\le e\le d-1} w'_{d-e} {{\overline B}}_{\frac{1}{2}(d-e),\frac{1}{2}(n-d+e)}(\cos^2\theta) \Biggr\},\end{aligned}$$ where $w_{d+1-e}$ and $w'_{d-e}$ are given in (\[w\]) and (\[w’\]), ${{\overline B}}_{a,b}(\cdot)$ is the upper probability of the beta distribution with parameter $(a,b)$.
The critical radius $\theta_{\mathrm{c}}$ can be evaluated using the following characterization (Theorem 4.18 of Federer [@Federer59], Proposition 4.3 of Johansen and Johnstone [@Johansen-Johnstone90], Lemma 2.2 of Takemura and Kuriki [@Takemura-Kuriki02]). For a proof, see Theorem 2.9 of Kuriki and Takemura [@Kuriki-Takemura09].
The critical radius $\theta_{\mathrm{c}}$ of $M$ is given by $$\label{critical}
\tan^2\theta_{\mathrm{c}} = \inf_{u,v\in M,\,u\ne v} \frac{(1-u^\top v)^2}{\Vert P^\perp_v(u-v)\Vert^2},$$ where $P^\perp_v$ is the orthogonal projection onto the normal cone $N_v({\mathrm{co}}(M))$ of ${\mathrm{co}}(M)$ at $v$.
The local critical radius $\theta_{\mathrm{c,loc}}$ is defined as $$\label{critical-local}
\tan^2\theta_{\mathrm{c,loc}} = \liminf_{u,v\in M,\,u\ne v,\,\Vert u-v\Vert\to 0} \frac{(1-u^\top v)^2}{\Vert P^\perp_v(u-v)\Vert^2}.$$ From the definition, it holds that $\theta_{\mathrm{c}}\le\theta_{\mathrm{c,loc}}$. In general, $\theta_{\mathrm{c,loc}}$ is easier to evaluate than $\theta_{\mathrm{c}}$.
We have summarized the volume-of-tube method to evaluate the upper tail probabilities of the maximum of random fields thus far. There is another method utilized for the same purpose, known as the expected Euler-characteristic heuristic ([@Adler-Taylor07], [@Worsley95]). When applied to the Gaussian random field $\xi^\top u$, $u\in M$, this method is stated as follows. For each $b$, define the excursion set by $$A_b = \{ u\in M \mid \xi^\top u \ge b \}.$$ Let $\chi(\cdot)$ be the Euler-Poincaré characteristic of a set, and ${\mathds{1}}(\cdot)$ be the indicator function for an event. The expected Euler-characteristic heuristic assumes that ${\mathds{1}}(A_b\ne\emptyset) \approx \chi(A_b)$ for large $b$, and $$\Pr\biggl( \max_{u\in M} \xi^\top u \ge b \biggr) = {{\mathrm{E}}}[{\mathds{1}}(A_b\ne\emptyset)] \approx {{\mathrm{E}}}[\chi(A_b)].$$ Note that $\chi(A_b)$ can be evaluated by Morse’s theorem, and is more tractable than ${\mathds{1}}(A_b\ne\emptyset)$. Takemura and Kuriki [@Takemura-Kuriki02] proved the equivalence of the volume-of-tube method and the expected Euler-characteristic heuristic as follows.
\[prop:equivalence\] $${{\mathrm{E}}}[\chi(A_b)] =
{{\overline P_{\mathrm{tube}}}}(b)
\ \ \mbox{for all $b\ge 0$}.$$
Using this, Takemura and Kuriki [@Takemura-Kuriki02] provided an alternative proof that the confidence band of Naiman [@Naiman86] is conservative.
Main results {#sec:main}
============
Recall that our aim is to derive the upper tail probability of the maximum of the Gaussian random field $Z(x,h)$ defined in (\[Z\]), or equivalently, the chi-square random process $Y(x)$ defined in (\[Y\]). The theorem below provides an answer, and proof is provided in the Appendix.
\[thm:main\] Let $\xi\sim {{\mathcal{N}}}_n(0,I)$, $n=p(k-1)$. Let $\Gamma\subset{\mathbb{S}}^{p-1}$ and $M\subset{\mathbb{S}}^{n-1}$ be defined by (\[Gamma\]) and (\[M\]), and let $|\Gamma|$ denote the length of $\Gamma$. Assume Assumption \[as:psi\] on $\psi$. Then, as $b\to\infty$, $$\begin{aligned}
\Pr\biggl(\max_{(x,h)\in\mathcal{X}\times{\mathbb{S}}^{k-2}} Z(x,h) \ge b \biggr)
=& \Pr\biggl(\max_{x\in\mathcal{X}} Y(x) \ge b^2 \biggr) \\
=& \Pr\biggl(\max_{u\in M}\xi^\top u \ge b \biggr) \\
=& {{\overline P_{\mathrm{tube}}}}(b) + O\bigl(b^{n-2} e^{-(1+\tan^2\theta_{\mathrm{c}})b^2/2}\bigr),\end{aligned}$$ where $$\begin{aligned}
\label{p-hat2}
{{\overline P_{\mathrm{tube}}}}(b)
= \frac{\Gamma(\frac{k}{2})}{\sqrt{\pi}\,\Gamma(\frac{k-1}{2})} |\Gamma| \bigl\{{{\overline G}}_k(b^2) - {{\overline G}}_{k-2}(b^2) \bigr\}
+ \chi(\Gamma) {{\overline G}}_{k-1}(b^{2}).\end{aligned}$$ Note that if $\Gamma$ (and hence $M$) has no boundary, then $\Gamma$ is homeomorphic to ${\mathbb{S}}^1$, and therefore $\chi(\Gamma)=0$. Otherwise, $\chi(\Gamma)$ is the number of connected components of $\Gamma$.
\[thm:conservative\] Assume Assumption \[as:psi\]. Suppose that $\Gamma$ has boundaries. The approximation formula given in Theorem \[thm:main\] is a conservative bound, specifically, $$\begin{aligned}
\Pr\biggl(\max_{u\in M}\xi^\top u \ge b \biggr) \le
{{\overline P_{\mathrm{tube}}}}(b)
\ \ \mbox{for all $b\ge 0$}.\end{aligned}$$
Arrange the $p(k-1)\times 1$ vector $\xi=(\xi_1^\top,\ldots,\xi_{k-1}^\top)^\top$, and define a $(k-1)\times p$ matrix $\Xi=(\xi_1,\ldots,\xi_{k-1})^\top$. Let $$\begin{aligned}
A_b
=& \{ u \in M \mid \xi^\top u \ge b \}
= \{ h\otimes q \mid (q,h) \in \Gamma\times{\mathbb{S}}^{k-2},\, h^\top \Xi q \ge b \}
\subset {\mathbb{S}}^{p(k-1)-1}, \\
\widetilde A_b
=& \{ (q,h) \in \Gamma\times{\mathbb{S}}^{k-2} \mid h^\top \Xi q \ge b \}
\subset {\mathbb{S}}^{p-1}\times {\mathbb{S}}^{k-2}, \\
B_b
=& \{ q \in \Gamma \mid q^\top\Xi^\top\Xi q \ge b^2 \} \subset {\mathbb{S}}^{p-1}.\end{aligned}$$ Note that $A_b$ is the excursion set of the Gaussian random field $\xi^\top u$, $u\in M$, $\widetilde A_b$ is the excursion set of the Gaussian random field $\sum_{i=1}^{k-1} h_i (\xi_i^\top q) = h^\top\Xi q$, $(q,h)\in \Gamma\times{\mathbb{S}}^{k-2}$, and $B_b$ is the excursion set of the chi-square random process $\sum_{i=1}^{k-1} (\xi_i^\top q)^2 = q^\top\Xi^\top\Xi q$, $q\in\Gamma$. We will prove that for each fixed $\xi$, ${\mathds{1}}(A_b\ne\emptyset)={\mathds{1}}(\widetilde A_b\ne\emptyset)={\mathds{1}}(B_b\ne\emptyset)$ and $\chi(A_b)=\chi(\widetilde A_b)=\chi(B_b)$.
First, note that owing to Assumption \[as:psi\], the map $(q,h)\mapsto h\otimes q$ is one-to-one. Hence, $A_b$ and $\widetilde A_b$ are homeomorphic and therefore ${\mathds{1}}(A_b\ne\emptyset)={\mathds{1}}(\widetilde A_b\ne\emptyset)$ and $\chi(A_b)=\chi(\widetilde A_b)$.
Moreover, noting that $\widetilde A_b\ne\emptyset$ $\Leftrightarrow$ $\max_h h^\top\Xi q\ge b$ for some $q$ $\Leftrightarrow$ $q^\top\Xi\Xi^\top q \ge b^2$ for some $q$ $\Leftrightarrow$ $B_b\ne\emptyset$, that is, ${\mathds{1}}(\widetilde A_b\ne\emptyset)={\mathds{1}}(B_b\ne\emptyset)$, we can write $$\widetilde A_b = \bigsqcup_{q\in B_b}\{(q,h) \mid h\in{\mathbb{S}}^{k-2},\,h^\top\Xi q \ge b \}.$$ Given $b\ge 0$, the set $\{ h\in{\mathbb{S}}^{k-2} \mid h^\top\Xi q \ge b \}$ is contractible and star-shaped about the point $h^*(q)=\Xi q/\Vert \Xi q\Vert$. That is, the map $$\varphi : \widetilde A_b\times[0,1]\to\widetilde A_b,\quad (q,h,t) \mapsto \biggl(q,\frac{(1-t) h+t h^*(q)}{\Vert (1-t) h+t h^*(q)\Vert}\biggr)$$ is continuous, and $\varphi\bigl(\widetilde A_b\times\{0\}\bigr)=\widetilde A_b$ is homotopy equivalent to the set $\varphi\bigl(\widetilde A_b\times\{1\}\bigr)=\bigsqcup_{q\in B_b}\{(q,h^*(q))\}$. This is homotopy equivalent to $\bigsqcup_{q\in B_b}\{q\} = B_b$. Hence, $\chi(\widetilde A_b)=\chi(B_b)$.
Recall that $B_b$ is the excursion set of the chi-square random process on the one-dimensional index set $\Gamma$. This means that $B_b$ is also one-dimensional, and $\chi(B_b)$ is only the number of connected components of $B_b$. Therefore ${\mathds{1}}(B_b\ne\emptyset)\le\chi(B_b)$. By taking expectations, $$\begin{aligned}
\Pr\biggl(\max_{u\in M}\xi^\top u \ge b \biggr) = {{\mathrm{E}}}[{\mathds{1}}(A_b\ne\emptyset)] & = {{\mathrm{E}}}[{\mathds{1}}(B_b\ne\emptyset)] \\
& \le {{\mathrm{E}}}[\chi(B_b)] = {{\mathrm{E}}}[\chi(A_b)] =
{{\overline P_{\mathrm{tube}}}}(b).\end{aligned}$$ The last equality is owing to Proposition \[prop:equivalence\].
Naiman [@Naiman86] proved that application of the volume-of-tube method to a Gaussian random process with a one-dimensional index set always provides a conservative band. Theorem \[thm:conservative\] is a generalization of Naiman’s [@Naiman86] inequality to a chi-square random process.
\[thm:critical\_radius\] The interior and boundary of $\Gamma$ are denoted by ${\mathrm{Int}}\Gamma$ and $\partial\Gamma$, respectively. The critical radius $\theta_{\mathrm{c}}$ of $M$ is given by $$\tan^2\theta_{\mathrm{c}} = \min\biggl\{ \inf_{x\ne \tilde x,\,\psi(x)\in{\mathrm{Int}}\Gamma} \frac{(1-\alpha s)^2}{1-s^2-\alpha^2 t^2},
\inf_{x\ne \tilde x,\,\psi(x)\in\partial\Gamma} \frac{(1-\alpha s)^2}{1-s^2-\max\{0,\varepsilon(x)\alpha t\}^2} \biggr\},$$ where the infima are taken over $x,\tilde x\in\mathcal{X}$, and $\alpha\in[-1,1]$ as well as additional conditions (arguments of $\inf$), and $$s=s(x,\tilde x)=\psi(x)^\top\psi(\tilde x), \quad
t=t(x,\tilde x)=\frac{\psi_x(x)^\top\psi(\tilde x)}{\Vert\psi_x(x)\Vert},$$ $\psi_x(x)=\partial\psi(x)/\partial x$, $$\varepsilon(x) = \begin{cases}
1 & (\mbox{$\psi_x(x)$ is inward to $\Gamma$}), \\
-1 & (\mbox{$\psi_x(x)$ is outward to $\Gamma$}).
\end{cases}$$ $\psi_x(x)$ is said to be inward or outward to $\Gamma$ if the support cone of $\Gamma$ at $\psi(x)$ is $S_{\psi(x)}\Gamma=\{\lambda\psi_x(s) \mid \lambda\ge 0\}$ or $\{\lambda\psi_x(s) \mid \lambda\le 0\}$, respectively.
\[thm:local\_critical\_radius\] Assume Assumption \[as:psi\]. Moreover, assume that $\psi:\mathcal{X}\to{\mathbb{S}}^{p-1}$ is of $C^4$-class. Then, the local critical radius $\theta_{\mathrm{c,loc}}$ is given by $$\tan^{2}\theta_{\mathrm{c,loc}} = \min\biggl\{
\inf_{x\in\mathcal{X}:\kappa(x)\le 2} \biggl\{1-\frac{\kappa(x)}{4}\biggr\},
\inf_{x\in\mathcal{X}:\kappa(x)\ge 2} \frac{1}{\kappa(x)} \biggr\}$$ with $$\label{kappa}
\kappa(x) = \frac{\psi_{xx}(x)^\top\psi_{xx}(x)}{\{\psi_x(x)^\top\psi_x(x)\}^2}
-\frac{\{\psi_{xx}(x)^\top\psi_x(x)\}^2}{\{\psi_x(x)^\top\psi_x(x)\}^3} -1,$$ where $\psi_x(x)=\partial\psi(x)/\partial x$ and $\psi_{xx}(x)=\partial^2\psi(x)/\partial x^2$.
The proofs of Theorems \[thm:critical\_radius\] and \[thm:local\_critical\_radius\] are included in the Appendix.
### A numerical example {#a-numerical-example .unnumbered}
At the end of this section, we provide a numerical example to determine the accuracy of the approximation formula given in Theorem \[thm:main\], and degree of conservativeness proved by Theorem \[thm:conservative\].
Suppose that $f(x)=(1,x,x^2)^\top$, $\mathcal{X}=[-1,1]$, and $$\Sigma = \begin{pmatrix} 1 & 0 & \frac{2}{3} \\ 0 & \frac{2}{3} & 0 \\ \frac{2}{3} & 0 & 1 \end{pmatrix}, \quad
\Sigma^{1/2} = \begin{pmatrix}
1 & 0 & \frac{2}{3} \\ 0 & \sqrt{\frac{2}{3}} & 0 \\ 0 & 0 & \frac{\sqrt{5}}{3} \end{pmatrix}.$$ Then, $$\begin{aligned}
& \psi(x) = \frac{1}{3(1+x^2)} \bigl(3+2x^2,\sqrt{6}x,\sqrt{5}x^2\bigr)^\top, \quad
|\Gamma| = \int_{\mathcal{X}} \Vert\dot\psi(x)\Vert \,{{\mathrm{d}}}x =
\int_{-1}^1 \sqrt{\frac{2}{3}}\frac{1}{1+x^2} \,{{\mathrm{d}}}x = \frac{\pi}{\sqrt{6}}.\end{aligned}$$ $\kappa(x)$ in (\[kappa\]) is always $5$. Hence, the local critical radius is $\theta_{\mathrm{c,loc}}=\tan^{-1}(1/\sqrt{5})=0.134 \pi$. Further, we can also confirm that the critical radius is the same as $\theta_{\mathrm{c}}=\theta_{\mathrm{c,loc}}$ using Mathematica [@Mathematica].
Under this setting, we suppose the case of $k=3$. The probability we need is $$\label{maxY}
\Pr\biggl( \max_{x\in[-1,1]}Y(x)\ge b^2 \biggr)
= 1-\Pr\bigl(T(x,c)\le b,\ \forall x\in[-1,1],\,\forall c\in\mathcal{C}\bigr),$$ where $$Y(x)=\sum_{i=1}^2 \{\xi_i^\top\psi(x)\}^2,\ \ \xi_1,\xi_2\sim {{\mathcal{N}}}_3(0,I)\ \rm i.i.d.$$ is a chi-square random process $Y(x)$ with two degrees of freedom. The tube formula for the upper tail probability (\[maxY\]) is $$\label{maxY-hat}
{{\overline P_{\mathrm{tube}}}}(b)
= \frac{\pi}{2\sqrt{6}} \bigl\{{{\overline G}}_3(b^2)-{{\overline G}}_1(b^2)\bigr\} + {{\overline G}}_2(b^2)
= \biggl(\frac{\sqrt{\pi}}{2\sqrt{3}} b +1\biggr) e^{-b^2/2}.$$ Figure \[fig:tail\_prob\] depicts the upper tail probability of the maximum (\[maxY\]) and its approximate value (\[maxY-hat\]). We can see that the tube formula approximates the true upper tail probability with sufficient accuracy in the moderate tail regions (for example, the upper probability is less than 0.2), and it provides a conservative bound as per Theorem \[thm:conservative\].
{width="0.65\linewidth"}
(solid line: tube formula, dashed line: Monte Carlo with 10,000 replications)
\[fig:tail\_prob\]
We have proposed that the threshold for the confidence band should be determined as the solution $b=b_{\mathrm{tube},1-\alpha}$ for ${{\overline P_{\mathrm{tube}}}}(b)=\alpha$. Figure \[fig:confidence\_coeff\] depicts the actual confidence coefficient (coverage probability) $$\Pr\biggl( \max_{x\in[-1,1]}Y(x)\ge b_{\mathrm{tube},1-\alpha}^2 \biggr), \ \ \alpha\in[0,1].$$ This further demonstrates that the confidence bands obtained by the tube method are always conservative and very accurate.
{width="0.65\linewidth"}
(solid line: actual confidence coefficient, dashed line: 45-degree line)
\[fig:confidence\_coeff\]
Simulation study under model misspecification {#sec:simulation}
=============================================
Throughout this paper, it is assumed that the nonlinear model has a finite number of basis functions $g_i(x)=\beta_i^\top f(x)$ in (\[beta\_f\]). However, we can only approximate the true model in practice. Under a slight misspecification of the model, Sun and Loader [@Sun-Loader94] estimated the bias of the coverage probability, and proposed an adjustment to the volume-of-tube formula. Although their approach may be applied to our model, the result would be more complicated. Instead, to investigate what happens under model misspecification, we conducted a Monte Carlo simulation study in the following setting.
The domain of explanatory variable is set to be $\mathcal{X}=[0,1]$. The data are generated from the model $$y_{ij} = g_i(x_j) + \varepsilon_{ij}, \ \ \varepsilon_{ij}\sim {{\mathcal{N}}}(0,1), \ \ i=1,\ldots,k,\ \ j=1,\ldots,n,$$ where $k=3$, $n=11$, and $x_j=(j-1)/n$, $j=1,\ldots,n$. As the true regression curve is $g_i(x)$, we assume three models.
Model 1: $$g_i(x) = \beta_i^\top f_{2,5,0,1}(x), \quad \beta_1=(0,\ldots,0)^\top, \ \ \beta_2= K(0,0,1/2,1,1)^\top, \ \ \beta_3= K(0,0,4/3,0,0)^\top,$$ where $K=1$, $3$, or $9$, $$f_{d,m,a,b} = \left( B_d\biggl(\frac{x-a}{b-a}(m-d)-(i-d-1)\biggr) \right)_{i=1,\ldots,m},$$ and $B_d(\cdot)$ is the B-spline function $$\label{Bd}
B_d(x) = \sum_{r=0}^{d+1} (-1)^{d+1-r} {d+1 \choose r} \frac{(r-x)_+^d}{d!}$$ ([@deBoor78], p.89).
Model 2: $$g_1(x) = 0, \quad g_2(x) = K\sin(x\pi/2), \quad g_3(x) = K\sin(x\pi), \quad K=1,3,9.$$
Model 3: $$g_1(x) = 0, \quad g_2(x) = K\frac{e^{-x/2}-e^{-x}}{e^{-1/2}-e^{-1}}, \quad g_3(x) = K\frac{\cosh(x-1/2)-1}{\cosh(1/2)-1}, \quad K=1,3,9.$$ For all models, $g_2(x)$ is unimodal, and $g_3(x)$ is increasing. $g_2(x)$ and $g_3(x)$ are designed to have the range $[0,K]$.
We fit the curve $\beta_i^\top f_{2,m,0,1}(x)$ to the generated data $y_{ij}$, where $m=3,\ldots,10$. Using these models, we constructed a $1-\alpha=0.95$ confidence band. Coverage probabilities were estimated based on Monte Carlo simulations with 1,000,000 replications, and are summarized in Table \[tab:coverage\_prob\]. In this table, $$\label{delta}
\delta = \max_{x\in\mathcal{X},\,c\in\mathcal{C}}
\left| \frac{\sum_{i=1}^k c_i \{(\beta_i^*)^\top f_{2,m,0,1}(x) - g_i(x)\}}{\sqrt{f_{2,m,0,1}(x)^\top\Sigma f_{2,m,0,1}(x)}} \right|,
\quad
\Sigma = \left(\sum_{i=1}^n f_{2,m,0,1}(x_i)^\top f_{2,m,0,1}(x_i)\right)^{-1}$$ is the bias of regression function, where $\beta_i^*$ is the best parameter in the assumed model $\beta_i^\top f_{2,m,0,1}(x)$. $$\label{Delta}
\Delta = \max\bigl\{ \alpha-{{\overline P_{\mathrm{tube}}}}(b_{\mathrm{tube},1-\alpha}+\delta), {{\overline P_{\mathrm{tube}}}}(b_{\mathrm{tube},1-\alpha}-\delta)-\alpha \bigr\}$$ is an approximate upper bound of the bias of coverage probability, where $b_{\mathrm{tube},1-\alpha}$ is the approximate value of $b_{1-\alpha}$ obtained by the tube method. (See \[sec:bound\] for the detail.)
[(prob: coverage probability, $\delta$: bias (\[delta\]), $\Delta$: bound for coverage probability bias (\[Delta\]))]{}
From this table, we first see that, for the true models ($m=5,8$ when model 1 is true), the coverage probabilities are more than, but approximately equal to, the nominal value 0.95, meaning that the proposed method is valid. The most remarkable point is that, throughout the study, the coverage probabilities are kept at approximately 0.95, unless the assumed model is too small, and the bias $\delta$ is large.
Table \[tab:width\] shows the average width of the confidence band defined by $$W = \frac{\int_{\mathcal{X}} b_{\mathrm{tube},1-\alpha} \sqrt{f(x)^\top\Sigma f(x)} \,{{\mathrm{d}}}x}{\int_{\mathcal{X}} \,{{\mathrm{d}}}x} = b_{\mathrm{tube},0.95} \int_0^1 \sqrt{f_{2,m,0,1}(x)^\top\Sigma f_{2,m,0,1}(x)} \,{{\mathrm{d}}}x.$$ When the model is increasing in size, $W$ is increasing in size. This suggests that a smaller model is preferable, unless it is too small to cause serious bias.
In summary, too small of a model should surely be avoided, whereas, a larger model has the disadvantage of having a wider confidence band. This trade-off is crucially important in practice, and a promising future research topic, although it is out of scope for this paper. For related topics, refer to Casella and Hwang [@Casella-Hwang12], for shrinkage confidence bands, and Leeb et al. [@Leeb-etal15], for confidence band post-model selection.
Growth curve analysis {#sec:growth}
=====================
As mentioned in Section \[sec:introduction\], the growth curve analysis is one of our research objectives to which we apply our method. In this section, we demonstrate the analysis of mouse growth as an illustration. Sun et al. [@Sun-Raz-Faraway99] proposed simultaneous confidence bands for a growth curve by virtue of the volume-of-tube method. Differently from their analysis, we focus on the contrast of several growth curves.
Mice are one of the most popular model organisms, and are often used in genomic research. Figure \[fig:growth\] depicts the average body weights of male mice from four different strains measured from 2 to 20 weeks after birth. The four strains are C57BL/6 (referred to as B6), MSM/Ms (MSM), B6-Chr17[$^{\rm MSM}$]{}(B6-17), and B6-ChrXT[$^{\rm MSM}$]{}(B6-XT). Among these, B6 is the most common laboratory strain and serves as the standard. MSM is a wild-derived strain having contrasting properties to B6 such as non-black color, small size, and aggressive behavior. B6-17 and B6-XT are artificial strains known as consomic mice made from B6 and MSM. B6-17 has all the chromosomes from B6, and only chromosome 17 from MSM; B6-XT has all the chromosomes from B6, and only half of the X chromosome from MSM. By comparing the consomic strains with B6, we expect to reveal the role of each chromosome.
----------------------------------------------------------------------------------------------------- -- --
{width="0.8\linewidth"}
----------------------------------------------------------------------------------------------------- -- --
\[fig:growth\]
The dataset we utilized is publicly available as Supplemental Table S1 of Takada et al. [@Takada-etal08]. In their experiments, the weight (unit: gram) $y_{ijh}$ of the $h$th individual from strain $i$ was measured at time point $x_j$. The measurement time points were $\{x_1,\ldots,x_{10}\}=\{2,4,\ldots,20\}$ ($n=10$). This dataset includes the average body weight $y_{ij}$ of strain $i$ at time $x_j$, and its standard error $$y_{ij} = \frac{1}{r_i} \sum_{h=1}^{r_i} y_{ij h}, \quad
\widehat{\mathrm{s.e.}}(y_{ij}) = \sqrt{\frac{1}{r_i^2}\sum_{h=1}^{r_i} (y_{ij h}-y_{ij})^2},$$ as well as the number $r_i$ of individuals of strain $i$.
In the following analysis, we use $k=3$ groups (strains) B6 ($i=1$), B6-17 ($i=2$), and B6-XT ($i=3$). The number of individuals are $r_1=12$, $r_2=24$, and $r_3=12$.
We fit the model (\[model\]) to these data. We estimate the variance as $$\widehat\sigma(x_j)^2
= \frac{1}{\sum_{i=1}^k (r_i-1)}\sum_{i=1}^k \sum_{h=1}^{r_i}(y_{ijh}-y_{ij})^2
= \frac{1}{\sum_{i=1}^k (r_i-1)}\sum_{i=1}^k r_i^2\widehat{\mathrm{s.e.}}(y_{ij})^2,$$ which is used as the true value $\sigma(x_j)^2$ hereafter. Figure \[fig:se\] plots the estimated standard error $\widehat\sigma(x_j)$. One particular feature of this dataset is that the experiment is well controlled and measurement errors are quite small.
![Estimated standard error $\widehat\sigma(x_j)$.[]{data-label="fig:se"}](se-d2m5a5.eps){width="0.65\linewidth"}
As the basis function $f(x)$, we consider a family of basis functions $$f(x) = f_{d,m,2,20}(x) =
\left(B_d\biggl(\frac{x-2}{20-2}(m-d)-(i-d-1)\biggr)\right)_{1\le i\le m},$$ with $B_d(x)$ given in (\[Bd\]). $f_{d,m,2,20}(x)$ consists of $m$ B-spline bases with equally-spaced knots at intervals of $(20-2)/(m-d)$. Note that $f_{d,m,2,20}(x)$ is piecewise of class $C^{d}$.
In the range $d=2,3,4$ and $m=d+1,d+2,\ldots,n\,(=10)$, we searched for the best model that minimizes AIC and BIC defined below: $$\mathrm{AIC}_{d,m} = L_{d,m} + 2 k m, \quad
\mathrm{BIC}_{d,m} = L_{d,m} + \sum_{i=1}^k \ln(r_i n) m, \quad
L_{d,m} = \sum_{i=1}^k r_i \sum_{j=1}^n \frac{(y_{ij}-\widehat y_{ij})^2}{\sigma(x_j)^2}$$ with $k=3$, $n=10$, where $\widehat y_{ij}=\widehat\beta_i^\top f_{d,m,2,20}(x_j)$. In both criteria, the minimizer was $(d,m)=(2,5)$, which we use as the true value hereafter.
Suppose that we are interested in the period $\mathcal{X}=[a,b]=[2,20]$. An approximate value of the length of $\Gamma$ in (\[Gamma\]) is given by $$|\Gamma| \approx
\sum_{t=1}^{N} \bigl\Vert\psi(x_{t})-\psi(x_{t-1})\bigr\Vert,$$ where $x_t = a+t(b-a)/N$, $t=0,1,\ldots,N$. When $N=10,000$, the approximate value of $|\Gamma|$ is $6.989=2.225 \pi$. Using this, the critical value is $b_{1-\alpha}=3.258$ ($\alpha=0.05$). To compare $k$ groups, various types of contrasts are used. For a pairwise comparison between group $i$ and group $j$, we choose $c = (\ldots,0,\underset{i\rm th}{1},0,\ldots,0,\underset{j\rm th}{-1},0,\ldots)$. For the comparison of groups $\{i,j\}$ and group $k$, we use $$c = \biggl(\ldots,0,\underset{i\rm th}{\frac{r_i}{r_i+r_j}},0,\ldots,0,\underset{j\rm th}{\frac{r_j}{r_i+r_j}},0,\ldots,0,\underset{k\rm th}{-1},0,\ldots\biggr).$$ Figure \[fig:difference\] depicts the difference curves of strains B6-17 vs. B6 (left) and B6-XT vs. B6 (right), and their 95% simultaneous confidence bands. In the left panel, the horizontal line representing zero difference is almost between the confidence bands. This indicates that there is no significant difference between B6-17 and B6. In contrast, in the right panel, after around week 14, the horizontal line is outside the confidence bands, thereby indicating that B6-XT and B6 are different during this period.
------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------
![Differences of body weights and 95% confidence bands.[]{data-label="fig:difference"}](17vsB6-d2m5a5.eps "fig:"){width="0.45\linewidth"} ![Differences of body weights and 95% confidence bands.[]{data-label="fig:difference"}](XTvsB6-d2m5a5.eps "fig:"){width="0.45\linewidth"}
------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------
For a fixed $x$, the test statistic for the null hypothesis $H_{0,x}:\beta_1^\top f(x) = \ldots = \beta_k^\top f(x)$ is $$\chi^2(x) = \frac{1}{f(x)^\top\Sigma f(x)}\sum_{i=1}^k r_i \Biggl\{\widehat\beta_i^\top f(x)-\frac{\sum_{i=1}^k r_i\widehat\beta_i^\top f(x)}{\sum_{i=1}^k r_i}\Biggr\}^2.$$ For a fixed $x$, the null distribution is the chi-square distribution with $k-1$ degrees of freedom. However, for the overall null hypothesis $H_0:\beta_1^\top f(x) = \ldots = \beta_k^\top f(x)$ for all $x\in\mathcal{X}$, the distribution of the maximum of the chi-square random process should be used. Figure \[fig:chi2\_process\] shows $\chi^2(x)$ and its upper 5% critical value $b_{0.95}^2$. As already shown in Figure \[fig:difference\], after around week 14, the hypothesis of equality is rejected.
![Chi-square process $\chi^2(x)$ and its upper 5% critical value.[]{data-label="fig:chi2_process"}](chi2-d2m5a5.eps){width="0.65\linewidth"}
Appendix: Proofs
================
Proof of Theorem \[thm:main\]
-----------------------------
### Contribution of the inner points ${\mathrm{Int}}\,M$ {#contribution-of-the-inner-points-mathrmintm .unnumbered}
Here, we obtain the coefficients $w_{d+1-e}$ in (\[w\]) when $M$ is given in (\[M\]).
Let $h=h(\theta)$, $\theta=(\theta_i)_{1\le k-2}$, be a local coordinate system of ${\mathbb{S}}^{k-2}$. For example, $$h = h(\theta) = \left(\begin{array}{l}
\cos\theta_1 \\
\sin\theta_1 \cos\theta_2 \\
\sin\theta_1 \sin\theta_2 \cos\theta_3 \\
\quad \vdots \\
\sin\theta_1 \cdots \sin\theta_{k-3} \cos\theta_{k-2} \\
\sin\theta_1 \cdots \sin\theta_{k-3} \sin\theta_{k-2}
\end{array}\right)_{(k-1)\times 1},$$ where $$\theta\in\Theta =
\{(\theta_1,\ldots,\theta_{k-2}) \mid 0\le\theta_i\le \pi\ (i=1,\dots,k-3),\ 0\le\theta_{k-2}< 2\pi \}.$$
Let $(x,\theta)\in\mathcal{X}\times\Theta$ be fixed, and let $\phi(x,\theta) = h(\theta)\otimes\psi(x) \in M$. We write $\psi=\psi(x)$, $h=h(\theta)$ and $\phi=\phi(x,\theta)$ for simplicity. We first assume that $x\in{\mathrm{Int}}\mathcal{X}$, hence, $\phi(x,\theta)\in{\mathrm{Int}}M$.
By applying the Gram-Schmidt orthonormalization to the sequence $\psi, \partial\psi/\partial x, \partial^2\psi/\partial x^2, \ldots$, we construct the orthonormal basis (ONB) $\psi_{(i)}$, $i=0,\ldots,p-1$, of ${\mathbb{R}}^p$. The first three bases are $$\psi_{(0)} = \psi,\ \ \psi_{(1)} = \frac{1}{\sqrt{g}}\frac{\partial\psi}{\partial x}, \ \ \psi_{(2)} = \frac{1}{\sqrt{\eta-\frac{\gamma^2}{g}-g^2}} \biggl(\frac{\partial^2\psi}{\partial x^2} + g \psi - \frac{\gamma}{g}\frac{\partial\psi}{\partial x}\biggr),$$ where $$g = g(x) = \biggl(\frac{\partial\psi}{\partial x}\biggr)^\top\biggl(\frac{\partial\psi}{\partial x}\biggr), \quad
\gamma = \gamma(x) = \biggl(\frac{\partial^2\psi}{\partial x^2}\biggr)^\top\biggl(\frac{\partial\psi}{\partial x}\biggr), \quad
\eta = \eta(x) = \biggl(\frac{\partial^2\psi}{\partial x^2}\biggr)^\top\biggl(\frac{\partial^2\psi}{\partial x^2}\biggr).$$ Similarly, from the sequence $h$, $\partial h/\partial\theta_i$, $i=1,\ldots,k-2$, we obtain ONB $h_{(i)}$, $i=0,\ldots,k-2$, of ${\mathbb{R}}^{k-1}$. We prepare a $(k-2)\times (k-2)$ upper triangle matrix $D$ such that $$\biggl(h,\frac{\partial h}{\partial\theta_1},\ldots,\frac{\partial h}{\partial\theta_{k-2}}\biggr) = \bigl(h_{(0)},h_{(1)},\ldots,h_{(k-2)}\bigr)
\begin{pmatrix} 1 & 0 \\ 0 & D \end{pmatrix}, \quad\mbox{or}\ \ D=\biggl(h_{(i)}^\top\frac{\partial h}{\partial\theta_j}\biggr)_{1\le i,j\le k-2}.$$ Now we have the ONB $h_{(i)}\otimes \psi_{(j)}$, $i=0,\ldots,k-2$, $j=0,\ldots,p-1$, of the ambient space ${\mathbb{R}}^n$ with $n=p(k-1)$. Note that $\phi=h_{(0)}\otimes\psi_{(0)}$.
The tangent space $T_\phi M$ is spanned by $$\frac{\partial\phi}{\partial x} = h\otimes\frac{\partial\psi}{\partial x}, \quad
\frac{\partial\phi}{\partial\theta_i} = \frac{\partial h}{\partial\theta_i} \otimes\psi, \ \ i=1,\ldots,k-2.$$ The metric matrix of $T_\phi M$ with respect to the parameter $x,\theta_1,\ldots,\theta_{k-2}$ is $$\label{metric}
\begin{pmatrix} g & 0 \\ 0 & G \end{pmatrix}, \quad\mbox{where}\ \ G=\biggl(\biggl(\frac{\partial h}{\partial\theta_i}\biggr)^\top\biggl(\frac{\partial h}{\partial\theta_j}\biggr)\biggr)_{1\le i,j\le k-2} = D^\top D.$$ $T_\phi M$ has the ONB $h_{(0)}\otimes\psi_{(1)}$, $h_{(i)}\otimes\psi_{(0)}$, $i=1,\ldots,k-2$.
The normal space perpendicular to $T_\phi({\mathrm{co}}(M))=T_\phi M\oplus{\mathrm{span}}\{\phi\}$ is $$\begin{aligned}
N_{\phi}({\mathrm{co}}(M))
= {\mathrm{span}}\bigl\{ & h_{(0)}\otimes\psi_{(j)},\,j=2,\ldots,p-1;\ \nonumber \\
& h_{(i)}\otimes\psi_{(j)},\,i=1,\ldots,k-2,\,j=1,\ldots,p-1 \bigr\}.
\label{onb-normal}\end{aligned}$$
The second order derivatives of $\phi=\phi(x,\theta)$ are $$\frac{\partial^2\phi}{\partial x^2} = h\otimes\frac{\partial^2\psi}{\partial x^2}, \quad
\frac{\partial^2\phi}{\partial x\partial\theta_i} = \frac{\partial h}{\partial\theta_i}\otimes\frac{\partial\psi}{\partial x}, \quad
\frac{\partial^2\phi}{\partial\theta_i\partial\theta_j} = \frac{\partial^2 h}{\partial\theta_i\partial\theta_j}\otimes\psi.$$ Taking the inner product of the second derivatives and the ONB of $N_{\phi}({\mathrm{co}}(M))$ listed in (\[onb-normal\]), we see that the nonzero elements of the second fundamental form are $$-\biggl(h\otimes\frac{\partial^2\psi}{\partial x^2}\biggr)^\top (h_{(0)}\otimes \psi_{(2)}) = -\biggl(\frac{\partial^2\psi}{\partial x^2}\biggr)^\top \psi_{(2)} = -\zeta,$$ where $$\zeta = \zeta(x) = \sqrt{\eta-\frac{\gamma^2}{g}-g^2}$$ and $$-\biggl(\frac{\partial h}{\partial\theta_i}\otimes\frac{\partial\psi}{\partial x}\biggr)^\top(h_{(j)}\otimes\psi_{(1)}) = -\biggl(\frac{\partial h}{\partial\theta_i}\biggr)^\top h_{(j)}\sqrt{g}=-D_{ji}\sqrt{g}.$$
We renumber the ONB of $N_{\phi}({\mathrm{co}}(M))$ as $$N_1 = h_{(0)}\otimes\psi_{(2)},\ \ N_i = h_{(i-1)}\otimes\psi_{(1)},\,i=2,\ldots,k-1,$$ and $N_k,\ldots,N_{pk-p-k}$ are the other vectors. Write $N(t) = \sum_{i=1}^{pk-p-k} N_i t_i$, where $t=(t_1,\ldots,t_{pk-p-k})$. Then, $$\begin{aligned}
-\biggl(\frac{\partial^2\phi}{\partial x^2}\biggr)^\top N(t) =& -\zeta t_1, \\
-\biggl(\frac{\partial^2\phi}{\partial x\partial\theta_i}\biggr)^\top N(t) =& -\sum_{j=1}^{k-2}D_{ji}t_{j+1}\sqrt{g}, \\
-\bigg(\frac{\partial^2\phi}{\partial\theta_i\partial\theta_j}\biggr)^\top N(t) =& 0.\end{aligned}$$ Therefore, the second fundamental form (unnormalized version) in the direction $N(t)$ is $$\label{2nd}
\left(\begin{array}{ccc}
-\zeta t_1 & -(t_2,\ldots,t_{k-1}) D \sqrt{g} \\
-D^\top \begin{pmatrix}t_{2}\\ \vdots \\ t_{k-1}\end{pmatrix} \sqrt{g} & 0
\end{array}\right).$$ Multiplication of the inverse of the metric (\[metric\]) enables us to obtain the normalized version of the second fundamental form. Noting that $$\begin{pmatrix} g & 0 \\ 0 & G \end{pmatrix}^{-1} =
\begin{pmatrix} g & 0 \\ 0 & D^\top D \end{pmatrix}^{-1} =
\begin{pmatrix} 1/\sqrt{g} & 0 \\ 0 & D^{-1} \end{pmatrix}
\begin{pmatrix} 1/\sqrt{g} & 0 \\ 0 & (D^\top)^{-1} \end{pmatrix},$$ we multiply $$\begin{pmatrix} 1/\sqrt{g} & 0 \\ 0 & (D^\top)^{-1} \end{pmatrix}
\quad\mbox{and}\quad
\begin{pmatrix} 1/\sqrt{g} & 0 \\ 0 & D^{-1} \end{pmatrix}$$ from the left and right to (\[2nd\]), respectively, to obtain $$\begin{aligned}
\begin{pmatrix} 1/\sqrt{g} & 0 \\ 0 & (D^\top)^{-1} \end{pmatrix}
&
\begin{pmatrix}
-\zeta t_1 & -(t_2,\ldots,t_{k-1}) D \sqrt{g} \\
-D^\top \begin{pmatrix} t_2 \\ \vdots \\ t_{k-1} \end{pmatrix} \sqrt{g} & 0
\end{pmatrix}
\begin{pmatrix} 1/\sqrt{g} & 0 \\ 0 & D^{-1} \end{pmatrix} \\
& =\begin{pmatrix}
-(\zeta/g) t_1 & -(t_2,\ldots,t_{k-1}) \\
-\begin{pmatrix} t_2 \\ \vdots \\ t_{k-1} \end{pmatrix} & 0
\end{pmatrix}
= H(x,\theta;N(t)).\end{aligned}$$ This is the second fundamental form with respect to the orthonormal coordinates. Now we have $$\label{tr_e}
{\mathrm{tr}}_e H(x,\theta;N(t)) =
\begin{cases}
1 & (e=0), \\
-(\zeta/g) t_1 & (e=1), \\
-\sum_{j=2}^{k-1} t_j^2 & (e=2), \\
0 & (\mbox{otherwise}).
\end{cases}$$
Next, we evaluate the integral $$\begin{aligned}
\label{integral}
\int_{v\in N_{\phi}({\mathrm{co}}(M))\cap{\mathbb{S}}^{n-1}}
{\mathrm{tr}}_e H(x,\theta;v) \,{{\mathrm{d}}}v,\end{aligned}$$ where $n=p(k-1)$, ${{\mathrm{d}}}v$ is the volume element of $N_{\phi}({\mathrm{co}}(M))\cap{\mathbb{S}}^{n-1}$, by following Section 4.2.2 of Kuriki and Takemura [@Kuriki-Takemura01]. Recall that $d=\dim M=k-1$.
Because $N_{\phi}({\mathrm{co}}(M))$ is a linear space of dimension $n-d-1 = p(k-1)-(k-1)-1 = pk-p-k$, $N_{\phi}({\mathrm{co}}(M))\cap{\mathbb{S}}^{n-1}$ is nothing but a ($pk-p-k-1$)-dimensional unit sphere. Hence, $$\int_{v\in N_{\phi}({\mathrm{co}}(M))\cap{\mathbb{S}}^{n-1}} \,{{\mathrm{d}}}v = {\mathrm{Vol}}({\mathbb{S}}^{pk-p-k-1}) = \Omega_{pk-p-k}.$$ Therefore, if $V$ is distributed as the uniform distribution on $N_{\phi}M\cap{\mathbb{S}}^{n-1}$, denoted by ${\mathcal{U}}(N_{\phi}M\cap{\mathbb{S}}^{n-1})$, then $\mbox{(\ref{integral})} = \Omega_{pk-p-k}\times {{\mathrm{E}}}[{\mathrm{tr}}_e H(x,\theta;V)]$.
Suppose that $T = (T_1,\ldots,T_{pk-p-k}) \sim {{\mathcal{N}}}_{pk-p-k}(0,I)$, and let $N(T) = \sum_{i=1}^{pk-p-k} N_i T_i$. Then, $$\Vert N(T)\Vert^2 = \sum_{i=1}^{pk-p-k} T_i^2 \sim \chi_{pk-p-k}^2
\quad\mbox{and}\quad
V = \frac{N(T)}{\Vert N(T)\Vert} \sim {\mathcal{U}}(N_{\phi} M\cap{\mathbb{S}}^{n-1})$$ are independently distributed. Hence, $${{\mathrm{E}}}[{\mathrm{tr}}_e H(x,\theta; N(T))]
= {{\mathrm{E}}}[\Vert N(T)\Vert^e {\mathrm{tr}}_e H(x,\theta,V)]
= {{\mathrm{E}}}[\Vert N(T)\Vert^e\} {{\mathrm{E}}}\{{\mathrm{tr}}_e H(x,\theta,V)],$$ and $${{\mathrm{E}}}[{\mathrm{tr}}_e H(x,\theta,V)] = \frac{{{\mathrm{E}}}[{\mathrm{tr}}_e H(x,\theta; N(T))]}{{{\mathrm{E}}}\bigl[(\chi_{pk-p-k}^2)^{e/2}\bigr]}.$$ From (\[tr\_e\]), $${{\mathrm{E}}}[{\mathrm{tr}}_e H(x,\theta; N(T))] =
\begin{cases}
1 & (e=0), \\
0 & (e=1), \\
-\sum_{i=2}^{k-1} {{\mathrm{E}}}[T_i^2] = -(k-2) & (e=2), \\
0 & (\mbox{otherwise}),
\end{cases}$$ hence, $${{\mathrm{E}}}[{\mathrm{tr}}_e H(x,\theta;V)] =
\begin{cases}
1 & (e=0), \\
\displaystyle
-\frac{k-2}{pk-p-k} & (e=2), \\
0 & (\mbox{otherwise}).
\end{cases}$$ Therefore, $$\int_{v\in N_{\phi}({\mathrm{co}}(M))\cap{\mathbb{S}}^{n-1}}
{\mathrm{tr}}_e H(x,\theta,v) \,{{\mathrm{d}}}v = \begin{cases}
\Omega_{pk-p-k} & (e=0), \\
\displaystyle
-\frac{k-2}{pk-p-k} \Omega_{pk-p-k} & (e=2), \\
0 & (\mbox{otherwise}). \end{cases}$$
Note that the results are independent of $x$ and $\theta$. This implies that the integral in (\[w\]) with respect to the volume element $${{\mathrm{d}}}u = \sqrt{g}\,{{\mathrm{d}}}x\times |G(\theta)|^{1/2} \,{{\mathrm{d}}}\theta_{1}\cdots {{\mathrm{d}}}\theta_{k-2}$$ is simply multiplying the constant $${\mathrm{Vol}}(M) = |\Gamma|\times{\mathrm{Vol}}({\mathbb{S}}^{k-2}) = |\Gamma|\Omega_{k-1}.$$ Finally, from (\[w\]) of Proposition \[prop:tube\], $$\begin{aligned}
w_{d+1} &= w_k =
\frac{1}{\Omega_k\Omega_{pk-p-k}}\times |\Gamma| \Omega_{k-1}\Omega_{pk-p-k},
\\
w_{d-1} &= w_{k-2} =
\frac{1}{\Omega_{k-2}\Omega_{pk-p-k+2}}\times |\Gamma| \Omega_{k-1}\times\biggl(-\frac{k-2}{pk-p-k}\biggr)\Omega_{pk-p-k},\end{aligned}$$ and the other $w$’s are zero. Simple calculations give $$\label{w2}
w_k = -w_{k-2}
= \frac{\Gamma (\frac{k}{2})}{\sqrt{\pi}\Gamma(\frac{k-1}{2})}|\Gamma|.$$
### Contribution of the boundary $\partial M$ {#contribution-of-the-boundary-partial-m .unnumbered}
Here, we obtain the coefficients $w'_{d-e}$ in (\[w’\]) when $M$ is given in (\[M\]).
Suppose that $\mathcal{X}=[a,b]$. Then, $$\partial M = \{\phi(a,\theta) \mid \theta \in \Theta\} \sqcup \{\phi(b,\theta) \mid \theta \in \Theta\}.$$ Let $x=a$ and $\theta\in\Theta$ be fixed. Then, $\phi(a,\theta)\in\partial M$. The metric of the boundary $\partial M$ at $\phi(a,\theta)$ is $$\biggl(\frac{\partial\phi}{\partial\theta_i}\biggr)^\top \biggl(\frac{\partial\phi}{\partial\theta_j}\biggr)\bigg|_{(a,\theta)} = (G(\theta))_{ij}.$$
Note that $\partial ({\mathrm{co}}(M)) = {\mathrm{co}}(\partial M)$. The support cone of ${\mathrm{co}}(M)$ at $\phi(a,\theta)\in\partial ({\mathrm{co}}(M))$ is $$S_{\phi(a,\theta)}({\mathrm{co}}(M)) = L \oplus K_1,$$ where $$L = {\mathrm{span}}\biggl\{h\otimes\psi,\ \frac{\partial h}{\partial\theta_i}\otimes\psi,\,i=1,\ldots,k-2\biggr\}, \quad
K_1 = \biggl\{\lambda\biggl({h\otimes\frac{\partial\psi}{\partial x}}\biggr) \mid \lambda \ge 0\biggr\}.$$ This is a direct sum (the Minkowski sum) of a linear subspace and a cone. To obtain its dual cone, the following lemma is useful.
Let $K_{1}$ be a cone, and $L$ be a linear subspace. Let $K = K_1 \oplus L$. Then, the dual cone of $K$ is $K^* = K_1^*\cap L^\perp$.
Because $$K_1^* =
\bigl\{ \lambda(h_{(0)}\otimes\psi_{(1)}) \mid \lambda\le 0 \bigr\} \oplus {\mathrm{span}}\bigl\{ h_{(0)}\otimes\psi_{(1)} \bigr\}^\perp$$ and $$L^\perp = {\mathrm{span}}\bigl\{
h_{(i)}\otimes\psi_{(j)},\,i=0,\ldots,k-2,\,j=1,\ldots,p-1 \bigr\},$$ we have $$\begin{aligned}
N_{\phi(a,\theta)}({\mathrm{co}}(M))
= K^* = K_1^*\cap L^\perp
= \bigl\{ &\lambda(h_{(0)}\otimes \psi_{(1)}) \mid \lambda\le 0 \bigr\} \oplus {\mathrm{span}}\bigl\{ \\
& h_{(0)}\otimes \psi_{(j)},\,j=2,\ldots,p-1;\, \\
& h_{(i)}\otimes\psi_{(j)},\,i=1,\ldots,k-2,\,j=1,\ldots,p-1 \bigr\},\end{aligned}$$ with $\dim N_{\phi(a,\theta)}({\mathrm{co}}(M)) = pk-p-k+1$.
The second fundamental form of ${\mathrm{co}}(\partial M)$ at $\phi(a,\theta)$ is $$\begin{aligned}
\label{2nd-boundary}
-\biggl(\frac{\partial^{2}\phi(a,\theta)}{\partial\theta_i\partial\theta_j}\biggr)^\top v =
-\biggl(\frac{\partial^2 h}{\partial\theta_i\partial\theta_j} \otimes \psi\biggr)^\top v,\ \ v\in N_{\phi(a,\theta)}({\mathrm{co}}(M)).\end{aligned}$$ We can easily see that the second fundamental form (\[2nd-boundary\]) is always zero. Therefore, the contribution of the boundary to $w'_{d-e}$ in (\[w’\]) is only to case of $e=0$. That is, all $w'_i$ except $w'_{k-1}$ are zero. The contribution of the boundary $\{\phi(a,\theta)\mid\theta\in\Theta\}$ to $w'_{k-1}$ is $$\begin{aligned}
\frac{1}{\Omega_{k-1-0}\Omega_{p(k-1)-(k-1)+0}}
\int_{\partial M}|G|^{1/2} \,{{\mathrm{d}}}\theta
\int_{N_{\phi(a,\theta)}({\mathrm{co}}(M))\cap{\mathbb{S}}^{n-1}} \,{{\mathrm{d}}}v
=& \frac{{\mathrm{Vol}}({\mathbb{S}}^{k-2}){\mathrm{Vol}}(\mbox{half of }{\mathbb{S}}^{pk-p-k})}{\Omega_{k-1}\Omega_{pk-p-k+1}} \\
=& \frac{1}{2}.\end{aligned}$$ The contribution of the other boundary $\{ \phi(b,\theta) \mid \theta\in\Theta \}$ to $w_{k-1}$ has the same value of 1/2. Moreover, if the number of connected components of $\Gamma$ exceeds one, we need to select all boundaries. Since the number of boundaries is $2\chi(\Gamma)$, we have $$\label{w'2}
w'_{k-1} = \frac{1}{2}\times 2\chi(\Gamma) = \chi(\Gamma).$$
Substituting (\[w2\]) and (\[w’2\]) into (\[p-hat\]) yields (\[p-hat2\]).
Proof of Theorem \[thm:critical\_radius\]
-----------------------------------------
We apply the formula (\[critical\]) for $\theta_{\mathrm{c}}$ to the case where $M$ is given in (\[M\]).
Let $$\begin{aligned}
u = \phi(\tilde x,\tilde\theta)
= h(\tilde\theta) \otimes \psi(\tilde x), \quad
v = \phi(x,\theta) = h(\theta) \otimes \psi(x),\end{aligned}$$ and write $h=h(\theta)$, $\tilde h=h(\tilde\theta)$, $\psi=\psi(x)$, $\tilde\psi=\psi(\tilde x)$. We discuss the two cases (i) $\psi\in{\mathrm{Int}}\Gamma$ and (ii) $\psi\in\partial\Gamma$ separately.
Case (i). Suppose that $\psi(x)\in{\mathrm{Int}}\Gamma$. Write $\phi_{\theta_i}=\{\partial h(\theta)/\partial\theta_i\}\otimes\psi(x)$, $\phi_x=h(\theta)\otimes\psi_x(x)$, $\psi_x=\partial\psi(x)/\partial x$.
The orthogonal projection matrix onto the space $T_{\phi}({\mathrm{co}}(M))={\mathrm{span}}\{ \phi,\phi_{\theta_i},\phi_x \}$ is $$\begin{aligned}
P_v
& = \begin{pmatrix} \phi & \phi_{\theta_i} & \phi_x \end{pmatrix}_{p(k-1)\times k}
\left( \begin{pmatrix} \phi^\top \\ \phi_{\theta_i}^\top \\ \phi_x^\top \end{pmatrix}
\begin{pmatrix} \phi & \phi_{\theta_i} & \phi_x \end{pmatrix} \right)^{-1}_{k\times k}
\begin{pmatrix} \phi^\top \\ \phi_{\theta_i}^\top \\ \phi_x^\top \end{pmatrix}_{k\times p(k-1)} \\
& = \begin{pmatrix} \phi & \phi_{\theta_i} & \phi_x \end{pmatrix}
\begin{pmatrix} 1 & 0 & 0 \\ 0 & G(\theta) & 0 \\ 0 & 0 & g(x) \end{pmatrix}^{-1}
\begin{pmatrix} \phi^\top \\ \phi_{\theta_i}^\top \\ \phi_x^\top \end{pmatrix} \\
& = \phi\phi^\top + (\phi_{\theta_i})G(\theta)^{-1}(\phi_{\theta_i})^\top
+ \frac{1}{g(x)}\phi_x\phi_x^\top \\
& = h h^\top \otimes\psi\psi^\top + (I_{k-1} - h h^\top) \otimes \psi\psi^\top
+ \frac{1}{g} h h^\top \otimes \psi_x\psi_x^\top \\
& = I_{k-1} \otimes \psi\psi^\top + \frac{1}{g} h h^\top \otimes \psi_x\psi_x^\top.\end{aligned}$$ As $P_v^\perp(w)=I(w)-P_v(w)$, $$\label{infarg}
\frac{(1-u^\top v)^2}{\Vert(I_n-P_v) u\Vert^2}
= \frac{\{1-(\tilde h\otimes\tilde\psi)^\top (h\otimes\psi)\}^2}
{\Vert(I_{k-1}\otimes I_p -I_{k-1}\otimes\psi\psi^\top -\frac{1}{g} h h^\top\otimes\psi_x\psi_x^\top)(\tilde h\otimes\tilde\psi)\Vert^2}.$$ Let $s = \psi^\top(x)\psi(\tilde x) = \psi^\top\tilde\psi$, $r = \psi_x^\top(x)\psi(\tilde x) = \psi_x^\top\tilde\psi$, and $\alpha = h(\theta)^\top h(\tilde\theta) = h^\top \tilde h$. In (\[infarg\]), the numerator is $(1-\alpha s)^2$, and the denominator is $$\begin{aligned}
\left\Vert \tilde h\otimes\tilde\psi-s \tilde h\otimes\psi-\frac{\alpha r}{g} h\otimes\psi_x\right\Vert^2
=& 1 + s^2 + \frac{\alpha^2 r^2}{g} - 2 s^2 - 2\frac{\alpha r}{g}\alpha r \\
=& 1 - s^2 - \frac{\alpha^2 r^2}{g} = 1 - s^2 - \alpha^2 t^2,\end{aligned}$$ where $t = r/\sqrt{g} = \psi_x(x)^\top\psi(\tilde x)/\Vert \psi_x(x)\Vert$. Hence, $$\mbox{(\ref{infarg})} = \frac{(1-\alpha s)^{2}}{1-s^2-\alpha^2 t^2}.$$
The infimum is taken over “$x\ne \tilde x$ or $\theta\ne\tilde\theta$”, or equivalently, “$x\ne \tilde x$ or $\alpha\ne 1$”. However, when $x=\tilde x$ and $\alpha\ne 1$, the argument of the infimum is $(1-\alpha)^2/0=\infty$. Therefore, we can exclude case $x=\tilde x$ from the infimum argument.
Case (ii). Suppose that $\psi(x)\in\partial\Gamma$. Fix a point on the boundary $$v = \phi(x,\theta) = h(\theta)\otimes\psi(x) \in \partial M.$$ The support cone of ${\mathrm{co}}(M)$ at $v$ is $$S_v({\mathrm{co}}(M)) =
{\mathrm{span}}\{ \phi,\phi_{\theta_i} \}\oplus \{\lambda\varepsilon\phi_x \mid \lambda\ge 0 \}$$ where $\varepsilon=\varepsilon(x)=1$ if $\psi_x$ is inward to $\Gamma$, $\varepsilon=-1$ if $\psi_x$ is outward to $\Gamma$.
The orthogonal projection operator onto the cone $S_v({\mathrm{co}}(M))$ is $w\mapsto P_v(w)$, where $$\begin{aligned}
P_v(w)
=& \phi\phi^\top w + \phi_{\theta}G^{-1}(\theta)\phi_{\theta}^{-1}w
+ \frac{\phi_x}{\Vert\phi_x\Vert^2}\max\{ 0,\varepsilon\phi_x^\top w \} \\
=& (I_{k-1} \otimes \psi\psi^\top)w
+ \frac{\phi_x}{\Vert\phi_x\Vert^2}\max\{0,\varepsilon\phi_x^\top w\}.\end{aligned}$$ Hence, $$\begin{aligned}
P_v^\perp(w)
= w-P_v(w)
&= w-(I_k \otimes \psi\psi^\top)w-\frac{h \otimes \psi_x}{g}\max\{0,\varepsilon(h \otimes \psi_x)^\top w\}.\end{aligned}$$ Substituting $u = \phi(\tilde x,\tilde\theta) = \tilde h\otimes\tilde\psi$, $$P_v^\perp(u-v) = \tilde h\otimes\tilde\psi - s \tilde h\otimes\psi
- \frac{\max\{0,\varepsilon\alpha r\}}{g} h \otimes \psi_x,$$ $$\begin{aligned}
\Vert P_v^\perp(u-v)\Vert^2
=& 1 + s^2 + \frac{\max\{0,\varepsilon\alpha r\}^2}{g} - 2 s^2 - 2\alpha r \frac{\max\{0,\varepsilon\alpha r\}}{g} \\
=& 1 - s^2 - \frac{\max\{0,\varepsilon\alpha r\}^2}{g} = 1 - s^2 - \max\{0,\varepsilon\alpha t\}^2,\end{aligned}$$ and we have $$\begin{aligned}
\frac{(1-u^\top v)^2}{\Vert P_v^\perp(u-v)\Vert^2}
= \frac{(1-\alpha s)^2}{1-s^2-\max\{0,\varepsilon \alpha t\}^2}.\end{aligned}$$ For the same reason as in case (i), the infimum is taken over the set $x\ne \tilde x$ and $\alpha\in[-1,1]$.
Proof of Theorem \[thm:local\_critical\_radius\]
------------------------------------------------
We use the same notations as in the proof of Theorems \[thm:main\] and \[thm:critical\_radius\]. The local critical radius $\theta_{\mathrm{c,loc}}$ defined by (\[critical-local\]) is rewritten as $$\begin{aligned}
\tan^2\theta_{\mathrm{c,loc}}
= \liminf_{|x-\tilde x|\to 0,\,\alpha\to 1}\frac{(1-\alpha s)^2}{1-s^2-\alpha^2 t^2}.\end{aligned}$$ Let $\tilde x=x+\Delta$ and $\alpha=1-\delta$, and consider $\Delta\to 0$ and $\delta\to 0$. Write $\psi_x=\partial\psi/\partial x$, $\psi_{xx}=\partial^2\psi/\partial x^2$, etc., and $g=\psi_x^\top\psi_x$, $\gamma=\psi_{xx}^\top\psi_x$, and $\eta=\psi_{xx}^\top\psi_{xx}$ as before. Noting that $$\begin{aligned}
& 0={{\mathrm{d}}}(\psi^\top\psi)/{{\mathrm{d}}}x=2\psi_x^\top\psi, \\
& 0={{\mathrm{d}}}^2(\psi^\top\psi)/{{\mathrm{d}}}x^2=2\psi_{xx}^\top\psi+2\psi_x^\top\psi_x, \\
& 0={{\mathrm{d}}}^3(\psi^\top\psi)/{{\mathrm{d}}}x^3=2\psi_{xxx}^\top\psi+6\psi_{xx}^\top\psi_x, \\
& 0={{\mathrm{d}}}^4(\psi^\top\psi)/{{\mathrm{d}}}x^4=2\psi_{xxxx}^\top\psi+8\psi_{xxx}^\top\psi_x+6\psi_{xx}^\top\psi_{xx},\end{aligned}$$ we have $$\psi_{xx}^\top\psi=-g, \quad \psi_{xxx}^\top\psi=-3\gamma, \quad \psi_{xxxx}^\top\psi+4\psi_{xxx}^\top\psi_x=-3\eta.$$ Substituting these, we have $$\begin{aligned}
s = \psi(x)^\top\psi(\tilde x)
=& \psi^\top \biggl( \psi + \psi_x\Delta + \frac{1}{2}\psi_{xx}\Delta^2 + \frac{1}{6}\psi_{xxx}\Delta^3 + \frac{1}{24}\psi_{xxxx}\Delta^4 \biggr) +o(\Delta^4) \\
=& 1 - \frac{1}{2}g\Delta^2 - \frac{1}{2}\gamma\Delta^3 + \frac{1}{24}\psi_{xxxx}^\top\psi\Delta^4 +o(\Delta^4),
\\
r = \psi_x(x)^\top\psi(\tilde x)
=& \psi_x^\top \biggl( \psi + \psi_x\Delta + \frac{1}{2}\psi_{xx}\Delta^2 + \frac{1}{6}\psi_{xxx}\Delta^3 \biggr) +o(\Delta^3) \\
=& g\Delta + \frac{1}{2}\gamma\Delta^2 + \frac{1}{6}\psi_{xxx}^\top\psi_x\Delta^3 +o(\Delta^3),
\\
t^2 = \frac{r^2}{g}
=& g\Delta^2 + \gamma\Delta^3 + \frac{1}{4g}\gamma^2\Delta^4 + \frac{1}{3}\psi_{xxx}^\top\psi_x\Delta^4 +o(\Delta^4),\end{aligned}$$ and $$\begin{aligned}
1-s^2-t^2
=& (1-s)\{2-(1-s)\}-t^2 = 2(1-s)-(1-s)^2-t^2 \\
=& 2 \biggl( \frac{1}{2}g\Delta^2 + \frac{1}{2}\gamma\Delta^3 - \frac{1}{24}\psi_{xxxx}^\top\psi\Delta^4 \biggr) - \biggl( \frac{1}{2}g\Delta^2 \biggr)^2 \\
& - \biggl( g\Delta^2 + \gamma\Delta^3 + \frac{1}{4g}\gamma^2\Delta^4 + \frac{1}{3}\psi_{xxx}^\top\psi_x\Delta^4 \biggr) +o(\Delta^4) \\
=& \frac{1}{4}\biggl( \eta-g^2-\frac{\gamma^2}{g} \biggr) \Delta^4 +o(\Delta^4)
= \frac{1}{4}\kappa g^2\Delta^4 +o(\Delta^4),\end{aligned}$$ where $$\kappa=\kappa(x)=\frac{\eta}{g^2}-\frac{\gamma^2}{g^3}-1.$$ Note that $\kappa$ is nonnegative because $$0\le \det\begin{pmatrix}\psi^\top \\ \psi_x^\top \\ \psi_{xx}^\top \end{pmatrix}
\begin{pmatrix}\psi & \psi_x & \psi_{xx} \end{pmatrix}
= \det\begin{pmatrix} 1 & 0 & -g \\
0 & g & \gamma \\
-g & \gamma & \eta \end{pmatrix} = \kappa g^3.$$
For the order of $\delta$, we consider two cases: (i) $\delta/\Delta^2 \sim g c$ ($0\le c<\infty$) and (ii) $\Delta^2/\delta \sim 0$.
For case (i), noting that $\alpha=1-\delta$, $1-s\sim g\Delta^2/2$, $t^2\sim g\Delta^2$, $$\begin{aligned}
& (1-\alpha s)^2 = \{1-s + \delta - \delta (1-s)\}^2 \sim (1-s + \delta)^2
\sim \biggl( \frac{1}{2}g\Delta^2+g c\Delta^2 \biggr)^2 = \frac{1}{4}g^2(1+2c)^2\Delta^4,
\\
& 1 -s^2-\alpha^2 t^2 = 1 -s^2 -t^2 +2\delta t^2 -\delta^2t^2
\sim \frac{1}{4}g^2(\kappa + 8 c)\Delta^4,\end{aligned}$$ hence $$\begin{aligned}
\frac{(1-\alpha s)^2}{1-s^2-\alpha^2 t^2}
\sim \frac{(1+2c)^2}{\kappa+8c}.
\label{c}\end{aligned}$$ We consider the minimum value of (\[c\]) for $c\ge 0$. For $c>-\kappa/8$, (\[c\]) has a unique minimum value $1-\kappa/4$ at $c=(2-\kappa)/4$. Therefore, $$\min_{c\ge 0}\frac{(1+2c)^2}{\kappa+8c} =
\begin{cases}
\displaystyle
1-\frac{\kappa}{4} & (\kappa\le 2), \\
\displaystyle
\frac{1}{\kappa} & (\kappa\ge 2).
\end{cases}$$
For case (ii), $$(1 -\alpha s)^2 = \{1-s + \delta - \delta (1-s)\}^2 \sim \delta^2,$$ $$1 -s^2 -\alpha^2 t^2 = 1 -s^2 -t^2 +2\delta t^2 -\delta^2 t^2
\sim 2 g\Delta^2\delta,$$ and $$\frac{(1-\alpha s)^2}{1-s^2-\alpha^2 t^2}
\sim \frac{\delta}{2g\Delta^2} \to \infty.$$
In summary, we have $$\tan^{2}\theta_{\mathrm{c,loc}} = \min\biggl\{
\inf_{x:\kappa(x)\le 2} \biggl\{1-\frac{\kappa(x)}{4}\biggr\},
\inf_{x:\kappa(x)\ge 2} \frac{1}{\kappa(x)} \biggr\}.$$
Coverage probability under model misspecification {#sec:bound}
-------------------------------------------------
First, note that the best parameter under the model $\beta_i^\top f(x)$ is given by $\beta_i^*= \Sigma X^\top g_i$, where $$X = \begin{pmatrix} f(x_1)^\top \\ \vdots \\ f(x_n)^\top \end{pmatrix}, \quad \Sigma = (X^\top X)^{-1} = \left( \sum_{i=1}^n f(x_i) f(x_i)^\top \right)^{-1}, \quad g_i = \begin{pmatrix} g_i(x_1) \\ \vdots \\ g_i(x_n) \end{pmatrix}.$$ Let $$y_i = \begin{pmatrix} y_{i1} \\ \vdots \\ y_{in} \end{pmatrix}, \quad \varepsilon_i = \begin{pmatrix} \varepsilon_{i1} \\ \vdots \\ \varepsilon_{i1} \end{pmatrix}.$$ The least square estimator for $\beta_i^*$ is $\widehat\beta_i^* = \Sigma X^\top y_i$, which is distributed as ${{\mathcal{N}}}_p(\beta_i^*,\Sigma)$. Let $b_{1-\alpha}$ be the threshold for $1-\alpha$ bands when the assumed model is the true model. The approximate value of $b_{1-\alpha}$ can be obtained by the tube method.
On the other hand, when the true model is $g_i(x)$, the coverage probability becomes $$\begin{aligned}
& \Pr\Biggl( \biggl|\sum_{i=1}^k c_i (\widehat\beta_i^*)^\top f(x) - \sum_{i=1}^k c_i g_i(x)\biggr| \le b_{1-\alpha} \sqrt{f(x)^\top\Sigma f(x)} \ \ \mbox{for all }x\in\mathcal{X},\ c\in\mathcal{C}\Biggr) \nonumber \\
&= \Pr\Biggl( \max_{x\in\mathcal{X},\,c\in\mathcal{C}}\frac{\sum_{i=1}^k c_i (\widehat\beta_i^*)^\top f(x) - \sum_{i=1}^k c_i g_i(x)}{\sqrt{f(x)^\top\Sigma f(x)}} \le b_{1-\alpha} \Biggr) \nonumber \\
&=
\Pr\Biggl( \max_{x\in\mathcal{X},\,c\in\mathcal{C}} \biggl[
\frac{\sum_{i=1}^k c_i (\widehat\beta_i^*-\beta_i^*)^\top f(x)}{\sqrt{f(x)^\top\Sigma f(x)}}+\frac{\sum_{i=1}^k c_i \{(\beta_i^*)^\top f(x) - g_i(x)\}}{\sqrt{f(x)^\top\Sigma f(x)}}\biggr] \le b_{1-\alpha} \Biggr).
\label{pr}\end{aligned}$$
Noting that, for the two functions $h_1(y)$ and $h_2(y)$ on $\mathcal{Y}$, if $\max_{y\in\mathcal{Y}} (-h_i(y)) = \max_{y\in\mathcal{Y}} h_i(y)$, then $$\max_y h_1(y) \le \max_y (h_1(y)+h_2(y)) + \max_y (-h_2(y)) = \max_y (h_1(y)+h_2(y)) + \max_y h_2(y),$$ hence, $$\max_y h_1(y) - \max_y h_2(y) \le \max_y (h_1(y)+h_2(y)) \le \max_y h_1(y) + \max_y h_2(y).$$ Therefore, (\[pr\]) is bounded below and above by $1-{{\overline P}}(b_{1-\alpha}-\delta)$ and $1-{{\overline P}}(b_{1-\alpha}+\delta)$, respectively, where $$\begin{aligned}
\delta
=& \max_{x\in\mathcal{X},\,c\in\mathcal{C}}\frac{\sum_{i=1}^k c_i \{(\beta_i^*)^\top f(x) - g_i(x)\}}{\sqrt{f(x)^\top\Sigma f(x)}} \nonumber \\
&= \max_{x\in\mathcal{X}}\sqrt{\frac{\sum_{i=1}^k [(\beta_i^*)^\top f(x) - g_i(x)
-\frac{1}{k}\sum_{i=1}^k \{(\beta_i^*)^\top f(x) - g_i(x)\}]^2}{f(x)^\top\Sigma f(x)}}\end{aligned}$$ as given in (\[delta\]), and $${{\overline P}}(b) = \Pr\Biggl( \max_{x\in\mathcal{X},\,c\in\mathcal{C}} \frac{\sum_{i=1}^k c_i (\widehat\beta_i^*-\beta_i^*)^\top f(x)}{\sqrt{f(x)^\top\Sigma f(x)}} \ge b \Biggr).$$ An upper bound for the bias of coverage probability for a $1-\alpha$ confidence band is $$\max\bigl\{ \alpha-{{\overline P}}(b_{1-\alpha}+\delta), {{\overline P}}(b_{1-\alpha}-\delta)-\alpha \bigr\},$$ which is approximated by $$\begin{aligned}
\Delta =
\max\bigl\{ \alpha-{{\overline P_{\mathrm{tube}}}}\bigl(b_{\mathrm{tube},1-\alpha}+\delta\bigr), {{\overline P_{\mathrm{tube}}}}\bigl(b_{\mathrm{tube},1-\alpha}-\delta\bigr)-\alpha \bigr\}\end{aligned}$$ in (\[Delta\]), where ${{\overline P_{\mathrm{tube}}}}(b)$ is the tube approximation formula for $P(b)$ given in (\[p-hat2\]), and $b_{\mathrm{tube},1-\alpha}$ is the solution of ${{\overline P_{\mathrm{tube}}}}(b)=\alpha$.
Note that (\[pr\]) is $$\Pr\Biggl( \max_{x\in\mathcal{X}} \frac{\sum_{i=1}^k [(\widehat\beta_i^*)^\top f(x) - g_i(x)
-\frac{1}{k}\sum_{i=1}^k \{(\widehat\beta_i^*)^\top f(x) - g_i(x)\}]^2}{f(x)^\top\Sigma f(x)} \le b_{1-\alpha}^2 \Biggr),$$ which is used for the simulation study.
### Acknowledgment {#acknowledgment .unnumbered}
The authors thank Professor Toshihiko Shiroishi and Dr. Toyoyuki Takada of the National Institute of Genetics, Japan, for providing the dataset of Takada et al. [@Takada-etal08].
[99]{}
Adler, R.J. and Taylor, J.E. (2007). [*Random Fields and Geometry*]{}, Springer.
de Boor, C. (1978). [*A Practical Guide to Splines*]{}, Springer.
Casella, G. and Hwang, J.T.G. (2012). Shrinkage confidence procedures, [*Statistical Science*]{}, [**27**]{} (1), 51–60.
Faraway, J.J. and Sun, J. (1995). Simultaneous confidence bands for linear regression with heteroscedastic errors, [*Journal of the American Statistical Association*]{}, [**90**]{} (431), 1094–1098.
Federer, H. (1959). Curvature measures, [*Transactions of the American Mathematical Society*]{}, [**93**]{} (3), 418–491. Hotelling, H. (1939). Tubes and spheres in $n$-spaces, and a class of statistical problems, [*American Journal of Mathematics*]{}, [**61**]{} (2), 440–460.
Jamshidian, M., Liu, W., and Bretz, F. (2010). Simultaneous confidence bands for all contrasts of three or more simple linear regression models over an interval, [*Computational Statistics and Data Analysis*]{}, [**54**]{} (6), 1475–1483.
Johansen, S. and Johnstone, I.M. (1990). Hotelling’s theorem on the volume of tubes: Some illustrations in simultaneous inference and data analysis, [*The Annals of Statistics*]{}, [**18**]{} (2), 652–684.
Johnstone, I. and Siegmund, D. (1989). On Hotelling’s formula for the volume of tubes and Naiman’s inequality, [*The Annals of Statistics*]{}, [**17**]{} (1), 184–194.
Krivobokova, T., Kneib, T., and Claeskens, G. (2010). Simultaneous confidence bands for penalized spline estimators, [*Journal of the American Statistical Association*]{}, [**105**]{} (490), 852–863.
Kuriki, S. and Takemura, A. (2001). Tail probabilities of the maxima of multilinear forms and their applications, [*The Annals of Statistics*]{}, [**29**]{} (2), 328–371.
Kuriki, S. and Takemura, A. (2009). Volume of tubes and the distribution of the maximum of a Gaussian random field, Selected Papers on Probability and Statistics, American Mathematical Society Translations Series 2, [**227**]{} (2), 25–48.
Leeb, H., Pötscher, B.M., and Ewald, K. (2015). On various confidence intervals post-model-selection, [*Statistical Science*]{}, [**30**]{} (2), 216–227.
Liu, W. (2010). [*Simultaneous Inference in Regression*]{}, Chapman & Hall/CRC.
Liu, W., Jamshidian, M., and Zhang, Y. (2004). Multiple comparison of several linear regression models, [*Journal of the American Statistical Association*]{}, [**99**]{} (466), 395–403.
Liu, W., Lin, S., and Piegorsch, W.W. (2008). Construction of exact simultaneous confidence bands for a simple linear regression model, [*International Statistical Review*]{}, [**76**]{} (1), 39–57.
Liu, W., Wynn, H.P., and Hayter, A.J. (2008). Statistical inferences for linear regression models when the covariates have functional relationships: polynomial regression, [*Journal of Statistical Computation and Simulation*]{}, [**78**]{} (4), 315–324.
Lu, X. and Chen, J.T. (2009). Exact simultaneous confidence segments for all contrast comparisons, [*Journal of Statistical Planning and Inference*]{}, [**139**]{} (8), 2816–2822.
Muirhead, R.J. (2005). [*Aspects of Multivariate Statistical Theory*]{}, 2nd ed., Wiley.
Naiman, D.Q. (1986). Conservative confidence bands in curvilinear regression, [*The Annals of Statistics*]{}, [**14**]{} (3), 896–906.
Spurrier, J.D. (1999). Exact confidence bounds for all contrasts of three or more regression lines, [*Journal of the American Statistical Association*]{}, [**94**]{} (446), 483–488.
Spurrier, J.D. (2002). Exact multiple comparisons of three or more regression lines: Pairwise comparisons and comparisons with a control, [*Biometrical Journal*]{}, [**44**]{} (7), 801–812.
Sun, J. (1993). Tail probabilities of the maxima of Gaussian random fields, [*The Annals of Probability*]{}, [**21**]{} (1), 34–71.
Sun, J., and Loader, C.R. (1994). Simultaneous confidence bands for linear regression and smoothing, [*The Annals of Statistics*]{}, [**22**]{} (3), 1328–1346.
Sun, J., Loader, C., and McCormick, W.P. (2000). Confidence bands in generalized linear models, [*The Annals of Statistics*]{}, [**28**]{} (2), 429–460.
Sun, J., Raz, J., and Faraway, J.J. (1999). Confidence bands for growth and response curves, [*Statistica Sinica*]{}, [**9**]{}, 679–698.
Takada, T., Mita, A., Maeno, A., Sakai, T., Shitara, H., Kikkawa, Y., Moriwaki, K, Yonekawa, H., and Shiroishi, T. (2008). Mouse inter-subspecific consomic strains for genetic dissection of quantitative complex traits, [*Genome Research*]{}, [**18**]{} (3), 500–508.\
<http://genome.cshlp.org/content/18/3/500> Takemura, A. and Kuriki, S. (2002). On the equivalence of the tube and Euler characteristic methods for the distribution of the maximum of Gaussian fields over piecewise smooth domains, [*The Annals of Applied Probability*]{}, [**12**]{} (2), 768–796.
Uusipaikka, E. (1983). Exact confidence bands for linear regression over intervals, [*Journal of the American Statistical Association*]{}, [**78**]{} (383), 638–644.
Weyl, H. (1939). On the volume of tubes, [*American Journal of Mathematics*]{}, [**61**]{} (2), 461–472.
Wolfram Research, Inc. (2016). Mathematica, Version 11.0. Working, H. and Hotelling, H. (1929). Applications of the theory of error to the interpretation of trends, [*Journal of the American Statistical Association*]{}, [**24**]{} (165), 73–85.
Worsley, K. (1995). Boundary corrections for the expected Euler characteristic of excursion sets of random fields, with an application to astrophysics, [*Advances in Applied Probability*]{}, [**27**]{} (4), 943–959. Wynn, H.P. and Bloomfield, P. (1971). Simultaneous confidence bands in regression analysis, [*Journal of the Royal Statistical Society, Series B*]{}, [**33**]{} (2), 202–217.
Zhou, S., Shen, X., and Wolfe, D.A. (1998). Local asymptotics for regression splines and confidence regions, [*The Annals of Statistics*]{}, [**26**]{} (5), 1760–1782.
[^1]: Department of Statistical Sciences, Graduate University for Advanced Studies, 10-3 Midoricho, Tachikawa, Tokyo 190-8562, Japan, Email: [[email protected]]{}
[^2]: The Institute of Statistical Mathematics, 10-3 Midoricho, Tachikawa, Tokyo 190-8562, Japan, Email: [[email protected]]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper considers learning deep features from long-tailed data. We observe that in the deep feature space, the head classes and the tail classes present different distribution patterns. The head classes have a relatively large spatial span, while the tail classes have significantly small spatial span, due to the lack of intra-class diversity. This uneven distribution between head and tail classes distorts the overall feature space, which compromises the discriminative ability of the learned features. Intuitively, we seek to expand the distribution of the tail classes by transferring from the head classes, so as to alleviate the distortion of the feature space. To this end, we propose to construct each feature into a “ feature cloud”. If a sample belongs to a tail class, the corresponding feature cloud will have relatively large distribution range, in compensation to its lack of diversity. It allows each tail sample to push the samples from other classes far away, recovering the intra-class diversity of tail classes. Extensive experimental evaluations on person re-identification and face recognition tasks confirm the effectiveness of our method.'
author:
- |
Jialun Liu$^1$[^1], Yifan Sun$^2$, Chuchu Han$^3$, Zhaopeng Dou$^4$, Wenhui Li$^1$[^2]\
[$^1$Jilin University]{} [$^2$Megvii Inc.]{} [$^3$Huazhong University of Science and Technology ]{} [$^4$Tsinghua University]{}\
[`` ]{} [``]{} [``]{}
bibliography:
- 'egbib.bib'
title: 'Deep Representation Learning on Long-tailed Data: A Learnable Embedding Augmentation Perspective'
---
Introduction
============
{width="0.9\linewidth"}
Large-scale datasets play a crucial role in training a model with good discriminability. However, in the real-world, large-scale datasets often exhibit extreme long-tailed distribution [@everingham2010pascal; @guo2016ms]. Some identities have sufficient samples, while for other massive identities, only very few samples are available. They are defined as the head classes and tail classes, respectively. With this distribution, deep neural networks have been found to perform poorly on tail classes [@buda2018systematic].
The issue is clearly shown in Fig. \[fig:1\]. Firstly, we select eight head classes from DukeMTMC-reID dataset [@zheng2017unlabeled; @ristani2016performance], and the visualization of features is shown in Fig. \[fig:1\] (a). It is observed that the head classes have relatively large spatial span. With larger inter-class distances, the head classes can be well distinguished. This observation is consistent with [@yin2019feature]. Further, we reduce the samples of some head classes so they are marked as tail classes. As shown in Fig. \[fig:1\] (b), we observe that samples from tail class distribute narrowly in the learned feature space, due to the lack of intra-class diversity. This uneven distribution between head and tail classes distorts the overall feature space and consequentially compromises the discriminative ability of the learned features. The phenomenon indicates that when the class-imbalance exists, the feature distribution is closely related to the number of class samples. Since the tail classes with scanty training samples cannot provide sufficient intra-class diversity for learning discriminative features, they cannot be accurately distinguished from other classes.
With this insight, we propose to transfer the intra-class distribution of head classes to tail classes in the feature space. We model the distribution of angles between features and the corresponding class center, which can reflect the distribution of the intra-class features. We make a statistical analysis of the intra-class angular variance. Under a setting of person re-identification $\left \langle H20,S4 \right \rangle$, where $H$ is the number of head classes and $S$ is the number of samples per tail class, in baseline([@wang2018cosface]), the variations of head classes are centered at 0.463 ($\pm 0.0014$), and that of tail classes are centered at 0.288 ($\pm 0.0089$). It clearly shows that 1) tail classes have smaller variance and 2) the sample number per class is the dominating factor on the variance. Our target is to encourage the tail classes to achieve similar intra-class angular variability with the head classes in training. Specifically, we first calculate the distribution of angles between the features of head class and their corresponding class center. By averaging the angular variances of all the head classes, we obtain the overall variance of head classes. Next, we consider transferring the variance of head class to each tail class. To this end, we build a feature cloud around each tail instance in the embedding layer, and several pseudo features can be sampled with the same identities. Each instance with the corresponding feature cloud will have a relatively large distribution range, making the tail classes have a similar angular distribution with head class. Our method enforces stricter supervision on the tail classes, and thus leads to higher within-class compactness. As Figure. \[fig:1\] (c) shows, with the compensation of intra-class diversity during training, the tail classes are separated from other classes by a clear margin. Under the setting of person re-identification: $\left \langle H20,S4 \right \rangle$, the intra-class angular variance of tail classes turn out over even lower(than the tail classes in baseline), which is cnetered at 0.201.
Moreover, to improve the flexibility of the method, we abandon the explicit definition of head class and tail class. Compared with some methods that divide the two classes, our approach makes the calculation entirely related to the distribution of dataset, and there is no human interference.
We summarize the contributions of our work as follows:
- We propose a learnable embedding augmentation perspective to alleviate the problem of discriminative feature learning on long-tailed data, which transfers the intra-class angular distribution learned from head classes to tail classes.
- Extensive ablation experiments on person re-identification and face recognition demonstrate the effectiveness of the proposed method.
Related Work
============
**Feature learning on imbalanced datasets.** Recent works for feature learning on imbalanced data are mainly divided into three manners: re-sampling [@buda2018systematic], re-weighting [@mahajan2018exploring], and data augmentation[@choi2018stargan]. The re-sampling technique includes two types: over-sampling the tail classes and under-sampling the head classes. Over-sampling manner samples the tail data repeatedly, which enables the classifier to learn tail classes better. But it may lead to over-fitting of tail classes. To reduce the risk of over-fitting, SMOTE [@chawla2002smote] is proposed to generate synthetic data of the tail class. It randomly places the newly created instances between each tail class data point and its nearest neighbor. The under-sampling manner [@drummond2003c4] reduces the amount of data from head classes while keeping the tail classes. But it may lose valuable information on head classes when data imbalance is extreme. The re-weighting approach assigns different weights for different classes or different samples. The traditional method re-weights classes proportionally to the inverse of their frequency of samples. Cui *et al.* [@cui2019class] improve the re-weighting by the inverse effective number of samples. Li *et al.* [@li2019gradient] propose a method which down-weights examples with either very small gradients or large gradients because examples with small gradients are well-classified and those with large gradients tend to be outliers. Recently, data augmentation methods based on Generative Adversarial Network (GAN) [@choi2018stargan] are popular. [@yin2019feature] and [@gao2018low] transfer the semantic knowledge learned from the head classes to compensate tail classes, which encourage the tail classes to have similar data distribution to the head classes. All the methods divide the classes into the head or tail class, while our method abandons the constraint.
{width="1.0\linewidth"}
**Loss function.** Loss function plays an important role in deep feature learning, and the most popular one is the Softmax loss [@sun2014deep]. However, it mainly considers whether the samples can be correctly classified and lacks the constraint of inter-class distance and intra-class distance. In order to improve the feature discrimination, many loss functions are proposed to enhance the cosine and angular margins between different classes. Wen *et al.* [@wen2016discriminative] design a center loss to reduce the distance between the sample and the corresponding class center. The L2-Softmax [@ranjan2017l2] and NormFace [@wang2017normface] add normalization to produce represented features and achieve better performance. Besides normalization, adding a margin can enhance the discrimination of features by inserting distance among samples of different classes. A-Softmax Loss [@liu2017sphereface] normalizes the weights and adds multiplicative angular margins to learn more divisible angular characteristics. CosFace [@wang2018cosface] adds an additive cosine margin to compress the features of the same class in a compact space, while enlarging the gap of features of different classes. ArcFace [@deng2019arcface] puts an additive margin into angular space so that the loss relies on both sine and cosine dynamically to learn more angular characteristics. Our baseline is CosFace [@wang2018cosface] and ArcFace [@deng2019arcface]. Although we model the intra-class angle, which is similar to them, our goal is to solve the problem of discriminative feature learning on long-tailed data.
The Proposed Approach
=====================
In this section, A brief description of our method is given in Section \[sec:3.1\]. We review the baseline in Section \[sec:3.2\]. We describe the updating process of the class center and the calculation of angular distribution in Section \[sec:3.3\]. The construction of the feature cloud for a tail instance is detailed in Section \[sec:3.4\].
Overview of Framework {#sec:3.1}
---------------------
The framework of our method is shown in Fig. \[fig:framework\]. First, the head data and tail data are fed into the deep model to extract high-dimensional features. And we consider to model the distribution of intra-class features by the distribution of angles between features and their corresponding class center. Then the center of each class is calculated, as to be detailed in Section \[sec:3.3\]. We build an angle memory for each class, which is used to store the angles between the features and their class center. Assuming the angles obey the Gaussian distribution, the angular distributions of head class and tail class can be denoted as $\theta_h \sim {\rm N} (\mu_h,\sigma_h^2)$ and $\theta_t \sim {\rm N} (\mu_t,\sigma_t^2)$, respectively. Next, we transfer the angular variance learned from the head class to every tail class. Consequently, the intra-class angular diversity of tail class is similar to the head class. Specifically, we build a feature cloud around each tail instance. An instance sampled from the feature cloud has the same identity with the tail instance. The angle between them is $\theta_\Delta$ and $\theta_\Delta \sim {\rm N} (0,\sigma_h^2 - \sigma_t^2)$. We assume the two distribution: $\theta_t \sim {\rm N} (\mu_t,\sigma_t^2)$ and $\theta_\Delta \sim {\rm N} (0,\sigma_h^2 - \sigma_t^2)$ are independent of each other. By transformation, the new intra-class angular distribution of tail class is built as $\theta_t +\theta_\Delta \sim {\rm N} (\mu_t,\sigma_h^2)$ in training process. Finally, we use the original features of head classes and the reconstructed features of tail classes to calculate the loss.
Baseline Methods {#sec:3.2}
----------------
The traditional softmax loss optimizes the decision boundary between two categories, but it lacks the constraint of inter-class distance and intra-class distance. CosFace [@wang2018cosface] effectively minimizes intra-class distance and maximums inter-class distance by the introducing a cosine margin to maximize the decision margin in the angular space. The loss function can be formulated as: $$\label{eq:3}
L_{1} = -\frac{1}{N}\sum_{n=1}^{N}{\log{\frac{e^{s (\cos(\theta_{{y}}) - m_c)}}{e^{s (\cos(\theta_{{y}}) - m_c)} + \sum_{j\neq y_n}^C{e^{s \cos(\theta_{j})}}}}},
\vspace{-1mm}$$ where $N$ and $C$ are the mini-batch size and the number of total classes, respectively. $y$ is the label of $n$-th image. We define the feature vector of $n$-th image and the weight vector of class $y$ as $f_n$ and $W_{y}$, respectively. $f_n$ and $W_{y}$ are normalized by $l_2$ normalisation and the norm of feature vector is rescaled to $s$. $\theta_{y}$ is the angle between the weight $W_{y}$ and the feature $f_n$. $m_c$ is a hyper-parameter controlling the magnitude of the cosine margin.
Different from CosFace [@wang2018cosface], ArcFace [@deng2019arcface] employs an additive angular margin loss, which is formulated as: $$\label{eq:4}
L_{2} = -\frac{1}{N}\sum_{n=1}^{N}{\log{\frac{e^{s (\cos(\theta_{{y}} + m_a))}}{e^{s (\cos(\theta_{{y}} + m_a))} + \sum_{j\neq y}^C{e^{s \cos(\theta_{j})}} }}},
\vspace{-1mm}\textbf{}$$ where $m_a$ is an additive angular margin penalty between feature vector $f_n$ and its corresponding $W_{y}$. It aims to enhance the intra-class compactness and inter-class distance simultaneously.
In this paper, we choose CosFace [@wang2018cosface] and ArcFace [@deng2019arcface] as baseline. The reasons are as follows:
- They have achieved the state-of-the-art performance in the face recognition task, which can be seen as strong baselines in the community of deep feature learning.
- They optimize the intra-class similarity by achieving much lower intra-class angular variability. Since our method employs intra-class angles to model the intra-class feature distribution, the two loss functions can be naturally combined with our method.
Learning the intra-class angular distribution {#sec:3.3}
---------------------------------------------
The intra-class angular diversity can intuitively show the diversity of intra-class features. In this section, we study the distribution of angles between the features and their corresponding class center. $c_i$ denotes the $i$-th class center of features. $f^k_i$ is the $k$-th instance feature of class $i$. $c_i$ has the same dimension as $f^k_i$. So, we can calculate the angle between $f^k_i$ and $c_i$ as follow: $$\label{eq:5}
{\beta _{i,k}} = arccos(\frac{{f_i^k{c_i}}}{{||f_i^k||||{c_i}||}}),$$
where the $c_i$ should be updated in the training processing. Ideally, we need to take the entire training samples into account and average the features of every class in each epoch. Obviously, this approach is impractical and inefficient. Inspired by [@wen2016discriminative], we also perform the update based on a mini-batch. In each mini-batch, the class center is computed by averaging the feature vectors of the corresponding class. To avoid the misleading by some mislabelled samples, we set a center learning rate $\gamma$ to update the class center. The updating method of $c_i$ is formulated as: $$\label{eq:6}
{c_i^l} = {(1-{\gamma} ){c_i^l}+\gamma{c_i^{l-1}}},$$ where $c_i^l$ is the center of class $i$ in $l$-th mini-batch. Each class center is updated by the center of current and previous mini-batch.
For the class $i$, we maintain an angle memory $\beta_i$ to storage the angles between the features and their corresponding class center $c_i$. The size of angle memory is formulated as: $$\label{eq:7}
{S_i} = K_i \times P.$$ $K_i$ is the sample number of the $i$-th class. $P$ is a hyper-parameter determining the angle memory per class. Then we calculate the mean $\mu_i$ and variance $\sigma_i^2$ of $\beta_i$. The angular distribution of the class $i$ is formulated as $\beta _i$ $\sim {\rm N} (\mu_i,\sigma_i^2)$.
Constructing the feature cloud for tail data {#sec:3.4}
--------------------------------------------
![We transfer the intra-class angular distribution learned from the head class to the tail class. Replace each tail instance with a feature cloud. []{data-label="fig:transfer"}](transfer.pdf){width="1.0\linewidth"}
**Feature cloud**: given a specified feature $f$ of a tail class, we generate several virtual feature vectors $f'$ around it (subject to the probability distribution learned from head class), yielding the so-called “feature cloud”.
In this section, we elaborate the process of constructing the feature cloud for a tail instance. First, like the previous works [@yin2019feature; @zhong2019unequal], we assign a label to mark the head and tail class, yielding the vanilla version of our method. On the other hand, we introduce a full version which abandons the explicit division of head and tail class. This manner is more flexible since it is only related to the distribution of the dataset.
**Vanilla version.** We strictly divide the head class and the tail class through a threshold **$T$**. If the number of samples belonging to class $i$ is larger than **$T$**, the $i$-th class is defined as a head class. Otherwise, it is defined as a tail class.
In the Section \[sec:3.3\], we have calculated the angular distribution of each class, which is assumed to lie in Gaussian distribution. By averaging the variance of all head classes, we obtain the overall variance of the head class. The mean is computed in the similar way. So the overall angular distribution of the head class is as follow: $$\label{eq:8}
{\mu_h} = \frac{{\sum\limits_{z = 1}^{{C_h}} {{\mu _z}} }}{{{C_h}}},\quad \quad \quad {\sigma _h^2} = \frac{{\sum\limits_{z = 1}^{{C_h}} {\sigma _z^2} }}{{{C_h}}},$$ where $C_{h}$ is the number of head classes. $\mu_z$ and $\sigma_z^2$ is the angular mean and variance of the $z$-th head class, respectively. $\mu_h$ and ${\sigma}_h^2$ describe the overall angular distribution of the head class. We can also obtain the class center for every tail class. The angular distribution of the $x$-th tail class is denoted as $ {\rm N} (\mu_t^x,{\sigma_t^x}^2)$.
For the head classes, they include sufficient samples which show the intra-class angular diversity. In general, ${\sigma_h}$ is greater than ${\sigma_t}$, so our target is to transfer $\sigma_h^2$ to each tail class. As in Fig. \[fig:transfer\], we construct a feature cloud around each feature of $x$-th tail class. By this way, the space spanned by tail class is enlarged, in training, and the real tail instances are pushed away from other classes. The angle between the feature belonging to the $x$-th tail class and a feature sampled from its corresponding feature cloud is ${\alpha_x}$, where ${\alpha_x} \sim {\rm N} (0,\sigma_h^2 - {\sigma_t^x}^2)$ and ${\alpha _x} \in \mathbb{R} {^{1 \times C}}$. In training, the feature sampled from the feature cloud shares the same identity with the real tail feature. We have assumed the two distributions: ${\rm N} (\mu_t^x,{\sigma_t^x}^2)$ and ${\rm N} (0,\sigma_h^2 - {\sigma_t^x}^2)$ are independent of each other in Section \[sec:3.1\]. So the original angular distribution of the $x$-th tail class is transferred from $ {\rm N} (\mu_t^x,{\sigma_t^x}^2)$ to $ {\rm N} (\mu_t^x,{\sigma_h}^2)$.
The new loss functions based on CosFace [@wang2018cosface] and ArcFace [@deng2019arcface] are defined as:
$$\label{eq:9}
\small
L_{3} = -\frac{1}{N}\sum_{n=1}^{N}{\log{\frac{e^{s(\cos(\theta_{{y}} + \alpha_{y}) -m_c) }}{e^{s(\cos(\theta_{{y}} + \alpha_{y}) - m_c)} + \sum_{j\neq y}^{C}{e^{s \cos(\theta_{{j}} + \alpha_{y})}}}}},$$
$$\label{eq:10}
\small
L_{4} = -\frac{1}{N}\sum_{n=1}^{N}{\log{\frac{e^{s(\cos(\theta_{y} + \alpha_{y} + m_a)) }}{e^{s(\cos(\theta_{{y}} + \alpha_{y} + m_a))} + \sum_{j\neq y}^C{e^{s \cos(\theta_{{j}} + t\alpha_{y})}}}}},$$
in Eq.\[eq:9\] and \[eq:10\], $\theta + \alpha$ and $\theta + \alpha + m_a$ are all clipped in the range $[0, \pi]$. $N$ and $C$ are the mini-batch size and class number, respectively. $\theta_{y}$ is the angle between the feature $f_n$ and the weight $W_{y}$ . $s$ is the scale, and $m_c$, $m_a$ are the cosine margin and the angular margin in CosFace [@wang2018cosface] and ArcFace [@deng2019arcface], respectively. If $y$ is a head class, $\alpha_y$ = 0. As the training progresses, the tail class has the rich angular diversity as head class.
Actually, we approximate the angle ($\theta'$) between the feature sampled from feature cloud and the weight. If $\alpha > 0$, we approximate $\theta'$ by the upper bound of it, and the lower bound when $\alpha \le 0$. The proof is given below.
#### Proposition.
We denote a feature in the tail class as ${f}$, and $W$ is the corresponding weight vector in the full connection layer. ${f'}$ is a feature randomly sampled from the feature cloud around $f$. $$\small{ \langle f, W \rangle = \theta, \quad \langle f, f' \rangle = \alpha_{+}, \quad \langle W, f'\rangle=\theta'},$$ $$\small{ \|f\| = \|w\| = \|f'\| = 1, \quad 0 \le \theta+\alpha_{+} \le \pi },$$ where $\langle a,b \rangle$ represents the angle between vector $a$ and $b$, and $\|a\|$ represent the norm of vector $a$. We want to prove: $|\theta-\alpha_{+}| \le \theta' \le \theta + \alpha_{+}$.
#### Proof.
Simply, we suppose that $f = [1, 0, \cdots, 0]$, then $W = [\cos{\theta}, w_2, \cdots, w_n]$. We use the Householder transformation [@householder1958unitary] to transform $W$ to $V$, where $V = [\cos{\theta}, \sin{\theta}, 0, \cdots, 0]$. Let $P = I - 2{U \cdot {U^T}}$, where $U = {W-V} / \|W-V\|$, then $f = Pf, V=PW, \hat{f'}= Pf'$. $P$ is an orthogonal transformation which preserves the inner product and norm. Therefore, we have $$\langle f, V \rangle = \theta, \quad \langle f, \hat{f'} \rangle= \alpha_{+}, \quad \langle V, \hat{f'} \rangle = \theta'.$$ Denote $\hat{f'}=[\hat{f_1'}, \hat{f_2'}, \cdots, \hat{f_n'}]$, then $$\cos\alpha_{+} = f \cdot \hat{f'} = \hat{f_1'}, \quad \hat{f_2'}^2 + \cdots + \hat{f_n'}^2 = {\sin ^2}\alpha_{+}.$$ We get $\hat{f_2'}\sin \theta \in [- sin\alpha_{+} sin\theta ,sin\alpha_{+} sin\theta ]$, where $\theta \in [0,\pi ]$. Further, we have $${\cos {\theta'}} = {{\hat{f}'} \cdot V} = \cos \alpha_{+} \cos \theta + \hat{f_2'} \sin \theta ,$$ $${\cos{\theta'}} \in {[cos(\theta + \alpha_{+}), cos(\theta - \alpha_{+})]}.$$ We get the conclusion: $|\theta-\alpha_{+}| \le \theta' \le \theta + \alpha_{+}$.
Although $\alpha \sim{N}(0, \sigma^2)$, we only need to focus on $\alpha \in [ - \pi ,\pi ]$, since $\theta +\alpha$ is clipped in the range $[0, \pi]$.
- when $0 \le \alpha \le \pi$, substituting $\alpha$ for $\alpha_{+}$, we have $|\theta-\alpha|\le \theta'\le\theta+\alpha$, in which $\theta+\alpha$ is the upper bound.
- when $-\pi \le \alpha \le 0$, substituting $-\alpha$ for $\alpha_{+}$, we have $|\theta - (-\alpha)| \le \theta' \le \theta + \alpha$, which is equivalent to $\theta+\alpha \le \theta' \le \theta-\alpha$, so $\theta+\alpha$ is the lower bound.
**Full version.** The distorted feature space is well repaired by constructing a feature cloud around a tail instance. But the process in the vanilla version is inflexible. We need to set a threshold $T$ to divide the head and tail classes, artificially. The overall angular distribution in Eq.\[eq:8\] only depends on the head classes. In the full version, the explicit definition is discarded. We have observed that the intra-class diversity is positively correlated with the number of samples, in general. Therefore, we calculate the overall variance by weighting the angular variance of each class. The weight is the number of samples in each class. The final variance is formulated as: $$\label{eq:11}
\vspace{-1mm}
{\sigma}^2 = \sum\limits_{i = 1}^C {\frac{{({K_i} - 1)\sigma _i^2}}{{\sum {{\rm{(}}{{\rm{K_i}}} - 1)} }}} ,$$ where $C$ is the number of classes, and $K_i$ is the number of samples belong to class $i$. $\sigma_i^2$ is the angular variance of the $i$-th class. A smaller $K_i$ means that the variance of the $i$-th class almost has no contribution to the final variance, so the final variance mainly depends on the classes with sufficient samples. For $i$-th class, if $\sigma _i^2 < \sigma ^2$, it means the class $i$ has poor intra-class diversity. Therefore $\alpha$ is available in Eq.\[eq:9\] and \[eq:10\], and we construct the feature cloud for each instance sampled from class $i$.
The advantage of the full version is that the calculation of feature cloud entirely depends on the distribution of the dataset. There is no human interference in the process.
Experiments
===========
In this section, we conduct extensive experiments to confirm the effectiveness of our method. First we describe the experimental settings. Then we show the performance on person re-identification and face recognition with different long-tailed settings.
Settings
--------
**Person re-identification.** Evaluations are conducted on three datasets: Market-1501 [@zheng2015scalable], DukeMTMC-reID [@ristani2016performance; @zheng2017unlabeled] and MSMT17 [@wei2018person]. To study the impact of the ratio between head classes and tail classes on training a person re-identification system, we construct several long-tailed datasets based on the original dataset. We rank the classes by their number of samples. The top $150, 100, 50$ and $20$ identities are marked as the head class, respectively. The rest is treated as the tail classes, and the number of samples is reduced to $5$ each class. In this way, we form the training sets of $\left \langle H150,S5 \right \rangle$, $\left \langle H100,S5 \right \rangle$, $\left \langle H50,S5 \right \rangle$, and $\left \langle H20,S5 \right \rangle$. For training, we train the model with a learning rate of $3.5e-4$ for ResNet-50 [@he2016deep]. The last layer of the network is followed by a Batch Normalization layer (BN). The optimizer is Adam. The scale $s$ and $m_c$ of CosFace [@wang2018cosface] are set to be $24$ and $0.2$, respectively. The scale $s$ and $m_a$ of ArcFace [@wang2018cosface] are set to be $16$ and $0.2$, respectively. The learning rate of class center $\gamma$ is set to be $0.1$. For testing, the 2048-d global features after BN are used for evaluation. The cosine distance of features is computed as the similarity score. We use two evaluation metrics: Cumulative Matching Characteristic (CMC) and mean average precision(mAP) to evaluate our method.
**Face recognition.** We adopt the widely used dataset MS-Celeb-1M for training. The original MS-Celeb-1M data is known to be very noisy, so we clean the dirty face images and exclude the $79K$ identities and $1M$ images. We rank the classes through the number of samples they have. The top $5K$ and $3K$ are selected as head classes. Among the rest classes, we select the first $10K$ and $20K$ as tail classes and randomly pick $5$ images per class. In this way, we form the training set of $\left \langle H5K,T20K \right \rangle$, $\left \langle H5K,T10K \right \rangle$, $\left \langle H3K,T20K \right \rangle$ and $\left \langle H3K,T10K \right \rangle$. The face images are resized to $112 \times 112$. For training, we choose the ResNet-18 [@he2016deep] as our backbones. We train the model for $30$ epoch by adopting the triangular learning rate policy[@smith2017cyclical], and construct feature cloud at the start of the third cycle. The scale $s$ and $m_c$ of CosFace [@wang2018cosface] are set to be $64$ and $0.35$. The scale $s$ and $m_a$ of ArcFace [@wang2018cosface] are set to be $64$ and $0.5$. We extract $512$-D features for model inference. For testing, we evaluate our method on LFW [@huang2008labeled], MegaFace challenge1 (MF1) [@kemelmacher2016megaface] and IJB-C [@maze2018iarpa]. We report our results on the Rank-1 accuracy of LFW and MF1, and different TPR@FPR of IJB-C TPR@FPR.
{width="1.0\linewidth"}
\[V&F\_comparison\]
---------------------------- ------ -------- ------ -------- ------ --------
mAP Rank-1 mAP Rank-1 mAP Rank-1
HA-CNN [@li2018harmonious] 75.7 91.2 63.8 80.5 - -
PCB [@sun2018beyond] 77.4 92.3 66.1 81.8 40.4 68.2
Mancs [@wang2018mancs] 82.3 93.1 71.8 84.9 - -
CosFace 79.5 92.4 73.0 85.6 49.2 75.3
ArcFace 81.1 92.5 73.2 85.8 50.5 75.5
---------------------------- ------ -------- ------ -------- ------ --------
: Comparison with the advanced methods on the Market-1501, DukeMTMC-reID and MSMT17 datasets[]{data-label="basel"}
\
-- -------------------------------------- -------- ---------- ---------- ----------
mAP Rank-1 mAP Rank-1
SVDNet [@sun2017svdnet] 62.1 82.3 56.8 76.7
BraidNet [@wang2018person] 69.5 83.7 69.5 76.4
CamStyle [@zhong2018camera] 71.6 89.5 57.6 78.3
Advesarial [@huang2018adversarially] 70.4 86.4 62.1 79.1
Dual [@du2018interaction] 76.6 91.4 64.6 81.8
Mancs [@wang2018mancs] 82.3 93.1 84.9 71.8
IANet [@hou2019interaction] 83.1 94.4 73.4 87.1
DG-Net [@zheng2019joint] 86.0 **94.8** **74.8** 86.6
AACN [@xu2018attention] 66.9 85.9 59.2 76.8
PSE [@saquib2018pose] 69.0 87.7 62.0 79.8
PCB [@sun2018beyond] 77.4 92.3 66.1 81.8
SPReID [@kalayeh2018human] 81.3 92.5 70.9 84.4
LEAP-CF 84.2 94.4 74.2 **87.8**
LEAF-AF 83.2 93.5 74.2 86.9
-- -------------------------------------- -------- ---------- ---------- ----------
: Comparison with state-of-the-art methods on Market-1501 and DukeMTMC-reID. Three groups: global features(GF), part features(PF) and ours. LEAP-CF and LEAP-AF are our full version combined with CosFace and ArcFace, respectively.[]{data-label="M&D"}
\
Experiments on person re-identification
---------------------------------------
**Performance of baseline.** Table \[basel\] reports the results of the baseline. We compare our baseline with the advanced methods. Our baseline achieves very competitive performance, which is reliable.
**Comparison with state-of-the-art approaches.** In the original Market-1501 dataset, the statistical result shows that 77.44% of the identities only include no more than $20$ images, and only 3.86% of identities have more than $40$ images. It indicates that the Market-1501 dataset shows the long-tailed distribution. The similar phenomenon also exists in DukeMTMC-reID. We compare our full version with the state-of-the-art methods on Market-1501 and DukeMTMC-reID. The comparisons are summarized in Table \[M&D\]. It shows that our baseline has surpassed many advanced methods. And our method further improve the performance compared with baseline. Specifically, LEPA-CF achieves 94.4% on rank-1 for Market-1501, and 87.8% on rank-1 for DukeMTMC-reID. We further evaluate our method on a recently released large scale dataset MSMT17 [@wei2018person]. MSMT17 also shows a long-tailed distribution, and the comparison is shown in Table \[MSMT17\]. Compared with DG-Net [@zheng2019joint], our performance is very close to it. However, our method is a simple but efficient method, which does not use GAN to generate many image-level samples.
Methods mAP Rank-1 Rank-5 Rank-10
-------------------------------------------- ---------- ---------- ---------- ----------
GoogleNet [@szegedy2015going] 23.0 47.6 65.0 71.8
Pose-driven [@su2017pose] 29.7 58.0 73.6 79.4
Verif-Identif [@zheng2018discriminatively] 31.6 60.5 76.2 81.6
GLAD [@wei2017glad] 34.0 61.4 76.8 81.6
PCB [@sun2018beyond] 40.4 68.2 81.2 85.5
IANet [@hou2019interaction] 46.8 75.5 85.5 88.7
DG-Net [@zheng2019joint] **52.3** **77.2** **87.4** **90.5**
LEAP-CF 50.8 76.7 86.9 90.0
LEAP-AF 51.3 76.3 86.5 89.8
: Comparison with advanced methods on the MSMT17.[]{data-label="MSMT17"}
\
-- --------------------- ---------- ---------- ---------- ----------
Method $\downarrow$ mAP Rank-1 mAP Rank-1
CosFace 67.3 86.3 57.3 75.6
LEAP-CV **70.6** **86.9** **59.4** **77.1**
ArcFace 70.6 87.3 60.2 77.6
LEAP-AV **71.3** **87.9** **60.6** **78.7**
CosFace 62.8 83.3 52.6 70.3
LEAP-CV **68.7** **86.5** **55.6** **74.8**
ArcFace 68.0 86.6 56.7 74.8
LEAP-AV **69.8** **87.3** **57.9** **76.5**
CosFace 60.5 80.7 48.0 67.7
LEAP-CV **67.3** **84.9** **53.1** **73.0**
ArcFace 64.2 83.8 51.1 71.1
LEAP-AV **67.1** **84.6** **54.4** **73.5**
CosFace 55.6 78.6 47.0 66.0
LEAP-CV **64.1** **83.2** **52.4** **72.7**
ArcFace 60.1 81.1 50.5 69.3
LEAP-AV **64.3** **82.2** **54.2** **73.7**
-- --------------------- ---------- ---------- ---------- ----------
: Controlled experiments by varying the ratio between head and tail data in training sets. $H$ is the number of head class. The sample size of head class remains the same. $S$ denotes that the number of samples for each tail class is clipped to a certain amount. CosFace and ArcFace are baselines. LEAP-CV and LEAP-AV are vanilla version combined with CosFace and ArcFace.[]{data-label="table:2"}
\
**Evaluation with the vanilla version.** We evaluate the effectiveness of the vanilla version. For comparison, we train the baseline model on the long-tailed person re-identification datasets under the supervision of CosFace [@wang2018cosface] and ArcFace [@deng2019arcface]. We compare our method with baseline methods. The results are shown in Table \[table:2\]. We have the following observations. First, compared with CosFace, ArcFace has higher Rank-1 and mAP accuracy on the same long-tailed setting. For example, on Market-1501 with $\left \langle H20,S5 \right \rangle$, ArcFace achieves the Rank-1 accuracy of $81.1\%$, while the Rank-1 accuracy of CosFace is $78.6\%$. This indicates that Arcface has a stronger robustness for the long-tailed person re-identification. Second, in different long-tailed settings, the proposed LEAP method combined with CosFace and ArcFace achieves consistently better results than the baseline with significant margins. This indicates that the LEAP is a robust method for long-tailed data distribution. Third, as the long-tailed distribution is more serious, the improvement of our method becomes even more obvious. For example, in the $\left \langle H20,S5 \right \rangle$ setting on DukeMTMC-reID, the improvement of LEAP-CV reaches $+6.7\%$ (from $66.0\%$ to $72.7\%$) in the Rank-1 accuracy.
**Comparison between vanilla version and full version.** We show the results comparison of vanilla version and full version under different long-tailed settings in Figure \[V&F\_comparison\]. We observe that the full version obtains the results very close to vanilla version, and even better results in some settings. By this experiment, we justify that compared with those methods which need a label to distinguish between head class and tail class, the full version is more flexible.
-- --------------------- ---------- ---------- ---------- ----------
Method $\downarrow$ mAP Rank-1 mAP Rank-1
CosFace 55.6 78.6 47.0 66.0
LEAP-CF **65.2** **83.4** **52.7** **72.8**
ArcFace 60.1 81.1 50.5 69.3
LEAP-AF **63.9** **83.2** **54.2** **73.6**
CosFace 43.1 67.7 36.0 53.7
LEAP-CF **54.7** **76.8** **42.6** **63.0**
ArcFace 49.4 73.8 39.7 58.8
LEAP-AF **56.5** **77.9** **44.2** **64.4**
CosFace 31.9 55.5 25.6 40.8
LEAP-CF **43.5** **67.2** **33.2** **51.1**
ArcFace 36.2 60.1 28.9 46.7
LEAP-AF **44.1** **66.1** **34.3** **53.3**
-- --------------------- ---------- ---------- ---------- ----------
: Impact analysis of different tail data for feature learning.[]{data-label="table:3"}
\
LFW MegaFace
-------------------- --------------------- ----------- ----------- ----------- ----------- -----------
Train $\downarrow$ Method $\downarrow$ Rank-1 Rank-1 1e-3 1e-4 1e-5
CosFace 98.73 81.41 83.35 73.32 63.42
LEAP-CV **98.88** **81.78** **83.83** **73.96** **64.64**
ArcFace 98.60 81.08 82.30 72.45 62.46
LEAP-AV **98.67** **81.69** **83.16** **72.97** **63.22**
CosFace 98.87 82.72 84.77 76.71 68.19
LEAP-CV **98.98** **83.16** **84.82** **77.21** **68.88**
ArcFace 98.73 82.76 84.45 76.22 66.93
LEAP-AV **99.10** **83.36** **85.70** **77.77** **68.05**
CosFace 97.65 72.27 79.08 68.06 56.52
LEAP-CV **97.97** **73.19** **79.60** **69.18** **58.89**
ArcFace 97.82 72.45 78.24 66.99 55.31
LEAP-AV **98.07** **73.43** **78.84** **67.82** **55.75**
CosFace 98.02 74.06 81.21 71.68 61.03
LEAP-CV **98.23** **75.18** **81.87** **72.16** **62.62**
ArcFace 98.28 75.24 81.09 71.36 61.60
LEAP-AV **98.73** **76.28** **82.61** **73.21** **62.72**
\[table:5\]
**The impact of tail data.** When the head class is reduced gradually and the tail data is increasing, the results are shown in Table \[table:3\], we observe the effect of tail data on performance. We gradually reduce the samples of each tail class, which results in insufficient training data, and the performance of the model drops dramatically. However, our method still makes a large margin improvement over the baseline. For example, in the $\left \langle H20,S3 \right \rangle$ setting on Market-1501, even the number of samples for each tail class is only $3$, the improvement of LEAP-CF reaches $+11.7\%$ (from $55.5\%$ to $67.2\%$) in the Rank-1 accuracy.
![Different timings of constructing the feature cloud for tail data. (a) Combined our method with CosFace [@wang2018cosface]. (b) Combined our method with ArcFace [@deng2019arcface] []{data-label="fig:timing"}](timing.pdf){width="1.0\linewidth"}
**Timing of feature cloud for tail data.** We investigate the effect of timing of constructing a feature cloud for tail data on Market-1501 and DukeMTMC-reID dataset. We take a long-tailed version: $\left \langle H20,S4 \right \rangle$ as an example. The varying curve of the results is shown in Figure \[fig:timing\]. (a) Combined our method with CosFace [@wang2018cosface]. It can be seen that, when epoch is in the range of $10$ to $30$, our results are just marginally impacted and the best results are achieved. (b) Combined our method with ArcFace [@deng2019arcface]. Our results are impacted just marginally and the best results are achieved from $20$-th to $30$-th epoch.
Experiments on face recognition
-------------------------------
To further verify the observations in the person re-identification task, we perform a similar set of experiments on the face recognition task. Different from person re-identification, the dataset of face recognition has a relatively large scale. Unlike the angle memory and class center are updated frequently, this will result in a huge storage burden and computing time. In order to improve the training efficiency, we reduce the update frequency to every 5 iterations. The result is shown in Table \[table:5\]. On LFW, our performance is improved slightly since LFW has been well solved. MF1 and IJB-C are the most challenging testing benchmark for face recognition. We report the Rank-1 accuracy of MF1 and TPR@FPR of IJB-C. Compared with the baseline, our method obtains consistency improvement. For example, in the $\left \langle H3K,T10K \right \rangle$ setting, we evaluate our method on IJB-C, the LEAP-CV improves TPR@FPR(1e-5) from 56.52% to 58.89%. in the $\left \langle H3K,T20K \right \rangle$ setting, we evaluate our method on MF1, the LEAP-CV improves Rank-1 accuracy from 74.06% to 75.18%.
Conclusions
===========
In this paper, we proposed a novel approach for feature learning on long-tailed data. We transfer the diversity of intra-class angles learned from the head class to the tail class. We replace each tail instance with a feature cloud which is determined by the head and tail classes. In this way, the space spanned by tail class is extended, in feature space. Moreover, instead of distinguishing between head and tail classes, we propose a flexible solution which can learn the intra-class angular distribution adaptively. Experiments on person re-identification and face recognition consistently show that our method achieves significant improvement in many types of long-tailed settings.
[^1]: Equal contribution.
[^2]: Corresponding author.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '> A natural and established way to restrict the constraint satisfaction problem is to fix the relations that can be used to pose constraints; such a family of relations is called a *constraint language*. In this article, we study arc consistency, a heavily investigated inference method, and three extensions thereof from the perspective of constraint languages. We conduct a comparison of the studied methods on the basis of which constraint languages they solve, and we present new polynomial-time tractability results for singleton arc consistency, the most powerful method studied.'
author:
- 'Hubie Chen[^1], Victor Dalmau[^2], Berit Grußien[^3]'
bibliography:
- 'local.bib'
title: Arc Consistency and Friends
---
\[theorem\] [Lemma]{} \[theorem\] [Remark]{} \[theorem\] [Fact]{} \[theorem\] [Claim]{} \[theorem\] [Corollary]{} \[theorem\] [Definition]{} \[theorem\] [Proposition]{} \[theorem\] [Warning]{} \[theorem\] [Example]{}
Introduction
============
Background
----------
The constraint satisfaction problem (CSP) involves deciding, given a set of variables and a set of constraints on the variables, whether or not there is an assignment to the variables satisfying all of the constraints. Cases of the constraint satisfaction problem appear in many fields of study, including artificial intelligence, spatial and temporal reasoning, logic, combinatorics, and algebra. Indeed, the constraint satisfaction problem is flexible in that it admits a number of equivalent formulations. In this paper, we work with the formulation as the relational homomorphism problem: given two similar relational structures ${\mathbf{A}}$ and ${\mathbf{B}}$, does there exist a homomorphism from ${\mathbf{A}}$ to ${\mathbf{B}}$? In this formulation, one can view each relation of ${\mathbf{A}}$ as containing variable tuples that are constrained together, and the corresponding relation of ${\mathbf{B}}$ as containing the permissible values for the variable tuples [@FederVardi].
The constraint satisfaction problem is in general NP-hard; this general intractability has motivated the study of restricted versions of the CSP that have various desirable complexity and algorithmic properties. A natural and well-studied way to restrict the CSP is to fix the value relations that can be used to pose constraints; in the homomorphism formulation, this corresponds to fixing the right-hand side structure ${\mathbf{B}}$, which is also known as the *constraint language*. Each structure ${\mathbf{B}}$ then gives rise to a problem ${{\mathsf{CSP}}}({\mathbf{B}})$, and one obtains a rich family of problems that include boolean satisfiability problems, graph homomorphism problems, and satisfiability problems on algebraic equations. One of the primary current research threads involving such problems is to understand for which finite-universe constraint languages ${\mathbf{B}}$ the problem ${{\mathsf{CSP}}}({\mathbf{B}})$ is polynomial-time tractable [@bulatov-valeriote]; there is also work on characterizing the languages ${\mathbf{B}}$ for which the problem ${{\mathsf{CSP}}}({\mathbf{B}})$ is contained in lower complexity classes such as L (logarithmic space) and NL (non-deterministic logarithmic space) [@dalmau-pathwidth; @larose-tesson]. With such aims providing motivation, there have been efforts to characterize the languages amenable to solution by certain algorithmic techniques, notably, representing solution spaces by generating sets [@IMMVW] and consistency methods [@LaroseZadori; @abd-affine; @bkbw], which we now turn to discuss.
Checking for *consistency* is a primary reasoning technique for the practical solution of the CSP, and has been studied theoretically from many viewpoints [@LaroseZadori; @abd-affine; @akv; @abd; @aft; @bkbw; @atserias-weyer]. The most basic and simplest form of consistency is *arc consistency*, which algorithmically involves performing inferences concerning the set of feasible values for each variable. The question of how to efficiently implement an arc consistency check has been studied intensely, and highly optimized implementations that are linear in both time and space have been presented. In general, a consistency check typically involves running an efficient method that performs inference on bounded-size sets of variables, and which can sometimes detect that a CSP instance is inconsistent and has no solution. While these methods exhibit one-sided error in that they do not catch all non-soluble CSP instances (as one expects from the conjunction of their efficiency and the intractability of the CSP), it has been shown that, for certain constraint languages, they can serve as complete decision procedures, by which is meant, they detect an inconsistency if (and only if) an instance has no solution. As an example, *unit propagation*, a consistency method that can be viewed as arc consistency specialized to SAT formulas, is well-known to decide the Horn-SAT problem in this sense.
Contributions
-------------
In this paper, we study arc consistency and three natural extensions thereof from the perspective of constraint languages. The extensions of AC that we study are look-ahead arc consistency (LAAC) [@SLAAC]; peek arc consistency (PAC) [@peek], and singleton arc consistency (SAC) [@db-sac; @bd-theoretical]. Each of these algorithms is natural, conceptually simple, readily understandable, and easily implementable using arc consistency as a black box. Tractability results for constraint languages have been presented for AC by Feder and Vardi [@FederVardi] (for instance); and for LAAC and PAC in the previously cited work. In fact, for each of these three algorithms, characterizations of the class of tractable languages have been given, as we discuss in the paper.
We give a uniform presentation of these algorithms (Section \[sect:algorithms\]), and conduct a comparison of these algorithms on the basis of which languages they solve (Section \[sect:comparison\]). Our comparison shows, roughly, that the algorithms can be placed into a hierarchy: solvability of a language by AC or LAAC implies solvability by PAC; solvability by PAC in turn implies solvability by SAC (see Section \[sect:comparison\] for precise statements). We also study the strictness of the containments shown. We thus contribute to a basic, foundational understanding of the scope of these algorithms and of the situations in which these algorithms can be demonstrated to be effective.
We then present new tractability results for singleton arc consistency (Section \[sect:tractability\]). We prove that languages having certain types of *2-semilattice polymorphisms* can be solved by singleton arc consistency; and, we prove that any language having a *majority polymorphism* is solvable by singleton arc consistency. The presence of a majority polymorphism is a robust and well-studied condition: majority polymorphisms were used to give some of the initial language tractability results, are known to exactly characterize the languages such that *3-consistency* implies *global consistency* (we refer to [@CCC] for definitions and more details), and gave one of the first large classes of languages whose constraint satisfaction problem could be placed in non-deterministic logarithmic space [@majority-nl]. While the languages that we study are already known to be polynomial-time tractable [@CCC; @twosemilattices], from the standpoint of understanding the complexity and algorithmic properties of constraint languages, we believe our tractability results to be particularly attractive for a couple of reasons. First, relative to a fixed language, singleton arc consistency runs in quadratic time [@bd-theoretical], constituting a highly non-trivial running time improvement over the cubic time bound that was previously known for the studied languages. Also, in showing that these languages are amenable to solution by singleton arc consistency, we demonstrate their polynomial-time tractability in an alternative fashion via an algorithm that is different from the previously used ones; the techniques that we employ expose a different type of structure in the studied constraint languages.
Preliminaries
=============
Our definitions and notation are fairly standard. For an integer $k \geq 1$, we use the notation $[k]$ to denote the set containing the first $k$ positive integers, that is, the set $\{ 1, \ldots, k \}$.
#### Structures.
A *tuple* over a set $B$ is an element of $B^k$ for a value $k \geq 1$ called the *arity* of the tuple; when ${\overline{t}}$ is a tuple, we often use the notation ${\overline{t}} = (t_1, \ldots, t_k)$ to denote its entries. A *relation* over a set $B$ is a subset of $B^k$ for a value $k \geq 1$ called the *arity* of the relation. We use $\pi_i$ to denote the operator that projects onto the $i$th coordinate: $\pi_i({\overline{t}})$ denotes the $i$th entry $t_i$ of a tuple ${\overline{t}} = (t_1, \ldots, t_k)$, and for a relation $R$ we define $\pi_i(R) = \{ \pi_i({\overline{t}}) ~|~ {\overline{t}} \in R \}$. Similarly, for a subset $I \subseteq [k]$ whose elements are $i_1 < \cdots < i_m$, we use $\pi_I({\overline{t}})$ to denote the tuple $(t_{i_1}, \ldots, t_{i_m})$, and we define $\pi_I(R) = \{ \pi_I({\overline{t}}) ~|~ {\overline{t}} \in R \}$.
A *signature* $\sigma$ is a set of symbols, each of which has an associated arity. A *structure* ${\mathbf{B}}$ over signature $\sigma$ consists of a universe $B$, which is a set, and a relation $R^{{\mathbf{B}}} \subseteq B^k$ for each symbol $R \in \sigma$ of arity $k$. (Note that in this paper, we are concerned only with relational structures, which we refer to simply as structures.) Throughout, we will use the bold capital letters ${\mathbf{A}}, {\mathbf{B}}, \ldots$ to denote structures, and the corresponding non-bold capital letters $A, B, \ldots$ to denote their universes. We say that a structure ${\mathbf{B}}$ is *finite* if its universe $B$ has finite size. Unless stated otherwise, we assume all structures under discussion in this paper to be finite. We say that a structure ${\mathbf{B}}$ *has all constants* if for each $b \in B$, there is a relation symbol $R_b$ with $R_b^{{\mathbf{B}}} = \{ (b) \}$.
When two structures ${\mathbf{A}}$ and ${\mathbf{B}}$ are defined over the same signature $\sigma$, we say that they are *similar*. We define the following notions on similar structures. For similar structures ${\mathbf{A}}$ and ${\mathbf{B}}$ over a signature $\sigma$, we say that ${\mathbf{A}}$ is an *induced substructure* of ${\mathbf{B}}$ if $A \subseteq B$ and for every $R \in \sigma$ of arity $k$, it holds that $R^{{\mathbf{A}}} = A^k \cap R^{{\mathbf{B}}}$. Observe that for a structure ${\mathbf{B}}$ and a subset $B' \subseteq B$, there is exactly one induced substructure of ${\mathbf{B}}$ with universe $B'$. For similar structures ${\mathbf{A}}$ and ${\mathbf{B}}$ over a signature $\sigma$, the product structure ${\mathbf{A}}\times {\mathbf{B}}$ is defined to be the structure with universe $A \times B$ and such that $R^{{\mathbf{A}}\times {\mathbf{B}}} =
\{ ((a_1, b_1), \ldots, (a_k, b_k)) ~|~ {\overline{a}} \in R^{{\mathbf{A}}}, {\overline{b}} \in R^{{\mathbf{B}}} \}$ for all $R \in \sigma$. We use ${\mathbf{A}}^n$ to denote the $n$-fold product ${\mathbf{A}}\times \cdots \times {\mathbf{A}}$.
We say that a structure ${\mathbf{B}}$ over signature $\sigma'$ is an *expansion* of another structure ${\mathbf{A}}$ over signature $\sigma$ if (1) $\sigma' \supseteq \sigma$, (2) the universe of ${\mathbf{B}}$ is equal to the universe of ${\mathbf{A}}$, and (3) for every symbol $R \in \sigma$, it holds that $R^{{\mathbf{B}}} = R^{{\mathbf{A}}}$. We will use the following non-standard notation. For any structure ${\mathbf{A}}$ (over signature $\sigma$) and any subset $S \subseteq A$, we define $[{\mathbf{A}}, S]$ to be the expansion of ${\mathbf{A}}$ with the signature $\sigma \cup \{ U \}$ where $U$ is a new symbol of arity $1$, defined by $U^{[{\mathbf{A}}, S]} = S$ and $R^{[{\mathbf{A}}, S]} = R^{{\mathbf{A}}}$ for all $R \in \sigma$. More generally, for a structure ${\mathbf{A}}$ (over $\sigma$) and a sequence of subsets $S_1, \ldots, S_n \subseteq A$, we define $[{\mathbf{A}}, S_1, \ldots, S_n]$ to be the expansion of ${\mathbf{A}}$ with the signature $\sigma \cup \{ U_1, \ldots, U_n \}$ where $U_1, \ldots, U_n$ are new symbols of arity $1$, defined by $U_i^{[{\mathbf{A}}, S_1, \ldots, S_n]} = S_i$ for all $i \in [n]$, and $R^{[{\mathbf{A}}, S_1, \ldots, S_n]} = R^{{\mathbf{A}}}$ for all $R \in \sigma$.
#### Homomorphisms and the constraint satisfaction problem.
For similar structures ${\mathbf{A}}$ and ${\mathbf{B}}$ over the signature $\sigma$, a *homomorphism* from ${\mathbf{A}}$ to ${\mathbf{B}}$ is a mapping $h: A \rightarrow B$ such that for every symbol $R$ of $\sigma$ and every tuple $(a_1, \ldots, a_k) \in R^{{\mathbf{A}}}$, it holds that $(h(a_1), \ldots, h(a_k)) \in R^{{\mathbf{B}}}$. We use ${\mathbf{A}}{\rightarrow}{\mathbf{B}}$ to indicate that there is a homomorphism from ${\mathbf{A}}$ to ${\mathbf{B}}$; when this holds, we also say that ${\mathbf{A}}$ *is homomorphic to* ${\mathbf{B}}$. It is well-known and straightforward to verify that the homomorphism relation ${\rightarrow}$ is transitive, that is, if ${\mathbf{A}}{\rightarrow}{\mathbf{B}}$ and ${\mathbf{B}}{\rightarrow}{\mathbf{C}}$, then ${\mathbf{A}}{\rightarrow}{\mathbf{C}}$.
The *constraint satisfaction problem (CSP)* is the problem of deciding, given as input a pair $({\mathbf{A}}, {\mathbf{B}})$ of similar structures, whether or not there exists a homomorphism from ${\mathbf{A}}$ to ${\mathbf{B}}$. When $({\mathbf{A}}, {\mathbf{B}})$ is an instance of the CSP, we will also call a homomorphism from ${\mathbf{A}}$ to ${\mathbf{B}}$ a *satisfying assignment*; say that the instance is *satisfiable* if there exists such a homomorphism; and, say that the instance is *unsatisfiable* if there does not exist such a homomorphism. We generally assume that in an instance of the CSP, the left-hand side structure ${\mathbf{A}}$ contains finitely many tuples. For any structure ${\mathbf{B}}$ (over $\sigma$), the *constraint satisfaction problem for ${\mathbf{B}}$*, denoted by ${{\mathsf{CSP}}}({\mathbf{B}})$, is the constraint satisfaction problem where the right-hand side structure is fixed to be ${\mathbf{B}}$, that is, the problem of deciding, given as input a structure ${\mathbf{A}}$ over $\sigma$, whether or not there exists a homomorphism from ${\mathbf{A}}$ to ${\mathbf{B}}$. In discussing a problem of the form ${{\mathsf{CSP}}}({\mathbf{B}})$, the structure ${\mathbf{B}}$ is often referred to as the *template* or *constraint language*. There are several equivalent definitions of the constraint satisfaction problem. For instance, in logic, the constraint satisfaction problem can be formulated as the model checking problem for primitive positive sentences over relational structures, and in database theory, it can be formulated as the containment problem for conjunctive queries [@ChandraMerlin].
#### Polymorphisms.
When $f: B^n \rightarrow B$ is an operation on $B$ and $${\overline{t_1}} = (t_{11}, \ldots, t_{1k}), \ldots, {\overline{t_n}} = (t_{n1}, \ldots, t_{nk}) \in B^k$$ are tuples of the same arity $k$ over $B$, we use $f({\overline{t_1}}, \ldots, {\overline{t_n}})$ to denote the arity $k$ tuple obtained by applying $f$ coordinatewise, that is, $$f({\overline{t_1}}, \ldots, {\overline{t_n}}) =
(f(t_{11}, \ldots, t_{n1}), \ldots, f(t_{1k}, \ldots, t_{nk})).$$ An operation $f: B^n \rightarrow B$ is a *polymorphism* of a structure ${\mathbf{B}}$ over $\sigma$ if for every symbol $R \in \sigma$ and any tuples ${\overline{t_1}}, \ldots, {\overline{t_n}} \in R^{{\mathbf{B}}}$, it holds that $f({\overline{t_1}}, \ldots, {\overline{t_n}}) \in R^{{\mathbf{B}}}$. That is, each relation $R^{{\mathbf{B}}}$ is closed under the action of $f$. Equivalently, an operation $f: B^n \rightarrow B$ is a polymorphism of ${\mathbf{B}}$ if it is a homomorphism from ${\mathbf{B}}^n$ to ${\mathbf{B}}$.
Algorithms {#sect:algorithms}
==========
In this section, we give a uniform presentation of the four algorithms under investigation in this paper: arc consistency, look-ahead arc consistency, peek arc consistency, and singleton arc consistency, presented in Sections \[ss:ac\], \[ss:laac\], \[ss:pac\], and \[ss:sac\], respectively. The results on the first three algorithms come from previous work, as we discuss in presenting each of these algorithms; for singleton arc consistency, we here develop results similar to those given for the other algorithms.
Our treatment of arc consistency, peek arc consistency, and singleton arc consistency is uniform: for each of these algorithms, we present a homomorphism-based consistency condition, we show that the algorithm checks precisely this consistency condition, and we give an algebraic condition describing the structures ${\mathbf{B}}$ such that the algorithm solves ${{\mathsf{CSP}}}({\mathbf{B}})$. These three algorithms give one-sided consistency checks: each either correctly rejects an instance as unsatisfiable or outputs “?”, which can be interpreted as a report that it is unknown whether or not the instance is satisfiable. The other algorithm, look-ahead arc consistency, has a somewhat different character. It attempts to build a satisfying assignment one variable at a time, using arc consistency as a filtering criterion; it either returns a satisfying assignment, or outputs “?”.
Throughout this section and in later sections, we will make use of a structure ${\wp}({\mathbf{B}})$ that is defined for every structure ${\mathbf{B}}$, as follows [@FederVardi; @DalmauPearson]. For a structure ${\mathbf{B}}$ (over $\sigma$), we define ${\wp}({\mathbf{B}})$ to be the structure with universe ${\wp}(B) \setminus \{ \emptyset \}$ and where, for every symbol $R \in \sigma$ of arity $k$, $R^{{\wp}({\mathbf{B}})} = \{ (\pi_1 S, \ldots, \pi_k S) ~|~ S \subseteq R^{{\mathbf{B}}}, S \neq \emptyset \}$. Here, ${\wp}(B)$ denotes the power set of the set $B$.
Arc Consistency {#ss:ac}
---------------
We now present the arc consistency algorithm. The main idea of the algorithm is to associate to each element $a \in A$ a set $S_a$ of values which, throughout the execution of the algorithm, has the property that for any solution $h$, it must hold that $h(a) \in S_a$. The algorithm continually shrinks the sets $S_a$ in a natural fashion until they stabilize; at this point, if some set $S_a$ is the empty set, then no solution can exist, and the algorithm rejects the instance.
[**Arc Consistency**]{}\
------------------------------------------------------------------------
Feder and Vardi [@FederVardi] have studied arc consistency, under an equivalent formulation in terms of Datalog Programs, for constraint languages. The results in this section are due to this reference. The connection of the results in Feder and Vardi with arc consistency was made explicit in Dalmau and Pearson [@DalmauPearson].
\[def:acc\] An instance $({\mathbf{A}}, {\mathbf{B}})$ has the *arc consistency condition (ACC)* if there exists a homomorphism from ${\mathbf{A}}$ to ${\wp}({\mathbf{B}})$.
\[prop:acc\] The arc consistency algorithm does not reject an instance $({\mathbf{A}}, {\mathbf{B}})$ if and only if the instance has the ACC.
\[def:ac-solves\] Let ${\mathbf{B}}$ be a structure. We say that arc consistency *solves* ${{\mathsf{CSP}}}({\mathbf{B}})$ if for all structures ${\mathbf{A}}$, the following holds: $({\mathbf{A}}, {\mathbf{B}})$ has the ACC implies that there is a homomorphism ${\mathbf{A}}\rightarrow {\mathbf{B}}$.
Note that the converse of the condition given in this definition always holds: if $h$ is a homomorphism from ${\mathbf{A}}$ to ${\mathbf{B}}$, then the mapping sending each $a \in A$ to the set $\{ h(a) \}$ is a homomorphism from ${\mathbf{A}}$ to ${\wp}({\mathbf{B}})$.
\[thm:ac-solves\] Let ${\mathbf{B}}$ be a structure. Arc consistency solves ${{\mathsf{CSP}}}({\mathbf{B}})$ if and only if there is a homomorphism ${\wp}({\mathbf{B}}) \rightarrow {\mathbf{B}}$.
Look-Ahead Arc Consistency {#ss:laac}
--------------------------
We now present the look-ahead arc consistency algorithm. It attempts to construct a satisfying assignment by setting one variable at a time, using arc consistency as a filter to find a suitable value for each variable.
[ **Look-Ahead Arc Consistency**]{}\
------------------------------------------------------------------------
Look-ahead arc consistency was introduced and studied by Chen and Dalmau [@SLAAC], and the theorem that follows is due to them. This algorithm can be viewed as a generalization of an algorithm for SAT studied by Del Val [@delval].
\[def:laac-solves\] Let ${\mathbf{B}}$ be a structure. We say that look-ahead arc consistency *solves* ${{\mathsf{CSP}}}({\mathbf{B}})$ if for all structures ${\mathbf{A}}$, the following holds: if there exists a homomorphism ${\mathbf{A}}\rightarrow {\mathbf{B}}$, then the look-ahead arc consistency algorithm, given $({\mathbf{A}}, {\mathbf{B}})$, outputs such a homomorphism.
\[thm:laac-solves\] Let ${\mathbf{B}}$ be a structure. Look-ahead arc consistency solves ${{\mathsf{CSP}}}({\mathbf{B}})$ if and only if there is a homomorphism $l: {\wp}({\mathbf{B}}) \times {\mathbf{B}}\rightarrow {\mathbf{B}}$ such that $l( \{ b \}, b') = b$ for all $b, b' \in B$.
Peek Arc Consistency {#ss:pac}
--------------------
We now present the peek arc consistency algorithm. It attempts to find, for each variable $a \in A$, a value $b \in B$ such that when $a$ is set to $b$, the arc consistency check is passed.
[**Peek Arc Consistency**]{}\
------------------------------------------------------------------------
Peek arc consistency was introduced and studied by Bodirsky and Chen [@peek]; the notions and results that follow come from them. In their work, the algorithm is shown to solve certain constraint languages, including some languages having infinite-size universes; such languages actually gave the motivation for introducing the algorithm. In this work, it is pointed out that peek arc consistency can be readily parallelized; by invoking the arc consistency checks independently in parallel, one can achieve a linear parallel running time.
\[def:pacc\] An instance $({\mathbf{A}}, {\mathbf{B}})$ has the *peek arc consistency condition (PACC)* if for every element $a \in A$, there exists a homomorphism $h$ from ${\mathbf{A}}$ to ${\wp}({\mathbf{B}})$ such that $h(a)$ is a singleton.
\[prop:pacc\] The peek arc consistency algorithm does not reject an instance $({\mathbf{A}}, {\mathbf{B}})$ if and only if the instance has the PACC.
\[def:pac-solves\] Let ${\mathbf{B}}$ be a structure. We say that peek arc consistency *solves* ${{\mathsf{CSP}}}({\mathbf{B}})$ if for all structures ${\mathbf{A}}$, the following holds: $({\mathbf{A}}, {\mathbf{B}})$ has the PACC implies that there is a homomorphism ${\mathbf{A}}\rightarrow {\mathbf{B}}$.
The converse of the condition given in this definition always holds. Suppose that $h$ is a homomorphism from ${\mathbf{A}}$ to ${\mathbf{B}}$; then, the mapping taking each $a \in A$ to the singleton $\{ h(a) \}$ is a homomorphism from ${\mathbf{A}}$ to ${\wp}({\mathbf{B}})$ and hence $({\mathbf{A}}, {\mathbf{B}})$ has the PACC.
We use the notation ${\mathsf{Sing}}({\wp}({\mathbf{B}})^n)$ to denote the induced substructure of ${\wp}({\mathbf{B}})^n$ whose universe contains an $n$-tuple of ${\wp}({\mathbf{B}})^n$ if and only if at least one coordinate of the tuple is a singleton.
\[thm:pac-solves\] Let ${\mathbf{B}}$ be a structure. Peek arc consistency solves ${{\mathsf{CSP}}}({\mathbf{B}})$ if and only if for all $n \geq 1$ there is a homomorphism ${\mathsf{Sing}}({\wp}({\mathbf{B}})^n) \rightarrow {\mathbf{B}}$.
Singleton Arc Consistency {#ss:sac}
-------------------------
We now present the singleton arc consistency algorithm. As with arc consistency, this algorithm associates to each element $a \in A$ a set $S_a$ of feasible values. It then continually checks, for pairs $(a, b)$ with $a \in A$ and $b \in S_a$, whether or not arc consistency can be established with respect to the sets $S_a$ and when $a$ is assigned to $b$; if for some pair $(a, b)$ it cannot, then $b$ is removed from the set $S_a$. As with arc consistency, this algorithm’s outer loop runs until the sets $S_a$ stabilize, and the algorithm rejects if one of the sets $S_a$ is equal to the empty set.
[**Singleton Arc Consistency**]{}\
------------------------------------------------------------------------
denote $A=\{a_1,\ldots ,a_n\}$
Singleton arc consistency was introduced by Debruyne and Bessiere [@db-sac]. We now give a development of singleton arc consistency analogous to that of arc consistency and peek arc consistency.
\[def:sacc\] An instance $({\mathbf{A}}, {\mathbf{B}})$ has the *singleton arc consistency condition (SACC)* if there exists a mapping ${s: A \rightarrow {\wp}(B) \setminus \{ \emptyset \}}$ such that for all $a \in A$, $b \in s(a)$ there exists a homomorphism $h_{a,b}: {\mathbf{A}}\rightarrow {\wp}({\mathbf{B}})$ where:
- $h_{a,b}(a)=\{b\}$, and
- for all $a' \in A$, it holds that $h_{a,b}(a') \subseteq s(a')$.
\[prop:sacc\] The singleton arc consistency algorithm does not reject an instance $({\mathbf{A}}, {\mathbf{B}})$ if and only if the instance has the SACC.
Suppose that the singleton arc consistency algorithm does not reject an instance $({\mathbf{A}}, {\mathbf{B}})$. Let $\{ S_a \}_{a \in A}$ denote the sets computed by the algorithm at the point of termination, and define $s$ to be the mapping where $s(a) =S_a$ for all $a\in A$. Let $a \in A$ and $b \in s(a)$. By the definition of the algorithm, the pair $([{\mathbf{A}}, \{a_1\},\ldots,\{a_n\},\{a\}],\allowbreak[{\mathbf{B}}, S_{a_1},\ldots, S_{a_n},\{b\}])$ has the ACC, and thus the desired homomorphism $h_{a,b}$ exists.
Now, suppose that the instance $({\mathbf{A}}, {\mathbf{B}})$ has the SACC, and let $s$ be a mapping with the described properties. We show that throughout the execution of the algorithm, it holds that $s(a)\subseteq S_a$ for all $a\in A$. First, $S_a$ is initialized with $B$ for every $a\in A$. Next, we show that when $a \in A$ and $b \in s(a)$, then $b$ is never removed from $S_a$ by the algorithm. This is because by definition of SACC, there exists a homomorphism $h_{a,b}: {\mathbf{A}}\rightarrow {\wp}({\mathbf{B}})$ with $h_{a,b}(a)=\{b\}$ such that for all $a' \in A$, it holds that $h_{a,b}(a') \subseteq s(a')$. Since $s(a')\subseteq S_{a'}$ by the inductive assumption, $([{\mathbf{A}}, \{a_1\},\ldots,\{a_n\},\{a\}],[{\mathbf{B}}, S_{a_1},\ldots, S_{a_n},\{b\}])$ has the ACC and hence the algorithm does not remove $b$ from $S_a$.
\[def:sac-solves\] Let ${\mathbf{B}}$ be a structure. We say that singleton arc consistency *solves* ${{\mathsf{CSP}}}({\mathbf{B}})$ if for all structures ${\mathbf{A}}$, the following holds: $({\mathbf{A}}, {\mathbf{B}})$ has the SACC implies that there is a homomorphism ${\mathbf{A}}\rightarrow {\mathbf{B}}$.
The converse of the condition given in this definition always holds: suppose that $h$ is a homomorphism from ${\mathbf{A}}$ to ${\mathbf{B}}$. Then, the instance $({\mathbf{A}}, {\mathbf{B}})$ has the SACC via the mapping $s$ where $s(a) = \{ h(a) \}$ for all $a \in A$ and the mappings $h_{a, b}$ defined by $h_{a, b}(a') = \{ h(a') \}$ for all $a' \in A$.
We use the notation ${\mathsf{UnionSing}}({\wp}({\mathbf{B}})^n)$ to denote the induced substructure of ${\wp}({\mathbf{B}})^n$ whose universe contains an $n$-tuple $(S_1, \ldots, S_n)$ of ${\wp}({\mathbf{B}})^n$ if and only if it holds that $\bigcup_{i \in [n]} S_i = \bigcup_{i \in [n], |S_i| = 1} S_i$.
\[thm:sac-solves\] Let ${\mathbf{B}}$ be a structure. Singleton arc consistency solves ${{\mathsf{CSP}}}({\mathbf{B}})$ if and only if for all $n \geq 1$ there is a homomorphism ${\mathsf{UnionSing}}({\wp}({\mathbf{B}})^n) \rightarrow {\mathbf{B}}$.
First we show that if singleton arc consistency solves ${{\mathsf{CSP}}}({\mathbf{B}})$, then there is a homomorphism from ${\mathsf{UnionSing}}({\wp}({\mathbf{B}})^{n})$ to ${\mathbf{B}}$ for all $n\geq 1$. Let $n \geq 1$; we show that $({\mathsf{UnionSing}}({\wp}({\mathbf{B}})^{n}),{\mathbf{B}})$ has the SACC. Then, there is a homomorphism from ${\mathsf{UnionSing}}({\wp}({\mathbf{B}})^{n})$ to ${\mathbf{B}}$, since the singleton arc consistency algorithm solves [CSP(${\mathbf{B}}$)]{}.
Let $s$ be the mapping $s(a):=\bigcup_{i \in [n]} S_i$ for all tuples $a=(S_1,\dots,S_n)$ of ${\mathsf{UnionSing}}({\wp}({\mathbf{B}})^{n})$. Now let us consider an arbitrary tuple $a=(S_1,\dots,S_n)$ of ${\mathsf{UnionSing}}({\wp}({\mathbf{B}})^{n})$ and an arbitrary $b \in s(a)$. Since $\bigcup_{i \in [n]} S_i = \bigcup_{i \in [n], |S_i| = 1} S_i$, there is an $i\in[n]$ such that $\{b\}=S_i$. Thus, the homomorphism $\pi_i\colon {\mathsf{UnionSing}}({\wp}({\mathbf{B}})^{n}) \rightarrow {\wp}({\mathbf{B}})$ that projects onto the $i$th coordinate satisfies $\pi_i(a)=\{b\}$, and for all tuples $a'$ of ${\mathsf{UnionSing}}({\wp}({\mathbf{B}})^{n})$, it holds that $\pi_i(a')\subseteq s(a')$. Hence, $({\mathsf{UnionSing}}({\wp}({\mathbf{B}})^{n}),{\mathbf{B}})$ has the SACC.
For the other direction, we show that if there is a homomorphism from ${\mathsf{UnionSing}}({\wp}({\mathbf{B}})^{n})$ to ${\mathbf{B}}$ for all $n \geq 1$, then singleton arc consistency solves ${{\mathsf{CSP}}}({\mathbf{B}})$. Thus, we have to show that there exists a homomorphism from ${\mathbf{A}}$ to ${\mathbf{B}}$ if $({\mathbf{A}},{\mathbf{B}})$ has the SACC. Let $s$ be the homomorphism from the definition of SACC, and let us use $\{h_1,\ldots ,h_n\}$ to denote the set ${\{h_{a,b}\mid a\in A,b\in s(a)\}}$ of homomorphisms. Further, let $g$ be the homomorphism $(h_1,\ldots ,h_n) \colon {\mathbf{A}}\to {\wp}({\mathbf{B}})^{n}$. Now, for every element $a\in A$ the image $g(a)=(h_1(a),\ldots ,h_n(a))$ is a tuple of ${\mathsf{UnionSing}}({\wp}({\mathbf{B}})^{n})$: for every $b\in \bigcup_{j \in [n]} h_j(a)$, it holds that $b \in s(a)$ and thus there exists a homomorphism $h_{a,b}=h_i$ that maps $a$ to the singleton $\{b\}$; so, we have $\bigcup_{j \in [n]} h_j(a) = \bigcup_{i \in [n], |h_i(a)| = 1} h_i(a)$. Since $g$ is a homomorphism from ${\mathbf{A}}$ to ${\mathsf{UnionSing}}({\wp}({\mathbf{B}})^{n})$, we can compose $g$ and a homomorphism from ${\mathsf{UnionSing}}({\wp}({\mathbf{B}})^{n})$ to ${\mathbf{B}}$, which we know to exist by assumption, to get a homomorphism from ${\mathbf{A}}$ to ${\mathbf{B}}$. Consequently, singleton arc consistency solves ${{\mathsf{CSP}}}({\mathbf{B}})$.
Strength Comparison {#sect:comparison}
===================
In this section, we investigate relationships among the sets of structures solvable by the various algorithms presented. We show that for the structures having all constants, AC solves a strictly smaller set of structures than LAAC does; on the other hand, we show that there is a structure (not having all constants) solvable by AC but not LAAC. We then show that the structures solvable by AC or LAAC are strictly contained in those solvable by PAC; and, in turn, that the structures solvable by PAC are strictly contained in those solvable by SAC. We also show that the structures solvable by SAC (and hence, those solvable by any of the studied algorithms) all fall into the class of structures having *bounded width*; bounded width is a well-studied condition admitting multiple characterizations [@FederVardi; @LaroseZadori; @bkbw].
\[prop:all-constants-ac-implies-laac\] Suppose that ${\mathbf{B}}$ is a structure having all constants. If ${{\mathsf{CSP}}}({\mathbf{B}})$ is solvable by AC, then it is solvable by LAAC.
By Theorem \[thm:ac-solves\], there is a homomorphism $f: {\wp}({\mathbf{B}}) \rightarrow {\mathbf{B}}$. Since the structure ${\mathbf{B}}$ has all constants, for each $b \in B$ there is a relation symbol $R_b$ with $R_b^{{\mathbf{B}}} = \{ (b) \}$. Since $(\{ b \}) \in R_b^{{\wp}({\mathbf{B}})}$, it must hold that $f( \{ b \} ) \in R_b^{{\mathbf{B}}}$, from which it follows that $f( \{ b \} ) = b$. The mapping $l$ defined by $l(S, b) = f(S)$ is then a homomorphism of the type described in Theorem \[thm:laac-solves\].
There exists a structure ${\mathbf{B}}$ having all constants such that ${{\mathsf{CSP}}}({\mathbf{B}})$ is solvable by LAAC but not by AC.
Take ${\mathbf{B}}$ to be the relational structure with universe $\{ 0, 1 \}$ over signature $\{ U_0, U_1, R_{(0, 0)}, R_{(1, 1)} \}$ where $$U_0^{\mathbf{B}}= \{ 0 \}$$ $$U_1^{\mathbf{B}}= \{ 1 \}$$ $$R_{(0, 0)}^{\mathbf{B}}= \{ 0, 1 \}^2 \setminus \{ (0, 0) \}$$ $$R_{(1, 1)}^{\mathbf{B}}= \{ 0, 1 \}^2 \setminus \{ (1, 1) \}.$$ It is straightforward to verify that the mapping $l$ defined by $l(\{ 0, 1 \}, b') = b'$, $l(\{ 0 \}, b') = 0$, and $l(\{ 1 \}, b') = 1$ for all $b' \in \{ 0, 1 \}$ is a homomorphism from ${\wp}({\mathbf{B}}) \times {\mathbf{B}}$ to ${\mathbf{B}}$ satisfying the condition of Theorem \[thm:laac-solves\]. Hence, the problem ${{\mathsf{CSP}}}({\mathbf{B}})$ is solvable by LAAC.
To show that the problem ${{\mathsf{CSP}}}({\mathbf{B}})$ is not solvable by AC, let $f$ be an arbitrary mapping from ${\wp}(B) \setminus \{ \emptyset \}$ to $B$. We show that $f$ cannot be a homomorphism from ${\wp}({\mathbf{B}})$ to ${\mathbf{B}}$, which suffices by Theorem \[thm:ac-solves\]. Let $b = f(\{ 0, 1 \})$. It holds that $( \{ 0, 1 \}, \{ 0, 1 \} ) \in R_{(b, b)}^{{\wp}({\mathbf{B}})}$, but $(f(\{ 0, 1 \}), f(\{ 0, 1 \})) = (b, b) \notin R_{(b, b)}^{{\mathbf{B}}}$, and we are done.
There exists a structure ${\mathbf{B}}$ (not having all constants) such that ${{\mathsf{CSP}}}({\mathbf{B}})$ is solvable by AC but not by LAAC.
Take ${\mathbf{B}}$ to be the relational structure with universe $\{ 0, 1 \}$ over signature $\{ R, S \}$ where $R^{{\mathbf{B}}} = \{ 0, 1 \}^3 \setminus \{ (0, 1, 1) \}$ and $S^{{\mathbf{B}}} = \{ 0, 1 \}^3 \setminus \{ (1, 0, 0) \}$. The mapping $p$ that sends each element of ${\wp}(B) \setminus \{ \emptyset \}$ to $0$ is a homomorphism from ${\wp}({\mathbf{B}})$ to ${\mathbf{B}}$, and hence AC solves ${{\mathsf{CSP}}}({\mathbf{B}})$ by Theorem \[thm:ac-solves\].
To show that the problem ${{\mathsf{CSP}}}({\mathbf{B}})$ is not solvable by LAAC, let $f$ be an arbitrary mapping from $({\wp}(B) \setminus \{ \emptyset \}) \times B$ to $B$ that satisfies $f(\{ b \}, b') = b$ for all $b, b' \in B$. We show that $f$ cannot be a homomorphism from ${\wp}({\mathbf{B}}) \times {\mathbf{B}}$ to ${\mathbf{B}}$, which suffices by Theorem \[thm:laac-solves\]. We consider two cases depending on the value of $f( \{ 0, 1 \}, 0)$.
- If $f( \{ 0, 1 \}, 0) = 1$, then we use the facts that $(\{ 0 \}, \{ 0, 1 \}, \{ 0, 1 \}) \in R^{{\wp}({\mathbf{B}})}$ and that $(0, 0, 0) \in R^{{\mathbf{B}}}$; we have that $(f( \{0 \}, 0), f( \{ 0, 1 \}, 0), f( \{ 0, 1 \}, 0)) = (0, 1, 1)$, which is not contained in $R^{{\mathbf{B}}}$, implying that $f$ is not a homomorphism of the desired type.
- If $f( \{ 0, 1 \}, 0) = 0$, then we use the facts that $(\{ 1 \}, \{ 0, 1 \}, \{ 0, 1 \}) \in S^{{\wp}({\mathbf{B}})}$ and that $(0, 0, 0) \in S^{{\mathbf{B}}}$; we have that $(f( \{ 1 \}, 0), f( \{ 0, 1 \}, 0), f( \{ 0, 1 \}, 0)) = (1, 0, 0)$, which is not contained in $S^{{\mathbf{B}}}$, implying that $f$ is not a homomorphism of the desired type.
We now proceed to study PAC, and in particular, show that the structures solvable by AC or LAAC are solvable by PAC.
\[prop:ac-in-pac\] Let ${\mathbf{B}}$ be a structure. If ${{\mathsf{CSP}}}({\mathbf{B}})$ is solvable by AC, then it is also solvable by PAC.
Proposition \[prop:ac-in-pac\] follows directly from the algebraic characterizations given in Theorems \[thm:ac-solves\] and \[thm:pac-solves\]; it can also be seen to follow from the corresponding algorithm descriptions.
Let ${\mathbf{B}}$ be a structure. If ${{\mathsf{CSP}}}({\mathbf{B}})$ is solvable by LAAC, then it is also solvable by PAC.
Suppose that look-ahead arc consistency solves CSP(${\mathbf{B}}$). By Theorem \[thm:laac-solves\] there exists a homomorphism $l\colon {\wp}({\mathbf{B}}) \times {\mathbf{B}}\to {\mathbf{B}}$ such that $l( \{ b \}, b') = b$ for all $b, b' \in B$. We want to show that peek arc consistency solves CSP(${\mathbf{B}}$) by using Theorem \[thm:pac-solves\]. Thus, we have to show that for all $n \geq 1$ there is a homomorphism $g_n\colon {\mathsf{Sing}}({\wp}({\mathbf{B}})^n) \rightarrow {\mathbf{B}}$.
Let $n\geq 1$. Let us consider the mapping $g_n$ with $$g_n(S_1,\ldots ,S_n)=l(S_1,l(S_2,\ldots l(S_{n-1},l(S_n,b))\ldots ))$$ defined for all tuples $T=(S_1,\ldots ,S_n)\in {\mathsf{Sing}}({\wp}(B)^n)$ and all $b\in B$. First we want to show that $g_n$ is well defined. Let $b_1,b_2\in B$ with $b_1\not = b_2$, let $(S_1,\ldots ,S_n)\in {\mathsf{Sing}}({\wp}(B)^n)$ and let $i\in[n]$ be an index such that $S_i$ is a singleton. Let $S_i=\{b_*\}$ for a $b_*\in B$. We obtain that $$\begin{aligned}
& l(S_1,\ldots l(S_{i-1},l(S_{i},\ldots l(S_{n-1},l(S_n,b_1))\ldots ))\ldots )\\
=\;& l(S_1,\ldots l(S_{i-1},l(S_{i},b'))\ldots )\\
=\;& l(S_1,\ldots l(S_{i-1},b_*)\ldots )
\intertext{with $b'=l(S_{i+1},\ldots l(S_{n-1},l(S_n,b_1))\ldots )\in B$, because $l$ is applied to
the singleton $S_i=\{b_*\}$ and $b'$.
Similarly, we obtain that}
& l(S_1,\ldots l(S_{i-1},l(S_{i},\ldots l(S_{n-1},l(S_n,b_2))\ldots ))\ldots )\\
=\;& l(S_1,\ldots l(S_{i-1},b_*)\ldots )\end{aligned}$$ Consequently, $g_n$ is well defined. Next, we prove that $g_n$ is a homomorphism. Let $R^{{\mathsf{Sing}}({\wp}({\mathbf{B}})^n)}$ be a $k$-ary relation and let $(T^1,\ldots ,T^k)$ be a tuple in this relation. Denote $T^i=(S^i_1,S^i_2,\ldots ,S^i_n)$ for all $i\in[k]$; then, $S'_j=(S^1_j,\ldots ,S^k_j)$ has to be in $R^{{\wp}({\mathbf{B}})}$ for all $j\in[n]$. Further, we know that there exists a tuple $\bar{b}=(b_1,\ldots ,b_k)\in R^{\mathbf{B}}$, because $R^{{\wp}({\mathbf{B}})}$ is not empty. Since $l$ is a homomorphism, the tuple $$g_n(S'_1, S'_2, \ldots, S'_n)=l(S'_1, l(S'_2, \ldots
l(S'_{n-1},l(S'_n,\bar{b}))\ldots ))$$ is in $R^{{\mathbf{B}}}$. Thus, $g_n$ is a homomorphism from ${\mathsf{Sing}}({\wp}({\mathbf{B}})^{n})$ to ${\mathbf{B}}$.
There exists a structure ${\mathbf{B}}$ having all constants such that ${{\mathsf{CSP}}}({\mathbf{B}})$ is solvable by PAC but not by LAAC nor AC.
Let us consider the structure with universe $\{0,1,2\}$ over the signature $\{U_0,U_1,U_2,R_1,R_2\}$ where $$U_0^{\mathbf{B}}=\{(0)\}$$ $$U_1^{\mathbf{B}}=\{(1)\}$$ $$U_2^{\mathbf{B}}=\{(2)\}$$ $$R_1^{\mathbf{B}}= \big(\{0,1\}\times\{0,1,2\}\big) \setminus \{(0,0)\}$$ $$R_2^{\mathbf{B}}=\{(0,0),(1,2),(2,1)\}.$$ First we show that there is no homomorphism $l: {\wp}({\mathbf{B}})\times{\mathbf{B}}\to{\mathbf{B}}$ such that $l(\{b\},b')=b$ for all $ b,b'$. Let us assume there is one. Since $(\{0\},\{1,2\})\in R_1^{{\wp}({\mathbf{B}})}$ and $(1,0)\in R_1^{\mathbf{B}}$ the tuple $( l(\{0\},1),l(\{1,2\},0))$, which is equal to $(0,l(\{1,2\},0))$, has to be contained in $R_1^{\mathbf{B}}$. Thus, $l(\{1,2\},0)$ cannot be equal to $0$. On the other hand, $(\{1,2\},\{1,2\})\in R_2^{{\wp}({\mathbf{B}})}$ and $(0,0)\in R_2^{\mathbf{B}}$ implies that $(l(\{1,2\},0),l(\{1,2\},0))$ is in $R_2^{\mathbf{B}}$. Therefore, $l(\{1,2\},0)$ has to be $0$, which is a contradiction. This establishes that the structure is not solvable by LAAC; by Proposition \[prop:all-constants-ac-implies-laac\], it follows that the structure is not solvable by AC.
Next we show that for all $n$, there exists a homomorphism $f$ from ${\mathsf{Sing}}({\wp}({\mathbf{B}})^n)$ to ${\mathbf{B}}$. Let $n$ be arbitrary and let $(S_1,\dots,S_n)$ be an arbitrary $n$-tuple of ${\mathsf{Sing}}({\wp}({\mathbf{B}})^n)$. Further, let $i$ be the minimal number such that $S_i$ is $\{1\}$, $ \{2\}$, $ \{0,1\}$ or $\{0,2\}$; if such an $S_i$ does not exists, then $i=0$. The homomorphism $f$ can be defined as follows: $$f(S_1,\dots,S_n)=
\begin{cases}
1 & \text{if $i>0$ and $S_i$ is $\{1\}$ or $\{0,1\}$}\\
2 & \text{if $i>0$ and $S_i$ is $\{2\}$ or $\{0,2\}$}\\
0 & \text{otherwise.}\\
\end{cases}$$ Let us verify that $f$ is indeed a homomorphism: First of all, it is easy to see that $f(S_1,\dots,S_n)$ is in $U_i^{{\mathbf{B}}}$ whenever $(S_1,\dots,S_n)$ is in $U_i^{{\mathsf{Sing}}({\wp}({\mathbf{B}})^n)}$. Next, let us consider $R_2$. Let $(S_1,\dots,S_n)$ and $(T_1,\dots,T_n)$ be arbitrary $n$-tuples of ${\mathsf{Sing}}({\wp}({\mathbf{B}})^n)$ such that $(S_l,T_l)$ is in $R_2^{{\wp}({\mathbf{B}})}$ for all $l$. Let $i$ be the minimal number such that $S_i$ is $\{1\}$, $\{2\}$, $\{0,1\}$ or $\{0,2\}$, and let $j $ be the minimal number such that $T_j$ is $\{1\}$, $\{2\}$, $\{0,1\}$ or $\{0,2\}$, and if such an $S_i $ or $T_j$ does not exists, then $i=0$ or $j=0$ respectively. If $i > 0$, then $T_i$ has to be $\{1\}$, $\{2\}$, $\{0,1\}$ or $\{0,2\}$ and hence $0<j \leq i$. Symmetrically, if $j > 0$, then $0<i \leq j$. Therefore, $i = j$. Now, if $i = j = 0$, then $(f(S_1,\dots,S_n), f(T_1,\dots,T_n)) = (0,0)$, which is in $R_2^{{\mathbf{B}}}$; if $i = j > 0$, then $(f(S_1,\dots,S_n), f(T_1,\dots,T_n)) \in R_2^{{\mathbf{B}}}$ follows directly from $(S_i,T_i)$ being in $R_2^{{\wp}({\mathbf{B}})}$. Finally, let us consider two arbitrary $n$-tuples $(S_1,\dots,S_n)$ and $(T_1,\dots,T_n)$ of ${\mathsf{Sing}}({\wp}({\mathbf{B}})^n)$ such that $(S_l,T_l)$ is in $R_1^{{\wp}({\mathbf{B}})}$ for all $l$. If $f(S_1,\dots,S_n)=2$, then $S_i=\{2\}$ or $\{0,2\}$ and $(S_i,T_i)$ cannot be in $R_1^{{\wp}({\mathbf{B}})}$. If $f(S_1,\dots,S_n)=1$, then $(f(S_1,\dots,S_n), f(T_1,\dots,T_n) )$ is in $\{1\}\times\{0,1,2 \}$ and, thus, in $R_1^{{\mathbf{B}}}$. If $j=0$, then let $k$ be an index such that $T_k=\{0\}$. Such an index has to exist, because $(T_1,\dots,T_n)$ is a tuple of ${\mathsf{Sing}}({\wp}({\mathbf{B}})^n)$. Since $(S_k,T_k)$ is in $R_1^{{\wp}({\mathbf{B}})}$, $S_k$ has to be $\{1\}$, and hence $f(S_1,\dots.,S_n) \in\{1,2\}$, and we appeal to one of the first two cases. The remaining case is $i = 0$ and $j>0$. In this case, $(f(S_1,\dots,S_n), f(T_1,\dots,T_n) )$ is in $\{0\}\times\{1,2 \}$ and therefore in $R_1^{{\mathbf{B}}}$.
We now move on to study SAC; we show that SAC is strictly more powerful than PAC.
\[prop:pac-in-sac\] Let ${\mathbf{B}}$ be a structure. If ${{\mathsf{CSP}}}({\mathbf{B}})$ is solvable by PAC, then it is also solvable by SAC.
Proposition \[prop:pac-in-sac\] follows directly from the algebraic characterizations given in Theorems \[thm:pac-solves\] and \[thm:sac-solves\]; it can also be seen to follow from the corresponding algorithm descriptions.
There exists a structure ${\mathbf{B}}$ having all constants such that ${{\mathsf{CSP}}}({\mathbf{B}})$ is solvable by SAC but not by PAC.
We will consider a structure that has as a polymorphism the idempotent binary commutative operation ${*}$ defined on the set $\{ 0, 1, 2, 3 \}$ by $1 {*}2 = 2$, $2 {*}3 = 3$, $3 {*}1 = 1$, and $0 {*}a = a$ for all $a \in \{ 1, 2, 3 \}$. We consider the structure ${\mathbf{B}}$ with universe $\{ 0, 1, 2, 3 \}$ over the signature $\{U_0, U_1, U_2, U_3, R_1, R_2 \}$ where we have $$U_0^{{\mathbf{B}}} = \{ (0) \}$$ $$U_1^{{\mathbf{B}}} = \{ (1) \}$$ $$U_2^{{\mathbf{B}}} = \{ (2) \}$$ $$U_3^{{\mathbf{B}}} = \{ (3) \}.$$ $$R_1^{{\mathbf{B}}} = \{ 0, 1, 2, 3 \}^2 \setminus \{ (0, 0) \},$$ $$R_2^{{\mathbf{B}}} = \{ (1, 2), (2, 3), (3, 1), (0, 0) \}$$ It is straightforward to verify that this structure ${\mathbf{B}}$ has the operation ${*}$ as a polymorphism. The solvability of ${\mathbf{B}}$ follows from Theorem \[thm:twosem-sac\], which is proved in the next section; see also the discussion in Example \[ex:twosem\].
To show that peek arc consistency does not solve ${{\mathsf{CSP}}}({\mathbf{B}})$, we prove that there is no homomorphism from ${\mathsf{Sing}}({\wp}({\mathbf{B}})^2)$ to ${\mathbf{B}}$, which is sufficient by Theorem \[thm:pac-solves\]. Define ${\overline{t_1}} = ( \{ 0 \}, \{ 1, 2, 3 \} )$ and ${\overline{t_2}} = (\{ 1, 2, 3 \}, \{ 0 \} )$. It is straightforward to verify that $({\overline{t_1}}, {\overline{t_2}}) \in R_1^{{\wp}({\mathbf{B}})^2}$; since each of the tuples ${\overline{t_1}}, {\overline{t_2}}$ contains a singleton, it holds that $({\overline{t_1}}, {\overline{t_2}}) \in R_1^{{\mathsf{Sing}}({\wp}({\mathbf{B}})^2)}$. Assume, for a contradiction, that $h$ is a homomorphism from ${\mathsf{Sing}}({\wp}({\mathbf{B}})^2)$ to ${\mathbf{B}}$. It then holds that $(h({\overline{t_1}}), h({\overline{t_2}})) \in R_1^{{\mathbf{B}}}$. Since $(0, 0) \notin R_1^{{\mathbf{B}}}$, we have that one of the values $h({\overline{t_1}}), h({\overline{t_2}})$ is not equal to $0$. Let us assume that $h({\overline{t_1}})$ is not equal to $0$; the other case is symmetric. Denote $h({\overline{t_1}})$ by $b$; we have $b \in \{ 1, 2, 3 \}$. Since each of the two tuples $( \{ 0 \}, \{ 0 \} )$, $( \{ 1, 2, 3 \}, \{ 1, 2, 3 \} )$ is contained in $R_2^{{\wp}({\mathbf{B}})}$, we have that $(( \{ 0 \}, \{ 1, 2, 3 \} ), ( \{ 0 \}, \{ 1, 2, 3 \} ))
\in R_2^{{\mathsf{Sing}}({\wp}({\mathbf{B}})^2)}$. It follows that $( b, b ) \in R_2^{{\mathbf{B}}}$, but since no tuple of the form $(c, c)$ with $c \in \{ 1, 2, 3 \}$ is contained in $R_2^{{\mathbf{B}}}$, we have reached our contradiction.
We close this section by showing that the structures solvable by SAC, and hence those solvable by any of the algorithms studied here, fall into the class of structures having *bounded width*. We begin by defining bounded width. A [*partial homomorphism*]{} from ${\mathbf{A}}$ to ${\mathbf{B}}$ is a mapping $f:A'\rightarrow B$, where $A'\subseteq A$, that defines a homomorphism to ${\mathbf{B}}$ from the substructure of ${\mathbf{A}}$ induced by $A'$. When $f$ and $g$ are partial homomorphisms we say that $g$ [*extends*]{} $f$, denoted by $f\subseteq g$, if ${\mathrm{Dom}}(f)\subseteq{\mathrm{Dom}}(g)$ and $f(a)=g(a)$ for every $a\in{\mathrm{Dom}}(f)$.
Let $k>1$. A [*$k$-strategy*]{} for an instance $({\mathbf{A}},{\mathbf{B}})$ is a nonempty collection $H$ of partial homomorphisms from ${\mathbf{A}}$ to ${\mathbf{B}}$ satisfying the following conditions:
1. (restriction condition) if $f\in H$ and $g\subseteq f$, then $g\in H$;
2. (extension condition) if $f\in H$, $|{\mathrm{Dom}}(f)|<k$, and $a\in A$, there is $g\in H$ such that $f\subseteq g$ and $a\in{\mathrm{Dom}}(g)$.
When $H$ is a $k$-strategy for $({\mathbf{A}},{\mathbf{B}})$ and $a_1,\dots,a_j\in A$ is a sequence, we define $H_{a_1,\dots,a_j}\subseteq B^j$ to be the relation $$\{(f(a_1),\dots,f(a_j)) ~|~ f\in H, {\mathrm{Dom}}(f)=\{a_1,\dots,a_j\}\}.$$
Let ${\mathbf{B}}$ be a structure and $k \geq 1$. We say that ${{\mathsf{CSP}}}({\mathbf{B}})$ has [*width*]{} $k$ if for all structures ${\mathbf{A}}$ the following holds: if there is a $(k+1)$-strategy for $({\mathbf{A}},{\mathbf{B}})$ then there is a homomorphism ${\mathbf{A}}\rightarrow{\mathbf{B}}$. We say that ${{\mathsf{CSP}}}({\mathbf{B}})$ has [*bounded width*]{} if it has width $k$ for some $k \geq 1$.
Let ${\mathbf{B}}$ be a structure. If ${{\mathsf{CSP}}}({\mathbf{B}})$ is solvable by SAC, then ${{\mathsf{CSP}}}({\mathbf{B}})$ has bounded width.
Let $r$ be the maximum of all the arities of the signature of ${\mathbf{B}}$, and set $k = \max(2, r+1)$. We shall show that for any instance ${\mathbf{A}}$ of ${{\mathsf{CSP}}}({\mathbf{B}})$, if $H$ is a $k$-strategy for $({\mathbf{A}},{\mathbf{B}})$, then the instance $({\mathbf{A}},{\mathbf{B}})$ has the SACC, which suffices.
Let us define the mapping $s:A\rightarrow {\wp}(B)\setminus\{ \emptyset \}$ as $s(a)=H_a$. Furthermore, for every $a\in A$, $b\in s(a)$, define $h_{a,b}: A \rightarrow {\wp}(B)\setminus\{\emptyset\}$ as the mapping $h_{a,b}(a')=\{b' ~|~ (b,b')\in H_{a,a'} \}$. Note that the extension property of $H$ guarantees that, for every $a'\in A$, $h_{a,b}(a')$ is, indeed, nonempty. It follows from the definition of $h_{a,b}$ that $h_{a,b}(a)=\{b\}$, and that for all $a' \in A$, $h_{a,b}(a') \subseteq s(a')$.
It is only necessary to show that $h_{a,b}$ defines a homomorphism from ${\mathbf{A}}$ to ${\wp}({\mathbf{B}})$. Let $R^{{\mathbf{A}}}$ be any relation in ${\mathbf{A}}$, let $(a_1,\dots,a_i)\in R^{{\mathbf{A}}}$, and let $S_j=h_{a,b}(a_j)$ for each $j \in [i]$. In order to prove that $(S_1,\dots,S_i)\in R^{{\wp}({\mathbf{B}})}$ it suffices to show that for every $j\in [i]$ and every $b_j\in S_j$, there exists some $(c_1,\dots,c_i)\in R^{{\mathbf{B}}} \cap (S_1\times\cdots\times S_i)$ with $c_j=b_j$. This is a direct consequence of the properties of the strategy. Indeed, by the definition of $h_{a,b}$ we know that $(b,b_j)\in H_{a,a_j}$ and then, by an iterative application of the extension property, we can show that there exists an extension $(b,c_1,\dots,c_i)\in H_{a,a_1,\dots,a_i}$ with $c_j=b_j$. The fact that $H$ contains only partial homomomorphisms guarantees that $(c_1,\dots,c_i)\in R^{\mathbf{B}}$. Finally, it follows from the restriction condition that for every $l \in [i]$, we have $c_l\in S_{l}$.
Tractability via singleton arc consistency {#sect:tractability}
==========================================
Majority operations
-------------------
An operation $m\colon B^3\to B$ is a *majority* operation if it satisfies the identity $m(x,y,y)=m(y,x,y)=m(y,y,x)=y$ for all $x,y\in B$. Relative to a majority operation $m \colon B^3 \to B$, when $I \subseteq J \subseteq B$, we say that $I$ is an *ideal* of $J$ if for every $x,y,z\in J$ such that $x,z\in I$ we have $m(x,y,z)\in I$. We will establish the following result.
\[the:SACCmajority\] If ${\mathbf{B}}$ is a structure that has a majority polymorphism, then singleton arc consistency solves ${{\mathsf{CSP}}}({\mathbf{B}})$.
The proof is obtained by using a strengthened version of the Prague strategy defined by Barto and Kozik [@bkbw].
In this section, for the sake of readability, we will typically use the notation $t[i]$ to denote the $i$th entry of a tuple $t$.
We introduce the following definitions relative to an instance $({\mathbf{A}}, {\mathbf{B}})$ with signature $\sigma$. A *pattern* $p$ of ${\mathbf{A}}$ is a sequence $a_1,e_1,a_2,\dots,e_{m-1},a_m$ such that $a_1,\dots,a_m$ are elements of $A$ and for every $n \in [m-1]$, we have that $e_n$ is a triple $(R,i,j)$ where $R$ is a symbol in $\sigma$ and $i,j$ are indices such that there is a tuple $t\in R^{{\mathbf{A}}}$ with $t[i]=a_n$ and $t[j]=a_{n+1}$. The [*length*]{} of pattern $p$ is defined to be $m$. A pattern is a *cycle* if $a_1=a_m$. By a *set system*, we mean any mapping $H: A \rightarrow {\wp}(B) \setminus \{ \emptyset \}$.
A pattern $q = b_1,e'_1,\dots,e'_{m-1},b_m$ of ${\mathbf{B}}$ having the same length as a pattern $p$ of ${\mathbf{A}}$ is a [*realization*]{} of $p$ if $e_n=e'_n$ for all $n \in [m-1]$. The pair $(b_1,b_m)$ is said to be a [*support*]{} of $p$. For a set system $H$, if it holds that $b_n\in H(a_n)$ for all $n\in [m]$ then $(b_1,b_m)$ is said to be a support of $p$ [*inside*]{} $H$.
A set system $H$ is a *weak strategy* if for every pattern $p=a_1,e_1,\dots,e_{m-1},a_m$ of ${\mathbf{A}}$, and every $b_1\in H(a_1)$ there exists some $b_m\in H(a_m)$ such that $(b_1,b_m)$ supports $p$ inside $H$. A set system $H$ is a *strong strategy* if for every cycle $p=(a=a_1,\dots,a_m=a)$ in ${\mathbf{A}}$ and every $b\in H(a)$, the pair $(b,b)$ supports $p$ inside $H$. Note that every strong strategy is a weak strategy and that the class of weak strategies remains the same if, in the definition of weak strategy, one replaces “every pattern $p=a_1,e_1,\dots,e_{m-1},a_m$” by “every pattern $p=a_1,e_1,\dots,e_{m-1},a_m$ of length $m=2$”.
\[obs:arcconsistency\] Every strong strategy is a weak strategy, relative to an instance $({\mathbf{A}}, {\mathbf{B}})$.
For a pattern $p=a_1,e_1,\dots,e_{m-1},a_m$ of ${\mathbf A}$, one needs only to apply the definition of strong strategy to the the pattern $a_1,e_1,\dots,e_{m-1},a_m,e_{m-1}^{-1},a_{m-1},\dots,e_1^{-1},a_1$, where $(R,i,j)^{-1}$ is defined to be $(R,j,i)$.
\[le:SACCtostrong\] There exists a strong strategy for an instance $({\mathbf{A}}, {\mathbf{B}})$ having the SACC.
Let $s: A \rightarrow {\wp}(B) \setminus \{ \emptyset \}$, $\{ h_{a, b} \}$ be the mappings witnessing that $({\mathbf{A}},{\mathbf{B}})$ has the SACC. We claim that the set system $H$ defined by $H(a)=s(a)$ for all $a\in A$ is a strong strategy. Indeed, let $p=a_1,e_1,\dots,a_m$ be a pattern of $A$ with $a_1=a_m=a$ and let $b\in H(a_1)$. We claim that there exists a realization $b_1,e_1,\dots,b_m$ of $p$ with $b_1=b_m=b$ such that for every $1\leq n\leq m$, $b_n\in h_{a,b}(a_n)$. The realization is constructed in an inductive manner. First, set $b_1$ to $b$. Assume now that $b_{n-1}$ is already set and let $e_{n-1}$ be $(R,i,j)$. There exists a tuple $(x_1,\dots,x_r)\in R^{{\mathbf{A}}}$ such that $x_i=a_{n-1}$ and $x_j=a_n$. Since $h_{a,b}$ is a homomorphism, the subset $S\subseteq B^r$ defined by $\pi_l S=h_{a,b}(x_l)$ for every $1\leq l\leq r$ is a subset of $R^{{\mathbf{B}}}$. From $b_{n-1}\in h_{a,b}(x_i)$ it follows that there exists a tuple $(y_1,\dots,y_r)\in S$ with $y_i=b_{n-1}$. Define $b_n$ to be $y_j$. Since, by definition of SACC strategy $h_{a,b}(a)=\{b\}$, it follows that $b_m=b$.
We now prove the following lemma, which, as we explain after the proof, essentially establishes the desired theorem. In the course of proving this lemma, we establish a number of observations.
By a minimal strong strategy, we mean minimal with respect to the ordering where for two strategies $H, H'$, we consider $H \subseteq H'$ if $H(a) \subseteq H'(a)$ for all $a \in A$.
\[le:singletonstrong\] If the relations of ${\mathbf{B}}$ are invariant under a majority operation $\phi$ and $H$ is a minimal strong strategy then for every $a \in A$, the set $H(a)$ is a singleton.
Towards a contradiction assume that $H$ is a minimal strong strategy and $a^* \in A$ is such that $H(a^*)$, is not a singleton. Consider the digraph $G$ whose nodes are of the form $(a,C)$ with $a\in A$ and $C\subseteq H(a)$, and there is an edge from $(a,C)$ to $(a',C')$ if there is a pattern $p=a_1,\dots,a_m$ with $a=a_1$ and $a'=a_m $ in ${\mathbf A}$ such that the following holds: $C'$ is the set containing all $b'\in H(a')$ such that $(b,b')$ is supported by $p$ inside $H$ for some $b\in C$.
\[observation3\] Let $p=a_1,e_1,\dots,a_m$ be a pattern, let $1<i<m$, let $q$ be the pattern $a_1,e_1,\dots,a_i$ and $r$ be the pattern $a_i,e_i\dots,e_m$. If $q$ defines an edge from $(a_1,C_1)$ to $(a_i,C_i)$ and $r$ defines an edge from $(a_i,C_i)$ to $(a_m,C_m)$ then $p$ defines an edge from $(a_1,C_1)$ to $(a_m,C_m)$. Hence, the graph $G$ is transitive.
The following observation follows from the definition of strong strategy.
\[observation1\] If there is an edge from $(a,C)$ to $(a,C')$ in $G$, then necessarily $C\subseteq C'$.
\[observation2\] If there is an edge from $(a,C)$ to $(a',C')$ in $G$, and $C$ is an ideal of $H(a)$, then $C'$ is an ideal of $H(a')$.
\[ (Observation \[observation2\])\] Let us prove the claim by induction on the length $m$ of the pattern that defines the edge.
Assume first that $m=2$. Let $a,(R,i,j),a'$ be any such pattern. Let $x_1,x_2,x_3\in H(a')$ and assume that two of them, say $x_1,x_3$, belong to $C'$. It follows, by the definition of edge, that for every $n\in\{1,3\}$ there exists tuple $t_n\in R^{{\mathbf{B}}}$ with $t_n[j]=x_n$ and $t_n[i]\in C$. Also, it follows by considering pattern $a',(R,j,i),a$ and from the fact that $H$ is a weak strategy that there exists a tuple $t_2\in R^{{\mathbf{B}}}$ with $t_2[i]\in H(a)$ and $t_2[j]=x_2$. Consider now tuple $t=\phi(t_1,t_2,t_3)$. Since $C$ is an ideal of $H(a)$ we have that $t[i]\in C$. Hence, we conclude that $\phi(x_1,x_2,x_3)=t[j]\in C'$.
The case $m>2$ follows from the inductive hypothesis and Observation \[observation3\].
Now, let $G'$ be the subgraph of $G$ induced by all nodes $(a,C)$ such that $C$ is an ideal of $H(a)$ and $C\neq H(a)$. Observe that as $H(a^*)$ is not a singleton, the graph $G'$ is nonempty, because every singleton is an ideal.
A subset $M$ of vertices of a directed graph is a *strongly connected component* if for every pair $(v,w)\in M^2$ there exists a path from $v$ to $w$ consisting only of vertices in $M$. It is a *maximal strongly connected component* if additionally, there is no edge $(v,w)$ with $v\in M$ and $w\not\in M$.
Let $M$ be a maximal strongly connected component of $G'$. The following observation is a direct consequence of Observations \[observation3\] and \[observation1\].
\[observation4\] The maximal strongly connected component $M$ cannot have two vertices $(a,C)$, $(a,C')$ with $C\neq C'$.
We shall construct a new strong strategy $H'$ as follows. If $(a,C)$ belongs to $M$, then set $H'(a)=C$ otherwise set $H'(a)=H(a)$. Clearly $H'$ is strictly smaller than $H$.
We shall start by showing that $H'$ is a weak strategy. By the note following the definition of weak strategy it is only necessary to show that for every pattern $p=a_1,e_1,a_2$ of length $2$ of ${\mathbf{A}}$ and every $b_1\in H'(a_1)$, there exists a support $(b_1,b_2)$ of $p$ inside $H'$.
We do a case analysis. If $(a_2,H'(a_2))$ does not belong to $M$ the claim follows from the fact that $H$ is a weak strategy. Assume now that $(a_2,H'(a_2))$ belongs to $M$. Consider the pattern $p=a_2,e_1^{-1},a_1$ where $(R,i,j)^{-1}=(R,j,i)$. This pattern defines an edge (in $G$) from $(a_2,H'(a_2))$ to a node $(a_1,C)$. Observe, that by the definition of the edges of $G$, we know that for every element $b\in C$ there is some $b'\in H'(a_2)$ such that $(b,b')$ is supported by $p$ inside $H$. Hence we only need to show that $H'(a_1)\subseteq C$.
If $(a_1,C)$ is in $G'$ then, since $M$ is a maximal strongly connected component of $G'$, we have that $(a_1,C)$ belongs to $M$ as well and hence $C=H'(a_1)$. If $(a_1,C)$ is not in $G'$ this must be because $C$ is not an ideal of $H(a_1)$ or because $C=H(a_1)$. We can rule out the first possibility in the following way: by the definitions of $G'$ and $H'$, $H'(a_2)$ is an ideal of $H(a_2)$. It follows by observation \[observation2\] that $C$ is an ideal of $H(a_1)$. In consequence $C=H(a_1)$ and the proof that $H'$ is a weak strategy is concluded.
It remains to show that $H'$ is a strong strategy. Let $p=a_1,e_1,\dots,e_{m-1},a_m$ be any cycle in $\mathbf A$ with $a_1=a_m=a$ and let $b$ be any element in $H'(a)$. Since $H'$ is a weak strategy we know that there is a realization $b_1,\dots,b_m$ of $p$ with $b_1=b$ inside $H'$. Notice that we do not necessarily have $b_m = b$. Symmetrically, by considering pattern $a_m,e^{-1}_{m-1}, \ldots, e^{-1}_1,a_1$ we know that there is a realization $d_m,e^{-1}_{m-1}, \ldots, e^{-1}_1,d_1$ of $p$ with $d_m=b$ inside $H'$. Also, since $H$ is a strong strategy we know that there exists a realization $c_1,e_1,\dots,c_m$ of $p$ such that $c_1=c_m=b$ inside $H$ (but not necessarily inside $H'$). Finally consider the sequence $x_1,\dots,x_m$ defined by $x_j=\phi(b_j,c_j,d_j), 1\leq j\leq m$. This sequence is a realization of $p$. Furthermore, we have that $x_1=x_m=b$. It remains to show that it is inside $H'$. Indeed, for every $1\leq j\leq m$, $\{b_j,d_j\}\subseteq H'(a_j)$ and $c_j\in H(a_j)$. Since $H'(a_j)$ is an ideal of $H(a_j)$ the claim follows.
(Theorem \[the:SACCmajority\]) Suppose that the instance $({\mathbf{A}}, {\mathbf{B}})$ has the SACC and that ${\mathbf{B}}$ has the majority polymorphism $\phi$. By Lemmas \[le:SACCtostrong\] and \[le:singletonstrong\] there exists a strong strategy $H$ for $({\mathbf{A}}, {\mathbf{B}})$ such that $H(a)$ is a singleton for every $a\in A$. Consider now the mapping $h: A\rightarrow B$ that maps every $a\in A$ to the only element in $H(a)$. We claim that $h$ is a homomorphism from ${\mathbf{A}}$ to ${\mathbf{B}}$. Indeed, let $R$ be any relation symbol, and $(a_1,\dots,a_r)$ be any tuple in $R^{{\mathbf{A}}}$. Fix any $1\leq i,j\leq r$ and consider pattern $a_i,(R,i,j),a_j$. It follows by the definition of strong strategy that there is a $t\in R^{{\mathbf{B}}}$ such that $t[i]=h(i)$ and $t[j]=h(j)$. Since $R^{{\mathbf{B}}}$ is necessarily $2$-decomposable [@CCC], $h$ is a homomorphism.
2-semilattice operations
------------------------
A 2-semilattice ${\mathbb{G}}= (G, {\star})$ consists of a set $G$, which in this paper we assume to be finite, and a binary operation ${\star}$ satisfying $x {\star}x = x$ (idempotency), $x {\star}y = y {\star}x$ (commutativity), and $x {\star}(x {\star}y) = (x {\star}x) {\star}y$ (restricted associativity).
Each 2-semilattice naturally induces a directed graph $(G, E)$ where $(a, b) \in E$ if and only if $a {\star}b = b$. When $(a,b)\in E$, we also write $a \leq b$. The graph $(G,E)$ is connected, since $a {\star}(a {\star}b) = b {\star}(a {\star}b) = a {\star}b$ for any $a,b\in G$, and therefore, $a,b \leq a {\star}b$. Each 2-semilattice has a unique maximal strongly connected component, that is a strongly connected component with no outgoing edges, denoted by ${\overline{G}}$. The component ${\overline{G}}$ is also the unique strongly connected component of $(G, E)$ such that for any $a \in G$, there exists $b \in {\overline{G}}$ such that $a \leq b$. In this section, we will prove that a certain class of 2-semilattices is tractable via singleton arc consistency. Our treatment of 2-semilattices is inspired and influenced by the study conducted by Bulatov [@twosemilattices], who proved that they are polynomial-time tractable.
A 2-semilattice ${\mathbb{G}}= (G, {\star})$ is an algebra. By an *algebra*, we mean a pair $(A, O)$ consisting of a set $A$, the *universe* of the algebra, and a set $O$ of operations on $A$. A *congruence* of an algebra is an equivalence relation preserved by the operation(s) of the algebra, and an algebra is *simple* if its only congruences are trivial (that is, if its only congruences are the equality relation on $A$ and $A \times A$, where $A$ is the universe of the algebra).
We will begin by proving some general results on singleton arc consistency. In the following discussion, a *subalgebra* is defined, with respect to a relational structure ${\mathbf{B}}$, as a subset $S \subseteq B$ that is preserved by all polymorphisms of ${\mathbf{B}}$. For an arbitrary subset $T \subseteq B$, we use ${\langle}T {\rangle}$ to denote the smallest subalgebra containing $T$.
\[prop:poly-ac\] Suppose that $g_1, \ldots, g_k: {\mathbf{A}}\rightarrow {\wp}({\mathbf{B}})$ are homomorphisms, and suppose that $f: B^k \rightarrow B$ is a polymorphism of ${\mathbf{B}}$. Then the map $g: A \rightarrow {\wp}(B) \setminus \{ \emptyset \}$ defined by $g(a) = f(g_1(a), \ldots, g_k(a))$ for all $a \in A$ is a homomorphism ${\mathbf{A}}\rightarrow {\wp}({\mathbf{B}})$.
For an operation $f: B^k \rightarrow B$ and a sequence of subsets $B_1, \ldots, B_k \subseteq B$, by the notation $f(B_1, \ldots, B_k)$, we denote the set $\{ f(b_1, \ldots, b_k) ~|~ b_1 \in B_1, \ldots, b_k \in B_k \}$. Regarding this notation, it is easy to verify that $f$ can be understood as a polymorphism of ${\wp}({\mathbf{B}})$ if $f$ is a polymorphism of ${\mathbf{B}}$. Proposition \[prop:poly-ac\] follows straightforwardly from the definitions.
\[prop:subalgebra-ac\] Suppose that $h: {\mathbf{A}}\rightarrow {\wp}({\mathbf{B}})$ is a homomorphism. Then the map $h'$ defined by $h'(a) = {\langle}h(a) {\rangle}$ for all $a \in A$ is also a homomorphism ${\mathbf{A}}\rightarrow {\wp}({\mathbf{B}})$.
Repeatedly apply Proposition \[prop:poly-ac\] with a polymorphism $f$ and $g_1 = \cdots = g_k = h$, each time taking the resulting $g$ and updating $h$ to be $h \cup g$. Note that at each step, the new $h$ is a homomorphism ${\mathbf{A}}\rightarrow {\wp}({\mathbf{B}})$, since the union operation $\cup$ is a polymorphism of ${\wp}({\mathbf{B}})$. When no changes can be made, the resulting $h$ is the desired $h'$.
Let us say that a CSP instance $({\mathbf{A}}, {\mathbf{B}})$ has the *subalgebra SACC* if $({\mathbf{A}}, {\mathbf{B}})$ has the SACC relative to mappings $s, \{ h_{a, b} \}$ such that for all $a \in A$, the set $s(a)$ is a subalgebra, and for all $a, a' \in A$, $b \in s(a)$, the set $h_{a, b}(a')$ is a subalgebra.
\[prop:sacc-implies-subalgebra-sacc\] If a pair $({\mathbf{A}}, {\mathbf{B}})$ of similar structures has the SACC, and all polymorphisms of ${\mathbf{B}}$ are idempotent, then it has the subalgebra SACC.
Suppose that $({\mathbf{A}}, {\mathbf{B}})$ has the SACC with respect to the mappings $s, \{ h_{a, b} \}$. Set $s'(a) = {\langle}s(a) {\rangle}$ for all $a \in A$, and $h'_{a, b}(a') = {\langle}h_{a, b}(a') {\rangle}$ for all $a, a' \in A$, $b \in s(a)$. Clearly, for all such $a, a', b$ we have $h'_{a, b}(a') \subseteq s'(a')$, and also, that $h'_{a, b}$ is a homomorphism ${\mathbf{A}}\rightarrow {\wp}({\mathbf{B}})$ (by Proposition \[prop:subalgebra-ac\]). Let $b$ be an element in $s'(a) \setminus s(a)$ for some $a \in A$. We need to show that there exists a homomorphism $h'_{a, b}$ that satisfies the two conditions of Definition \[def:sacc\] with respect to $s'$, and that also satisfies the subalgebra condition. As $s'(a)$ is defined as ${\langle}s(a) {\rangle}$, it holds that $s'(a) = \{ f(b_1, \ldots, b_k) ~|~ f \mbox{ a polymorphism of } {\mathbf{B}}; b_1, \ldots, b_k \in s(a) \}$; the containment $\supseteq$ is clear by definition of subalgebra, and the containment $\subseteq$ follows from the fact that the right hand side is a subalgebra, which in turn follows from the fact that the set of polymorphisms of ${\mathbf{B}}$ forms a clone and is closed under composition [@szendrei86-clones]. Thus, there exists a polymorphism $f$ of $B$ and elements $b_1, \ldots, b_k \in s(a)$ such that $b = f(b_1, \ldots, b_k)$. Let $g'_{a, b}$ be the homomorphism obtained from Proposition \[prop:poly-ac\] with $g_i = h_{a, b_i}$ and $f$. Set $h'_{a, b}(a') = {\langle}g'_{a, b}(a') {\rangle}$ for all $a' \in A$. The homomorphism $h'_{a, b}$ has the desired properties.
We now turn to prove our tractability result. We will now use the term subalgebra to refer to a *subalgebra* of a 2-semilattice $(B, {\star})$, that is, a subset of $B$ preserved by ${\star}$. Note, however, that we will be working with a relational structure ${\mathbf{B}}$ assumed to have ${\star}$ as a polymorphism, so a subalgebra in the previous sense (that is, with respect to ${\mathbf{B}}$) will also be a subalgebra in this sense. An algebra $(B, {\star})$ having a binary operation is *conservative* if for all $b, b' \in B$, it holds that $b {\star}b' \in \{ b, b' \}$. The following is the statement of our tractability result.
\[thm:twosem-sac\] Let $(B, {\star})$ be a conservative 2-semilattice such that every strongly connected subalgebra is simple. If ${\mathbf{B}}$ is a structure having ${\star}$ as a polymorphism, then singleton arc consistency solves ${{\mathsf{CSP}}}({\mathbf{B}})$.
\[ex:twosem\] We consider the binary operation ${*}$ on $\{ 0, 1, 2, 3 \}$ defined by the following table.
$
\begin{tabular}{|c|c|c|c|c|}
\hline
{*}& 0 & 1 & 2 & 3 \\ \hline
0 & 0 & 1 & 2 & 3 \\ \hline
1 & 1 & 1 & 2 & 1 \\ \hline
2 & 2 & 2 & 2 & 3 \\ \hline
3 & 3 & 1 & 3 & 3 \\ \hline
\end{tabular}
$
It is straightforward to verify that this operation is commutative and conservative, and is a 2-semilattice. The graph induced by this operation has edges $(0, 1), (0, 2), (0, 3), (1, 2), (2, 3), (3, 1)$, as well as self-edges on each of the vertices. There is thus just one strongly connected component of size strictly greater than one, namely, the component $\{ 1, 2, 3 \}$. This is a subalgebra of the algebra $(\{ 0, 1, 2, 3 \}, {*})$ and is readily verified to be simple. Hence, the tractability via singleton arc consistency of any structure preserved by the operation ${*}$ follows from Theorem \[thm:twosem-sac\].
We will make use of the following results. For our purposes here, a *subdirect product* of algebras ${\mathbb{A}}_1, \ldots, {\mathbb{A}}_k$ is a subalgebra $S$ of $A_1 \times \cdots \times A_k$ such that for each $i \in [k]$, it holds that $\pi_i S = A_i$.
\[lemma:subdir-scc\] Suppose that $S$ is a subdirect product of 2-semilattices $S_1, \ldots, S_n$. Then $S \cap ({\overline{S_1}} \times \cdots \times {\overline{S_n}})$ is a subdirect product of ${\overline{S_1}}, \ldots, {\overline{S_n}}$.
Immediate from [@twosemilattices Lemma 3.2].
\[def:almost-trivial\] A relation $S \subseteq B^n$ is *almost trivial* if there exists a partition $I_1, \ldots, I_k$ of $[n]$ such that
- $t \in S$ if and only if for all $i \in [k]$, it holds that ${\pi}_{I_i} t \in {\pi}_{I_i} S$; and,
- for each $j \in [k]$, it holds that ${\pi}_{I_j} S$ has the form $\{ (\pi_1(p), \pi_2(p), \ldots, \pi_m(p)) ~|~ p \in [q] \}$ for some $q \geq 1$ and where each mapping $\pi_i$ is a bijection from $[q]$ to a subset of $B$.
\[prop:subdir-of-simple-sc\] A subdirect product of simple strongly connected 2-semilattices is an almost trivial relation, and is hence itself strongly connected.
Immediate from [@twosemilattices Proposition 3.1].
\[prop:sacc-almost-trivial\] Let $({\mathbf{A}}, {\mathbf{B}})$ be an instance that has the SACC with respect to $s: A \rightarrow {\wp}({\mathbf{B}}) \setminus \{ \emptyset \}$. If for each tuple $(a_1, \ldots, a_k) \in R^{{\mathbf{A}}}$, it holds that $R^{{\mathbf{B}}} \cap (s(a_1) \times \cdots \times s(a_k))$ is almost trivial, then there is a homomorphism from ${\mathbf{A}}$ to ${\mathbf{B}}$.
Consider the following graph $G=(A,E)$, where $\{a,b\}\in E$ if and only if there is a relation $R^{\mathbf{A}}$ in ${\mathbf{A}}$, and, if $I_1,\dots,I_k$ is its partition regarding almost triviality of $R^{{\mathbf{B}}} \cap (s(a_1) \times \cdots \times s(a_k))$, there further is an $l\in [k]$ and a tuple $(a_1,\dots,a_m)\in R^{{\mathbf{A}}}_{I_l}$ such that there are $i,j$ with $a=a_i$ and $b=a_j$. For each connected component $C$ of the graph $G$ arbitarily choose $a\in C$ and $b\in s(a)$. Since arc consistency can be established when $a$ is set to $b$ and using the structure of the projected relations $R^{\mathbf{B}}_{I_l}$, there exists a unique extension of $a\mapsto b$ to a homomorphism on $C$. Because of the first property of Definition \[def:almost-trivial\] the homomorphisms on the single components can be combined to a homomorphism on ${\mathbf{A}}$.
The following is the main result used to prove Theorem \[thm:twosem-sac\].
\[thm:sacc-on-max-components\] Suppose that ${\mathbf{B}}$ satisfies the hypotheses of Theorem \[thm:twosem-sac\], and suppose that $({\mathbf{A}}, {\mathbf{B}})$ has the subalgebra SACC via $s: A \rightarrow {\wp}(B) \setminus \{ \emptyset \}$. Then, $({\mathbf{A}}, {\mathbf{B}})$ has the SACC via the map $s': A \rightarrow {\wp}({\mathbf{B}}) \setminus \{ \emptyset \}$ defined by $s'(a) = {\overline{s(a)}}$ for all $a \in A$.
Let $a \in A$ and $b \in {\overline{s(a)}}$. By hypothesis, there exists a homomorphism $h: {\mathbf{A}}\rightarrow {\wp}({\mathbf{B}})$ where $h(a) = \{ b \}$ and for all $a' \in A$, it holds that $h(a') \subseteq s(a')$. We want to show that there exists a homomorphism $h': {\mathbf{A}}\rightarrow {\wp}({\mathbf{B}})$ where $h'(a) = \{ b \}$ and for all $a' \in A$, it holds that $h'(a') \subseteq s'(a')$. Define $h'(a)$ as ${\overline{h(a)}}$ if $h(a) \cap s'(a) \neq \emptyset$, and as $s'(a)$ otherwise. Observe that in the first case, we have $h'(a) = {\overline{h(a)}} \subseteq s'(a)$, and that in both cases, the subset $h'(a)$ is a subalgebra.
We claim that $h'$ is a homomorphism from ${\mathbf{A}}$ to ${\wp}({\mathbf{B}})$. Let $a \in R^{{\mathbf{A}}}$ be a tuple in ${\mathbf{A}}$. For the sake of notation, we assume that $a = (a_1, \ldots, a_{k+l})$, $I = \{ 1, \ldots, k \}$, $J = \{ k+1, \ldots, k+l \}$, and that $I$ contains exactly the coordinates $i \in [k+l]$ such that $h(a_i) \cap s'(a_i) \neq \emptyset$, so that $h'(a_i) = {\overline{h(a_i)}}$ for all $i \in I$ and $h'(a_j) = s'(a_j)$ for all $j \in J$. Let ${T = ({\pi}_I R^{{\mathbf{B}}} \cap (s(a_1) \times \cdots \times s(a_k))) \cap ({\overline{s(a_1)}} \times \cdots \times {\overline{s(a_k)}})}$. By Lemma \[lemma:subdir-scc\], we have that relation $T$ is a subdirect product of ${\overline{s(a_1)}}, \ldots, {\overline{s(a_k)}}$. Further, let $W = { (R^{{\mathbf{B}}} \cap (s(a_1) \times \cdots \times s(a_{k+l}))) \cap ({\overline{s(a_1)}} \times \cdots \times {\overline{s(a_{k+l})}}) }$. By Lemma \[lemma:subdir-scc\], we have that $W$ is a subdirect product of ${\overline{s(a_1)}}, \ldots, {\overline{s(a_{k+l})}}$. Clearly, ${\pi}_I W \subseteq T$. We show that $T \subseteq {\pi}_I W$ (and hence that $T = {\pi}_I W$), as follows. Let $t$ be a tuple in $T$. Let $w$ be a tuple in $W$ (such a tuple can be obtained, for instance, by ${\star}$-multiplying together all tuples of $R^{{\mathbf{B}}}\cap (s(a_1) \times \cdots \times s(a_{k+l}))$, in any order). By our assumption on ${\mathbf{B}}$ and by Proposition \[prop:subdir-of-simple-sc\], there is a sequence of tuples $u_1, \ldots, u_m$ in $T$ such that ${\pi}_I w \leq u_1 \leq \cdots \leq u_m = t$. We hence have tuples $v_1, \ldots, v_m$ with $v_i \in R^{{\mathbf{B}}}\cap (s(a_1) \times \cdots \times s(a_{k+l}))$ and ${\pi}_I v_i = u_i$ for each $i \in [m]$. The product $(\cdots ((w {\star}v_1) {\star}v_2) {\star}\cdots {\star}v_m)$ gives a tuple in $W$ whose projection onto $I$ is $t$.
By Proposition \[prop:subdir-of-simple-sc\], it holds that $W$ is almost trivial with respect to a partition $\{ I_i \}$. Remove from the $I_i$ any coordinates $l$ such that $s'(a_l)$ has just one element. We now show that the resulting partition $\{ I_i \}$ has the property that each $I_i$ is a subset of either $I$ or $J$. By the homomorphism $h$ and its subalgebra property, there exists a tuple $(t, x) \in R^{{\mathbf{B}}}$ such that $t \in T$ and $x_j \notin {\overline{s(a_j)}}$ for all $j \in J$ (just multiply all tuples in $R^{\mathbf{B}}\cap (h(a_1) \times \cdots \times h(a_{k+l}))$). By the fact that $T \subseteq {\pi}_I W$, we have a tuple $(t, u) \in W$. By the strong connectedness of $W$ (Proposition \[prop:subdir-of-simple-sc\]), there is a tuple $(t', u') \in W$ that is distinct from $(t, u)$ at each coordinate in $\cup I_i$ and such that $(t', u') {\star}(t, u) = (t, u)$. We also have $(t', u') {\star}(t, x) = (t, u')$; note that $u' {\star}x = u'$ by conservativity of ${\star}$. As $u$ and $u'$ differ at every coordinate in $J\cap(\cup I_i)$, the claim follows.
As a consequence of this last result, for any tuple $t \in {\pi}_I W$ and any tuple $u \in {\pi}_J W$, it holds that $(t, u) \in W$. Further it holds that $({\pi}_I R^{{\mathbf{B}}}\cap (h(a_1) \times \cdots \times h(a_k)))\cap (h'(a_1) \times \cdots \times h'(a_k))$ is a subdirect product of $h'(a_1), \ldots, h'(a_k)$ (Lemma \[lemma:subdir-scc\]), and we have that $h'$ is a homomorphism from ${\mathbf{A}}$ to ${\wp}({\mathbf{B}})$.
(Theorem \[thm:twosem-sac\]) Suppose that $({\mathbf{A}}, {\mathbf{B}})$ has the SACC. By Proposition \[prop:sacc-implies-subalgebra-sacc\], the instance $({\mathbf{A}}, {\mathbf{B}})$ has the subalgebra SACC. By Theorem \[thm:sacc-on-max-components\], $({\mathbf{A}}, {\mathbf{B}})$ has the SACC via a mapping $s'$ where for all $a \in A$, it holds that $s'(a)$ is a strongly connected subalgebra. By assumption, each such $s'(a)$ is simple, and it follows from Propositions \[prop:subdir-of-simple-sc\] and \[prop:sacc-almost-trivial\] that there is a homomorphism from ${\mathbf{A}}$ to ${\mathbf{B}}$.
Discussion
==========
In this work, we performed a systematic study of arc consistency and three simple, natural extensions thereof. We performed a comparison of the studied consistency notions based on constraint languages, and proved positive tractability results for singleton arc consistency.
Atserias and Weyer [@atserias-weyer] gave a uniform treatment of AC, PAC, SAC, and general consistency. Among other results, they show that it can be decided, given a constraint language and any pair of the previous consistency methods, whether it is true that the set of instances that passes one of the consistency tests coincides with the set of instances that passes the other. Their results combined with the fact that general consistency/bounded width is decidable [@bkbw] implies that it can be decided whether or not a given constraint language is solvable by any of the other methods.
We conclude by posing one question for future work. Barto and Kozik [@bkbw] have recently characterized all languages solvable by bounded width. Can all such languages be solvable by singleton arc consistency, or are there bounded width languages not solvable by singleton arc consistency? Resolving this question in the positive would seem to yield an interesting alternative characterization of the bounded width languages.
#### Acknowledgements.
The authors thank Manuel Bodirsky for his comments and collaboration. The authors also thank Johan Thapper for his many useful comments.
[^1]: Chen is supported by the Spanish program “Ramon y Cajal”.
[^2]: Both Chen and Dalmau are supported by MICINN grant TIN2010-20967-C04-02.
[^3]: Work by Grußien was supported by the Deutsche Forschungsgemeinschaft (DFG) within the research training group “Methods for Discrete Structures” (GrK 1408).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We analyse *RXTE* and *Swift* observations of SWIFT J1729.9$-$3437 after its outburst from 2010 July 20 to 2010 August 12. We calculate a spin frequency and spin frequency derivative of $1.8734(8) \times 10^{-3}$ Hz and $6.42(6) \times 10^{-12}$ Hz/s respectively from the quadratic fit of the pulse arrival times.The quadratic fit residuals fit well to a circular orbital model with a period of $15.3(2)$ days and a mass function of about $1.3M_{\odot}$, but they can also be explained by a torque noise strength of $6.8 \times 10^{-18}$ Hz sec$^{-2}$. Pulse profiles switch from double-peaked to single-peaked as the source flux continues to decrease. We find that the pulse shape generally shows no strong energy dependence. The hardness ratios reveal that the source becomes softer with decreasing flux. We construct a single spectrum from all the available *RXTE* and *Swift* observations. We find that adding an *Fe* line complex feature around 6.51 keV slightly improves the spectral fit. We also find that *Fe* line flux correlates with X-ray flux which might indicate the origin of the *Fe* emission as the source itself rather than the Galactic ridge.'
author:
- |
Ş. Şahiner$^{1}$[^1], S. Ç. İnam$^{2}$, M. M. Serim$^{1}$ and A. Baykal$^{1}$\
$^{1}$Physics Department, Middle East Technical University, 06531 Ankara, Turkey\
$^{2}$Department of Electrical and Electronics Engineering, Başkent University, 06530 Ankara, Turkey
date: Received 2013
title: ' *RXTE* and *Swift* observations of SWIFT J1729.9$-$3437'
---
\[firstpage\]
accretion, accretion discs - stars:neutron - pulsars:individual: SWIFT J1729.9-3437
Introduction
============
SWIFT J1729.9$-$3437 is a transient X-ray pulsar discovered during all-sky monitoring with the *Swift* Burst Alert Telescope (BAT) on 2010 July 13 (Markwardt et al. 2010a). At the same time, *Rossi X-ray Timing Explorer* (*RXTE*) Proportional Counter Array (PCA) monitoring scans of the Galactic centre region confirmed a gradual increase in flux from a position consistent with the *Swift* position of the source (RA = 262.53746, Dec. = $-$34.61239). Consecutive *RXTE*$-$PCA pointings identified SWIFT J1729.9$-$3437 as a pulsar with $\sim$530 s pulsations (Markwardt et al. 2010b). Markwardt et al.(2010b) also suggested that the X-ray spectrum of the source is compatible with basic X-ray pulsar spectra, modelled by an absorbed power law with a high energy cut-off.
In this paper, we analyse *RXTE* and *Swift* observations of SWIFT J1729.9$-$3437 and examine the spectral and timing properties of its outburst from 2010 July 20 to 2010 August 12 (MJD 55397 $-$ MJD 55420). In Section 2, we describe the observations that we analyse. In Section 3 and 4, we present our timing and spectral analysis results. In Section 5, we discuss our results and conclude.
Observations
============
*RXTE*
------
11 pointing *RXTE*$-$PCA observations of SWIFT J1729.9$-$3437 were performed between 2010 July 20 and 2010 August 8. The total exposure time is about 42 ks. PCA consisted of five co-aligned identical proportional counter units (PCUs) (Jahoda et al. 1996) each having an effective area of approximately 1300 $cm^{2}$. The energy range of PCA was from 2 to 60 keV, with an energy resolution of 18 per cent at 6 keV. The field of view (FOV) at full width at half-maximum (FWHM) was about 1$\degr$. Generally, various PCUs were turned off during PCA observations. The responses of PCU0 and PCU1 were not well known due to loss of their propane layers. Furthermore, the responses of the top anode layers were modeled better than the other layers. Although the number of active PCUs during the observations of SWIFT J1729.9$-$3437 varied between one and three, we select only data from the top anode layer of PCU2 because of the reasons mentioned above.
The standard software tools of `HEASOFT v.6.11` are used for the PCA data analysis. Data is filtered by selecting times when the elevation angle is greater than 10$\degr$, the offset from the source is less than 0.02$\degr$ and the electron contamination of PCU2 is less than 0.1. The background extraction for spectra and light curves are done by using the latest PCA background estimator models supplied by the RXTE Guest Observer Facility (GOF), `EPOCH 5C`. During the extraction of spectra, Standard2f mode data is considered with 128 energy channels and 16-s time resolution. Relevant response matrices for spectra are created by `PCARSP V.11.7.1`. Furthermore we construct pulse phase resolved spectra with the tool `FASEBIN Rev.1.8` by using Good Xenon mode event data and 125 $\mu$s time resolution event files (`E_125us_64M_0_1s`) when Good Xenon mode data is not available. From these spectra, we generate energy resolved pulse profiles by obtaining the count rates per phase bin with the tool `FBSSUM`. The 1 s binned 3-20 keV light curves are produced from Good Xenon and `E_125us_64M_0_1s` events.
*SWIFT*
-------
After the discovery of SWIFT J1729.9$-$3437 by the *Swift*$-$BAT (Barthelmy et al. 2005) all-sky monitoring on 2010 July 13 (Markwardt et al. 2010a), a total of 11 follow-up X-ray Telescope (XRT; Burrows et al. 2005) observations (each $\sim$3 ks) were carried out between 2010 July 20 and 2010 August 12. XRT is a focusing instrument on board the *Swift* satellite (Gehrels et al. 2004) which operates between 0.2 and 10 keV energy range. XRT has an effective area of 110 $cm^{2}$ at 1.5 keV with a spatial resolution of 18$\arcsec$ and the FOV of $23.6\arcmin \times 23.6\arcmin$. Operation mode of XRT switches between photon-counting (PC), imaging and timing depending on the brightness of the observed source. Pointing observations of SWIFT J1729.9$-$3437 are in the PC mode. Screened event files are produced by the `XRTDAS v.2.7.0` standard data pipeline package `XRTPIPELINE v.0.12.6`. Standard grade filtering (0-12) is applied to PC mode data.
The spectral extraction is carried out by filtering appropriate regions using `XSELECT v.2.4b`. For the observations in which the XRT count rate is high enough (above 0.5 cts/s) to produce event pile-up, source region is selected to be annular. We compared the observed and nominal point spread function (PSF) (Vaughan et al. 2006) and decided the size of the core affected in order to determine the radius of the inner circle of the annulus. For this purpose, we first modelled the outer wings ($>15 \arcsec$) of the PSF with a King function which has typical parameter values for XRT (Moretti et al. 2005). Then the model is extrapolated to the inner region for comparison with the observed PSF and the size of the piled-up region is determined from the deviation point between the data and the model. For the brightest observations an exclusion region of radius $\sim9\arcsec$ is sufficient to correct pile-up. For low count rate observations a circular region of radius 47$\arcsec$ is selected for the source spectral extraction. Source regions are centred on the position determined with `XRTCENTROID v.0.2.9`. A circular source-free region with 140$\arcsec$ radius is selected for the background spectral extraction. Resulting spectral files are grouped to require at least 50 counts per bin using the ftool `GRPPHA v.3.0.1`, for the $\chi^{2}$ statistics to be applicable. We used the latest response matrix file (version v013) and created individual ancillary response files using the tool `XRTMKARF v.0.5.9` with the exposure map produced by `XRTEXPOMAP v.0.2.7`. Spectral analysis is performed using `XSPEC v.12.7.0`.
For the timing analysis, arrival times of XRT events are first corrected to the Solar system barycenter by using the tool `BARYCORR v.1.11`. Then, background-subtracted XRT light curves in the 0.2$-$10 keV energy range; corrected for pile-up, PSF losses and vignetting; are extracted by using the highest possible time resolution (2.51 s) for PC mode data.
Timing Analysis
===============
Timing Solution
---------------
For the timing analysis, we use 1 s binned *RXTE*$-$PCA and 2.51 s binned *Swift*$-$XRT light curves of the source. To illustrate the temporal variation of the pulse phase averaged count rate of the source, in Fig. \[bothlc\], we present 530 s binned versions of these light curves. These background subtracted light curves are also corrected to the barycenter of the Solar system.
In order to estimate pulse periods of the source, the *RXTE*$-$PCA time series is folded on statistically independent trial periods (Leahy et al. 1983). Template pulse profiles are formed from these observations by folding the data on the period giving maximum $\chi^2$. Then the barycentred *Swift*$-$XRT time series are also folded over the best period obtained from *RXTE*. Pulse profiles consisting of 20 phase bins are represented by their Fourier harmonics (Deeter & Boynton 1985). Using cross-correlation of pulses between template and in each observation, we obtain the pulse arrival times.
We are able to connect all the pulse arrival times of the observations in phase over the whole observation time span. Following the approach of Deeter et al. (1981), we find that it is possible to fit the pulse arrival times to the second order polynomial, $$\delta \phi = \delta \phi_{o} + \delta \nu (t-t_{o})
+ \frac{1}{2}\dot{\nu} (t-t_{o})^{2}
\label{polyn}$$ where $\delta \phi$ is the pulse phase offset found from the pulse timing analysis, $t_{o}$ is the epoch for folding; $\delta \phi_{o}$ is the residual phase offset at t$_{o}$; $\delta \nu$ is the correction to the pulse frequency at time $t_0$; $\dot{\nu}$, being the pulse frequency derivative, is the second derivative of the pulse phase. In Fig. \[arrt\], we present the pulse phases and the residuals of this quadratic fit.
In Fig. \[pulexp\], we show typical *Swift*$-$XRT and *RXTE*$-$PCA pulse profiles. As seen from this figure, the *Swift* pulse has larger error bars and less signal to noise ratio relative to the *RXTE* pulse. In the pulse timing analysis we find that the error bars of the pulse phases of *RXTE* and *Swift* are inversely correlated with the signal to noise ratio of the pulses of each observation.
\[soln\]
The residuals shown in Fig. \[arrt\] fit well to a sinusoidal function with a period of 15.3(2) days. This corresponds to a circular orbital model, parameters of which are listed in Table \[soln\]. This circular model has a reduced $\chi^2$ of 1.0. Using an elliptical model we find an upper limit for the eccentricity as 0.6 (see bottom panel of Fig. \[arrt\]). In this case, reduced $\chi^2$ is found to be 0.4. This small reduced $\chi^2$ value indicates that the elliptical orbital model “over-fits” data.
The residuals of the quadratic can alternatively express the noise process due to random torque fluctuations (Bildsten et al. 1997, Baykal et al. 2007). In order to estimate the noise strength, we fit a cubic polynomial to the residuals of the pulse arrival times. The observed time series is simulated by the Monte Carlo technique for a unit white noise strength defined as $P_{\dot \nu}({f})=1$ and fitted to a cubic polynomial in time. Then the square of the third order polynomial term is divided into the value from Monte Carlo simulations (Deeter 1984, Cordes 1980). The torque noise strength is obtained as $6.8 \times 10^{-18}$ Hz sec$^{-2}$. This value of noise strength estimate is comparable with those of other accretion powered sources such as wind accretors e.g. Vela X$-$1, 4U 1538$-$52 and GX 301$-$2 which has the values changing between $10^{-20}$ and $10^{-18}$ Hz s$^{-2}$ (Bildsten et al. 1997). Her X$-$1 and 4U 1626$-$67, which are disc accretors with low mass companions, have shown pulse frequency derivatives consistent with noise strengths $10^{-21}$ to $10^{-18}$ Hz sec$^{-2}$ (Bildsten et al. 1997). Therefore residuals of quadratic fit can also be associated with torque noise fluctuation of the source.
Pulse Profiles and Hardness Ratios
----------------------------------
A double-peaked pulse profile is observed in the first five *RXTE*$-$PCA observations of SWIFT J1729.9$-$3437 (see panel (A) of Fig. \[lcex\]); however one peak loses its intensity starting from the middle of the observation on MJD 55406.7 (see panel (B) Fig. \[lcex\]) as the source flux continues to decrease after the burst. The peak value of 2$-$10 keV unabsorbed flux is $3.04 \times 10^{-10}\,$erg$\,$cm$^{-2}\,$s$^{-1}$ on MJD 55398.7, during its gradual decrease it reaches $1.96 \times 10^{-10}\,$erg$\,$cm$^{-2}\,$s$^{-1}$ on MJD 55408.5, when the shape of the pulse profile changes (see panel (C) Fig. \[lcex\]). The minimum flux observed for the last *RXTE*$-$PCA observation on MJD 55416.4 is $1.36 \times 10^{-10}\,$erg$\,$cm$^{-2}\,$s$^{-1}$. These flux values are calculated from the model flux of individual spectral fitting of each observation.
To search for a possible energy dependence of the pulse shape change, we construct pulse profiles in five energy bands, i.e. 3$-$8, 8$-$13, 13$-$18, 18$-$25 and 25$-$60 keV. Two examples for the energy resolved pulse profiles, one for double-peaked and one for single-peaked, are given in Fig. \[enrespp\]. We find that 8$-$13 keV and 13$-$18 keV pulses are stronger (they have higher pulse fraction) than the 3$-$8 keV and 18$-$25 keV pulses in all observations. In 25$-$60 keV the pulse fraction noticeably drops. The pulse shape shows no strong energy dependence except for the two observations on MJD 55400.78 (ID:95044-05-02-00 see Fig. \[enrespp\]) and MJD 55404.63 (ID:95044-05-02-02). In these two exceptions the secondary peak around pulse phase 0.3 loses its intensity at the 18$-$25 keV energy band.
We construct hardness ratio plots from the energy resolved light curves. Daily averaged count rates from 3$-$8, 8$-$13, 13$-$18 and 18$-$25 keV light curves are used to plot hardness ratios HR1: 8$-$13 keV/3$-$8 keV and HR2: 18$-$25 keV/13$-$18 keV over time (see left panels of Fig. \[hard\]). HR1 shows a noticeable evolution with respect to time, it remains almost constant at a value $\sim1.1$ until MJD 55406.7 and a gradual decrease starts after the sixth observation. The interval with constant HR1 correspond to times when the pulse profile is double-peaked, where as the decreasing interval of HR1 coincides with the times when the pulse profile is single-peaked. Nevertheless, it should be noted that the pulse profiles show no significant variation in the corresponding energy bands (see Fig. \[enrespp\]). Furthermore, 530 s binned hardness ratios are plotted over the total count rate in 3$-$25 keV band (see right panels of Fig. \[hard\]). A noticeable correlation is again observed for HR1. As the source flux decreases during the ongoing decline of the outburst, the emitted radiation becomes softer.
Spectral Analysis
=================
Overall Spectrum
----------------
\
A preliminary spectral analysis for the first pointed observations of SWIFT J1729.9$-$3437 was performed by Markwardt et al. (2010b). In this paper, we extend the spectral study by using all the available observations of the source defined in Section 2. Basically, the spectra can be modelled by a power law with a high energy cut-off and photoelectric absorption as it is suggested by Markwardt et al. (2010b). Among the several models that describe the cut-off power law, the best fit is achieved by `CUTOFFPL` model in `XSPEC`. An additional Gaussian component is also required for a weak *Fe* emission line around 6.4 keV. During the simultaneous fitting of *Swift*$-$XRT and *RXTE*$-$PCA spectra we included a multiplicative constant factor in the model to account for the normalization uncertainty between the two instruments. The data and its best fit are plotted in Fig. \[bothspe\] and the corresponding spectral parameters are given in Table \[simspe\].
The energy ranges for the simultaneous spectral analysis are initially selected to be 0.3$-$9.3 keV for the XRT spectrum and 3$-$25 keV for the PCA spectrum. Individually these spectra have similar shapes that can be modelled with the same models, apart from an offset in absolute flux calibration. However individual modelling of the data from different instruments yields different absorption parameters due to different instrumental band-passes (Markwardt et al. 2010b). During the simultaneous fitting trials we observe large residuals for the first energy bins of the PCA spectrum although the fit is adequate for the XRT spectrum. Therefore we exclude energies below 5 keV for the PCA spectrum, since the XRT spectrum has more spectral energy bins in soft X-rays.
The FOV of PCA is large and SWIFT J1729.9$-$3437 is near the Galactic ridge. The thermal emission from the ridge is usually known to contaminate the spectral count rates of the source. During the analysis of the first PCA observation, the weak line at $\sim$6.6 keV reported by Markwardt et al. (2010b) was suggested to be a contamination, since it could not be resolved in the spectrum of the first XRT observation. We tried to handle this issue in overall PCA spectrum with fixed additive Galactic ridge models. We confirm that the flux of the source is two orders of magnitude bigger than the flux of the Galactic ridge (Valinia & Marshall 1998) therefore we conclude that the contamination could be small enough to be neglected. Furthermore; addition of the Gaussian model component for the *Fe* line improves the individual fit of the 0.3$-$9.3 keV XRT spectrum after combining data from all XRT observations by reducing the reduced $\chi^{2}$ from 1.08 ($\chi^{2}/$d.o.f. $ = 159.0 / 147$) to 0.98 ($\chi^{2}/$d.o.f. $ = 143.8 / 146$). The F-test probability that this improvement is achieved just by chance is $1.3 \times 10^{-4}$. Therefore we suggest that the *Fe* emission originates from SWIFT J1729.9$-$3437. The line energy found for the source is shifted from the neutral value of *Fe* K$\alpha$ ($\sim$6.4 keV), which may be a consequence of an excess of ionization.
In accretion powered pulsars, the *Fe* K$\alpha$ line is produced by the reprocessing of the hard X-ray emission in the relatively low ionized and cool matter surrounding the pulsar. The correlation of the iron line intensity and the continuum intensity in many sources (e.g. Vela X$-$1: Ohashi et al 1984, GX 301$-$2: Leahy et al. 1989) is specified to be the evidence for the fluorescence of X-ray continuum by the cold material. As the continuum intensity increases the illumination of the matter increases and the line intensity strengthens. We measure *Fe* line flux values from the individual *RXTE* observations of SWIFT J1729.9$-$3437 by using the `CFLUX` model in `XSPEC` and find that they correlate with the source flux values (see Fig. \[Fe\]). In Fig. \[Fe\] the dotted line represents the linear fits to the data, with a slope of $0.46\pm0.12$. The Pearson product-moment correlation coefficient between *Fe* line flux and unabsorbed source flux at 2$-$10 keV is 0.96. The null-hypothesis probability calculated from the Student’s t-distribution (two-tailed) is $1.2 \times 10^{-5}$. This correlation also confirms the origin of the *Fe* emission.
Pulse Phase Resolved Spectra
----------------------------
We construct pulse phase resolved spectra of SWIFT J1729.9$-$3437 from the first three RXTE$-$PCA observations. There are two motivations to the data selection. First, the brightest of the observations are selected to ensure the best signal-to-noise ratio possible. Second, the selected observations are the only ones that have 256 channels Good Xenon event mode data available, which is required for the `FASEBIN` tool. The timing solution found for the source in Section 3.1 is appended to the timing files of the `FASEBIN` tool for correct phase-binning. The pulse period is divided into five equal segments resulting in $\sim$2.1 ks exposure time for each phase bin spectrum.
We model 3$-$25 keV spectra with the same models used for the overall spectrum in the previous section. As we do not observe any significant change in the Hydrogen column density ($n_H$) and the *Fe* line energy during the preliminary fitting trials, we fixed these parameters during the analysis in order to better constrain the other parameters of the fits. The photon index and the folding energy of the high energy cut-off ($E_{fold}$) are found to be varying within pulse phases (see Fig. \[phasespe\]). The phase dependence of photon index is in anti-correlation with the pulsed flux. The Pearson product-moment correlation coefficient between pulsed flux and photon index values is $-0.87$ and the corresponding null-hypothesis probability (two-tailed) is 0.06. The null-hypothesis probability of this anti-correlation is quite significant but our results still indicate that the softest emission is observed at the un-pulsed phase around 0.05; from which we might suggest a possible physical relation between the parameters. The $E_{fold}$ values show a similar trend as the photon index values, which means softer spectra have higher cut-off energies. The Pearson product-moment correlation coefficient between $E_{fold}$ and photon index values is 0.96 and the corresponding null-hypothesis probability (two-tailed) is 0.01. Although the correlation analysis indicates a strong dependence between the parameters, one should note that the uncertainties in the parameters are not taken into account during this analysis. When the large uncertainties in $E_{fold}$ values are taken into account the data is consistent with a constant value of $\sim11.7$ keV. Therefore it is difficult to infer a clear variation of the $E_{fold}$ with the pulse phase, suggesting a rather marginal detection.
Summary and Discussion
======================
In this paper, we study timing and spectral properties of SWIFT J1729.9$-$3437 using *RXTE* and *Swift* observations of its outburst from 2010 July 20 through 2010 August 12 (MJD 55397 $-$ MJD 55420). From these observations with a time span of $\simeq 23$ days, we find that arrival times can be fit to a quadratic. From this fit, we calculate a spin frequency and spin frequency derivative of $1.8734(8) \times 10^{-3}$ Hz and $6.42(6) \times 10^{-12}$ Hz/s respectively. The residuals of the quadratic is further found to fit a circular orbital model with ${{a} \over {c}}\sin i=65(3)$ lt.s and an orbital period of 15.3(2) days. We also try an elliptical model and find an upper limit for eccentricity as 0.60. However, this model over-fits the data with a reduced $\chi^2$ of 0.40. Future observations might help to refine orbital parameters of the source.
Using this ${{a} \over {c}}\sin i$ value, we find the mass function (${{4\pi^2} \over {G}}{{(a\sin i)^3} \over {P_{orbital}^2}}$) to be about $1.3M_{\odot}$. An orbital period of 15.3 days and a spin period of 533.76 s puts the source in line with the accretion powered pulsars with supergiant companions, in the Corbet diagram (Drave et al. 2012, Corbet 1984). On the other hand, the small mass function obtained from the circular orbital model should be an indication of a small orbital inclination angle. If the circular orbital model is preferred as the model fitting the residuals of the quadratic fit, this indicates that we observe the binary system nearly edge-on.
Alternatively the residuals of the quadratic fit can be explained by a torque noise strength of $6.8 \times 10^{-18}$ Hz sec$^{-2}$. This value is quite consistent with other accreting X-ray binaries (Baykal & Ögelman 1993, Bildsten et al. 1997). Future observations are needed to understand the exact nature of the source.
Initially, a double-peaked pulse profile is observed in the light curves of the source. Then we find that one peak loses its intensity starting from the middle of the observation on MJD 55406.7 as the source flux continues to decrease after the burst. To study the energy dependence and temporal variability of the pulse profiles, we construct pulse profiles with five different energy bands shown in Fig. \[enrespp\]. We observe stronger pulses in the 8$-$13 keV and 13$-$18 keV energy bands but generally the pulse shape shows no strong energy dependence. Double to single peak transition seen in this source might be due to a sharp decline in the intensity of the radiation coming from the fan beam, since the formation of fan beam strongly depends on the luminosity of the source as for EXO 2030+375 (Parmar et al. 1989) and GX 1+4 (Paul et al. 1997). However lack of significant energy dependence of the pulse profiles makes the fan beam explanation implausible since fan beams are expected to be spectrally harder than pencil beams.
In order to have a basic understanding of spectral variability of the whole observations, we study hardness ratios of the source. From the hardness ratio plots (see Fig. \[hard\]), we suggest that the emitted radiation becomes softer as the source flux decreases during the ongoing decline of the outburst. The similar spectral softening with decreasing flux was reported before for 1A 0535+262, A 1118$-$616, SWIFT J1626.6$-$5156, XTE J0658$-$073 and GRO J1008$-$57 using hardness-intensity relations (Reig & Nespoli, 2013).
To extend the preliminary spectral analysis of Markwardt et al. (2010b), we also construct a single spectrum from all the available *RXTE* and *Swift* observations (see Fig. \[bothspe\] and Table \[simspe\]). We find that adding an *Fe* line complex feature with a peak at 6.51 keV slightly improves the spectral fit. We discuss that this *Fe* line feature is more likely originated from the source as the galactic ridge emission is too weak to explain this emission alone. We also measure *Fe* line flux values from the individual *RXTE* observations of the source and find that they correlate with the source flux values (see Fig. \[Fe\]). This correlation confirms the origin of the *Fe* emission as the source itself rather than the galactic ridge.
We perform pulse phase resolved spectral analysis of the source using the first three *RXTE* observations. From this analysis, we find a marginal evidence of a variation of the photon index and the folding energy of the high energy cut-off with the pulse phase (see Fig. \[phasespe\]).
Acknowledgements {#acknowledgements .unnumbered}
================
We acknowledge support from TÜBİTAK, the Scientific and Technological Research Council of Turkey through the research project TBAG 109T748.
[99]{}
Barthelmy S. D., Barbier L. M., Cummings J. R. et al., 2005, Space Sci. Rev., 120, 143 Baykal A., İnam S. Ç., Stark M. J., Heffner C. M., Erkoca A. E., Swank J. H., 2007, MNRAS, 374, 1108 Baykal, A.; Ögelman, H., 1993, A&A, 267, 119 Bildsten L. et al., 1997, ApJS, 113, 367 Burrows D. N., Hill J. E., Nousek J. A. et al., 2005, Space Sci. Rev., 120, 165 Corbet R. H. D., 1984, A&A, 141, 91 Cordes J. M., 1980, ApJ, 237, 216 Deeter J. E., Boynton P. E., Pravdo S. H., 1981, ApJ, 247, 1003 Deeter J. E., 1984, ApJ, 281, 482 Deeter J. E., Boynton P. E., 1985, in Hayakawa S., Nagase F., eds, Proc. Inuyama Workshop: Timing Studies of X-Ray Sources. Nagoya Univ., Nagoya, p. 29 Drave S.P., Bird A.J., Townsend L.J., Hill A.B., McBride V.A., Sguera V., Bazzano A., Clark D.J., 2012, A&A, 539, 21 Gehrels N., Chincarini G., Giommi P. et al., 2004, ApJ, 611, 1005 Jahoda K., Swank J. H., Giles A. B., Stark M. J., Strohmayer T., Zhang W., Morgan E. H., 1996, Proc. SPIE, 2808, 59 Leahy D. A., Darbro W., Elsner R. F., Weisskopf M. C., Kahn S., Sutherland P. G., Grindlay J. E., 1983, ApJ, 266, 160 Leahy D. A., Matsuoka M., Kawai N. and Makino F., 1989, MNRAS, 236, 603 Markwardt C. B., Krimm H. A., Swank J. H., 2010a, Atel 2747 Markwardt C. B., Krimm H. A., Swank J. H., 2010b, Atel 2749 Moretti A., Campana S., Mineo, T. et al., 2005, SPIE, 5898, 360 Ohashi T., Inoue H., Koyama K. et al., 1984, PASJ, 36, 699 Parmar A.N., White N.E., Stella L., 1989, ApJ, 338, 373 Paul B., Agrawal P.C., Rao A.R., Manchanda R.K., 1997, A&A, 319, 507 Reig P., Nespoli E., 2013, A&A, 551, A1 Valinia A., Marshall F. E., 1998, ApJ, 505, 134 Vaughan S., Goad M. R., Beardmore A. P. et al., 2006, ApJ, 638, 920
\[lastpage\]
[^1]: E-mail: [email protected] (ŞŞ); [email protected] (SÇİ); [email protected] (AB); [email protected] (MMS)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Motivated by recent interest in spin-triplet superconductors, we investigate the vortex lattice structures for this class of unconventional superconductors. We discuss how the order parameter symmetry can give rise to U(1)$\times$U(1) symmetry in the same sense as in spinor condensates, making half-quantum vortices (HQVs) topologically stable.We then calculate the vortex lattice structure of HQVs, with particular attention on the roles of the crystalline lattice, the Zeeman coupling and Meissner screening, all absent in spinor condensates. Finally, we consider how spin.orbit coupling leads to a breakdown of the U(1)$\times$U(1) symmetry in free energy and whether the HQV lattice survives this symmetry breaking. As examples, we examine simpler spin-triplet models proposed in the context of Na$_x$CoO$_2 \cdot y$H$_2$O and Bechgaard salts, as well as the better known and more complex model for Sr$_2$RuO$_4$.'
author:
- 'Suk Bum Chung$^1$, Daniel F. Agterberg$^2$, and Eun-Ah Kim$^3$'
title: Fractional vortex lattice structures in spin triplet superconductors
---
Introduction
============
A half quantum vortices (HQV) with vorticity $h/4e$, which is half that of usual Abrikosov vortex with vorticity $\Phi_0\equiv h/2e$, presents an exciting example of fractionalized topological defects. Quantization of collective topological defects provides clear cut access to the nature of the ground state. For instance, the vorticity $h/2e$ of Abrikosov vortex in a type II superconductor clearly shows that the circulation associated with the vortex is that of charge $2e$ Cooper pairs. Since the vorticity in a condensate is determined by the requirement of single-valuedness of the order parameter describing the condensate, “fractionalization of vorticity” is possible with a multi-component order parameter when different components are allowed to wind separately. For instance, in a triplet superconductor, the additional Cooper pair spin degree of freedom can be free to rotate in plane [@PhysRevLett.55.1184; @PhysRevLett.89.067001; @chung:197002] giving rise to additional U(1) symmetry and an associated spin winding number; cases where vortex fractionalization is due to the U(1)$\times$U(1) symmetry of different physical origin has also been studied [@PhysRevLett.94.137001]. Hence observation of fractionalization of vortices can serve as an indicator of the structure of the order parameter in a given condensate. In addition, the recent proposals predicting non-Abelian fractional statistics for the composite of a HQV and the Majorana fermions bound at its core in the chiral triplet superconductors brought in recent rise in the attention and interest to the possibility of HQV’s in triplet superconductors [@PhysRevB.61.10267; @PhysRevLett.86.268; @PhysRevB.70.205338; @PhysRevB.73.014505]. This type of non-Abelian statistics was first studied for quasiholes in the spin-polarized $\nu = 5/2$ quantum Hall state [@moore-read; @nayak-wilczek]. Therefore, if we want to obtain the same statistics for vortices in a spinful superconductor, the vortices should be HQVs so that there would be phase winding only for a single component.
However, while there are a number of candidate triplet superconductors such as the single layer ruthenate Sr$_2$RuO$4$ [@Maeno:1994fk; @RevModPhys.75.657], the cobaltate Na$_x$CoO$_2 \cdot y$H$_2$O [@Nature.Physics.1.91] and organic superconductors [@organic], HQV’s have never been observed in bulk systems, in line with the energetic stability issues raised by two of us in Ref. . It was pointed out in Ref. due to the absence of screening for the spin supercurrent circulation required for HQV in triplet superconductors, HQV’s can be energetically unstable in bulk samples towards combining into full Abrikosov vortices despite their advantage in magnetic energy. Related considerations have appeared in the context of spin-triplet superconductivity in UPt$_3$ by Zhitomirsky [@Zhitomirskii_UPt3].
The main motivation of this work is to investigate the possibility of using high enough fields to generate a HQV lattice in triplet superconductors where the vortex lattice serves two purposes at once: 1) stabilizing HQV’s at finite separation 2) providing an unambiguous signature of its formation (halving of the vortex lattice unit cell). Experiments have already determined the vortex lattice structure successfully at low fields in Sr$_2$RuO$_4$ [@Nature.396.242; @PhysRevLett.84.6094] and the observed square lattice geometry was consistent with the theoretical prediction by one of the present authors based on a chiral triplet order parameter in Ref. . However, Ref. considered the limit of strong spin-orbit coupling, which leads to vortex lattices of full quantum Abrikosov vortices. Recently, measurements of the Knight shift for the field along the $c$-axis [@PRL.93.167004] as well as ARPES data [@haverkort:026406; @liu:026408] indicate that that spin-orbit coupling is perhaps not so strong. Therefore, in this paper, we extend the studies of Ref. to allow for weak spin-orbit coupling, leading to the possibility of HQV lattices for fields along the $c$-axis. Additionally, the organic superconductor (TMTSF)$_2$ClO$_4$ naturally has weak spin-orbit coupling and Knight shift measurements provide evidence for a spin-triplet state at high magnetic fields [@organic], which is precisely the situation we consider here. We also provide an analysis of this case and the closely related case for cobaltate spin-triplet superconductors [@Nature.Physics.1.91].
In this paper, we study the energetics of different vortex lattice configurations. The key additional physical ingredient is the U(1) spin-rotational invariance of the Cooper pairs that arises in a magnetic field. This generically leads to two different species of fractional vortices whose fractional fluxes sum to $\Phi_0$. When stable, these fractional vortices form interlacing lattices analogous to vortex.antivortex lattice configurations proposed by Refs. @PhysRevLett.71.2138 [@PhysRevLett.71.2142] in the context of a two-dimensional (2D) superfluid and the configuration in two-component Bose condensates proposed by Ref. @PhysRevLett.88.180403.
The rest of the paper consists of the following. In section \[sec:OP\], we give a pedagogical introduction to the symmetry properties of triplet OP. In particular, we will show how U(1)$\times$U(1) symmetry can arise in the OP of such systems. In section \[sec:GibbsFree\] we discuss the form of Gibbs free energy that is allowed by various symmetries in the problem when spin-orbit coupling is not included. In section \[sec:G\] we provide the general theoretical framework for the vortex lattice phases. In section \[sec:LLL\] we show that the lowest Landau level solution often provides an adequate description and we discuss this solution for the lattice of HQV’s. In section \[sec:SO\] we consider the effect of U(1)$\times$U(1) symmetry breaking driven by spin-orbit coupling. In section \[sec:observation\] we lay out predictions for how to detect the proposed HQV lattice structures and we conclude with a summary and outlook in section \[sec:conclusion\].
The triplet order parameter {#sec:OP}
===========================
The order parameter of a triplet superconductor takes a matrix form in the spin space[@RevModPhys.75.657; @RevModPhys.63.239]: $$\hat{\Delta} ({\bf k}) \!=\! \left [
\begin{array}{cc}
\Delta_{\uparrow\uparrow}({\bf k}) & \Delta_{\uparrow\downarrow}({\bf k}) \\
\Delta_{\downarrow\uparrow}({\bf k}) & \Delta_{\downarrow\downarrow}({\bf k})\\
\end{array}
\right ] \!\equiv\! \left[
\begin{array}{cc}
-d_x + id_y & d_z \\
d_z & d_x + id_y \\
\end{array}
\right ], \label{EQ:dVec}$$ where the spin quantization axis is along the $z$ direction. The triplet pairing requires $\Delta_{\uparrow\downarrow}\!=\!\Delta_{\downarrow\uparrow}$ and a set of three complex functions of ${\bf k}$, namely $(d_x({\bf k}),d_y({\bf k}), d_z({\bf k}))$, were introduced to parameterize the gap matrix. When the three functions are collectively represented using a vector notation, the “unit vector” ${\bf\hat d}({\bf k})$ represents the symmetry direction (zero projection direction) with respect to the rotation of Cooper pair spin. In the presence of the sufficiently high field along the $c$-axis, Zeeman splitting between electrons with opposite spins prohibits pairing, leading to $\Delta_{\uparrow\downarrow}= \Delta_{\downarrow\uparrow}=0$. In the $\bf{d}$-vector notation, this implies that the $\bf{d}$-vector lies in-plane (perpendicular to the applied field). In the rest of the paper, we assume that the field is sufficiently large so that this is the case. In the context of strontium ruthenate, our results apply for the field along the $c$-axis (this is also true for the cobaltates when we are discussing spin-orbit coupling in hexagonal systems). For organic and cobaltate superconductors, our results apply for the field along any two-fold or higher symmetry axis of symmetry.
For non-chiral triplet order parameter symmetry, which is expected of the cobaltate Na$_x$CoO$_2 \cdot y$H$_2$O [@Nature.Physics.1.91] and organic superconductors [@organic], the spin pairing gap matrix takes the form $$\hat{\Delta}({\bf k})=f({\bf k})\left[
\begin{array}{cc}
\Delta_{\uparrow\uparrow}& 0 \\
0 & \Delta_{\downarrow\downarrow}
\end{array}
\right] \label{EQ:OP-nonchiral}$$ where the function $f({\bf k})$ depends on the specific odd angular momentum channel. The key simplifying feature is the the orbital dependence is described by a one-dimensional representation encoded by $f({\bf k})$. There has been suggestions that the cobaltate Na$_x$CoO$_2 \cdot y$H$_2$O has a spin-triplet pairing through $f$-wave channel [@Nature.Physics.1.91], although data from the Knight shift experiments remain controversial [@PhysRevB.73.180503; @JPSJ.77.073702]. In this case a common choice is $f({\bf k}) = k_x(k_x^2-3k_y^2)$. However, the precise form of $f({\bf k})$ is not needed for our results. For the organic superconductor (TMTSF)$_2$ClO$_4$, there is a strong case that the system becomes a triplet superconductor under sufficient field $H\gtrsim $20kOe. In this case, there are many proposals for $f({\bf k})$. However, again, the specific form is not needed for our results.
For chiral order parameter symmetry expected of the Sr$_2$RuO$_4$, with $\bf{d}$ in the basal plane and no spin-orbit coupling, the order parameter has four complex degrees of freedom: $$\hat{\Delta}({\bf k})=
\sum_{\sigma=\pm}\tilde{f}(k_\sigma,k_z)\left[
\begin{array}{cc}
\Delta_{\uparrow\uparrow,\sigma}& 0 \\
0 & \Delta_{\downarrow\downarrow,\sigma}
\end{array}
\right], \label{EQ:OP-chiral}$$ where the function $\tilde{f}$, like the non-chiral case discussed above, depends on the specific odd angular momentum channel and $k_\sigma = k_x + i\sigma k_y$. This is equivalent to $${\bf d} ({\bf k}) = \Delta_+ {\bf {\hat d}}_+ \exp (i
n\varphi_{\bf{\hat k}})+\Delta_- {\bf {\hat d}}_- \exp (-i
n\varphi_{\bf{\hat k}}), \label{EQ:OPp+ip}$$ where $n$ is an integer ($n=1$ for $p$-wave, $n=3$ for $f$-wave) and $\varphi_{\bf{\hat k}}$ is the azimuthal angle associated with a unit vector $\hat{\bf k}$ in the 2D plane (assuming quasi-2D setting with the angular momentum along the $c$-axis: ${\bf \hat
l} = {\bf \hat z}$). Although $\Delta_- = 0$ for homogeneous chiral superconductor, we will show that $\Delta_- \neq 0$ often plays an important role in describing the vortex lattice structure of chiral superconductor.
Eqs.(\[EQ:OP-nonchiral\]) and (\[EQ:OPp+ip\]) clearly shows that under these circumstances, the order parameter symmetry takes the $U(1)\times U(1)$ form, which can allow for HQV’s with $h/4e$ vorticity associated with $\pi$ orbital phase winding and $\pi$ ${\bf d}$-vector winding.
The Gibbs free energy {#sec:GibbsFree}
=====================
In order to identify stable vortex type and the lattice structure itself, we start with the Gibbs free energy including all the terms allowed by symmetry up to quartic order. As usual, the quartic terms determine the vortex lattice structure. Due to additional spin degrees of freedom, the full expression for the Gibbs free energy involves a number of additional terms compared to singlet superconductor case and it is instructive to consider different contributions separately: $$f = f_{mag} + f^{(2)}_0 + f^{(2)}_{Z} + f^{(2)}_{in} +
f^{(2)}_{SO} + f^{(4)}_{hom} + f^{(4)}_{in}. \label{EQ:freeTot}$$ where $f_{mag} = h^2/8\pi - hH/4\pi$ is the magnetic energy (the field $h$ is the sum of the external field $H$ and the screening field), the superscript (2) indicates terms quadratic in OP and (4) quartic in OP
Other than the conventional quadratic term $f^{(2)}_0$: $$f^{(2)}_0 = -\alpha \sum_i |\Delta_i|^2 \label{EQ:f0}$$ the remaining quadratic terms in Eq. are consequences of additional spin degree of freedom for the triplet superconductors. The Zeeman coupling term $$f^{(2)}_{Z} = -\tilde{\kappa} h (|\Delta_{\uparrow\!\uparrow}|^2 -
|\Delta_{\downarrow\!\downarrow}|^2),$$ plays an important role for the HQV vortex lattice by introducing a slight spin-polarization. This slight spin-polarization gives rise to two phase transitions as in the case of the A$_1$/A$_2$ phase of $^3$He[@RevModPhys.47.331; @VollhardtHe3]. The inhomogeneous part of the quadratic terms $f^{(2)}_{in}$ are of the form $$K_{ij;kl} (D_i \Delta_k) (D_j \Delta_l)^* + {\rm c.c.},$$ where $D_i = \nabla_i + (2\pi i/\Phi_0)A_i$. For these terms, we require rotational invariance up to the lattice symmetry with respect to orbital degrees of freedom only, which means that we require invariance with respect to rotating $D_i$’s and the orbital component of $\Delta_i$’s. The complex structure of $f^{(2)}_{in}$ can result in the condensate wave function of a different form than that of conventional SC. $f^{(2)}_{SO}$ is the quadratic spin-orbit coupling term assuming the spin-orbit coupling to be small and is discussed in Sec. \[sec:SO\]. In the presence of spin-orbit coupling the free energy have to be invariant under the combined discrete rotation of the orbital and spin degree of freedom specific for the given lattice symmetry. For lattices with orthogonal or tetragonal symmetry, spin-orbit coupling may reduce the symmetry of the Gibbs free energy and tend to suppress HQV formation by introducing a length scale beyond which the HQV’s cannot exist (this length scale diverges as the spin-orbit coupling vanishes). This implies that the vortex lattice spacing must be less than this length scale for the HQV lattice to appear. However, we show that for spin-triplet hexagonal materials (specifically the two-dimensional $\Gamma^-_6$ and $\Gamma^-_5$ representations in the notation of Sigrist and Ueda [@RevModPhys.63.239]), even large spin-orbit coupling still allows for the existence of a fractional vortex lattice. This consideration may apply to Na$_x$CoO$_2\cdot$yH$_2$O.
Among the quartic terms, $f^{(4)}_{hom}$ represents the usual set of homogeneous terms. As is shown in Appendix \[app:weakcoupling\], certain quartic terms vanish in the weak-coupling theory. The terms that vanish are those that lift the energy degeneracy between the full quantum vortex (QV) and the HQV lattice. For this reason, we also include the inhomogeneous quartic term $f^{(4)}_{in}$. This term accounts for the difference between the spin phase stiffness $\rho_{sp}$ and the overall phase stiffness $\rho_s$. Not only does this difference play an important role in the stability of isolated HQV’s as it was shown in Ref. it plays the role of tuning parameter for the vortex lattice structure.
With multiple systems in mind, we consider contributions to the Gibbs free energy specific for non-chiral and chiral superconductors respectively.
Non-chiral Triplet Superconductor
---------------------------------
As mentioned earlier, here we assume the orbital dependence of the gap function to be the same for all spin-triplet components. Formally, this means that the orbital degree of freedom belongs to a one-dimensional irreducible representation of the point group. We apply our analysis to materials that have orthorhombic, tetragonal, or hexagonal point groups. One relevant example is a non-chiral triplet $f$-wave superconductor with hexagonal symmetry, which has been proposed in the context of the the cobaltates Na$_x$CoO$_2\cdot y$H$_2$O; in this case $f({\bf k}) = k_x(k_x^2-3k_y^2)$ from Eq.. When the ${\bf d}$-vector lies in the xy plane, the inclusion of spin-orbit coupling implies that formally this order parameter belongs to the $\Gamma_6^-$ representation of the hexagonal point group (the consequences of spin-orbit coupling for this representation is discussed in more detail in Section \[sec:SO\]).
With the in-plane spin rotational invariance, the relevant free energy within the assumptions stated above is given by $$\begin{aligned}
&f^{(2)}_{in} = \sum_{i=x,y,z} K_i(|D_i\Delta_{\uparrow\uparrow}|^2+
|D_i\Delta_{\downarrow\!\downarrow}|^2),\label{EQ:f2in-NX}\\
&f^{(4)}_{hom}=\beta_1(\sum_{i}|\Delta_{i}|^2)^2+\beta_2|\Delta_{\uparrow\!\uparrow}|^2
|\Delta_{\downarrow\!\downarrow}|^2,\label{EQ:f2hom-NX}\\
&f^{(4)}_{in} =
\gamma[\Delta_{\uparrow\!\uparrow}^*\Delta_{\downarrow\!\downarrow}(\textbf{D}_{\perp}\Delta_{\uparrow\!\uparrow})\!\cdot\!(\textbf{D}_{\perp}\Delta_{\downarrow\!\downarrow})^*+{\rm c.c}],\label{EQ:f4-NX}\end{aligned}$$ For tetragonal and hexagonal point groups $K_x=K_y$ while for orthorhombic point groups, $K_x\ne K_y$. For the high field limit we are considering, it is possible to re-scale lengths in two directions perpendicular to applied field such that $\tilde{K}_i=\tilde{K}_j$ for $i\ne j$ (where $\tilde{K}_i$ refers to the new coefficient in the re-scaled coordinates). We will therefore ignore the difference between the $K_i$ and assume that for orthorhombic point groups we are working in re-scaled co-ordinates. The term $f_{in}^{(4)}$ is not the most general such term allowed by symmetry. However, it is this term that allows the GL theory to give the same physics as in Ref. . Indeed, one can gain more insight into the vortex lattice solutions that minimizes Eq. and Eqs.(\[EQ:f2hom-NX\]-\[EQ:f4-NX\]) by relating the coefficient of the inhomogeneous quartic term $\gamma$ to the stiffness ratio $\rho_{sp}<\rho_s$ which controls the energetic stability of a pair of HQV’s[@chung:197002]. Within the London approximation, the gradient terms in Eq. and Eq. amounts to phase bending energy which will be proportional to $(\rho_{s}+\rho_{sp})$ and $(\rho_{s}-\rho_{sp})$ respectively for $\Delta_{\uparrow\!\uparrow}$ and $|\Delta_{\downarrow\!\downarrow}|$. Combining Eq. and Eq. with the homogeneous solution $|\Delta_{\uparrow\!\uparrow}|^2=|\Delta_{\downarrow\!\downarrow}|^2 = \alpha/(\beta_1-\beta_2)$, we obtain the following relation between $\gamma$ and $\rho_{sp}/\rho_s$: $$\gamma = \frac{K_1(\beta_1-\beta_2)}{\alpha}
\frac{1-\rho_{sp}/\rho_s}{1+\rho_{sp}/\rho_s}.$$ Hence $\gamma >0$ would imply stability of HQV’s and double transitions into two possible vortex phases: a lattice of ordinary Abrikosov vortices and a lattice of HQV’s. This transition is determined by the $\beta_2$ term of Eq. (\[EQ:f2hom-NX\]) and the $\gamma$ term of Eq. (\[EQ:f4-NX\]).
Chiral Triplet Superconductor
-----------------------------
With the ruthenate Sr$_2$RuO$_4$ in mind, we consider a chiral triplet $p$-wave superconductor with square symmetry for which $\tilde{f}(k_\sigma) = k_x + i\sigma k_y$ in the Eq. with $$\hat{\Delta}({\bf k})= \sum_{\sigma=\pm}(k_x + i\sigma k_y)\left[
\begin{array}{cc}
\Delta_{\uparrow\uparrow,\sigma}& 0 \\
0 & \Delta_{\downarrow\downarrow,\sigma}
\end{array}
\right], \label{EQ:spinChiral1}$$ where $\Delta_{s,\sigma}$ ($s =\uparrow\!\uparrow,\downarrow\!\downarrow$ and $\sigma=\pm$) form expansion parameters for the Gibbs free energy. In terms of the $\bf{d}$-vector notation ${\bf d} \equiv {\bf \hat{x}}(\eta_{xx} k_x + \eta_{xy} k_y) + {\bf \hat{y}}(\eta_{yx} k_x + \eta_{yy} k_y)$, $$\begin{aligned}
\Delta_{\uparrow\!\uparrow,+} = &-& (\eta_{xx} - i\eta_{xy} - i\eta_{yx} - \eta_{yy})/2,\nonumber\\
\Delta_{\uparrow\!\uparrow,-} = &-& (\eta_{xx} + i\eta_{xy} - i\eta_{yx} + \eta_{yy})/2,\nonumber\\
\Delta_{\downarrow\!\downarrow,+} = &\,& (\eta_{xx} - i\eta_{xy} + i\eta_{yx} + \eta_{yy})/2,\nonumber\\
\Delta_{\downarrow\!\downarrow,-} = &\,& (\eta_{xx} + i\eta_{xy} +
i\eta_{yx} - \eta_{yy})/2. \label{EQ:spinChiral2}\end{aligned}$$
Formally, without spin-orbit coupling, this order parameter is a direct product of a $E_u$ orbital representation of the tetragonal point group and the in-plane vector representation for spin rotations. When spin-orbit is included the order parameter contains the four different one-dimensional representations of the tetragonal point group. In the case without spin-orbit coupling, the relevant free energy for this representation can be constructed using the known free energy for the $E_u$ representation [@RevModPhys.63.239], we list below $f^{(2)}_{in}$, $f^{(4)}_{hom}$ and $f^{(4)}_{in}$. Before listing $f^{(2)}_{in}$, we note that this free energy term should respect the $C_4$ symmetry on the $xy$ plane only for the orbital degrees of freedom: $$(D_x, D_y, \Delta_{s,+}, \Delta_{s,-}) \to (D_y, -D_x,
i\Delta_{s,+}, -i\Delta_{s,-}).$$ However, for simplicity, we consider cylindrical symmetry in the orbital degrees of freedom, this does not significantly alter the arguments below. This symmetry gives us $f^{(2)}_{in} = \sum_s f^{(2,s)}_{in}$ where $$\begin{aligned}
f^{(2,s)}_{in} &=& K_1(|{\bf D}\Delta_{s,+}|^2 + |{\bf D}\Delta_{s,-}|^2)\nonumber\\
&+& K_2[\!\{\!(D_x \Delta_{s,+})\!(D_x \Delta_{s,-})^*\!-\!(D_y \Delta_{s,+})\!(D_y \Delta_{s,-})^*\!\}\!/2\nonumber\\
&+& \!\{\!(D_x \Delta_{s,-})\!(D_x \Delta_{s,+})^*\!-\!(D_y \Delta_{s,-})\!(D_y \Delta_{s,+})^*\!\}\!/2\nonumber\\
&+& i\!\{\!(D_x \Delta_{s,-})\!(D_y \Delta_{s,+})^*\!+\!(D_y \Delta_{s,-})\!(D_x \Delta_{s,+})^*\!\}\!/2\nonumber\\
&-& i\!\{\!(D_x \Delta_{s,+})\!(D_y \Delta_{s,-})^*\!+\!(D_y \Delta_{s,+})\!(D_x \Delta_{s,-})^*\!\}\!/2]\nonumber\\
&+& K_4 (|D_z\Delta_{s,+}|^2 + |D_z\Delta_{s,-}|^2).
\label{EQ:quadDeriv2}\end{aligned}$$ In addition, the following term is also allowed by symmetry $$\delta K \frac{2\pi}{\Phi_0}h \sum_s (-|\Delta_{s+}|^2 + |\Delta_{s-}|^2),
\label{EQ:chiralSplit}$$ and it stabilizes this in-plane chiral phase for strong enough magnetic field (note the similarity to the Zeeman term for the condensate spin degrees of freedom). As for the homogeneous quartic terms, $$\begin{aligned}
f^{(4)}_{hom} &=& \sum_s [\beta_1(|\Delta_{s+}|^4+|\Delta_{s-}|^4)/2 + \beta'_1|\Delta_{s+}|^2 |\Delta_{s-}|^2]\nonumber\\
&-&\sum_{\sigma=\pm}
(\beta_2|\Delta_{\uparrow\!\uparrow,\sigma}|^2
|\Delta_{\downarrow\!\downarrow,\sigma}|^2 + \beta'_2
|\Delta_{\uparrow\!\uparrow,\sigma}|^2
|\Delta_{\downarrow\downarrow,-\sigma}|^2)\nonumber\\
&-&\beta_3[(\Delta_{\uparrow\!\uparrow,+}\Delta_{\downarrow\!\downarrow,-})
(\Delta_{\uparrow\!\uparrow,-}\Delta_{\downarrow\!\downarrow,+})^*
+ {\rm c.c.}]. \label{EQ:int1}\end{aligned}$$ $\beta_2$, $\beta'_2$ and $\beta_3$ terms originate from interaction between spin up-up pairs and down-down pairs. Again, for simplicity, we written the free energy in the limit of a cylindrical Fermi surface. Lastly, we have $$\begin{aligned}
f^{(4)}_{in} &=& \gamma \sum_{\sigma=\pm} [\Delta^*_{\uparrow\uparrow,\sigma} \Delta_{\downarrow\downarrow,\sigma} ({\bm D}\Delta_{\uparrow\uparrow,\sigma})\cdot ({\bm D} \Delta^*_{\downarrow\downarrow,\sigma}) + {\rm c.c.}]\nonumber\\
&+& \gamma' \sum_{\sigma=\pm} [\Delta^*_{\uparrow\uparrow,\sigma} \Delta_{\downarrow\downarrow,-\sigma} ({\bm D}\Delta_{\uparrow\uparrow,\sigma})\cdot ({\bm D} \Delta^*_{\downarrow\downarrow,-\sigma}) + {\rm c.c.}].\nonumber\\
\label{EQ:int2}\end{aligned}$$ Note that the form of Eq. (\[EQ:int2\]) is consistent with the form of the interaction terms in Eq. (\[EQ:int1\]). Again, this is not the most general term allowed by symmetry, but it is the minimal term that captures the physics in the London limit described in Ref. .
Determining the vortex lattice structure - general considerations {#sec:G}
=================================================================
We consider the vortex lattice phases near the upper critical field to map out the stability condition for HQV lattice phases. As usual, the first step towards determining the vortex lattice structure is to identify the eigenstates of the linearized Ginzburg-Landau (GL) equations. In order to obtain a linearized GL equation we take a variation of the quadratic terms in the free energy, for example: $$\begin{aligned}
f^{(2)}_0 + f^{(2)}_{in} + f^{(2)}_Z &=& -\alpha \sum_j
|\Delta_j|^2 - \tilde{\kappa} h (|\Delta_{\uparrow\!\uparrow}|^2 -
|\Delta_{\downarrow\!\downarrow}|^2)\nonumber\\&+& [K_{jk;lm} (D_j \Delta_l) (D_k \Delta_m)^* + {\rm c.c.}]\nonumber\\\end{aligned}$$ with respect to one component of the order parameter $\Delta_{s}^*$. This gives an equation of the form $$\alpha \Delta_s = K_{kl;ss'}^* D_k D_l \Delta_s' -
\tilde{\kappa}H\Delta_s \label{EQ:genLinGL}$$ (note that we are ignoring the difference between $h$ and $H$ in this approximation). Since the gradient terms in Eq. cannot in general be reduced into a $D_x^2 + D_y^2$ form, the lowest Landau level wave functions are not sufficient for calculating the condensate wave function in general. However, the solution of this equation can still be expressed in terms of Landau level wave functions: $$\begin{aligned}
\phi_n ({\bf r}) &=& [2^n \pi^{1/2} (n!)]^{-1/2}\nonumber\\ &\times& \sum_m q_m e^{i k_m x'} e^{-(y'-k_m)^2/2} H_n (y'-k_m),\nonumber\\\end{aligned}$$ where $H_n$ is the Hermite polynomial of n$^{th}$ order, $x'$ and $y'$ are $x$, $y$ coordinates in the unit of the magnetic length $l = (\Phi_0/2\pi H)^{1/2}$. This is because $D_i$’s can be expressed as a linear combination of the raising and lowering operators of the Landau levels $\Pi_\pm$, since $\Pi_\pm = l(D_x \pm
iD_y)/\sqrt{2}$.
This wave function describes a vortex lattice when $|\phi_n ({\bf r})|$ is periodic in the lattice vectors ${\bf a_1} = al (1,0)$ and ${\bf a_2} = bl (\cos \theta,\sin \theta)$, and $\phi_n ({\bf r})$ vanish at $m_1 {\bf a_1} + m_2 {\bf a_1}$ when $m_1$ and $m_2$ are integers. This requires $$\begin{aligned}
k_m &=& 2\pi (m-1/2)/a = (m-1/2)\sqrt{2\pi \sigma}\nonumber\\
q_m &=& e^{i\pi m (\varsigma + 1 - m\varsigma)},\end{aligned}$$ where $\sigma = (b/a)\sin \theta$ and $\varsigma = (b/a)\cos\theta$. Note that we used the flux quantization condition $ab \sin\theta = 2\pi$. For a lattice of HQV’s we need to consider a second lattice that is translated by $l{\bm \tau} = l (\tau_x, \tau_y)$ with respect to the first lattice. For the wave function of this lattice, we can use [@superRev] $$\tilde{\phi}_n ({\bf r}) = e^{i \tau_y x} \phi_n ({\bf r} - {\bm
\tau}),$$ the phase factor being chosen so that $\Pi_- \tilde{\phi}_0 ({\bf r}) = 0$. This latter condition ensures that the translated eigenstates have the same gauge as the untranslated eigenstates.
The formalism considered here gives us not only the energy due to interaction between vortices but also the core energy of vortices as well. This is because our vortex lattice wave function gives full description of the core regions. From the linearized GL equation we use here, a full quantum vortex is merely two HQV’s of opposite spins coinciding at a same point. This means that the full quantum vortex core energy, if we ignore the cross term between two spin components in $f^{(4)}_{hom}$, is approximately twice the core energy of a HQV. If the HQV core energy is actually larger than this, that would make stabilization of the HQV more difficult, i.e., the largest allowed value for $\rho_{sp}/\rho_s$ for the HQV lattice would be smaller than what we obtain through the formalism used here.
The lattice structure can be determined by finding $(\sigma,\varsigma, {\bm \tau})$ that minimize the free energy expectation value. For this, we first set the amplitude of OP to minimize the energy for given $(\sigma,\varsigma, {\bm \tau})$ (the amplitude depends on $H_{c2}-H$), and then compare energy for different values of $(\sigma,\varsigma, {\bm
\tau})$. To determine these structures we will need to take the spatial integral of the product of four Landau level wavefunctions. We have computed these integrals in the Appendix \[app:corr\].
Lowest Landau level solution {#sec:LLL}
============================
In the bulk of this section we provide detailed analysis of the lowest Landau level solution for the non-chiral triplet superconductors and briefly comment on the chiral case in subsection \[subsec:chiral\]. In this case the relevant free energy (for orthorhombic, tetragonal, and hexagonal materials) is $$\begin{aligned}
f &=& \sum_{s=\uparrow\!\uparrow, \downarrow\!\downarrow}
\left[-\alpha|\Delta_s|^2 + \beta_1 |\Delta_s|^4/2 + \left(\sum_{i=x,y,z} K_i|D_i\Delta_s|^2\right)\right]\nonumber\\
&-& \beta_2\!|\Delta_{\uparrow\!\uparrow}|^2 |\Delta_{\downarrow\!\downarrow}|^2
- \tilde{\kappa} h (|\Delta_{\uparrow\!\uparrow}|^2 - |\Delta_{\downarrow\!\downarrow}|^2)\nonumber\\
&+& \gamma[\Delta_{\uparrow\!\uparrow}^*\Delta_{\downarrow\!\downarrow}
(\textbf{D}\Delta_{\uparrow\!\uparrow})\!\cdot\!(\textbf{D}\Delta_{\uparrow\!\uparrow})^*+{\rm c.c}]
+ \frac{h^2}{8\pi} - \frac{H h}{4\pi}.\nonumber\\
\label{EQ:freeLLL1}\end{aligned}$$ As mentioned before, we assume that we have re-scaled lengths so that we can take $K_z=K_x=K_y=K$. First look at the upper critical field problem. The linearized GL equation is: $$\frac{\alpha l^2}{K} \Delta_{s}= \left(1+2\Pi_+\Pi_- -s\frac{Hl^2}{K}\tilde{\kappa}\right) \Delta_{s}.$$ The largest $H_{c2}$ occurs when the $\Delta_s$ are in the lowest Landau level. This leads to two possible values for the upper critical field, $$H_{c2}^{\pm}=\frac{\alpha \Phi_0}{2\pi(K\pm Hl^2\tilde{\kappa})}$$ We assume that $\tilde{\kappa}>0$ so that $\Delta_{\uparrow\!\uparrow}$ has the larger $H_{c2}$. We do not assume that the splitting between these two critical fields is large since it is given by the small Zeeman term $\tilde{\kappa}$.
Now consider the expectation value of $f$ in terms of the Landau level wave functions. Applying Eq. (\[EQ:genLinGL\]) to the first two terms of Eq. (\[EQ:freeLLL1\]) gives the lowest Landau level solutions $\Delta_{\uparrow\!\uparrow} = C_{\uparrow\!\uparrow} \phi_0$ and $\Delta_{\downarrow\!\downarrow} = C_{\downarrow\!\downarrow} e^{i 2\alpha} \tilde{\phi}_0$. Inserting this solution as determined at $H=H_{c2}$ gives $$\begin{aligned}
\langle f \rangle &=& -\frac{1}{c}\langle\textbf{j}\cdot\delta\textbf{A}\rangle + \beta_1 \langle |\phi_0|^4 \rangle (C^4_{\uparrow\!\uparrow} + C^4_{\downarrow\!\downarrow})/2\nonumber\\
&+& \left[\frac{2\gamma}{l^2}\!(\!\langle\!|\phi_0|^2\!|\tilde{\phi}_0|^2\!\rangle\!-\!\langle\!|\phi_0|^2\!|\tilde{\phi}_1|^2\!\rangle\!)
- \beta_2\!\langle\!|\phi_0|^2\!|\tilde{\phi}_0|^2\!\rangle\right]C^2_{\uparrow\!\uparrow} C^2_{\downarrow\!\downarrow}\nonumber\\
&+& \frac{\langle h^2_s \rangle}{8\pi} -
\frac{H^2}{8\pi}+\alpha\left[1-\frac{K+H_{c2}l^2\tilde{\kappa}}{K-H_{c2}l^2\tilde{\kappa}}\right]C_{\downarrow\!\downarrow}^2,
\label{EQ:LLLfree}\end{aligned}$$ where $\bf{j}$ is the supercurrent, $\delta\bf{A}$ is deviation of the vector potential from what we would have for the $h = H_{c2}$, and $h_s$ is the screening field of superconductor. Note that in this approximation $\bf{j}$ is calculated solely from quadratic terms, ignoring $\gamma$ terms, and by the Maxwell equation $\nabla \times \bm{h_s} = 4\pi \bm{j} /c$. More specifically $$\begin{aligned}
\bm{j}=&2eK[\Delta^*_{\uparrow\!\uparrow}(\bm{D}\Delta_{\uparrow\!\uparrow})+
\Delta^*_{\downarrow\!\downarrow}(\bm{D}\Delta_{\downarrow\!\downarrow})+c.c]\nonumber\\
&-c\tilde{\kappa}\nabla\times\hat{z}(|\Delta_{\uparrow\!\uparrow}|^2-|\Delta_{\downarrow\!\downarrow}|^2)\end{aligned}$$ Since $\nabla \times \delta \bm{A} = \bm{\hat{z}} (h_s + H - H_{c2})$ and, from Maxwell’s equations, $\nabla \times \bm{h_s} = 4\pi \bm{j} /c$, partial integration leads to [@deGenneSC] $$\begin{aligned}
\frac{1}{c}\langle\textbf{j}\cdot\delta\textbf{A}\rangle &=& \frac{1}{4\pi}
\langle \bm{h_s}\cdot (\bm{h_s} + \bm{H} - \bm{\hat{z}} H_{c2}) \rangle\nonumber\\
&=& \frac{\langle h^2_s \rangle}{4\pi} + \frac{H_{c2} - H}{4\pi}\langle h_s \rangle.\end{aligned}$$ Meanwhile, when we calculate the expectation value of $\gamma$, we set $h = H_{c2}$. This leads to the free energy of $$\begin{aligned}
\langle f \rangle &=& -\frac{H_{c2}\!-\!H}{4\pi}\langle h_s\rangle - \frac{\langle h^2_s \rangle}{8\pi} - \frac{H^2}{8\pi}\nonumber\\
&+& \alpha\left[1-\frac{K+H_{c2}l^2\tilde{\kappa}}{K-H_{c2}l^2\tilde{\kappa}}\right]C_{\downarrow\!\downarrow}^2
+ \frac{\beta}{2} \langle |\phi_0|^4 \rangle (C^4_{\uparrow\!\uparrow} + C^4_{\downarrow\!\downarrow})\nonumber\\
&+& \left[\frac{2\gamma}{l^2}\!(\!\langle\!|\phi_0|^2\!|\tilde{\phi}_0|^2\!\rangle\!-\!\langle\!|\phi_0|^2\!|\tilde{\phi}_1|^2\!\rangle\!)
- \beta_2\!\langle\!|\phi_0|^2\!|\tilde{\phi}_0|^2\!\rangle\right]C^2_{\uparrow\!\uparrow} C^2_{\downarrow\!\downarrow}\nonumber\\
&\equiv& - \frac{H^2}{8\pi} +\langle \tilde{f} \rangle.
\label{EQ:freeExpect}\end{aligned}$$
The screening field $h_s$ can be calculated by assuming $|\hat{\Delta}| \propto (H_{c2} - H)^{1/2}$ near the second order phase transition at $H = H_{c2}$. Here we will deal only with $O(|\hat{\Delta}|^4)$ (or equivalently, $O(1-H/H_{c2})^2$) and ignore higher order terms. This allows us to calculate $\bf{j}$, and consequently $h_s$, solely from quadratic terms. In the lowest Landau level, this yields the screening field: $$\begin{aligned}
h_s &=& \frac{8\pi^2 K}{\Phi_0} (|\Delta_{\uparrow\!\uparrow}|^2 + |\Delta_{\downarrow\!\downarrow}|^2)
-4\pi \tilde{\kappa}(|\Delta_{\uparrow\!\uparrow}|^2 -|\Delta_{\downarrow\!\downarrow}|^2)\nonumber\\
&=&\!\left(\!\frac{8\pi^2 K}{\Phi_0}\!-\!4\pi\tilde{\kappa}\!\right)\!C^2_{\uparrow\!\uparrow}\!|\phi_0|^2\!
+\!\left(\!\frac{8\pi^2 K}{\Phi_0}\!+\!4\pi\tilde{\kappa}\!\right)\!C^2_{\downarrow\!\downarrow}\!|\tilde{\phi}_0|^2.\nonumber\\
\label{EQ:screenField}\end{aligned}$$
Inserting Eq. (\[EQ:screenField\]) into Eq. (\[EQ:freeExpect\]), the free energy takes the following form: $$\langle \tilde{f} \rangle = -\tilde{\alpha_1} C_{\uparrow\!\uparrow}^2 -\tilde{\alpha_2} C_{\downarrow\!\downarrow}^2
+\tilde{\beta_1} C_{\uparrow\!\uparrow}^4 +\tilde{\beta_2} C_{\downarrow\!\downarrow}^4 + \tilde{\beta_3} C_{\uparrow\!\uparrow}^2 C_{\downarrow\!\downarrow}^2
\label{eq:simplef}$$ with terms quadratic or quartic in $C_{\uparrow\!\uparrow}$ and $C_{\downarrow\!\downarrow}$ with coefficients that are independent in general (we will further specify these coefficients is the next two subsections). In the absence of screening fields, Zeeman-fields and the term proportional to $\gamma$, the form of the free energy in Eq. is similar to that examined in Ref. in the context of two-component Bose condensates (spin-half spinor condensate).In that case, the vortex lattice structure is solely determined by the competition between the $\beta_1$ and $\beta_2$ terms of Eq. (\[EQ:freeLLL1\]); the $\beta_1$ term determines the interaction energy within each vortex lattices, and the $\beta_1$ term the interaction energy between two fractional vortex species each forming lattices. Specifically, the quartic term $-\beta_2\langle |\phi_0|^2|\tilde{\phi}|^2\rangle C_{\uparrow\uparrow}^2C_{\downarrow\downarrow}^2$ determines the stability of the HQV lattice. If $\beta_2<0$, then a HQV lattice is the ground state. If $\beta_2>0$, then full quantum vortex lattice is the ground state. In Appendix \[app:weakcoupling\], we show that $\beta_2=0$ in weak-coupling theories, so that the two lattice structures are degenerate. In the rest of the paper, we will focus on aspects that are unique to triplet superconductors in subsections \[subsec:screening\] and \[subsec:zeeman\]: the effects of the screening fields $f_{in}^{(4)}$, and the Zeeman-field.
The effects of screening and $f_{in}^{(4)}$ {#subsec:screening}
-------------------------------------------
![Real space vortex lattice structure[]{data-label="fig:1a"}](lattice_structure.pdf){width="48.00000%"}
Here we look into the effect of screening. Ignoring the Zeeman field (in weak-coupling theories, the Zeeman field is vanishing in the clean limit), we have symmetry between $C_{\uparrow\!\uparrow}$ and $C_{\downarrow\!\downarrow}$, giving us $$\tilde{\alpha_1} = \tilde{\alpha_2} \equiv
\tilde{\alpha} = \frac{2\pi K(H_{c2}-H)}{\Phi_0}\langle |\phi_0|^2 \rangle
\label{EQ:quadCoeff}$$ and $$\tilde{\beta}_1=\tilde{\beta}_2 \equiv
\tilde{\beta} = \left(\frac{\beta}{2} - \frac{8\pi^3 K^2}{\Phi_0^2}\right)\langle |\phi_0|^4 \rangle,
\label{EQ:quartCoeff}$$ which means that the free energy in Eq. takes a simpler form: $$\langle \tilde{f} \rangle = -\tilde{\alpha} (C^2_{\uparrow\!\uparrow} + C^2_{\downarrow\!\downarrow}) +
\tilde{\beta}(C^4_{\uparrow\!\uparrow} + C^4_{\downarrow\!\downarrow}) + \tilde{\beta_3} C^2_{\uparrow\!\uparrow} C^2_{\downarrow\!\downarrow},
\label{EQ:freeScreen}$$ with $$\tilde{\beta_3} = \frac{2\gamma}{l^2} \langle |\phi_0|^2 (|\tilde{\phi}_0|^2- |\tilde{\phi}_1|^2)\rangle
- \left(\frac{16\pi^3 K^2}{\Phi_0^2} + \beta_2 \right) \langle |\phi_0|^2 |\tilde{\phi}_0|^2 \rangle.
\label{EQ:quartCoeff2}$$ Now the free energy of Eq. can be minimized by choosing $C^2_{\uparrow\!\uparrow} = C^2_{\downarrow\!\downarrow} = \tilde{\alpha}/(2\tilde{\beta} + \tilde{\beta_3})$ (which gives $|\hat{\Delta}| \propto (H_{c2} - H)^{1/2}$ as mentioned), giving us the free energy expectation value $$\langle f \rangle = -\frac{H^2}{8\pi} -
\frac{\tilde{\alpha}^2}{2\tilde{\beta}+\tilde{\beta_3}}.
\label{EQ:freeFormulaLLL}$$ Eqs. and allows for understanding the role of both screening and $f_{in}^{(4)}$. The screening affect the vortex lattice structure through the dependence of terms proportional to $K^2$ in Eqs. (\[EQ:quartCoeff\]) and (\[EQ:quartCoeff2\]) on the lattice structure parameters $\varsigma$, $\sigma$ and ${\bm
\tau}$. Since all quartic expectation values depend on the lattice structure, the lattice structure will be determined through minimizing $(2\tilde{\beta}+\tilde{\beta_3})$. Since the magnitude of $K^2$ term in Eq. is larger for full quantum vortices (Eq. is not affected), screening tend to disfavor HQV lattices, in line with earlier observation for isolated HQV’s[@chung:197002]. However, since weak-coupling theories lie near the point $\tilde{\beta_3}=0$ we expect the screening effect will put the physical system at a fine balance between interlacing lattices of HQV’s and the ordinary Abrikosov vortex lattice. Hence it should be possible to observe the transition between the two phases upon small change of field and temperature. Indeed, the interaction $f_{in}^{(4)}$ plays this role. In particular, we numerically find that for $\gamma>0$, then this term tends to favor HQV lattices. A positive sign of $\gamma$ occurs when $\rho_{sp},\rho_s$ and this is to be expected in spin-triplet superconductors [@chung:197002]. Note, that unlike the other contributions in $\tilde{\beta}_3$, the contribution from $f_{in}^{(4)}$ vanishes as $T \rightarrow T_c$. Consequently, this term can drive a field and temperature dependent transition between a HQV lattice and a full quantum lattice. Figures 1 and 2 show the phase diagram for $\gamma=0$. Note the similarity between the calculated phase diagram and that found in the context of two-component Bose condensates [@PhysRevLett.88.180403; @barnett:240405].
![Transition between different vortex lattice structures when $\gamma=0$. (a)-(e) label the vortex lattice structure shown in Fig. (\[fig:1a\]).[]{data-label="fig:1b"}](lattice_phase.pdf){width="45.00000%"}
The effects of the Zeeman term {#subsec:zeeman}
------------------------------
For simplicity, we consider here the effect of Zeeman field alone ignoring screening and setting $\gamma=0$ (ignoring $f_{in}^{(4)}$). In the presence of Zeeman field, the free energy Eq. would have different coefficients for two quartic terms: $$\langle \tilde{f} \rangle = -\tilde{\alpha_1}
C^2_{\uparrow\!\uparrow} -\tilde{\alpha_2}
C^2_{\downarrow\!\downarrow} +
\tilde{\beta}(C_{\uparrow\!\uparrow}^4 +
C_{\downarrow\!\downarrow}^4) + \tilde{\beta_3}
C_{\uparrow\!\uparrow}^2 C_{\downarrow\!\downarrow}^2
\label{eq-free-1}$$ where $\tilde{\alpha}_1=\alpha+\frac{K}{l^2}+H\tilde{\kappa}$, $\tilde{\alpha}_2=\alpha+\frac{K}{l^2}-H\tilde{\kappa}$, $\tilde{\beta}=\beta_1\langle |\phi_0|^4\rangle$, and $\tilde{\beta}_3=-\beta_2\langle|\phi_0|^2|\tilde{\phi}_0|^2|^2\rangle$. The Zeeman field has two main consequences: (i) it typically leads to two phase transitions. In the first phase $C_{\uparrow\!\uparrow}\ne 0$ and $C_{\downarrow\!\downarrow} = 0$ and in the second phase, both components are non-zero. The first phase is analogous to the $^3$He A$_1$ phase, with a non-unitary spin-triplet order parameter. However, weak-coupling theories prefer unitary spin-triplet states and this drives the second transition. (ii) In the fractional vortex lattice phase where both components are non-zero, the magnetic flux contained by isolated fractional vortices is no longer a half-integral flux quanta. Instead, two types of vortices each carry fractional flux values of $$\Phi_i=\Phi_0\frac{|c_i|^2}{|c_1|^2+|c_2|^2}.$$
The double transition is possible if $2\tilde{\beta}+\tilde{\beta_3}>0$ (note again that weak-coupling theories yield $\tilde{\beta_3}=0$ and $\tilde{\beta}>0$), there can be two transitions with a second transition appearing at a temperature $$T_{c2}-T_{c1}=\frac{4\tilde{\beta}}{2\tilde{\beta}+\tilde{\beta}_3}\frac{K+H_{c2}l^2\tilde{\kappa}}{K_1-H_{c2}l^2\tilde{\kappa}}T_{c1}.$$ In the high temperature phase, the vortex lattice is hexagonal and, at the second transition, the lattice will remain hexagonal and the second component will either coincide with first or be displaced half a hexagonal vortex lattice vector from the first. As temperature is further reduced below the second transition, the lattice will continuously deform, asymptotically approaching the phases presented in the subsection \[subsec:screening\] (those shown in Fig. 2). The resulting phase diagram is qualitatively shown in Fig. 3.
Both consequences of the Zeeman field stem from breaking the additional $\mathbb{Z}_2$ symmetry that is present when $\tilde{\alpha}_1=\tilde{\alpha}_2$. In general, the existence of fractional vortices is the result of the $U(1)\times U(1)$ symmetry of the free energy. When there is additional $\mathbb{Z}_2$ symmetry due to $\tilde{\alpha}_1=\tilde{\alpha}_2$, the flux contained in each fractional vortices are restricted to be half the flux quantum since the two components of the order parameter are no longer degenerate in a magnetic field. In Section VII, we will see that this helps us distinguish a lattice of HQVs from a lattice of full quantum vortex.
![The role of the Zeeman field is to cause two transitions and to change the HQV lattice phase to a a phase with two types of fractional vortices in which the two fractional fluxes sum to $\Phi_0$. Far below the second transition it is expected that these fractions will be well approximated by $1/2$.[]{data-label="fig2"}](fig2.pdf){width="45.00000%"}
Chiral Triplet superconductors: lowest Landau level solution {#subsec:chiral}
------------------------------------------------------------
The chiral triplet superconductor with tetragonal symmetry, because of the inhomogeneous quadratic terms we have already seen in Eq. (\[EQ:quadDeriv2\]), $$\begin{aligned}
f^{(2,s)}_{in} &=& K_1(|{\bf D}\Delta_{s,+}|^2 + |{\bf D}\Delta_{s,-}|^2)\nonumber\\
&+&K_2[ \!\{\!(D_x \Delta_{s,+})\!(D_x \Delta_{s,-})^*\!-\!(D_y \Delta_{s,+})\!(D_y \Delta_{s,-})^*\!\}\!/2\nonumber\\
&+& \!\{\!(D_x \Delta_{s,-})\!(D_x \Delta_{s,+})^*\!-\!(D_y \Delta_{s,-})\!(D_y \Delta_{s,+})^*\!\}\!/2\nonumber\\
&+& i\!\{\!(D_x \Delta_{s,-})\!(D_y \Delta_{s,+})^*\!+\!(D_y \Delta_{s,-})\!(D_x \Delta_{s,+})^*\!\}\!/2\nonumber\\
&-& i\!\{\!(D_x \Delta_{s,+})\!(D_y \Delta_{s,-})^*\!+\!(D_y \Delta_{s,+})\!(D_x \Delta_{s,-})^*\!\}\!/2]\nonumber\\
&+& K_4 (|D_z\Delta_{s,+}|^2 + |D_z\Delta_{s,-}|^2),\end{aligned}$$ has a much more complicated quadratic free energy, $$f^{(2)}_0 = \sum_{s=\uparrow\!\uparrow,\downarrow\!\downarrow}[-\alpha(|\Delta_{s,+}|^2+|\Delta_{s,-}|^2)+f^{(2,s)}_{in}],
\label{EQ:quadChiral0}$$ even when we exclude the Zeeman field and any spin-orbit coupling.
Due to these inhomogenous quadratic terms, we cannot put both chirality components in the lowest Landau level. This is due to the presence of $[(D_x \Delta_{s,\sigma})(D_y \Delta_{s,-\sigma})^* + {\rm c.c}]$ terms in $f^{(2,s)}_{in}$. The above quadratic free energy of Eq. (\[EQ:quadChiral0\]), together with Eq. (\[EQ:chiralSplit\]) that gives us the energy splitting between two chiralities, leads to the linearized GL equation $$\begin{aligned}
&\,&\alpha l^2\!\left(\!\begin{array}{c}
\Delta_{s+} \\ \Delta_{s-}\\
\end{array}\!\right) =\nonumber\\
&\,&\left[\!\begin{array}{cc}
K_1\!(\!1\!+\!2\Pi_+\Pi_-\!)\!-\!\delta K & K_2\Pi_-^2\\
K_2\Pi_+^2 & K_1\!(\!1\!+\!2\Pi_+\Pi_-\!)\!+\!\delta K\\ \end{array}\!\right]\!\left(\!\begin{array}{c}
\Delta_{s+} \\ \Delta_{s-}\\
\end{array}\!\right).\nonumber\\
\label{EQ:matrixGL}\end{aligned}$$
However, a lowest Landau level solution that satisfies Eq. (\[EQ:matrixGL\]) may still have the highest $H_{c2}$ and therefore be possible [@Zhitomirskii_UPt3]. This would lead to a nonzero order parameter for only one chirality - $(\Delta_{s+},\Delta_{s-}) = C(0,\phi_0)$ - and requires $\delta K
< -\frac{K_2^2}{4K_1}$. In this case, the vortex energetics of the chiral triplet superconductor is identical to that of the nonchiral triplet superconductor, for inserting this lowest Landau level solution into the full Gibbs free energy leads us back to Eq. (\[EQ:LLLfree\]).
However, Eq. (\[EQ:matrixGL\]) can also give us a solution with Landau level mixing; in this case, both $\Delta_{s+}$ and $\Delta_{s-}$ are nonzero. We will present discussion on this Landau level mixing in Appendix \[sec:mixLandau\].
Role of spin-orbit Coupling {#sec:SO}
===========================
Generic Case
------------
If the material has orthogonal or tetragonal symmetry (though not necessarily true for hexagonal symmetry, which is discussed in the next subsection), then there will exist spin-orbit coupling terms of the type $$\epsilon \Delta_{\uparrow\uparrow}\
\Delta^*_{\downarrow\downarrow}.$$ Such terms break the $U(1)\times U(1)$ symmetry and consequently isolated fractional flux vortices are no longer stable. Nevertheless, a fractional flux quantum vortex lattice can still exist, provided that the separation between vortices is less than $\xi_{so}$ defined through $\xi_{so}^2=K/\epsilon$.
For completeness, we write here the spin-orbit coupling terms that appear in the context of a chiral spin-triplet superconductor. While we do not include these terms in calculations, they may be useful in other contexts. Due to the tetragonal $C_4$ symmetry, homogeneous spin-orbit coupling terms should be invariant under the transformation $$\begin{aligned}
(\Delta_{\uparrow\!\uparrow,+},\Delta_{\uparrow\!\uparrow,-},\Delta_{\downarrow\!\downarrow,+},\Delta_{\downarrow\!\downarrow,-})
\to \nonumber \\
(-\Delta_{\uparrow\!\uparrow,+},\Delta_{\uparrow\!\uparrow,-},\Delta_{\downarrow\!\downarrow,+},-\Delta_{\downarrow\!\downarrow,-}).
\nonumber \end{aligned}$$ To the quadratic order, this condition is satisfied by [^1] $$\begin{aligned}
f^{(2)}_{SO} &=& \!\epsilon_1(\!|\Delta_{\uparrow\!\uparrow,+}|^2\!+\!|\Delta_{\downarrow\!\downarrow,-}|^2\!-\!|\Delta_{\uparrow\!\uparrow,-}|^2\!-\! |\Delta_{\downarrow\!\downarrow,+}|^2\!)\nonumber\\
&+& \epsilon_2[(\!\Delta_{\uparrow\!\uparrow,-}\!)\!(\!\Delta_{\downarrow\!\downarrow,+}\!)^* + {\rm c.c.}]\nonumber\\
&+&
\epsilon_3[(\!\Delta_{\uparrow\!\uparrow,+}\!)\!(\!\Delta_{\downarrow\!\downarrow,-}\!)^*
+ {\rm c.c.}].\end{aligned}$$ We note here that $\epsilon_i$’s can be estimated from the recent ARPES data [@haverkort:026406; @liu:026408].
Hexagonal Materials
-------------------
For hexagonal materials, there exist spin-triplet pairing states for which no such terms such as that in the above equation appear. These states belong to the two-dimensional representations labelled $\Gamma_{5}^-$ and $\Gamma_6^-$ in the review article by Sigrist and Ueda[@RevModPhys.63.239]. Consequently, these materials need to be considered more carefully.
We will now show that in hexagonal materials, a little away from $H_{c2}$, spin-orbit coupling does not break $U(1) \times U(1)$ symmetry. For hexagonal materials, the only term that exists in the GL free energy that is due to spin-orbit coupling is (note that the inclusion of this term gives rise the complete free energy found that is found in Sigrist and Ueda for the $\Gamma_{5,6}^-$ representations): $$\begin{aligned}
f_{SO} &=& K_{so}\Big \{(D_x \Delta_{\downarrow\downarrow})(D_x
\Delta_{\uparrow \uparrow})^*-(D_y \Delta_{\downarrow\downarrow})
(D_y \Delta_{\uparrow\uparrow})^* \nonumber\\
&+&(D_x \Delta_{\uparrow \uparrow})(D_x
\Delta_{\downarrow\downarrow})^*-
(D_y \Delta_{\downarrow\downarrow})(D_y \Delta_{\uparrow\uparrow})^*\nonumber\\
&-& i[(D_x \Delta_{\downarrow\downarrow})(D_y
\Delta_{\uparrow\uparrow})^*+
(D_y \Delta_{\downarrow\downarrow})(D_x \Delta_{\uparrow\uparrow})^*]\nonumber\\
&+& i[(D_x \Delta_{\uparrow\uparrow})(D_y
\Delta_{\downarrow\downarrow})^*+
(D_y \Delta_{\uparrow\uparrow})(D_x \Delta_{\downarrow\downarrow})^*]\Big \}/2\nonumber\\\end{aligned}$$ With the field along the $c$-axis, the solution to the quadratic problem satisfies $$\frac{\alpha l^2}{K}\!\left( \begin{array}{c}
\Delta_{\uparrow \uparrow} \\ \Delta_{\downarrow \downarrow}\\
\end{array} \right)\!=\!
\!\left( \begin{array}{cc}
1+2N-K_z & \tilde{K}_{so}\Pi_-^2\\
\tilde{K}_{so}\Pi_+^2 & 1+2N+K_z\\
\end{array} \right) \left(
\begin{array}{c}
\Delta_{\uparrow \uparrow} \\ \Delta_{\downarrow \downarrow}\\
\end{array} \right)
\nonumber$$ where $K_z=\frac{\tilde{K}Hl^2}{K}$ and $\tilde{K}_{so}=K_{so}/K$. All the eigenstates for this problem can be found analytically[@superRev]. Typically, $|\tilde{K}_{so}|<<1$, so we will be interested in the eigenstates that contain the lowest Landau level (which will minimize the free energy when $\tilde{K}_{so}=0$). The two relevant eigenstates that we wish to keep are: $(\Delta_{\uparrow \uparrow}, \Delta_{\downarrow,\downarrow})=(\phi_0,\epsilon \phi_2)$ and $(\Delta_{\uparrow \uparrow},\Delta_{\downarrow,\downarrow})=(0,\phi_0)$ where $\epsilon$ is proportional to $\tilde{K}_{so}$. Note that unlike in the subsection \[subsec:chiral\] we can keep both solutions because while we examined $H \sim H_{c2}$ in that subsection, we are a little away from $H_{c2}$ in this subsection. We therefore write $\Delta_{\uparrow\uparrow}=\gamma_1 \phi_0$ and $\Delta_{\downarrow\downarrow}=\gamma_1 \epsilon\phi_2+\gamma_2\tilde{\phi}_0$ to include these two eigenstates. For simplicity, we ignore screening and the Zeeman field to find the following free energy $$\begin{aligned}
\langle f \rangle =&&-(1-H/H_{c2,1})|\gamma_1|^2-(1-H/H_{c2,2})|\gamma_2|^2\nonumber\\
&&+\beta_1[|\gamma_1|^4\langle |\phi_0|^4\rangle+\langle|\gamma_1\epsilon\phi_2+\gamma_2\tilde{\phi}_0|^4\rangle]\nonumber\\
&&-\beta_2\langle |\gamma_1\phi_0|^2|\gamma_1\epsilon\phi_2+\gamma_2\tilde{\phi}_0|^2\rangle\end{aligned}$$ where $H_{c2,i}$ ($i=1,2$) is the upper critical field for eigenstate $i$. Since spin-orbit coupling is expected to be small, this implies that $\epsilon<<1$, so keeping to linear order in $\epsilon$ yields: $$\begin{aligned}
\langle f \rangle =&&-(1-H/H_{c2,1})|\gamma_1|^2-(1-H/H_{c2,2})|\gamma_2|^2\nonumber\\
&&+\beta_1\langle|\phi_0|^4\rangle(|\gamma_1|^4+|\gamma_2|^4)+\beta_2\langle|\phi_0|^2|\tilde{\phi}_0|^2\rangle|\gamma_1|^2|\gamma_2|^2\nonumber\\
&&+\epsilon[\beta_1|\gamma_2|^2\gamma_2\gamma_1^*\langle|\tilde{\phi}_0|^2\tilde{\phi}_0\phi_2^*\rangle+{\rm c.c}]\nonumber\\
&&-\epsilon[\beta_2|\gamma_1|^2\gamma_1\gamma_2^*\langle|\phi_0|^2\phi_2\tilde{\phi}_0^*\rangle + {\rm c.c}]
\label{eq-so}\end{aligned}$$ Without the last two terms, this theory is the same as that found for non-chiral spin-triplet superconductors with a Zeeman field but without any spin-orbit coupling. At the upper critical field, one of the two components $\gamma_1$ or $\gamma_2$ order and the vortex lattice will be hexagonal (this conclusion is correct even when including terms that are second order in $\epsilon$). As the temperature or magnetic field is reduced, the last two terms in Eq. (\[eq-so\]) can play an important role. These two terms break the $U(1)\times U(1)$ symmetry of the theory and therefore will tend to remove any HQV lattice phases. However, the spatial averages $\langle|\tilde{\phi}_0|^2\tilde{\phi}_0\phi_2^*\rangle$ and $\langle|\phi_0|^2\phi_2\tilde{\phi}_0^*\rangle$ [*vanish*]{} for a hexagonal vortex lattice (loosely speaking, this follows from noting that $\phi_n$ picks up a factor $e^{in\phi}$ under a rotation about $\hat{z}$ and that a hexagonal vortex lattice is symmetric under rotations of $\pi/3$) . The hexagonal symmetry of the materials conspires to remove this form of $U(1)\times U(1)$ symmetry breaking and the HQV lattice structures are still possible (indeed the theory is the same as that given for the non-chiral spin-triplet superconductors with a Zeeman field, but without spin-orbit coupling). Note that if ${\bm \tau}\ne 0$ (signaling the existence of the fractional vortex lattice), then $\langle|\tilde{\phi}_0|^2\tilde{\phi}_0\phi_2^*\rangle$= $\langle|\phi_0|^2\phi_2\tilde{\phi}_0^*\rangle$=0 for any lattice geometry. Consequently, the last two terms of Eq. ) do not play any role in the theory of the fractional vortex lattices.
It is reasonable to ask if there are any other $U(1)\times U(1)$ symmetry breaking terms that we have neglected in the above analysis. Indeed there is one that appears at order $\epsilon^2$: $\epsilon^2\beta_1\gamma_1^2(\gamma_2^2)^*\langle\phi_2^2(\tilde{\phi}_0^2)^*\rangle$. This term allows for the existence of a fractional vortex lattice phase subject to the constraint that ${\bm \tau}$ is half a vortex lattice translation vector[@superRev]. There are also $U(1)\times U(1)$ that appear at order $\epsilon^3$, but these vanish for the same reason as the order $\epsilon$ term. Consequently, the spin-orbit coupling for the hexagonal two-dimensional representations plays essentially the same role as the Zeeman field.
Observation of the Vortex Lattice {#sec:observation}
=================================
{width=".27\textwidth"} {width=".27\textwidth"} {width=".27\textwidth"}
The best way to determine both the vortex lattice structure and the vortex type is to observe the magnetic field distribution through the small angle neutron scattering. What we will see in this experiment is the Fourier transform $f({\bf G})$ of the screening field of Eq. (\[EQ:screenField\]), $$\begin{aligned}
h_s({\bf r}) &=& (\frac{8\pi^2 K}{\Phi_0}-4\pi\tilde{\kappa}) C^2_{\uparrow\!\uparrow} |\phi_0({\bf r})|^2\nonumber\\
&+& (\frac{8\pi^2 K}{\Phi_0}+4\pi\tilde{\kappa})C^2_{\downarrow\!\downarrow} |\tilde{\phi}_0({\bf r})|^2\end{aligned}$$ - that is $h_s({\bf r}) = \sum_{\bf G} f({\bf G}) \exp(i {\bf G} \cdot {\bf r})$, where ${\bf G}$ is the reciprocal lattice vectors of the vortex lattice in the unit of the inverse magnetic length.
The characteristic feature of the vortex lattice with half-vortices in the small angle neutron scattering experiment is the modulation of the Bragg peaks. The form factor of the Bragg peaks would be the $f({\bf G})$ of the last paragraph. Using $$|\phi_0({\bf r})|^2 = \sum_{\bf G} (-1)^{m_1 + m_2 + m_1 m_2} e^{-{\bf G}^2/2}$$ where ${\bf G} = m_1 {\bf G_1} + m_2 {\bf G_2}$, and ${\bf G_i}$’s are the basis vector of the reciprocal lattice, we obtain the form factor $$\begin{aligned}
f({\bm G}) &=& (-1)^{m_1 + m_2 + m_1 m_2} e^{-{\bm G}^2/2}[(\frac{8\pi^2 K}{\Phi_0}-4\pi\tilde{\kappa})|C_{\uparrow\!\uparrow}|^2\nonumber\\
&+& (\frac{8\pi^2 K}{\Phi_0}+4\pi\tilde{\kappa})|C_{\downarrow\!\downarrow}|^2 e^{i{\bm G}\cdot {\bm \tau}}].\end{aligned}$$ This equation implies that the intensity $|f({\bm G})|^2$ for our Bragg peaks does not come out same for all ${\bf G}$’s. This is because for almost all vortex lattice structure (the single exception being not very robuts honeycomb lattice) ${\bm \tau}$ is half a vortex translation vector so that we have $e^{i{\bm G}\cdot {\bm \tau}}=-1$ for half of ${\bf G}$’s and $e^{i{\bm G}\cdot {\bm \tau}}=1$ for the other half. When there is no Zeeman field, $e^{i{\bm G}\cdot {\bm \tau}}=-1$ peaks disappear completely; natural given that magnetic field cannot distinguish the spin up and the spin down HQV’s at all and thus sees the unit lattice vector halved. However, when the Zeeman field breaks down the $\mathbb{Z}_2$ symmetry between the spin up-up pairs and down-down pairs, we now see a secondary peak for $e^{i{\bm G}\cdot {\bm \tau}}=-1$ as shown on Fig. (\[fig:structure\]).
Another promising direction for detecting fractional vortex lattice would be to use spin-polarized STM to probe the vortex cores. The key point is that the low energy quasi-particle spins have opposite polarization in the two different HQV’s. This is because for half of HQV cores, we have $\Delta_{\uparrow\uparrow}=0$ and $\Delta_{\downarrow\downarrow}\ne 0$, so that only spin-down quasi-particles are gapped. On the other hand, for the other half of HQV cores, only spin-up quasi-particles are gapped. This spin-polarization of the subgap core modes should be readily detected through spin-polarized STM.
Conclusion {#sec:conclusion}
==========
In this paper we explored various possibilities for fractional vortex lattice structures in spin triplet superconductors starting from the most general from of Gibbs free energy that is allowed by the symmetry of the order parameter and that of the lattice symmetries relevant for three candidate spin triplet superconductors, namely single layer ruthenate Sr$_2$RuO$4$[@Maeno:1994fk; @RevModPhys.75.657] cobaltate Na$_x$CoO$_2 \cdot y$H$_2$O[@Nature.Physics.1.91] and organic[@organic] (TMTSF)$_2$ClO$_4$. The focus of our analysis was on the role of aspects unique to triplet superconductors, such as (i) Cooper pair Zeeman field, (ii) spin-orbit coupling, (iii) screening, and (iv) interaction effects in the energetics of the vortex lattice structure. (i) The Cooper pair Zeeman field breaks a $\mathbb{Z}_2$ symmetry of the free energy whose presence constrains the fractional vortices to contain half integral flux quanta. The resulting structure is that of two interlacing lattices of vortices containing arbitrary fraction of flux quanta that adds up to one flux quanta. Such fractional vortex lattices will have interesting field distributions in vortex lattice unit cell due to internal structures within the unit cell. (ii) The effect of spin-orbit coupling is lattice symmetry specific. In hexagonal lattices systems such as cobaltates Na$_x$CoO$_2\cdot$yH$_2$O, spin-orbit coupling has the same effect as the Cooper pair Zeeman field, supporting fractional vortex lattices. However, for tetragonal or orthorhombic lattices, sufficiently strong spin-orbit coupling generally favors ordinary Abrikosov vortex lattice over HQV’s. However, such an effect is relatively mild in a dense vortex lattice, provided that the separation between the HQV vortices is less that a length set by the spin-orbit coupling. (iii) The Meissner screening effectively generates attraction between two HQV’s with opposite winding of the spin phase and weakly destablizes the HQV’s. (iv) The interaction effects clearly support energetic stability of HQV’s within the GL theory. The interaction effects represented by inhomogeneous (unique to triplet superconductors) quartic terms can drive difference in effective superfluid stiffness $\rho_{sp} < \rho_s$ which stabilizes HQV’s in the London limit. When the above effects are put together, all weak coupling theories we examined appears to lie at the point of fine balance between ordinary Abrikosov vortex lattice and lattices of HQV’s. Hence it should be possible to observe transitions between these structures with small changes of parameters. This further motivates experimental search for these fractional vortex lattices. We have sketched possible routes for such searches using neutron scattering or spin polarized STM.
[**Acknowledgements**]{} We are grateful to H. Bluhm for helpful discussions regarding inhomogeneous quartic term and M.Sigrist for discussions on spin-orbit coupling. We thank M.Stone, D. Podolsky, S. Mukerjee, E. Berg, S. Raghu for numerous useful discussions. E-AK was supported in part by the Cornell Center for Materials Research (CCMR) through NSF Grant No. DMR 0520404. SBC was supported by the Stanford Institute of Theoretical Physics and NSF Grant No. DMR 06-03528. We acknowledge the KITP for its hospitality through the miniprogram “Sr$_2$RuO$_4$ and Chiral p-wave Superconductivity" during initial stages of this work.
Ginzburg Landau Energy: Fourth order terms from weak-coupling theory {#app:weakcoupling}
====================================================================
The GL free energy can be determined in the weak-coupling limit. In the context of the existence of 1/2 qv lattice structures, the result for the fourth order terms in the free energy turns out to be highly relevant. As shown here, this reveals that weak-coupling theories sit at a point in which the 1/2 qv and the full qv lattices are degenerate. This indicates that interactions beyond the weak-coupling limit are essential to determining the which lattice structure actually appears (screening plays a role here as well as shown earlier).
The portion of the free energy we calculate here is given in Eq. 3.6 $$f^{(4)}_{hom}=\beta_1(\sum_{i}|\Delta_{i}|^2)^2+\beta_2|\Delta_{\uparrow\!\uparrow}|^2
|\Delta_{\downarrow\!\downarrow}|^2.$$ The weak-coupling limit (without spin-orbit coupling) yields (this follows from Ref. ) $$f^{(4)}_{hom}\propto \langle|{\bf d}({\bf
k})|^4\rangle+\langle{\bf q}^2({\bf k})\rangle$$ where ${\bf q}({\bf k})=i{\bf d}({\bf k})\times {\bf d}^*({\bf
k})$, $\langle h({\bf k}) \rangle$ means average $h({\bf k})$ over all ${\bf k}$ on the Fermi surface, and the proportionality constant can be found but it is not important for our considerations. When ${\bf q}$ is non-zero, then the superconducting state is called non-unitary. In weak-coupling theories, non-unitary states cost energy and typically do not appear. Using the gap structure of Eq. 2.2, we find $$\begin{aligned}
f^{(4)}_{hom}&& \propto \langle|f({\bf
k})|^4\rangle[(|\Delta_{\uparrow\uparrow}|^2+|\Delta_{\downarrow\downarrow}|^2)^2+(|\Delta_{\uparrow\uparrow}|^2-|\Delta_{\downarrow\downarrow}|^2)]
\nonumber\\ && = 2\langle|f({\bf
k})|^4\rangle(|\Delta_{\uparrow\uparrow}|^4+|\Delta_{\downarrow\downarrow}|^4).\end{aligned}$$ This implies that $\beta_2=-2\beta_1$, independent of the shape of the Fermi surface. The lack of interaction between the two components of the gap function leads to the degeneracy between the 1/2 qv and the full qv lattice structures.
Ruthenate - the Landau level mixing {#sec:mixLandau}
===================================
We show here how we can have the Landau level mixing in a chiral triplet superconductor. The case we are considering here is in the weak pairing regime and has tetragonal crystalline symmetry and a cylindrical Fermi surface. Let us consider again the linearized GL equation: $$l^2\!\left( \begin{array}{c}
\Delta_{s+} \\ \Delta_{s-}\\
\end{array} \right)\!=\!
\frac{K}{\alpha}\!\left( \begin{array}{cc}
1+2\Pi_+\Pi_- & \Pi_-^2\\
\Pi_+^2 & 1+2\Pi_+\Pi_-\\ \end{array} \right) \left(
\begin{array}{c}
\Delta_{s+} \\ \Delta_{s-}\\
\end{array} \right),$$ where $s = \uparrow\!\uparrow, \downarrow\!\downarrow$. (Note that, though otherwise same as Eq. (\[EQ:matrixGL\]), we now ignore the energy splitting between the $\pm$ chiralities and set $K_1 = K_2 = K$.) This matrix equation as a solution in the form $$\begin{aligned}
\left(
\begin{array}{c}
\Delta_{s+} \\ \Delta_{s-}\\
\end{array} \right) = C_s \left(
\begin{array}{c}
\phi_0 \\ -\delta \phi_2 \\
\end{array} \right),
\label{EQ:landauMixingSol}\end{aligned}$$ where $\delta = \sqrt{3} - \sqrt{2}$.
When we ignore the Zeeman field, much of the vortex lattice energetics of the lowest Landau level case remains valid with the Landau level mixing. For instance, the two main formulas of Section V. A, Eqs., $$\langle \tilde{f} \rangle = -\tilde{\alpha} (C^2_{\uparrow\!\uparrow} + C^2_{\downarrow\!\downarrow}) +
\tilde{\beta}(C^4_{\uparrow\!\uparrow} + C^4_{\downarrow\!\downarrow}) + \tilde{\beta_3} C^2_{\uparrow\!\uparrow} C^2_{\downarrow\!\downarrow},$$ and ), $$\langle f \rangle = -\frac{H^2}{8\pi} - \frac{\tilde{\alpha}^2}
{2\tilde{\beta}+\tilde{\beta_3}},$$ remains valid, mainly due to $|\hat{\Delta}| \propto (H_{c2} - H)^{1/2}$. This means we can still calculate $h_s$, solely from quadratic terms. For quadratic terms, we simply have two copies (for $s = \uparrow\!\uparrow$ and $\downarrow\!\downarrow$) of what was obtained for the case of ${\bf d} = (k_x + ik_y){\bf \hat{z}}$ by one of us [@PhysRevB.58.14484], we can use the formula for $h_s$ for that case: $$\begin{aligned}
h_s &=& \frac{8\pi^2 K}{\Phi_0} [C^2_{\uparrow\uparrow} \{(1-3\delta/\sqrt{2} + 2\delta^2)|\phi_0|^2\nonumber\\
&+& (2\delta^2 - \delta/\sqrt{2})|\phi_1|^2 + \delta^2|\phi_2|^2\}\nonumber\\
&+& C^2_{\downarrow\downarrow}\{(1-3\delta/\sqrt{2} + 2\delta^2)|\tilde{\phi}_0|^2\nonumber\\
&+& (2\delta^2 - \delta/\sqrt{2})|\tilde{\phi}_1|^2
+ \delta^2|\tilde{\phi}_2|^2\}].
\label{EQ:screenField2}\end{aligned}$$ The spatial average of this equation is still proportional to $(C^2_{\uparrow\uparrow}+ C^2_{\downarrow\downarrow})$ just like Eq. (\[EQ:screenField\]). Also, $\tilde{\alpha} \propto (H_{c2} - H)$ still stands: $$\begin{aligned}
\tilde{\alpha} &=& \frac{2\pi K (H_{c2} - H)}{\Phi_0} [(1-3\delta/\sqrt{2} + 2\delta^2)\langle|\phi_0|^2\rangle\nonumber\\
&+& (2\delta^2 - \delta/\sqrt{2})\langle|\phi_1|^2\rangle +
\delta^2\langle|\phi_2|^2\rangle].\end{aligned}$$
However, the formula for $2\tilde{\beta}+\tilde{\beta_3}$ are much more complicated here, especially when we include all terms of Eqs. and for these coefficients. For sake of convenience, instead of directly writing down $\tilde{\beta}$ and $\tilde{\beta_3}$, we will list $\overline{h_s^2}$ (terms that are proportional to $K^2$), $\overline{f^{(4)}_{hom}}$ (terms involving coefficients of Eq.), and $\overline{f^{(4)}_{in}}$ (terms involving coefficients of Eq. ); to obtain $2\tilde{\beta}+\tilde{\beta_3}$, we can use the relation $$2\tilde{\beta}+\tilde{\beta_3} = \overline{f^{(4)}_{hom}} + \overline{f^{(4)}_{in}} - \frac{\overline{h_s^2}}{8\pi}.$$ The following is the full listing of $\overline{h_s^2}/8\pi$, $\overline{f^{(4)}_{hom}}$, and $\overline{f^{(4)}_{in}}$ (note that we have set $\beta_3$ of Eq. to be zero): $$\begin{aligned}
\frac{\overline{h_s^2}}{8\pi} &=& \frac{8\pi^3 K^2}{\Phi_0^2} [(1-3\delta/\sqrt{2} + 2\delta^2)^2 \langle |\phi_0|^4 \rangle\nonumber\\
&+& 2 (2\delta^2 - \delta/\sqrt{2})(1-3\delta/\sqrt{2} + 2\delta^2)\langle |\phi_0|^2|\phi_1|^2 \rangle\nonumber\\
&+& 2\delta^2 (1-3\delta/\sqrt{2} + 2\delta^2)\langle |\phi_0|^2|\phi_2|^2 \rangle\nonumber\\
&+& (2\delta^2 - \delta/\sqrt{2})^2 \langle |\phi_1|^4 \rangle\nonumber\\
&+& 2\delta^2 (2\delta^2 - \delta/\sqrt{2}) \langle |\phi_1|^2|\phi_2|^2 \rangle + \delta^4 \langle |\phi_2|^4 \rangle\nonumber\\
&+& (1-3\delta/\sqrt{2} + 2\delta^2)^2 \langle |\phi_0|^2|\tilde{\phi}_0|^2 \rangle\nonumber\\
&+& 2 (2\delta^2 - \delta/\sqrt{2})(1-3\delta/\sqrt{2} + 2\delta^2)\langle |\phi_0|^2|\tilde{\phi}_1|^2 \rangle\nonumber\\
&+& 2\delta^2 (1-3\delta/\sqrt{2} + 2\delta^2)\langle |\phi_0|^2|\tilde{\phi}_2|^2 \rangle\nonumber\\
&+& (2\delta^2 - \delta/\sqrt{2})^2 \langle |\phi_1|^2|\tilde{\phi}_1|^2 \rangle\nonumber\\
&+& 2\delta^2 (2\delta^2 - \delta/\sqrt{2}) \langle |\phi_1|^2|\tilde{\phi}_2|^2 \rangle\nonumber\\
&+& \delta^4 \langle \langle |\phi_2|^2|\tilde{\phi}_2|^2 \rangle],\\
\overline{f^{(4)}_{hom}} &=& \beta_1 (\langle|\phi_0|^4\rangle + \delta^4 \langle |\phi_2|^4 \rangle) + 2\beta'_1 \delta^2 \langle|\phi_0|^2\rangle \langle |\phi_2|^2 \rangle\nonumber\\
&-& \beta_2 (\langle |\phi_0|^2|\tilde{\phi}_0|^2 + \delta^4
\langle |\phi_2|^2|\tilde{\phi}_2|^2 \rangle)\nonumber\\ &-&
2\delta^2 \beta'_2 \langle |\phi_0|^2|\tilde{\phi}_2|^2 \rangle,\end{aligned}$$ and $$\begin{aligned}
\overline{f^{(4)}_{in}} &=& \frac{2\gamma}{l^2} (\langle |\phi_0|^2|\tilde{\phi}_0|^2 \rangle - \langle |\phi_0|^2|\tilde{\phi}_1|^2 \rangle\nonumber\\
&+& 3\delta^2 \langle |\phi_0|^2|\tilde{\phi}_2|^2 \rangle - 3\delta^2 \langle |\phi_0|^2|\tilde{\phi}_3|^2 \rangle)\nonumber\\
&+& \frac{2\gamma'}{l^2}(3\delta^2 \langle |\phi_0|^2|\tilde{\phi}_2|^2 \rangle - 3\delta^2 \langle |\phi_0|^2|\tilde{\phi}_3|^2 \rangle\nonumber\\
&+& 2\delta^4 \langle |\phi_0|^2|\tilde{\phi}_1|^2 \rangle + \delta^4 \langle |\phi_0|^2|\tilde{\phi}_2|^2 \rangle - 3\delta^4 \langle |\phi_0|^2|\tilde{\phi}_3|^2 \rangle\nonumber\\
&+& 2\delta^4 \langle |\phi_1|^2|\tilde{\phi}_1|^2 \rangle + \delta^4 \langle |\phi_1|^2|\tilde{\phi}_2|^2 \rangle - 3\delta^4 \langle |\phi_1|^2|\tilde{\phi}_3|^2 \rangle\nonumber\\
&+& 3\delta^4 \langle |\phi_2|^2|\tilde{\phi}_2|^2 \rangle -
3\delta^4 \langle |\phi_2|^2|\tilde{\phi}_3|^2 \rangle).\end{aligned}$$
Correlation functions {#app:corr}
=====================
In calculating $\langle f^{(4)}_{in} \rangle$, note $$({\bm D} f)\cdot ({\bm D} g)^* = \frac{1}{l^2}[(\Pi_+ f)(\Pi_+
g)^* + (\Pi_- f)(\Pi_- g)^*]$$ and $$\begin{aligned}
\Pi_+ \phi_n &=& \sqrt{n+1}\phi_{n+1}\nonumber\\
\Pi_- \phi_n &=& \sqrt{n}\phi_{n-1}.\end{aligned}$$ Together with partial integration $$\begin{aligned}
\langle (\Pi_+ \phi_n) \tilde{\phi}^*_m \tilde{\phi}_p \phi^*_q
\rangle &=& \langle \phi_n (\Pi_-\tilde{\phi}_m)^* \tilde{\phi}_p
\phi^*_q \rangle\nonumber\\ &-& \langle \phi_n \tilde{\phi}^*_m
(\Pi_+\tilde{\phi}_p) \phi^*_q \rangle\nonumber\\ &+& \langle
\phi_n \tilde{\phi}_m^* \tilde{\phi}_p (\Pi_-\phi_q)^* \rangle,
\label{EQ:partialInt}\end{aligned}$$ these equation gives $$\begin{aligned}
\frac{1}{l^2}\langle\phi_0^*\tilde{\phi}_0({\bm D}
\phi_0)\cdot({\bm D} \tilde{\phi}_0)^* \rangle &=&
\langle |\phi_0|^2|\tilde{\phi}_0|^2 \rangle - \langle |\phi_0|^2|\tilde{\phi}_1|^2 \rangle\nonumber\\
\frac{1}{l^2}\langle\phi_0^*\tilde{\phi}_2({\bm D}
\phi_0)\cdot({\bm D} \tilde{\phi}_2)^* \rangle &=&
\langle |\phi_0|^2|\tilde{\phi}_2|^2 \rangle - \langle |\phi_0|^2|\tilde{\phi}_3|^2 \rangle\nonumber\\
\frac{1}{l^2}\langle\phi_2^*\tilde{\phi}_2({\bm D}
\phi_2)\cdot({\bm D} \tilde{\phi}_2)^* \rangle &=&
2 \langle |\phi_0|^2|\tilde{\phi}_1|^2 \rangle + \langle |\phi_0|^2|\tilde{\phi}_2|^2 \rangle \nonumber\\
&+& 2 \langle |\phi_1|^2|\tilde{\phi}_1|^2 \rangle + \langle |\phi_1|^2|\tilde{\phi}_2|^2 \rangle\nonumber\\
&-& 3 \langle |\phi_0|^2|\tilde{\phi}_3|^2 \rangle - 3 \langle |\phi_1|^2|\tilde{\phi}_3|^2 \rangle\nonumber\\
&+& 3 \langle |\phi_2|^2|\tilde{\phi}_2|^2 \rangle - 3 \langle |\phi_2|^2|\tilde{\phi}_3|^2 \rangle.\nonumber\\\end{aligned}$$
These can be evaluated using $$\begin{aligned}
\frac{\langle|\phi_p|^2|\phi_q|^2\rangle}{\langle|\phi_0|^2\rangle^2} &=& \sum_{r,s} L^0_p({\bm k}_{rs}^2/2) L^0_q({\bm k}_{rs}^2/2) e^{-{\bm k}_{rs}^2/2},\nonumber\\
\frac{\langle|\phi_p|^2|\tilde{\phi}_q|^2\rangle}{\langle|\phi_0|^2\rangle^2} &=& \sum_{r,s} L^0_p({\bm k}_{rs}^2/2) L^0_q({\bm k}_{rs}^2/2) e^{-{\bm k}_{rs}^2/2}\cos({\bm k}_{rs}\cdot {\bm \tau})\nonumber\\\end{aligned}$$ where $L^0_n$ is a Laguerre polynomial of $n$th order and ${\bm
k}_{rs} = (\sqrt{2\pi\sigma}r,\sqrt{2\pi/\sigma}(s-\varsigma r))$
[26]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , , , , , ****, ().
, ****, ().
, ****, ().
, , , , , , , , ****, ().
.
, , , , , , , , , , , ****, ().
, , , , , , , , , , , ****, ().
, ****, ().
, , , , , ****, ().
, , , , , ****, ().
, , , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , , ****, ().
, , , , ****, ().
, ****, ().
, ** (, ).
.
.
, , , ****, ().
.
[^1]: In the absence of out-of-plane magnetic field, such spin-orbit interaction support superconductor with order parameter symmetry analogous to that of $^3$He-B phase. This class of phase has been recently discussed as a time-reversal invariant topological superconductor [@topoSC].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '[****]{} We introduce four algorithms for packet transport in complex networks. These algorithms use deterministic rules which depend, in different ways, on the degree of the node, the number of packets posted down each edge, the mean delivery time of packets sent down each edge to each destination and the time since an edge last transmitted a packet. On scale-free networks all our algorithms are considerably more efficient and can handle a larger load than the random walk algorithm. We consider in detail various attributes of our algorithms, for instance we show that an algorithm that bases its decisions on the mean delivery time jams unless it incorporates information about the degree of the destination node.'
author:
- Bernard Kujawski
- 'G.J. Rodgers'
- Bosiljka Tadić
title: |
Local Information Based Algorithms for Packet\
Transport in Complex Networks
---
Introduction
============
Complex networks can be used to model a wide range of physical and technological systems. One of the most interesting dynamical problems on network is transport, which can give us some insight into the transport of information in technology based communication networks like the internet [@faloutsos], the World Wide Web [@albert],[@Huberman] or phone call networks [@Adamic]. Here we use the term *transport* to mean transport of particles, which are packets in a network. Thus our model falls within the Network Layer of the OSI Reference Model and the algorithms described in section $3$ are routing algorithms that belong to the Network Layer of the OSI Reference Model. Of particular interest is the phenomenon of load in a network, as a function of the rate of packet creation R, which has been investigated for models of communication networks [@Tadic],[@Arenas],[@Sole] and in real networks [@Jacobson].
Typically the problem of transport is investigated using either a random walk algorithm [@Tadic], or the shortest path algorithm used by most internet protocols. The difficulty with these approaches is that random walk algorithm is very inefficient for transport in technology based communication networks and shortest path algorithm requires, for its implementation, information about all connections in network. In this paper we focus on algorithms that use local information about the topology, along with information about the flux of packets between neighbors, the link load and the time taken to deliver packets. We propose four algorithms that use some or all of these properties to deliver packets in a network.
In section 2 we describe the algorithm that we use to perform numerical simulations of our models. In section 3 we discuss the algorithms that packets use to find their destinations and in section 4 we show our results. In section 5 we summarise our results.
The Program
===========
A program was written to simulate packet transport on a network that does not depend on the size of the network or its topology. At the beginning of the program an external file with the adjacency matrix of the network is read in. We focus on the internet and consequently we treat nodes in our network as if they were routers. The connections between the routers have the same capacity for all networks. Such a model can not only be used to model internet packet transport but also for a range of transport networks in which the nodes have local routing information.
**Each node**:
- Generates a new packet with probability $r=R/N$ and with a randomly chosen destination, where $R$ is a fixed rate for the whole network, and $N$ is the number of nodes in network.
- Stores packets in a queue, which has maximum length is $L=1000$. Packets are despatched from the queue in a first in first out (FIFO) order.
- Sends packets to its neighbours.
**Each node has information about**:
- The address of all its neighbours (they have unique indices $j$).
- The degree of its neighbours - $k(i)$.
- Flow through all its neighbours, which is measured by
- The number of packets posted down each edge to neighbour $i$ - the Link Load - $C(i)$.
- The number of packets sends through neighbour $i$, which have reached their destination - $N_P(i)$.
- The sum of the delivery times of all the packets sent through neighbour $i$ that have reached their destination - $T_P(i)$.
- The time interval since an edge last transmitted a packet to neighbour $i$ and current time step - $\Delta T(i)$.
The index $i$ enumerates each neighbour of node $k$ and each node keeps all the statistics about its neighbours. Quantities $C(i)$, $N_P(i)$, $T_P(i)$ and $\Delta T(i)$ describe node $i$ from the perspective of node $k$. Each node is described by its neighbours and all properties can be different for all neighbours that describe node $i$.
The initialization part of the program sets up the network topology, the nodes and all the tables used by them. Inside the main loop a time step is incremented, and within that a loop over all nodes calculates and updates the statistics. The loop over all nodes includes three basics routines, which are run for each node; generating new packets, checking its queue for packets with its address and sending packets to its neighbours. Each node generates a packet with a randomly chosen destination with probability $R/N$. The node checks its own queue for packets addressed to itself. When it finds one of these it deletes it from the queue and updates the statistics $N_P(i)$ and $T_P(i)$ for all the nodes on the packet’s path. Each packet keeps track of its own path. The node sends packets to its neighbours by taking the first packet in its queue and checking the packet destination address. If the packet is addressed to one of its neighbour, the node will send it to the neighbour. If it is not, the node will use the *algorithm* to find where to send the packet. During this posting step the $C(i)$ property is updated. When node $k$ sends packets to node $i$, the number of sent packets $C(i)$ increases. After this loop over all the nodes is completed the quantities $\Delta T(i)$ and the mean delivery time of packets sent down each edge $N_P(i)/T_P(i)$ are updated for all nodes.
Algorithms
==========
The most important element in transport is the rule that determines the direction in which a packet is sent. A transport network without a rule is a random walk network. We call this rule the *algorithm*. It describes how nodes deal with packets and should help packets to get to their destination. Not all algorithms help packets to reach destinations, poor algorithms can easily be worse than the random walk algorithm. All algorithms considered in this paper work with‘deterministic rules.
The *shortest time*(ST) algorithm is our basic algorithm that uses information about the mean delivery time $T_P(i)/ N_P(i)$ and the time interval between the last packet that came to node $i$ and actual time step. The ST algorithm finds the minimum value $$\label{ST}
S_k=\min\left[\frac{T_P(i)}{N_P(i)}\frac{1}{\Delta
T(i)}\right]_{i=1\ldots n}$$ in order to determine which node to send the packet to. The idea of this algorithm is to try and find the minimum travel time for each packet between source and destination. At the start of the simulation $S$ is equal to $0$ for all neighbours. Because the update of $T_P(i)/ N_P(i)$ only occurs when a packet arrives at its destination, it can take a number of time steps before $T_P(i)/
N_P(i)$ becomes non-zero. The inclusion of the reciprocal of $\Delta
T(i)$ in $S$ ensures that the algorithm does not get into a state where it never sends a packet down certain links which have a large mean delivery time. This state is particularly likely to occur at the start of the simulation. The inclusion of the reciprocal of $\Delta T(i)$ in $S$ also prevents overcrowding when a node finds a node which is clearly better than all its other neighbours. Hence, because of the inclusion of $\Delta T(i)$ more nodes take part in the transport and in this way the large node do not become overcrowded. Because the algorithm with $T_P(i)/ N_P(i)$ is looking for minimum delivery time we call it the *shortest time* (ST) algorithm. To start this algorithm, and the STD algorithm, which we will introduce shortly, we use the random walk algorithm. We only use the deterministic algorithms at a node when all the values of $S$ of its neighbours are greater than 0. Without this initial random walk procedure both the ST and the STD algorithms would jam almost immediately. The *shortest time and degree* (STD) algorithm is a modification of the ST algorithm. It uses information about the local topology, the degree. This helps packets avoid the nodes with the largest degree, which are mostly overcrowded. The idea of incorporating information about the degree of nodes in the transport algorithm was discussed in [@Yan] and [@Wang]. In these papers models were introduced in which nodes were selected at a rate proportional to a power of their degree. It was found that the most efficient algorithm was one in which the the probability of selecting a node of degree $k$ was proportional to $1/k$ [@Yan] and [@Wang]. The STD algorithm is defined by $$\label{STD}
S_k=\min\left[\frac{T_P(i)}{N_P(i)}\frac{1}{\Delta
T(i)}k(i)\right]_{i=1\ldots n}$$ where $k(i)$ is a degree of node $i$ and $k(i)>1$. This last assumption allows the algorithm to avoid dead-end nodes. A node with degree $k=1$ can only receive a packet that is addressed to itself. The STD algorithm uses both temporal properties and also information about the local connectivity. For transport in a scale-free network the most important nodes are those with the largest degree. But because their neighbours send these nodes a large number of packets the queues at these nodes can become overcrowded. Information about the degree helps the algorithm to avoid these nodes, but it does not mean than they are not used.
The *connections and degree* (CD) algorithm and the *connections, degree and shortest time* (CDT) algorithm use information about the link load $C(i)$. Because of this the random walk starting procedure used in the ST and STD algorithms is not required for the CD and CDT algorithms. The CD algorithm uses only information about the link load and the degree. The CD algorithm is defined by $$\label{CD}
S_k=\min[C(i)k(i)]_{i=1\ldots n}$$ where $C(i)$ is a number of packets that node $k$ sends to node $i$.
For this algorithm $S$ equals $0$ at the start, but $C(i)$ is updated almost immediately. When node $k$ sends a packet then it automatically increases the value of $C(i)$. There is no need to wait for information from the destination about the delivery time like in the ST and STD algorithms. In this way CD algorithm improves very quickly and the random walk is not needed. The link load, $C(i)$, quantity helps the algorithm to deliver packets and ensures that almost all nodes take part in the transport. The degree quantity helps to prevent the largest nodes from becoming overcrowded. In this algorithm there is no property that can be optimised, unlike in the ST and STD algorithms where the delivery time is optimised.
The CDT algorithm is intermediate between the CD and the ST algorithms. It optimises the delivery time and does not need the random the walk starting procedure because it includes a dependance on the link load, $C(i)$. The dependence on degree prevents large nodes becoming overcrowded. For the CDT algorithm, the starting procedure is the same as for the CD algorithm except that we set $$\label{TimeElem}
\frac{T_P(i)}{N_P(i)}\frac{1}{\Delta
T(i)}$$ equal to $1$ at the start to avoid $0$ value. This means that we do not need to start off with a random walk algorithm as in the ST and STD algorithms. The CDT algorithm is defined by $$\label{CDT}
S_k=\min\left[\frac{T_P(i)}{N_P(i)}\frac{1}{\Delta
T(i)}C(i)k(i)\right]_{i=1\ldots n}\qquad\mbox{with }k(i)>1.$$
We use the learning property to describe behavior of an algorithm in the beginning. By learning we mean the proportion of links whose value of $S$ has changed since $t=0$. The CD and CDT algorithms learn the most quickly. After $5000$ time steps they tried $95\%$ of links. This is because the link load, $C(i)$, changes when a packet is sent down it whereas $T_P(i)/ N_P(i)$, used by the ST and STD algorithms, only changes when a packet sent down it gets to its destination. That is way the ST and STD algorithms need the random walk starting procedure. With this procedure after $5000$ time steps $35\%$ of links were tried. For the ST algorithm without the random walk starting procedure it was $5\%$. The speed of learning is important because when a network learns slowly, the network only uses a small proportion of its links for transport over a long period of time, which means that the network is easily jammed when a region of the network becomes overcrowded.
Results
=======
We consider transport on the Barabasi and Albert model of a network [@Albert_base] with $N=1000$ nodes and $m=2$. The parameter $m$ is the number of links of a new node that is added to network. When $m=2$ the network includes loops and has relative small number of connections. Our research show that this network jams for lower values of the posting rate than networks with $m=1$ or $m=3$ and higher. In this work we use a posting rate of $R=0.1$. This means that each node creates a packet with probability $R/N$. The number of time steps for all our simulations is $500,000$. We present results for the STD, CD and CDT algorithms. We do not consider the ST algorithm any further because it isn’t stable and always jams.
In figure \[LoadProp\]a we show the load in the network, the number of packets that are still in the network. All three algorithms are stable. We compared the level of load by finding the mean value of the number of packets in the network. The best algorithm with smallest mean value is the STD algorithm. For the CD and CDT the values are almost the same.
The number of packets in network can be treated as a noise in the network. Measuring the power spectrum of this noise shows that there are correlations in the number of packets in network. For all our algorithms the power spectrum (Fig.\[LoadProp\]b) is the same and the slope has $-2$. It means that the noise in network is like $1/f^2$; uncorrelated noise with short-range correlations only.
We measured the distribution of the time interval $\Delta T(i)$, the time that nodes wait for packets, and the results are shown in figure \[TIDdt\]. This is an important quantity for the SDT and CDT algorithms as without the $\Delta T(i)$ term these networks easily jam. For the STD algorithm the distribution of $\Delta T(i)$ has a tail and on a double logarithmic scale has a slope $b=-3/2$. The cut-off comes from the finite time of the simulation. The first part of the distribution for all algorithms is flat. For the CDT algorithm the function falls faster than for the STD. This is connected with the inclusion of the link load in the CDT algorithm, which means that more links are used and long time intervals of $\Delta T(i)$ do not occur as frequently as in the STD algorithm. The CD algorithm does not use $\Delta T(i)$ but we measured it to compare it to the other models.
The distribution of packet delivery time (Fig.\[TimeProp\]a) is similar for all the algorithms. However the distribution shows that the number of packets delivered in a short time is different for each algorithm For the STD algorithm packets are delivered quickly more frequently than for the CD and CDT algorithms. The STD algorithm finds the paths with the shortest delivery time because, whilest the CD and CDT algorithms are distributing the transport across the network, because their algorithms use the link load $C(i)$, the STD algorithm is looking for shortest delivery times. The distribution for the CDT algorithm is intermediate between the STD and CD algorithms because the CDT algorithm depends on the link load $C(i)$ and the shortest time statistics.
The time series for the overall mean delivery time (Fig.\[TimeProp\]b) show that algorithms involving the statistics for $T_P(i)/N_P(i)$ do not learn. The mean delivery time for CD and CDT is almost the same. The algorithms reach a stable mean delivery time and do not optimise it. Obviously for the CD algorithm no optimisation is possible because there is no quantity that could be optimized.
The result for the STD and CDT algorithms arise through two effects. The first is the inclusion of $\Delta T(i)$ in $S$ that send packets to rarely used links that often are not the best ones for transport. On the other hand without $\Delta T(i)$ all the algorithms with the mean time property start jamming. Secondly is the inclusion of the degree in $S$, which means that algorithms prefer to send packets to nodes with a small degree which makes the delivery time long.
Conclusions
===========
The algorithms STD, CD and CDT work well; for the same network and for the same value of $R$ the random walk algorithm jams, and these algorithms do not. One might expect that including the mean delivery time of packets sent to node $i$, $T_P(i)/N_P(i)$, in $S$ would optimize the delivery time. This does not happen because of the dependence of $S$ on the delivery time, link load and degree. But on the other hand without dependence on these terms the algorithms cannot work properly. This the case in the ST algorithm, which works better than the random walk algorithm, but much worse than the other algorithms. When the shortest time property is used in the scale-free network it needs to be balanced be degree quantity. The existence of nodes with large degrees causes traffic congestion for the shortest time algorithm. Using an algorithm which depends on local degree information but without dependence on the mean time (CD algorithm) works correctly but an algorithm without local degree dependence and with the mean time dependence (ST algorithm) jams easily. The biggest problem in implementing the STD and CDT algorithms is in finding accurate value for the edge dependent properties. A node needs a lot of connections through one link to find it proper time statistics. Because the mean delivery time is very long, it takes a lot of time to set up the edges dependent properties for all nodes. In particular, the algorithms that depend on the time $\Delta T(i)$ and the degree $k(i)$ do not jam but the cost is in learning and the mean delivery time. The inclusion of the $\Delta T(i)$ quantity in $S$, avoids jamming but destroys the learning behavior promoted by the inclusion of the mean time property in $S$. The degree property helps the algorithm to avoid nodes with large degree, and hence helps prevent overcrowding, but it also results in long delivery times. Our results show that in scale free networks we cannot avoid using nodes with large degree.
In future work, it may be possible to develop an algorithm that uses information on the mean local delivery time to find the optimal path for transport. One possible extension of this work would be to use an algorithm that allows a number of packets to be sent to a node in one time step, depending on the degree of the node. This is realistic because normally routers can use all their outputs almost in a parallel way. The biggest problem in networks is that nodes with a very high degree can receive as many packets as they have inputs in one time step, but they usually send only one packet. When we allow them to use all their outputs in one time step then jamming will disappear.
[11]{} M. Faloutsos, P. Faloutsos and C. Faloutsos, (1999) Comp. Comm. Rev. [**29**]{}, 251. R. Albert, H. Jeong and A.-L. Barabasi, (1999) Nature [**401**]{}, 130 . B. Huberman and L. Adamic, (1999) Nature [**401**]{}, 131. L. A. Adamic, R. M. Lukose, A. R. Puniyani and B. A. Huberman, Phys. Rev. E (2001) [**64**]{}, 046135. B. Tadic, S. Thurner and G. J. Rodgers, (2004) Phys. Rev. E [**69**]{}, 036102. A. Arenas, A. Diaz-Guilera and R. Guimera, (2001) Phys. Rev. Lett. [**86**]{}, 3196. R. Sole and S. Valverde, (2001) Physica A [**289**]{}, 595. V. Jacobson, in Proceedings of SIGCOMM ’88 (ACM, Standford, CA, 1988). G. Yan, T. Zhuo, B. Hu, Z.-Q. Fu and B.-H. Wang, (2005) cond-mat/0505366. C.-Y. Yin, B.-H. Wang, W.-X. Wang, T. Zhou and H.-J. Yang, (2005) cond-mat/0506204. R. Albert and A.-L. Barabasi, (2002) Rev. Mod. Phys. [**74**]{}, 47.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We study how to estimate a nearly low-rank Toeplitz covariance matrix $T$ from compressed measurements. Recent work of Qiao and Pal addresses this problem by combining sparse rulers (sparse linear arrays) with frequency finding (sparse Fourier transform) algorithms applied to the Vandermonde decomposition of $T$. Analytical bounds on the sample complexity are shown, under the assumption of sufficiently large gaps between the frequencies in this decomposition.
In this work, we introduce *random ultra-sparse rulers* and propose an improved approach based on these objects. Our random rulers effectively apply a random permutation to the frequencies in $T$’s Vandermonde decomposition, letting us avoid frequency gap assumptions and leading to improved sample complexity bounds. In the special case when $T$ is circulant, we theoretically analyze the performance of our method when combined with sparse Fourier transform algorithms based on random hashing. We also show experimentally that our ultra-sparse rulers give significantly more robust and sample efficient estimation then baseline methods.
bibliography:
- 'ICASSPsqrtk.bib'
title: |
Low-Rank Toeplitz Matrix Estimation\
via Random Ultra-Sparse Rulers
---
Introduction {#sec:intro}
============
We study the problem of estimating the $d \times d$ covariance matrix $T \in {\mathbb{R}}^{d \times d}$ of a distribution $\mathcal{D}$ over $d$-dimensional vectors given independent samples $x^{(1)},x^{(2)},\ldots, x^{(n)} \in {\mathbb{R}}^d$ drawn from $\mathcal{D}$. In particular, we focus on the case when the covariance matrix $T$ is Toeplitz, which arises when the vectors are wide-sense stationary: the covariance $t_{|j-k|}$ between the $j^{th}$ and $k^{th}$ entries only depends on the distance $|j -k|$. We let $t_s$ denote the covariance at distance $s$ for $s \in \{0,\ldots,d-1\}$.
Toeplitz covariance estimation arises in a range of applications, including direction of arrival (DOA) estimation [@KrimViberg:1996; @wax1997joint; @BogaleLe:2016], spectrum-sensing for cognitive radio [@MaLiJuang:2009; @CohenTsiperEldar:2018], medical and radar imaging [@SnyderOSullivanMiller:1989; @fuhrmann1991application; @BrookesVrbaRobinson:2008], [@CohenEldar:2018; @RufSwiftTanner:1988; @AslMahloojifar:2012] and Gaussian process regression (kriging) and kernel machine learning [@dietrich1991estimation; @wilson2015kernel]. We focus on estimation methods with *low sample complexity*, can be measured in two ways [@toeplitz]:
#### Entry Sample Complexity.
How many entries of each sample $x^{(i)} \sim \mathcal{D}$ must be read? Minimizing entry sample complexity typically corresponds to minimizing sensor cost, as, in many applications, each entry of $x^{(i)}$ is measured with a different sensor in a spatial grid. We consider algorithms where the same entries are read in each $x^{(i)}$ (i.e., the active sensors remain fixed).
#### Vector Sample Complexity.
How many $d$-dimensional samples $x^{(i)}$ must be drawn from $\mathcal{D}$? Vector sample complexity corresponds to minimizing acquisition time or measurement cost and is the classic notion of sample complexity in statistics and machine learning.
Typically there is a trade-off between these two measures. In this work, we seek to *minimize entry sample complexity*, while keeping vector sample complexity reasonably low.
Sparse Ruler Based Sampling
---------------------------
Our work centers on the powerful idea of sparse rulers (also known as sparse linear arrays), which let one perform covariance estimation with significantly reduced entry sample complexity. A sparse ruler is a subset of indices $R \subseteq \{1,\ldots, d\}$, such that for every distance $s \in \{0,1,\ldots, d-1\}$, there is some pair $i,j \in R$ with distance $| i - j| = s$. The set of distances measured by $R$ is $R$’s *difference coarray* or *difference set* [@ErdosGal:1948; @Leech:1956; @Wichmann:1963; @Moffet:1968; @PillaiBar-NessHaber:1985]. It is clear that to represent $d$ distances, we must have $|R| \ge \sqrt{d}$ so that ${R \choose 2} \ge d$ and it is well known that for any $d$, there exists a sparse ruler matching this optimal size up to constants. A large body of work has studied the design of sparse rulers under various additional objectives [@caratelli2011novel; @qin2015generalized; @cohen2019sparse]. We note that in some cases, which will arise later in this work, we may allow $R$ to be any set of integers, including those outside $\{1,\ldots, d\}$.
Sparse rulers have received significant attention in covariance estimation applications [@LexaDaviesThompson:2011; @ArianandaLeus:2012; @romero2016compressive; @WuZhuYan:2017; @toeplitz]. Given a sample $x \sim \mathcal{D}$ with Toeplitz covariance matrix $T$, if we read the $|R|$ entries of $x$ corresponding to indices in a ruler $R$, we obtain an estimate of the covariance $t_s$ *at every distance $s$*. So in principle, with enough samples, $x^{(1)},\ldots, x^{(n)} \sim \mathcal{D}$ we can accurately estimate $T$ while measuring just $|R| = O(\sqrt{d})$ entries in each sample (i.e., with $O(\sqrt{d})$ entry sample complexity). In fact, recent work has shown that, with sparse ruler measurements, $\tilde O(d/\epsilon^2)$ vector samples suffice to recover any Toeplitz matrix to accuracy $\epsilon$ in the spectral norm[@toeplitz].
Improved Bounds for Low-Rank Matrices {#sec:piya}
-------------------------------------
For general Toeplitz covariance matrices it is impossible to improve on the entry sample complexity achieved by sparse rulers: without reading at least $O(\sqrt{d})$ entries, we can never estimate the covariance at some distances. However in many applications, such as DOA estimation, when the number of sources is smaller than the number of sensors, the Toeplitz covariance matrix of the received signal snapshots is *low-rank*, or close to low-rank. This *additional structure* can be leveraged to recover $T$ with a smaller subset of its entries [@abramovich1996positive; @ChenChiGoldsmith:2015]. Recent work of Qiao and Pal [@qiao2017gridless] shows that, if $T$ is approximately rank $k$ for any $k < d$, an entry sample complexity of just $O(\sqrt{k})$ can be achieved using sparse rulers. The high-level idea is easily understood: if $T$ is exactly rank-$k$, then it can be decomposed uniquely using the Carathéodory-Fejér-Pisarenko decomposition (the Vandermonde decomposition) [@caratheodory1911zusammenhang] as $T = F_T D F_T^*$, where $D \in {\mathbb{R}}^{k \times k}$ is a diagonal matrix and $F_T \in {\mathbb{C}}^{d \times k}$ is a Fourier matrix, with $F_T({m,\ell}) = e^{2 \pi i f_\ell \cdot (m-1)}$ for some set of frequencies $f_1,\ldots, f_k \in [0,1]$.
We can see immediately that the top left $k + 1 \times k+1$ principal submatrix of $T$, denoted $T_{k+1,k+1}$ (which is also Toeplitz, positive semidefinite, and rank $k$) admits a Vandermonde decomposition with the same frequencies – obtained by simply restricting $F$ to its first $k+1$ rows. Further, it can be shown that this decomposition is unique. Thus, we can recover the frequencies $f_1,\ldots, f_k$ and their weights $D$ just from a decomposition of $T_{k+1,k+1}$. Thus, from this small submatrix, we can recover all of $T$!
With this observation in hand, Qiao and Pal apply sparse ruler methods to $T_{k+1,k+1}$ to obtain entry sample complexity just $O(\sqrt{k})$. The key difficulty is that the Vandermonde decomposition is notoriously unstable: noise in approximating $T_{k+1,k+1}$ and any deviation of $T$ from being exactly rank-$k$ (i.e., truly having just $k$ frequencies in its Vandermonde decomposition) can entirely change the frequency content of this decomposition. Nevertheless, Qiao and Pal prove a bound on reconstruction error, under the assumption that $f_1,\ldots, f_k$ have spacing at least $\Theta(1/k)$ and that the underlying MUSIC frequency-finding routine [@Schmidt:1981; @Schmidt:1986] is exact. They give a vector sample complexity bound of roughly $O \left ({d^4}/{k^2 \epsilon^2} \right )$ to approximate all entries of $T$ up to error $\epsilon \cdot t_0$, where $t_0$ is covariance at distance $0$ (and therefore the largest entry of $T$ since it is positive semidefinite).
Our Contributions {#sec:contributions}
-----------------
We propose the idea of *random ultra-sparse rulers* to avoid the frequency gap assumption of Qiao and Pal, while simultaneously giving much lower vector sample complexity with similar entry sample complexity. In the special case when $T$ is circulant (corresponding to frequencies in its Vandermonde decomposition being ‘on-grid’ multiples of $1/d$), we prove a sample complexity bound that depends only *logarithmically on the ambient dimension $d$*, and polynomially on the rank $k$ and error parameter $\epsilon$. Broadly, our *random ultra-sparse rulers* open the door to achieving low entrywise sample complexity for circulant Toeplitz covariance estimation via a wider class of randomized sparse FFTs, providing more robust frequency recovery than deterministic techniques.
\[thm:main\] Algorithm \[alg:alg\] takes $\tilde O(k/\epsilon^2)$[^1] independent samples from any sub-Gaussian distribution $\mathcal{D}$ on ${\mathbb{R}}^d$ with circulant covariance matrix $T$. The algorithm reads $\tilde O(\sqrt{k})$ entries from each sample and returns with probability at least $2/3$, $\tilde T \in {\mathbb{R}}^{d \times d}$ satisfying: $$\begin{aligned}
{\|T - \tilde T\|}_F \le \epsilon {\|T\|}_F + 2 \min_{\text{rank-$k$}\ B} {\|T - B\|}_F.\end{aligned}$$ Throughout, ${\|\cdot \|}_F$ denotes the Frobenius norm.
#### Random Hashing for Ruler Design.
Algorithm \[alg:alg\] (Section \[sec:analysis\]) is inspired by work on random hashing based sparse Fourier transform methods [@sft; @gilbert2014recent]. The idea is to *transform $T$* in way that is equivalent to applying a random hash function ${h}: [0,1] \rightarrow [0,1]$ to the frequencies $f_1,\ldots, f_d$ in $T$’s Vandermonde decomposition. When $T$ is nearly rank-$k$, there may be up to $d$ such frequencies, but only $k$ will significantly contribute to the decomposition. After hashing, we expect the $k$ dominant frequencies to be well separated (without small gaps), and thus recoverable via a frequency finding approach like that proposed by Qiao and Pal. Even if some small gaps remain, by applying repeated random hash functions we can eventually recover all $k$ significant frequencies.
As utilized in sparse Fourier transform methods, when all frequencies $f_1,\ldots, f_d$ are on-grid integer multiples of $1/d$ (i.e., $T$ is circulant), it suffices to chose $h$ from the family of random hash functions ${h}_{a,b}(x) = a (x-b) \mod d $, where $a, b$ are randomly chosen integers [@gilbert2005improved; @gilbert2014recent]. $h_{a,b}(x)$ is applied to $x = fd$ when $f$ is an on-grid frequency in $\{0,1/d, \ldots, (d-1)/d\}$, and the hashed frequency is taken as $\frac{h_{a,b}(x)}{d} \in [0,1]$. Critically, the random hash function $ {h}_{a,b}$ in frequency domain can be implemented simply via a transformation to $T$. For a random integer $c$, let ${g}_{a,c}(x) = a (t- c) \mod d$. If $a$ is coprime to $d$, $ {g}_{a,c}(\cdot)$ is a permutation of $\{0,\ldots, d-1 \}$. Let $T_{a,b,c}$ be a transformed covariance matrix obtained by permuting $T$’s rows and columns with ${g}_{a,c}(x)$ and multiplying the $j,k$ entry by $e^{\frac{2 \pi i ab |j-k|}{d}}$. Let $\tilde{f}_1, \ldots, \tilde{f_d}$ and $\tilde{D}$ denote the frequencies and diagonal matrix in $T_{a,b,c}$’s Vandermonde decomposition. One can check that $T$’s Vandermonde decomposition can be obtained by setting $f_j = \frac{h_{a,b}^{-1}(\tilde{f_j}d)}{d}$ and $D = P\tilde{D}$ where $P$ is diagonal with $j^{th}$ entry $e^{2 \pi i ac f_j}$.
Accordingly, estimating $T$ reduces to estimating $T_{a,b,c}$, which we will do by estimating $T_{a,b,c}$’s top $O(k) \times O(k)$ submatrix and applying a strategy similar to [@qiao2017gridless]. Naively, if the permutation $ {g}_{a,c}(\cdot)$ were truly random, it would destroy the possibility of using a sparse ruler to measure this top submatrix of $T_{a,b,c}$: a general ruler construction is not known for a ruler with the *arbitrary* difference set a random permutation would require. However, by leveraging $ {g}_{a,c}(\cdot)$’s simple structure, we show that we can still construct a ruler to read this submatrix. The ruler is 1) *random*: based on randomly chosen $a,c$ and 2) *ultra-sparse*: measuring the covariance at $O(k)$ random distances using just $O(\sqrt{k})$ entry sample complexity.
Random Ultra-Sparse Rulers {#sec:ultra}
==========================
We start with a simple random ultra-sparse ruler construction that will suffice for circulant matrix estimation.
\[def:rus\] For any $d$ and $k \le d$, let $a,c\in \mathbb Z$ be chosen randomly such that $a$ is coprime to $d$. Let $Q_{a,c} = \{q_1,q_2,\ldots, q_m\}$ be any ruler for the distance set $\{a(0-c), a(1-c), \ldots, a(k-c)\}$ and let $R_{a,c} = \{r_1,\ldots, r_m\}$ where $r_i = q_i \mod d$. We call $R_{a,c}$ a *random ultra-sparse ruler*.
\[clm:cyclic\] Let $R_{a,c}$ be constructed from any valid $Q_{a,c}$ as in Definition \[def:rus\] and $ {g}_{a,c}(x) = a(x-c) \mod d$ be the random permutation corresponding to $a,c$. Then the following hold:
1. $R_{a,c}$ is a *cyclic ruler* for $\{ {g}_{a,c}(0), {g}_{a,c}(1),\ldots, {g}_{a,c}(k)\}$. I.e., for any $s \in \{0,\ldots,k\}$ there are $r_i,r_j \in R_{a,c}$ with either $r_i - r_j = g_{a,c}(s)$ or $r_i - r_j = d - g_{a,c}(s)$.
2. There exists $Q_{a,c}$, a ruler for the difference set $\{a(0-c), a(1-c), \ldots, a(k-c)\}$, with $|Q_{a,c}| = O(\sqrt{k})$. Correspondingly, $ |R_{a,c}| = O(\sqrt{k})$.
Since $Q_{a,c}$ is a ruler for the distance set $\{a(0-c), a(1-c), \ldots, a(k-c) \}$, for any $s \in \{0,\ldots, k\}$ there is some pair $q_i,q_j$ with $q_i - q_j = a(s - c)$ Thus $q_i - q_j \equiv a(s - c) \ \mod d$ and so $r_i - r_j \equiv a(s - c) \ \mod d$ and so $r_i - r_j \equiv g_{a,c}(s)\ \mod d$. Since $r_i,r_j$, and $g_{a,c}(s)$ are in $\{0,\ldots,d-1\}$, this equivalence can only hold if $r_i - r_j = g_{a,c}(s)$ or $r_i - r_j = g_{a,c}(s) -d$ and so $r_j - r_i = d - g_{a,c}(s)$. This completes the first claim. For the second claim, set $Q_{a,c} = \{0,a,\ldots, \lceil \sqrt{k}\rceil \cdot a\} \bigcup \{a \lceil \sqrt{k} \rceil - ac,2a \lceil \sqrt{k} \rceil -ac,\ldots, \lceil \sqrt{k}\rceil \cdot a \lceil \sqrt{k} \rceil-ac\}$, as shown in Figure \[fig:ruler\]. We can see that $|Q_{a,c}| = |R_{a,c}| \leq 2\lceil \sqrt{k}\rceil +1$. Just considering distances between the first and second halves of the ruler, $Q_{a,c}$’s difference coarray includes $as - ac$ for all nonnegative $s \le \lceil \sqrt{k}\rceil^2$. So $Q_{a,c}$ is an ultra-sparse ruler for the distance set $\{a(0-c), a(1-c), \ldots, a(k-c)\}$, as required.
![Illustration of rulers in Lemma \[clm:cyclic\]. The top line shows $Q_{a,c}$, with the first set $\{0,a,\ldots, \lceil \sqrt{k}\rceil \cdot a\}$ shown in red and the latter $\{a \lceil \sqrt{k} \rceil - ac,2a \lceil \sqrt{k} \rceil -ac,\ldots, \lceil \sqrt{k}\rceil \cdot a \lceil \sqrt{k} \rceil-ac\}$ in blue. Considering pairwise distances between red and blue markers demonstrates that the difference set of $Q_{a,c}$ is as claimed. Note that the elements of $Q_{a,c}$ may be far greater than $d$, and they may even be negative (for simplicity, in this illustration we assume $\lceil k \rceil > c$). The bottom line visualizes $R_{a,c}$, which is $Q_{a,c}$ “wrapped around" $\mod d$.[]{data-label="fig:ruler"}](ruler_construction.jpg){width="0.5\linewidth"}
Note that in a circulant matrix $T$, we have $t_i = t_{d - i}$. Thus a cyclic ruler of the form guaranteed by Lemma \[clm:cyclic\] suffices to measure the covariance at the full set of random distances $\{{g}_a(0), \ldots, {g}_a(k)\}$. In Section 3, we will show how this precise structure of difference set is just what’s needed by an efficient existing sparse FFT for on-grid frequencies, and we derive corresponding error guarantees for circulant covariance estimation. However, for general Toeplitz matrices (i.e., not cyclically symmetric), we require a true ruler. In this case, we can restrict the range of $a$ to prevent wrap around. For simplicity, in the following definition we also do not implement a random shift $c$.
\[def:rus2\] For any dimension $d$ and $k \le d$, let $a\in \mathbb Z$ be chosen randomly such that $a$ is coprime to $d$ and $a \le bd/k$ for some $b \leq 1$. Let $R_a = \{r_1,r_2,\ldots, r_m\}$ be any ruler for $\{0, a, 2a , \ldots, k a\}$.
Again, it is clear that we can find $R_a$ with $m = O(\sqrt{k})$. While in this manuscript we do not fully cover how to recover a non-circulant $T$ from a Type 2 ultra-sparse ruler, we give a short sketch here. If we set $k' = O(k)$ and estimate the $k' \times k'$ principal submatrix of $T$ indexed by $\{0, a, \ldots, k'a \mod d\}$, we are equivalently measuring the top-left $k' \times k'$ submatrix of a *transformed* matrix $\tilde{T}$ whose Vandermonde decomposition frequencies are $\{\tilde{f}_1, \ldots, \tilde{f}_d\}$ where $\tilde{f}_j = a\cdot {f}_j \mod 1$. Ideally, we would estimate the frequencies of $\tilde{T}$, which are separated by larger gaps, and use them to recover the frequencies of $T$. However, this cannot be done directly because there is *ambiguity* in inverting each $\tilde{f}_j$: there are up to $a$ different solutions $f_j \in [0,1]$ to the equation ${f}_j \equiv a\tilde{f}_j \mod 1$, as shown in Figure \[fig:ambiguity\].
Fortunately, this issue can be combated with simple repetition. Each time we draw a different random $a$, we collect potential candidate dominant frequencies for $T$’s Vandermonde decomposition. Since we restrict $a \le bd/k$, there will be $a \cdot k \leq b \cdot d$ such candidates: $k$ will be the true dominant frequencies in $T$’s Vandermonde decomposition and the remainder will be nearly random. Roughly, any frequency outside the set of dominant frequencies will appear in the set with probability $b < 1$. Thus, setting $b$ small enough, after roughly $O(\log d)$ repetitions, by observing which frequencies appear as candidates the largest number of times, we can determine the true dominant $k$ frequencies with high probability.
Analysis for Circulant Covariance {#sec:analysis}
=================================
We now apply the random ultra-sparse ruler construction of Definition \[def:rus\] to circulant covariance matrix estimation. For the remainder of the section let $F \in {\mathbb{C}}^{d \times d}$ be the discrete Fourier transform matrix with $F(j,k) = \frac{1}{\sqrt{d}} \cdot e^{\frac{2\pi i (j-1)(k-1)}{d}}$. For $x \in {\mathbb{R}}^d$ let $F^* x = \hat x$ denote its Fourier transform. Let $\operatorname{diag}(x)$ be the diagonal matrix with $x$ on its diagonal, and let ${\mathrm{Toep}}(x)$ be the symmetric Toeplitz matrices with first column $x$. Our algorithm makes blackbox use of a random hashing based sparse Fourier transform (SFT, or sparse FFT), with output guarantees as follows:
\[thm:sft\] Consider $x \in {\mathbb{R}}^d$ with Fourier transform $\hat{x}$. Assume that $d$ is a power of $2$. Let $\delta>0$ be a fixed error parameter. Algorithm 4.1 of [@sft] $SFT(x)$ outputs $k$-sparse $\hat{z}$ satisfying with $\frac{2}{3}$ probability: $$||\hat{x}-\hat{z}||_2 \leq 2 \min_{\text{k-sparse }y}|| \hat{x}-y||_2 + \delta||\hat{x}||_2.$$ The algorithm reads $O(\log^2 \frac{d}{k})$ blocks of $O\left(k \log \frac{d}{\delta} \right)$ entries. Each block consists of the first entries of $x$ after applying a different random permutation $g_{a,c}(\cdot)$ for $a,c$ chosen uniformly from $\{1,\ldots,d\}$ with $a$ odd (and thus coprime to $d$).
By Lemma \[clm:cyclic\], the sparse Fourier transform algorithm of Theorem \[thm:sft\] can be implemented via random ultra-sparse rulers with low entry sample complexity in the covariance estimation setting:
\[cor:cyclic\] There is a ruler $R$ with $O\left (\sqrt{k \log\frac{d}{\delta}} \cdot \log^2\frac{d}{k}\right )$ elements measuring all distances required for the algorithm of Theorem \[thm:sft\] to be applied to the first column of any circulant matrix $T$.
Let $R$ be the union of random ultra-sparse rulers $R_{a,c}$, each of which measures the entries in a permuted block read by the algorithm of Theorem \[thm:sft\] [@sft]. By Lemma \[clm:cyclic\], each $R_{a,c}$ is guaranteed to exist with just $O(\sqrt{k \log \frac{d}{\delta}})$ entries and there are $O(\log^2\frac{d}{k})$ blocks. Thus $|R| \leq O\left (\sqrt{k \log\frac{d}{\delta}} \cdot \log^2\frac{d}{k}\right )$.
With Corollary \[cor:cyclic\] in place, we present our main algorithm (Algorithm \[alg:alg\]). Note that in this algorithm, $\bar t_s$ is only estimated at the $\tilde O(k)$ positions represented by the ruler $R$ (i.e. in the difference set of $R$). Since $SFT(\bar t)$ only requires reading $\bar t$ at these positions, its output *does not depend on the other positions*. We have:
\[lem:5\] Consider circulant covariance matrix $T \in {\mathbb{R}}^{d \times d}$ for a sub-Gaussian distribution $\mathcal{D}$. Let $t \in {\mathbb{R}}^d$ be the first column of $T$ and $\bar t \in {\mathbb{R}}^d$ be the estimate computed by Algorithm \[alg:alg\] (line 2). Let $w \in {\mathbb{R}}^d$ match $\bar t$ on all entries read by SFT and match $t$ elsewhere. Letting $m = O \left (k \cdot \log^2 \frac{d}{k} \cdot \log\frac{d}{\epsilon} \right)$ (then number of entries of $\bar t$ that are read by SFT in line 3), and $n = O \left (\frac{m \sqrt{\log m}}{\epsilon^2} \right )$, we have with probability at least $2/3$, ${\|t - w\|}_2 \le \epsilon {\|T\|}_2$ and further, Algorithm \[alg:alg\] outputs $\hat z$ with: $$\begin{aligned}
{\|\hat z -
\hat t\|}_2 \le 5 \epsilon \cdot t_0 + 2 \min_{\text{k-sparse }y}|| \hat{t}-y||_2.\end{aligned}$$ Note that Algorithm \[alg:alg\] has entry sample complexity $\tilde O(\sqrt{k \log \frac{d}{\epsilon}} \cdot \log^2 \frac{d}{k})$ (see Corollary \[cor:cyclic\]) and vector sample complexity $n$.
[**input**]{}: i.i.d. samples $x^{(1)},\ldots,x^{(n)} \sim \mathcal{D}$ with Toeplitz covariance $T$. Random ultra-sparse ruler $R$ from Corollary \[cor:cyclic\] with parameters $d,k$, and $\delta = \frac{\epsilon}{\sqrt{k}}$.\
[**output**]{}: $\tilde T \in {\mathbb{R}}^{d \times d}$ approximating $T$.
\
\[alg:alg\]
We apply Theorem \[thm:sft\] to input $w$ with $\delta = \epsilon/\sqrt{d}$, which requires reading $m =O \left (k \cdot \log^2(\frac{d}{k}) \cdot \log(\frac{d}{\epsilon} )\right) $ entries of $w$. Since $\mathcal{D}$ is sub-Gaussian and since $t_0$ is the largest entry to $T$ by positive semidefiniteness, for $n = O \left (\frac{m \sqrt{\log m}}{\epsilon^2} \right )$, we have $|t_s - \bar t_s| \le \frac{\epsilon t_0}{\sqrt{m}}$ for each $s$ measured by $R$ with probability $1/\Theta(m)$. By a union bound, this approximation then holds for all $s$ with good probability. We thus have ${\|w - t\|}_2 \le \sqrt{m \cdot \frac{\epsilon^2 t_0^2}{m}} = \epsilon t_0$. By Parseval’s theorem, it follows that ${\|\hat w - \hat t\|}_2 \le \epsilon t_0$, and by the triangle inequality: $$\min_{\text{k-sparse }y}|| \hat{w}-y||_2 \le ||\hat{w}-\hat{t}|| + \min_{\text{k-sparse }y}|| \hat{t}-y||_2 \le \epsilon t_0 + \min_{\text{k-sparse }y}|| \hat{t}-y||_2.$$ By Theorem \[thm:sft\] with $\delta = \epsilon/\sqrt{d}$: $$\begin{aligned}
{\|\hat z -
\hat w\|}_2 \le 2\min_{\text{k-sparse }y}||\hat{w}-y|| + \delta ||\hat{w}|| \leq 2\epsilon t_0 + 2 \min_{\text{k-sparse }y}|| \hat{t}-y||_2 + \frac{\epsilon}{\sqrt{d}} ({\|\hat t\|}_2 + \epsilon t_0).\end{aligned}$$ Again noting that by positive semidefiniteness, $t_0$ is the largest entry in $t$, we have $\frac{\epsilon}{\sqrt{d}} {\|\hat t\|}_2 = \frac{\epsilon}{\sqrt{d}} {\|t\|}_2 \le \epsilon t_0$. This gives ${\|\hat z - \hat w\|}_2 \le 4 \epsilon t_0 + 2 \min_{\text{k-sparse }y}|| \hat{t}-y||_2$. The claim follows by applying the triangle inequality one more time to bound ${\|\hat z - \hat t\|}_2 \le {\|\hat z - \hat w\|}_2 + {\|\hat w - \hat t\|}_2 \le {\|\hat z - \hat w\|}_2 + \epsilon t_0$.
Finally, we prove Theorem \[thm:main\] by using the above bound on ${\|\hat z -
\hat t\|}_2$ to bound ${\|T - \tilde T\|}_F = {\|{\mathrm{Toep}}(t)-{\mathrm{Toep}}(z)\|}_F$.
If the bound of Lemma \[lem:5\] holds: $
{\|T - \tilde T\|}_F\le 5\epsilon {\|T\|}_F + 2 \min_{\operatorname{rank}-k\ B} {\|T - B\|}_F
$
Both ${\mathrm{Toep}}(t)$ and ${\mathrm{Toep}}(z)$ are circulant, and so can be written in their eigendecompositions as $F D F^*$ where $D = \sqrt{d} \operatorname{diag}(F^* t) = \sqrt{d} \operatorname{diag}(\hat t)$ and $\tilde D = \sqrt{d} \operatorname{diag}(F^* z) = \sqrt{d}\operatorname{diag}(\hat z)$. Thus: $$\begin{aligned}
{\|{\mathrm{Toep}}(t) - {\mathrm{Toep}}(z)\|}_F &= {\|FDF^* - F\tilde{D}F^*\|}_F\\
&= {\| D - \tilde D\|}_F \\
&= \sqrt{d}{\|\operatorname{diag}(F^*t)-\operatorname{diag}(F^*z)\|}\\
&= \sqrt{d}{\|\hat t- \hat z\|}_2 \le \sqrt{d} \cdot 5 \epsilon \cdot t_0 + \sqrt{d} \cdot 2 \min_{\text{k-sparse }y}|| \hat{t}-y||_2,\end{aligned}$$ where the last bound follows from Lemma \[lem:5\]. We can see that $\sqrt{d} \cdot 5\epsilon t_0 \le 5\epsilon {\|T\|}_F$. Further, the best rank-$k$ approximation of $T$ is given by projecting onto its top $k$-eigenvectors (equivalently, setting to zero all but the largest $k$ entries of $D$ to obtain $D_k$, or approximating $\hat t$ with its best $k$-sparse approximation, $t_k$). We thus have:
$$\begin{aligned}
\sqrt{d} \cdot \min_{\text{k-sparse }y}|| \hat{t}-y||_2 = \sqrt{d} \cdot ||\hat{t}-\hat{t_k}|| = {\|D-D_k\|} = {\|FDF^*-FD_kF^*\|} = \min_{\operatorname{rank}-k\ B} {\|T - B\|}_F,\end{aligned}$$
which yields $${\|{\mathrm{Toep}}(t) - {\mathrm{Toep}}(z)\|}_F \le 5\epsilon {\|T\|}_F + 2 \min_{\operatorname{rank}-k\ B} {\|T - B\|}_F,$$ completing the proof.
Experimental Validation {#sec:experiments}
=======================
We conclude by experimentally evaluating the driving intuition behind random ultra-sparse rulers: when there is a small frequency gap in $T$’s Vandermonde decomposition, it can be very advantageous to randomly permute the frequencies to remove this gap. To do so, we generate a low-rank, positive semidefinite real Toeplitz matrix with on-grid but clustered frequencies, add entrywise noise $\eta \sim N(0,\nu)$ and apply the following simple reconstruction procedure. Given a subset of noise-corrupted measurements of $T$’s first column $t$, we use the `pmusic` and `findpeaks` functions in Matlab to identify $k$ estimated frequencies $\tilde{f}_1,\dots,\tilde{f}_{k}$, and solve the appropriate linear regression problem to recover diagonal $\tilde D$ so that $T$ is approximated by $\tilde T = \tilde{F}\tilde D\tilde{F}^*$, where $\tilde{F}$ is the $n \times k$ Fourier matrix corresponding to $\tilde{f}_1,\dots,\tilde{f}_k$. We note that this simple reconstruction approach matches that of [@qiao2017gridless] *up to a preliminary denoising step*. This step could be applied to all sampling schemes and should preserve their relative performance. The sampling schemes compared are:
1. *First $O(k)$ samples*: Input to `pmusic` $4k$ noisy estimates of $t_0,\dots,t_{4k-1}$, which can be measured from samples $x^{(i)} \sim \mathcal{D}$ via existing ruler constructions with entrywise sample complexity $O(\sqrt{k})$. This corresponds to the approach of [@qiao2017gridless].
2. *Permuted $O(k)$ samples*: Input to `pmusic` $4k$ noisy samples of $t_{g^{-1}_{a,c}(0)},\ldots,t_{g^{-1}_{a,c}(4k-1)}$, where $g_{a,c}$ is as described in Section \[sec:contributions\]. For simplicity we take $c = 0$. These samples can be obtained with entrywise sample complexity $\tilde{O}(\sqrt{k})$ using a random ultra-sparse ruler by Lemma \[clm:cyclic\].
3. *All samples*: As a baseline, input to `pmusic` all $d$ noisy measurements of $t_0,\dots,t_{d-1}$.
Experimental results and validation are shown in Figure 2. As expected, sampling scheme (3) (which requires $O(\sqrt{d})$ entrywise sample complexity) performs best, but is closely followed by our proposed permutation-based sampling method. More elaborate reconstructions following the full algorithm of [@sft] would likely improve further on this simple algorithm. Nonetheless, it is clear that when T is circulant with some frequencies clustered, the permutation approach enabled by random ultra-sparse rulers can vastly improve robustness to noise while retaining low, $\tilde{O}(\sqrt{k})$ entrywise sample complexity.
\[fig:results\]
![Normalized estimation error $\frac{||t-\tilde{t}||_2}{||t||_2}$ of different ruler-enabled sampling schemes as a function of the variance $\nu$. Here, $k=6$ and $d=2400$, with minimum frequency gap $\approxeq 0.01 < \frac{1}{k}$. Results are averaged over 10 random permutations, each of which is further averaged over 20 trials. In subfigures (b) and (c), we demonstrate for a single iteration at $\nu=0.5$ how nearby frequencies are conflated without permutation, but likely to be separated and accurately identified with a permutation. As the frequencies are symmetric (to ensure T is real), only the first $\frac{k}{2}=3$ are shown.[]{data-label="fig:res"}](finalfiga){width="10cm"}
\(a) Estimation error as a function of noise variance.
![Normalized estimation error $\frac{||t-\tilde{t}||_2}{||t||_2}$ of different ruler-enabled sampling schemes as a function of the variance $\nu$. Here, $k=6$ and $d=2400$, with minimum frequency gap $\approxeq 0.01 < \frac{1}{k}$. Results are averaged over 10 random permutations, each of which is further averaged over 20 trials. In subfigures (b) and (c), we demonstrate for a single iteration at $\nu=0.5$ how nearby frequencies are conflated without permutation, but likely to be separated and accurately identified with a permutation. As the frequencies are symmetric (to ensure T is real), only the first $\frac{k}{2}=3$ are shown.[]{data-label="fig:res"}](figb){width="7cm"}
\(b) With permutation
![Normalized estimation error $\frac{||t-\tilde{t}||_2}{||t||_2}$ of different ruler-enabled sampling schemes as a function of the variance $\nu$. Here, $k=6$ and $d=2400$, with minimum frequency gap $\approxeq 0.01 < \frac{1}{k}$. Results are averaged over 10 random permutations, each of which is further averaged over 20 trials. In subfigures (b) and (c), we demonstrate for a single iteration at $\nu=0.5$ how nearby frequencies are conflated without permutation, but likely to be separated and accurately identified with a permutation. As the frequencies are symmetric (to ensure T is real), only the first $\frac{k}{2}=3$ are shown.[]{data-label="fig:res"}](figc){width="7cm"}
\(c) Without permutation
Acknowledgements {#acknowledgements .unnumbered}
----------------
We thank Yonina Eldar for many valuable conversations on sparse rulers and covariance estimation. We also thank Piya Pal for bringing the related work of [@qiao2017gridless] to our attention.
[^1]: $\tilde O(\cdot)$ hides log factors in the input parameters. For more precise bounds see Section \[sec:analysis\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The formula for the lens is derived based on the information of instantaneous focal function. Focal function is an important tool in designing lenses with extended depth of focus (EDoF) because this allows **EDoF** lens designers to try out various mathematical curves using computers to optimize their design. Once an optimal focal function information is obtained, the corresponding physical **EDoF** lens can be fabricated using the lens equation formulated in this presentation.'
author:
- Sung Nae Cho
date: 'Prepared on January 30, 2009'
title: Lens design based on instantaneous focal function
---
Introduction
============
Optical imaging system with large depth of field (**DOF**) is required to produce sharp images@DOF_wikipedia_site. In photography, the **DOF** is the portion of a scene that appears sharp in the image, for example, the region denoted by **A** in Fig. \[fig:DOF\]. Ordinarily, a lens focuses parallel rays of light at one distance known as the focal point, as illustrated in Fig. \[fig:ordinary\_lens\]. Therefore, not all points within the **DOF** can be claimed as focused per se. However, due to a gradual decrease in the sharpness of the image from the focused spot, the amount of blurring within the **DOF** is imperceptible to human eyes under normal viewing conditions. As such, in particularly for films and photography, the image region can be subdivided into two, where one lies within and the other lies external to the **DOF**. In the photograph of Fig. \[fig:DOF\], the region **A**, wherein the image appears sharp and focused, is said to lie within the **DOF**; whereas the region **B**, in which region the image is blurred, is said to lie external to the **DOF**.
![\[fig:DOF\] Illustration of depth of field. Region **A** is within the depth of field (**DOF**), whereas the region **B** is outside of **DOF**. The image is blurred drastically in region **B**.](hyun)
Alternatively, but equivalently, the **DOF** in an imaging system is defined as the distance in the object space in which objects are considered to be in focus. The distance over which objects appear sharp can be increased by extending the **DOF** of an imaging system. Traditionally, the **DOF** of an imaging system can be increased by either decreasing the size of lens aperture or by increasing the shutter speed, or through tweaking of the both. These methods, however, drastically reduces the amount of light passing through the lens and require extra lighting. For developing still images, the required extra lighting can be accommodated by the use of a flash. For motion pictures, however, this approach proves to be inadequate, as typical video involves about thirty frames of images per second.[^1] In addition to this difficulty, the smaller lens aperture increases diffraction and this places a practical limit on the extent to which the **DOF** of an optical imaging system can be enhanced by the aforementioned methods. That being said, can the **DOF** of an imaging system be increased (A) without sacrificing the intensity of light passing through the lens and (B) without increasing the diffraction? The answer to this question is yes and this approach involves some sort of digital filtering.
![\[fig:ordinary\_lens\] An ideal ordinary lens is characterized by a single focal point.](Ordinary_Lens)
The digital filtering method requires a scheme for the image reconstruction algorithm based on the principles of wave optics@Portney [@Mendlovic; @Ben-Eliezer; @Dowski; @Bradburn; @Bradburn2]. The image reconstruction code is often hardcoded in the accompanying digital processing unit (**DPU**) and, for this reason, the imaging method based on digital filtering is coined as the software assisted imaging technology or **SAIT** for short.[^2] The **SAIT** solution for an imaging system is a promising technology in that it has potential to increase the **DOF** while encompassing altogether the processes of which (A) drastically reduce the amount of light passing through the lens and (B) those problems associated with increased diffraction due to the reduced diameter of the lens aperture.
The software assisted imaging technology is not as perfect as it sounds and it has some of its own problems to offer. Among them, one directly relates to the image processing speed. In the traditional imaging solutions based on lenses, which I refer to as the analog technology, complex arrangement of lenses function to focus image at the focal plane from wherein the image is developed. As such systems based on analog technology by-pass the digital filtering stage altogether, images are developed instantaneously in analog imaging systems. However, due to the aforementioned problems and limitations of the traditional method for increasing the **DOF**, the analog technology provides only a limited solution when concerning the image quality. The systems based on **SAIT** principle face problems that are exactly the opposite in nature from that of the analog systems. In principle, the reconstructed image by a way of image reconstruction algorithm can be made to resemble the original image to any level or degree of resemblance, provided the image processing speed is of no concern. But, such methodology would limit systems based on **SAIT** to still imagery applications and the video sector of the market must be discarded, which is a bad idea for business. Naturally, for systems based on **SAIT** principle, a trade off must be made between the image quality and the image processing speed.
Among the early pioneers to successfully commercialize imaging system based on **SAIT** are Cathay and Dowski, who did much of their work at the University of Colorado@Dowski [@Bradburn; @Bradburn2]. In the modern literature, their work is cited as “wavefront coding.” The **SAIT** solution based on wavefront coding has been trade marked by CDM Optics, Inc., and it is known as the $\textnormal{Wavefront }\textnormal{Coding}^{\textup{TM}}.$ The imaging solution based on wavefront coding basically involves two stages: the input and the output stages. The input stage involves the optical element and this represents the hardware contribution side of the **SAIT** solution. The output stage involves the **DPU**, wherein the image reconstruction code base is hardcoded, and this represents the software contribution side of the **SAIT** solution. The optical element in **SAIT** solution is distinguished from the input stage of traditional imaging system, which is just complex series of ordinary lenses, in that it produces many focal points along the optical axis instead of just one at the focal plane. Such optical element is referred to as lens having extended depth of focus (or **EDoF** for short) and this is illustrated in Fig. \[fig:EDoF\_lens\].
Both the quality of reconstructed image and the image reconstruction speed are critically important in **SAIT**. In principle, the ordinary lens, such as the one illustrated in Fig. \[fig:ordinary\_lens\], can just as well serve as the input stage for the **SAIT** imaging systems. However, this must be done at the cost of overly complicated algorithm routines for the software side of the system. The length of image reconstruction code directly relates to the number of transistors in a **DPU**. As a simple rule of a thumb, more transistors there are in a **DPU**, more energy it requires to operate. And, more lines of coding for the image reconstruction algorithm implies the slower processing speed for the reconstructed image output. The full-**HD** quality of a video involves sixty image frames every second. This implies, the application of SAIT system to full-HD video processing would require the image reconstruction time span of $16\,\textup{ms}$ or less. The demand for very fast processing speed and low power consumption make ordinary lenses inadequate for the input stage of the **SAIT** system, which leave open for an alternate solution for the optical element to be used as the input stage.
![\[fig:EDoF\_lens\] Lens with extended depth of focus has many focal points, $f_{1},$ $f_{2},$ $f_{3},$ and so on. An ideal **EDoF** lens has infinitely many focal points and all light rays are confined within the cylindrical tube very small diameter $w$ and very long length $l.$ ](EDoF_lens)
The image reconstruction code base can be optimized for image processing speed and quality if the **EDoF** lens is used as the optical element for the input stage of **SAIT** system. The **EDoF** lens is characterized by parameters $l$ and $w.$ The parameter $l$ represents the depth of focal points along the optical axis and the parameter $w$ represents the width of bundle containing light rays associated with each focal points, as illustrated in Fig. \[fig:EDoF\_lens\]. The ideal EDoF lens has the parameter $l$ of which is infinite in length and the parameter $w$ of which is infinitely thin. For the realistic **EDoF** lenses, however, the parameters $l$ and $w$ are typically on the order of microns.
The idea of **EDoF** lens as an optical element which focuses light into longitudinally directed line along the optical axis was first proposed by Golub, et. al. @Golub. The idea was later adopted by others and it has found applications in various imaging systems, such as microscopes, cameras, and lithography to list a few, @Portney [@Mendlovic; @Ben-Eliezer; @Dowski; @Forster; @Liu; @Getman]. Alexander and Lukyanov have recently proposed a conceptual scheme for the **EDoF** lens@Getman. Their scheme for **EDoF** lens consists of zones that are axially symmetric about the optical axis and this is illustrated in Fig. \[fig:Getman\_EDoF\_lens\]. The idea behind their concept is as follows. The light ray crossing each zone gets focused to a unique spot on the optical axis and this spot is within in the **EDoF** lens parameter $l.$ If $\beta_{i}$ is the function which describes the focal length for the $i\,\textup{th}$ concentric zone in Fig. \[fig:Getman\_EDoF\_lens\], the lens aperture, in principle, can be tailored to behave like **EDoF** lens by tweaking $\beta_{i}.$ Borrowing their terminology, the $\beta_{i}$ is referred to as the instantaneous focal function. The idea behind **EDoF** lens is to make $l$ as large and $w$ as small as possible. By experimenting with different functions for $\beta_{i},$ the **EDoF** lens parameters $l$ and $w$ can be engineered to the acceptable range for the imaging systems based on **SAIT** principle. This is exactly just what Alexander and Lukyanov did, and their focal function is summarized in Fig. \[fig:lens-data-Getman\]. In the figure, the vertical lines represent discontinuities and each of the twelve zones has been indicated appropriately by a number. For each of the zones in Fig. \[fig:Getman\_EDoF\_lens\], the incident parallel rays are focused at different points on the optical axis within $l;$ and, the focusing is described by the instantaneous focal function, $\beta=\beta_{i},$ summarized in Fig. \[fig:lens-data-Getman\].
![\[fig:Getman\_EDoF\_lens\] Schematic of conceptual **EDoF** lens proposed by Alexander and Lukyanov. The lens aperture has an axial symmetry about the optical axis.](concentric_lens)
The imaging system based on **SAIT** is a package solution in which both the hardware (optical element or **EDoF** lens) and the software (image reconstruction code base) contributions must be optimized. Alexander and Lukyanov experimented with various mathematical curves for the focal function $\beta$ in an attempt to make the **EDoF** lens parameters $l$ and $w$ as ideal as possible, but did not attempt to provide any solutions concerning the physical shape or the profile for their conceptual **EDoF** lens. To test the image reconstruction code for the processing speed and the quality of generated image output, the information on the point spread function (or **PSF** for short) of the optical element input stage is required. Since no information on the physical profile of their conceptual **EDoF** lens was available, the test for their image reconstruction algorithm had to be deferred. It was my job to design a physical **EDoF** lens. Since Alexander’s image reconstruction code was based on the input from a lens aperture satisfying the focal function described in Fig. \[fig:lens-data-Getman\], the physical **EDoF** lens to be designed had a constraint of satisfying the same focal function characteristics. To end the story, the physical **EDoF** lens with such characteristics for the focusing behavior was found@Cho. The obtained physical profile of lens was entered into CODE V[^3] to generate the needed **PSF** information for the lens aperture.[^4] This information was in turn used by Alexander to test for the performance of his algorithm.
![\[fig:lens-data-Getman\] Focal function $\beta$ proposed by Alexander and Lukyanov @Getman. ](lensdata_0)
This work concerns the result of my role in the project, which was to design a physical **EDoF** lens from the focal function characteristics. As such, this work concerns the hardware contribution side of the optical imaging system based on **SAIT**.
Theory
======
Axial symmetry
--------------
By definition, an axial symmetry is a symmetry about a given given axis. The object has an axial symmetry if its appearance is unchanged with the rotation about some axis. Illustrated in Fig. \[fig:3D\_sphere\] is a schematic of conceptual lens, which shows an axial symmetry about the optical axis, i.e., the $x$ axis. Such a lens can be dissected through the origin with the $xy$ plane as shown in Fig. \[fig:3D\_sphere\]. The curve traced on the $xy$ plane, which is a set of points on the surface of lens, can be revolved about the optical axis for the three dimensional shape of the lens. The design of axially symmetric lens, therefore, simplifies to the problem of finding the set of points on the surface of lens of which gets traced on the $xy$plane.
![\[fig:3D\_sphere\] Schematic of lens symmetric about optical axis. ](3D_sphere)
The physical law of which governs the bending of light is the Snell’s law. I shall apply Snell’s principle to derive the equation for the cross-sectional profile of the lens. The terminology, “cross-sectional profile of the lens,” implies the lens curvature on the $xy$ plane, which is illustrated in Fig. \[fig:3D\_sphere\].
Derivation from Snell’s law
---------------------------
When a ray of light passes across media of different refractive indices, its path is governed by the Snell’s law, $$n_{\phi}\sin\phi=n_{\theta}\sin\theta,\label{eq:Snell-law}$$ as illustrated in Fig. \[fig:Law\_of\_refraction\]. Here, $n_{\phi}\equiv n_{\phi}\left(\omega\right)$ and $n_{\theta}\equiv n_{\theta}\left(\omega\right)$ are frequency dependent refractive indices with $\omega$ denoting the angular frequency of the light. The parameters $\phi$ and $\theta$ represent the angle of incidence and angle of refraction, respectively.
If $\mathbf{N}$ denotes the normal vector to the local point $y=\gamma$ on the curve $x=h\left(y\right),$ then it can be shown $$\left\Vert -\mathbf{N}\times\left(-\mathbf{e}_{1}\right)\right\Vert =\left\Vert -\mathbf{N}\right\Vert \left\Vert -\mathbf{e}_{1}\right\Vert \sin\phi=N\sin\phi$$ and the expression for $\sin\phi$ becomes $$\begin{aligned}
\sin\phi=\frac{\left\Vert \mathbf{N}\times\mathbf{e}_{1}\right\Vert }{N}, & & N\equiv\left\Vert \mathbf{N}\right\Vert ,\label{eq:sin-of-phi}\end{aligned}$$ where $\mathbf{e}_{1}$ is the unit basis for the $x$ axis.
Similarly, the expression for $\sin\theta$ may be obtained by considering vectors $\mathbf{A},$ $\mathbf{B},$ and $\mathbf{C}$ of Fig. \[fig:Law\_of\_refraction\]. The vectors $\mathbf{A},$ $\mathbf{B},$ and $\mathbf{C}$ satisfy the relation, $$\mathbf{A}+\mathbf{B}=\mathbf{C}.\label{eq:C_pre1}$$ In explicit form, vectors $\mathbf{A}$ and $\mathbf{B}$ are defined as $$\mathbf{A}=-\gamma\mathbf{e}_{2},\quad\mathbf{B}=\left(\beta-\alpha\right)\mathbf{e}_{1},\label{eq:C_pre2}$$ where $\mathbf{e}_{2}$ is the unit basis for the $y$ axis. With Eqs. (\[eq:C\_pre1\]) and (\[eq:C\_pre2\]), the vector $\mathbf{C}$ becomes $$\mathbf{C}=\left(\beta-\alpha\right)\mathbf{e}_{1}-\gamma\mathbf{e}_{2}.\label{eq:C}$$ The vector cross product $\mathbf{N}\times\mathbf{C}$ is given by $$\mathbf{N}\times\mathbf{C}=\left(\beta-\alpha\right)\mathbf{N}\times\mathbf{e}_{1}-\gamma\mathbf{N}\times\mathbf{e}_{2}$$ and its magnitude becomes $$\begin{aligned}
\left\Vert \mathbf{N}\times\mathbf{C}\right\Vert & =\left\Vert \left(\beta-\alpha\right)\mathbf{N}\times\mathbf{e}_{1}-\gamma\mathbf{N}\times\mathbf{e}_{2}\right\Vert \nonumber \\
& =NC\sin\theta,\label{eq:sin-of-theta-pre}\end{aligned}$$ where $N\equiv\left\Vert \mathbf{N}\right\Vert $ and $C=\left\Vert \mathbf{C}\right\Vert .$ Utilizing Eq. (\[eq:C\]), $C$ may be expressed as $$C=\left(\mathbf{C}\cdot\mathbf{C}\right)^{1/2}=\left[\left(\beta-\alpha\right)^{2}+\gamma^{2}\right]^{1/2}$$ and the Eq. (\[eq:sin-of-theta-pre\]) is solved for $\sin\theta$ to yield $$\sin\theta=\frac{\left\Vert \left(\beta-\alpha\right)\mathbf{N}\times\mathbf{e}_{1}-\gamma\mathbf{N}\times\mathbf{e}_{2}\right\Vert }{N\left[\left(\beta-\alpha\right)^{2}+\gamma^{2}\right]^{1/2}}.\label{eq:sin-of-theta}$$
![\[fig:Law\_of\_refraction\] Illustration of Snell’s law. The scalar quantity $\beta$ denotes the focal length. ](x_vs_y_vector)
Insertion of Eqs. (\[eq:sin-of-phi\]) and (\[eq:sin-of-theta\]) into the Snell’s law of Eq. (\[eq:Snell-law\]) gives $$\frac{n_{\phi}}{n_{\theta}}=\frac{\left\Vert \left(\beta-\alpha\right)\mathbf{N}\times\mathbf{e}_{1}-\gamma\mathbf{N}\times\mathbf{e}_{2}\right\Vert }{\left\Vert \mathbf{N}\times\mathbf{e}_{1}\right\Vert \left[\left(\beta-\alpha\right)^{2}+\gamma^{2}\right]^{1/2}}.\label{eq:Snell-alt-1}$$ By definition, the normal vector $\mathbf{N}$ satisfies the relation, $$g\left(x,y\right)=x-h\left(y\right),$$ where $g\left(x,y\right)$ is a function whose gradient gives $\mathbf{N},$ $$\mathbf{N}=\nabla g=\frac{\partial g}{\partial x}\mathbf{e}_{1}+\frac{\partial g}{\partial y}\mathbf{e}_{2}=\mathbf{e}_{1}-\frac{\partial h}{\partial y}\mathbf{e}_{2}.$$ Because $\mathbf{N}$ is the normal vector at the location $\left(x=\alpha,y=\gamma\right),$ I write $$\mathbf{N}=\mathbf{e}_{1}-\left.\frac{\partial h}{\partial y}\right|_{y=\gamma}\mathbf{e}_{2}.\label{eq:N-explicit-xa}$$ The following vector cross products are valid, $$\begin{aligned}
\mathbf{N}\times\mathbf{e}_{1} & =\mathbf{e}_{1}\times\mathbf{e}_{1}-\left.\frac{\partial h}{\partial y}\right|_{y=\gamma}\mathbf{e}_{2}\times\mathbf{e}_{1},\\
\mathbf{N}\times\mathbf{e}_{2} & =\mathbf{e}_{1}\times\mathbf{e}_{2}-\left.\frac{\partial h}{\partial y}\right|_{y=\gamma}\mathbf{e}_{2}\times\mathbf{e}_{2},\end{aligned}$$ where Eq. (\[eq:N-explicit-xa\]) was used to replace $\mathbf{N}.$ Since $\mathbf{e}_{1}\times\mathbf{e}_{1}=\mathbf{e}_{2}\times\mathbf{e}_{2}=0,$ the previous relations reduce to $$\mathbf{N}\times\mathbf{e}_{1}=\left.\frac{\partial h}{\partial y}\right|_{y=\gamma}\mathbf{e}_{3},\quad\mathbf{N}\times\mathbf{e}_{2}=\mathbf{e}_{3},\label{eq:N-cross-ei}$$ where $\mathbf{e}_{3}$ is the unit basis for the $z$ axis of which satisfies the relation, $$\mathbf{e}_{1}\times\mathbf{e}_{2}=\mathbf{e}_{3},\quad\mathbf{e}_{2}\times\mathbf{e}_{1}=-\mathbf{e}_{3}.$$ Insertion of Eq. (\[eq:N-cross-ei\]) into Eq. (\[eq:Snell-alt-1\]) gives $$\frac{n_{\phi}}{n_{\theta}}=\frac{\left(\beta-\alpha\right)\left.\frac{\partial h}{\partial y}\right|_{y=\gamma}-\gamma}{\left.\frac{\partial h}{\partial y}\right|_{y=\gamma}\left[\left(\beta-\alpha\right)^{2}+\gamma^{2}\right]^{1/2}},$$ which expression can be rearranged to yield $$\left.\frac{\partial h}{\partial y}\right|_{y=\gamma}=\frac{\gamma}{\beta-\alpha-\frac{n_{\phi}}{n_{\theta}}\left[\left(\beta-\alpha\right)^{2}+\gamma^{2}\right]^{1/2}},\label{eq:Snell-alt-2}$$ where $\alpha$ and $\gamma$ are constants of which are depicted in Fig. \[fig:Law\_of\_refraction\].
For Alexander and Lukyanov’s optical element, the instantaneous focal function $\beta\equiv\beta\left(y\right)$ in Eq. (\[eq:Snell-alt-2\]) is as defined in Fig. \[fig:lens-data-Getman\]. The $\gamma$ for the $y$ axis is not anything special, of course. Any $y$ belonging to the domain of $h$ satisfies the Eq. (\[eq:Snell-alt-2\]). The generalization of Eq. (\[eq:Snell-alt-2\]) for all $y$ belonging to the domain of $h$ is done by making the following replacements: $$\alpha\rightarrow x,\quad\gamma\rightarrow y,\quad\left.\frac{\partial h}{\partial y}\right|_{y=\gamma}\rightarrow\frac{\partial h}{\partial y}=\frac{dx}{dy}.$$ With these replacements, Eq. (\[eq:Snell-alt-2\]) gets re-expressed in form as $$\frac{dx}{dy}=\frac{y}{\beta-x-\frac{n_{\phi}}{n_{\theta}}\left[\left(\beta-x\right)^{2}+y^{2}\right]^{1/2}}.\label{eq:dxdy0}$$
How is the instantaneous focal function, $\beta,$ restricted? The $\beta$ in Eq. (\[eq:dxdy0\]) is restricted so that the expression for $dx/dy$ does not blow up. Equation (\[eq:dxdy0\]) is well defined if and only if the denominator satisfies the condition, $$\beta-x-\frac{n_{\phi}}{n_{\theta}}\left[\left(\beta-x\right)^{2}+y^{2}\right]^{1/2}\neq0.\label{eq:b-x-start-comp}$$ Contrarily, but equivalently, the previous statement can be reworded as follows. Equation (\[eq:dxdy0\]) is ill defined if and only if the denominator satisfies the condition, $$\beta-x-\frac{n_{\phi}}{n_{\theta}}\left[\left(\beta-x\right)^{2}+y^{2}\right]^{1/2}=0.\label{eq:b-x-start}$$ For reasons to follow, I shall proceed with the latter. To solve for $\beta,$ I shall first rearrange Eq. (\[eq:b-x-start\]) as $$\beta-x=\frac{n_{\phi}}{n_{\theta}}\left[\left(\beta-x\right)^{2}+y^{2}\right]^{1/2}.\label{eq:b-x-first}$$ Squaring of both sides give $$\left(\beta-x\right)^{2}=\frac{n_{\phi}^{2}}{n_{\theta}^{2}}\left(\beta-x\right)^{2}+\frac{n_{\phi}^{2}}{n_{\theta}^{2}}y^{2}.$$ This expression can be rearranged to become $$\left(\beta-x\right)^{2}\left(1-\frac{n_{\phi}^{2}}{n_{\theta}^{2}}\right)=\frac{n_{\phi}^{2}}{n_{\theta}^{2}}y^{2}$$ or $$\left(\beta-x\right)^{2}\left(\frac{n_{\theta}^{2}-n_{\phi}^{2}}{n_{\theta}^{2}}\right)=\frac{n_{\phi}^{2}}{n_{\theta}^{2}}y^{2}.$$ And, solving for $\left(\beta-x\right),$ I obtain $$\begin{aligned}
\beta-x & =\pm\frac{n_{\phi}y}{\sqrt{n_{\theta}^{2}-n_{\phi}^{2}}},\label{eq:b-x-second}\end{aligned}$$ where the $\pm$ came from the action of taking the square root on both sides, of course. Now, one of the signs in Eq. (\[eq:b-x-second\]) can be eliminated by comparing with Eq. (\[eq:b-x-first\]). This is the reason why I proceeded with Eq. (\[eq:b-x-start\]) instead of Eq. (\[eq:b-x-start-comp\]). The $n_{\phi}$ and $n_{\theta}$ in Eq. (\[eq:b-x-first\]) are both real refractive indices, which cannot be negative numbers. The instantaneous focal function, $\beta,$ and the lens thickness, $x,$ must be real, which implies $\left[\left(\beta-x\right)^{2}+y^{2}\right]$ must be non-negative else $\left[\left(\beta-x\right)^{2}+y^{2}\right]^{1/2}$ becomes an imaginary term. As real refractive indices cannot be negative numbers, the term $\left(\beta-x\right)$ is also a non-negative real in Eq. (\[eq:b-x-first\]), provided $n_{\phi}>0,$ $n_{\theta}>0,$ and $\left[\left(\beta-x\right)^{2}+y^{2}\right]>0,$ of course. Therefore, the right hand side of Eq. (\[eq:b-x-second\]) must be positive; and, this gives $$\beta=x+\frac{n_{\phi}y}{\sqrt{n_{\theta}^{2}-n_{\phi}^{2}}}.$$ Now, this is precisely the condition for $\beta$ which makes Eq. (\[eq:dxdy0\]) ill defined. Equivalently, then Eq. (\[eq:dxdy0\]) becomes well behaved for $\beta$ satisfying the condition given by $$\beta\neq x+\frac{n_{\phi}y}{\sqrt{n_{\theta}^{2}-n_{\phi}^{2}}}.\label{eq:beta-restriction}$$ Equation (\[eq:beta-restriction\]) defines the restriction for the instantaneous focal function, $\beta.$
What can be concluded of the restriction so defined in Eq. (\[eq:beta-restriction\]) for the instantaneous focal function? To answer this, recall that terms such as $\beta,$ $x,$ $y,$ $n_{\phi},$ and $n_{\theta}$ are all real values. And, there are no restrictions on $n_{\phi}$ and $n_{\theta}$ to speak of which of the two must be bigger or smaller in value. Interesting per se, the choice of $n_{\phi}>n_{\theta}$ results in the statement, $$\begin{aligned}
\beta\neq x+\frac{in_{\phi}y}{\sqrt{n_{\phi}^{2}-n_{\theta}^{2}}}, & & n_{\phi}>n_{\theta},\label{eq:beta-rest-test}\end{aligned}$$ where the $i$ denotes the imaginary symbol and the term $\sqrt{n_{\theta}^{2}-n_{\phi}^{2}}$ in Eq. (\[eq:beta-restriction\]) has been modified to $\sqrt{n_{\phi}^{2}-n_{\theta}^{2}}.$ But, this condition defined in Eq. (\[eq:beta-rest-test\]) is always satisfied, as $\beta$ is a real function. Therefore, it is concluded that Eq. (\[eq:dxdy0\]) is well behaved everywhere for $n_{\phi}>n_{\theta}.$
Result
======
Lens surface equation
---------------------
The profile of axially symmetric lens about its optical axis is obtained by solving the initial-value differential equation, Eq. (\[eq:dxdy0\]), $$\frac{dx}{dy}=\frac{y}{\beta-x-\frac{n_{\phi}}{n_{\theta}}\left[\left(\beta-x\right)^{2}+y^{2}\right]^{1/2}},\quad x\left(y_{\textup{0}}\right)=x_{\textup{0}},$$ where $x\left(y_{\textup{0}}\right)=x_{\textup{0}}$ is the initial condition to be specified and the instantaneous focal function $\beta$ satisfies the constrain defined in Eq. (\[eq:beta-restriction\]). Without loss of generality, one may choose $x\left(y=y_{\textup{0}}=0\right)=0$ for the initial condition and the lens profile satisfies the differential equation, $$\begin{aligned}
\frac{dx}{dy} & =\frac{y}{\beta-x-\frac{n_{\phi}}{n_{\theta}}\left[\left(\beta-x\right)^{2}+y^{2}\right]^{1/2}},\nonumber \\
\label{eq:ODE}\\ & x\left(0\right)=0,\quad\beta\neq x+\frac{n_{\phi}y}{\sqrt{n_{\theta}^{2}-n_{\phi}^{2}}}.\nonumber \end{aligned}$$ The quantities $n_{\phi}$ and $n_{\theta}$ are the two refractive indices in which one represents the lens and the other representing the surrounding medium. Which of the two refractive indices corresponds to the lens depends on the configuration of the problem, as demonstrated in the proceeding sections.
Alexander and Lukyanov lens
---------------------------
### Instantaneous focal function
The instantaneous focal function proposed by Alexander and Lukyanov has been discussed previously in Fig. \[fig:lens-data-Getman\]. The instantaneous focal function for each of the twelve zones can be curve fitted and represented by a quadratic polynomial of the form given by $$\begin{aligned}
\beta\equiv\beta_{i}=ay_{i}^{2}+by_{i}+c, & & y_{i,\textup{min}}\leq y_{i}\leq y_{i,\textup{max}},\label{eq:beta-i}\end{aligned}$$ where the subscript $i$ of $\left(\beta_{i},\, y_{i},\, y_{i,\textup{min}},\, y_{i,\textup{max}}\right)$ denotes the $i\,\textup{th}$ concentric zone. The coefficients $a,$ $b,$ and $c,$ and the range for $y,$ which defines the width for each of the axially symmetric concentric zones, are summarized in Table \[tab:table1\]. Since the instantaneous focal function, $\beta,$ and the lens radius, $y,$ have units of length measured in meters $\left[\textup{m}\right],$ the coefficient $a$ must have a unit of $\left[\textup{m}^{-1}\right],$ $c$ a unit of $\left[\textup{m}\right],$ and $b$ must be a unit-less scalar. To reduce the width of the table, couple columns were represented in millimeter units, $\left[\textup{mm}\right].$
$y_{i,\textup{min}},\, y_{i,\textup{max}}\,$ [\[]{}mm\] $a\,$[\[]{}1/m\] $b$ $c\,$[\[]{}mm\]
--------------------------------------------------------- ------------------ --------- -----------------
0.0, 0.19182692 -313.07 0.0235 3.5137034
0.19519231, 0.27259615 534.53 -0.2472 3.527877626
0.27596154, 0.33317308 -309.02 0.0818 3.5088062
0.33653846, 0.38701923 536.05 -0.3275 3.5493232
0.39038462, 0.43413462 -306.12 0.1182 3.502912672
0.4375, 0.47451923 539.03 -0.3891 3.569538239
0.47788462, 0.51153846 -303.68 0.1463 3.496845208
0.51490385, 0.54855769 542.21 -0.4417 3.589312176
0.55192308, 0.58221154 -301.27 0.1695 3.49080193
0.58557692, 0.6125 545.96 -0.4895 3.609151039
0.61586538, 0.64278846 -298.81 0.1893 3.484870978
0.64615385, 0.67307692 179.08 -0.0474 3.469596542
: Domain $y_{i}$ and coefficients ($a,b,c$) of $\beta_{i}=ay_{i}^{2}+by_{i}+c$ for Fig. \[fig:lens-data-Getman\] \[tab:table1\]
### Lens profile
Equation (\[eq:ODE\]) was solved by the Runge-Kutta routine coded in **FORTRAN 90**@Derrick-Grossman. For the computation, refractive indices, $n_{\phi}$ and $n_{\theta},$ were chosen as follows, $$\begin{aligned}
n_{\phi}=1.5311, & & n_{\theta}=1.0.\end{aligned}$$ In this configuration, $n_{\phi}$ denotes the refractive index for the lens and $n_{\theta}$ denotes the refractive index for air.[^5] The resulting cross-sectional profile of the lens curvature is shown in Fig. \[fig:lens\_cross\_section\]. The profile in Fig. \[fig:lens\_cross\_section\] was revolved about the optical axis, which is the $x$ axis in the figure, to generate the three dimensional profile of the physical lens. This result is shown in Fig. \[fig:3D\_lens\_profile\], where the optical axis is located at $\left(\lambda y=200,\lambda z=200\right).$ The scaling factor of $\lambda=3.365385\times10^{-6}$ was introduced for graphing purpose only.
In spite of the non constant $\beta$ for the instantaneous focal function (see Fig. \[fig:lens-data-Getman\]), the resulting **EDoF** lens shown in Figs. \[fig:lens\_cross\_section\] and \[fig:3D\_lens\_profile\] seems to resemble the parabolic curve, which configuration is known to have only one focal point, i.e., $\beta=\textup{constant}.$@thomas-finney. Is the result portrayed in Fig. \[fig:lens\_cross\_section\] (or Fig. \[fig:3D\_lens\_profile\]) correct? To give a qualitative answer to this, I shall recall the **EDoF** lens parameter $l,$ which was previously illustrated in Fig. \[fig:EDoF\_lens\].
![\[fig:lens\_cross\_section\] The cross-sectional profile of lens curvature corresponding to the instantaneous focal function proposed by Alexander and Lukyanov, Fig. \[fig:lens-data-Getman\]. ](lens_profile_nt_air)
![\[fig:3D\_lens\_profile\] Three dimensional profile of lens satisfying the focal function $\beta$ proposed by Alexander and Lukyanov, Fig. \[fig:lens-data-Getman\]. The $y$ and $z$ axes have been multiplied by $\lambda=3.365385\times10^{-6}.$ ](3D_lens_profile)
In explicit form, the **EDoF** lens parameter $l$ is defined as $$\begin{aligned}
l=\left\Vert \beta_{\textup{max}}-\beta_{\textup{min}}\right\Vert , & & \left\{ \begin{array}{c}
\beta_{\textup{max}}>0,\\
\beta_{\textup{min}}>0,\end{array}\right.\label{eq:EDoF_parameter_L}\end{aligned}$$ where $\beta_{\textup{max}}$ and $\beta_{\textup{min}}$ denote, respectively, the maximum and the minimum values in the profile of instantaneous focal function. In the focal function profile of Alexander and Lukyanov, Fig. \[fig:lens-data-Getman\], $\beta_{\textup{max}}\approx3.5188\,\textup{mm}$ and $\beta_{\textup{min}}\approx3.4831\,\textup{mm}.$ Plugging the information into Eq. (\[eq:EDoF\_parameter\_L\]), this roughly gives $l\approx36\,\textup{um}.$ The resemblance of the **EDoF** lens to the parabolic curve can be attributed to the small value for $l.$ Considering that Alexander and Lukyanov’s lens has a maximum radius of $y\approx0.67\,\textup{mm},$ which can be identified from Table \[tab:table1\], the lens diameter comes out to be about $d=2y\approx1.34\,\textup{mm}.$ This implies, the lens diameter is larger than the **EDoF** lens parameter, $l,$ by a factor of thirty seven. Under such circumstance, the lens could be perceived as having a single focal point from the perspective of human eye. In spite of the existence of number of very closely spaced focal points within the length of $l$ along the optical axis, the human eyes do not have sufficient resolving power to distinguish those focal points. Even less so, the human eyes cannot distinguish the actual cross-sectional profile of the lens of which is only slightly perturbed from the cross-sectional profile of the parabolic lens. To prove that this is indeed the case, I shall use the derived lens formula, Eq. (\[eq:ODE\]), to generate various parabolic lenses.
Validation of the result
-------------------------
How does one know that Eq. (\[eq:ODE\]) generates the correct profile for the lens? The easiest way to settle this dilemma is to actually apply Eq. (\[eq:ODE\]) to the well known types, i.e., the parabolic lenses.
### Simple parabolic lens
The parabolic curves are known to merge parallel rays of incidence light to a unique focal point called a focus@thomas-finney. As a consequence of this, the parabolic curves flatten in the curvature with the focus positioned at distances further from the vertex. Such property of parabolic curves make them ideal for testing and validating the lens formula defined in Eq. (\[eq:ODE\]). The constant focus of $\beta=1\,\textup{m},$ $\beta=5\,\textup{m},$ and $\beta=10\,\textup{m}$ were considered to generate curves using Eq. (\[eq:ODE\]); and, the result is summarized in Fig. \[fig:lens\_cross\_section\_c1\_c2\_c3\]. As expected, the generated curves are that of parabolic curves in which each curves corresponds to focal points $\beta=1\,\textup{m},$ $\beta=5\,\textup{m},$ and $\beta=10\,\textup{m}.$ The curve corresponding to $\beta=10\,\textup{m}$ is more flat in curvature than the ones corresponding to $\beta=1\,\textup{m}$ or $\beta=5\,\textup{m},$ as expected. The result corresponding to $\beta=1\,\textup{m}$ was revolved about the optical axis to illustrate the three dimensional profile of the parabolic lens. This is shown in Fig. \[fig:3D\_lens\_profile\_c1\]. Again, in the figure, the optical axis is at $\left(\lambda y=200,\lambda z=200\right),$ where the scaling factor $\lambda$ is $\lambda=3.365385\times10^{-6}.$
![\[fig:lens\_cross\_section\_c1\_c2\_c3\] The cross-sectional profile of lens curvature for $\beta=1\,\textup{m},$ $\beta=5\,\textup{m},$ and $\beta=10\,\textup{m}.$](lens_profile_nt_air_c1_c2_c3)
![\[fig:3D\_lens\_profile\_c1\] Three dimensional profile of a lens with a constant instantaneous focal function, $\beta=1\,\textup{m}.$ The $y$ and $z$ axes have been multiplied by $\lambda=3.365385\times10^{-6}.$ ](result3D_c1)
### Concentric parabolic lens
Now I consider a slightly more complicated profile for the instantaneous focal function, $\beta.$ To show that, indeed, the likeliness of Alexander and Lukyanov’s **EDoF** lens to the parabolic lens is attributed to the small value for the **EDoF** lens parameter $l,$ I shall modify only the $\beta$ portion of Alexander and Lukyanov’s profile for the instantaneous focal function, while leaving the size of lens diameter unmodified. The modified focal function for this test is shown in Fig. \[fig:lens-data-0\_test\]. And, the coefficients $a,$ $b,$ and $c,$ and the range of $y$ corresponding to the $\beta\equiv\beta_{i}$ of Eq. (\[eq:beta-i\]) for each of the twelve zones is summarized in Table \[tab:table-test\]. Besides the increased focal length for each of twelve zones, the focal point is unique within each zones in this test configuration. In Alexander and Lukyanov’s focal profile, $l$ was much smaller than the lens diameter $d,$ i.e., $d\approx37l.$ The $\beta_{\textup{max}}$ and $\beta_{\textup{min}}$ for the test configuration are, respectively, $\beta_{\textup{max}}=2.5\,\textup{m}$ and $\beta_{\textup{min}}=0.01\,\textup{m},$ which can be verified from Table \[tab:table-test\]. Using the formula for $l$ defined in Eq. (\[eq:EDoF\_parameter\_L\]), this gives $l\approx2.49\,\textup{m}.$ Therefore, in this test configuration, $l$ is much larger than $d,$ i.e., $d\approx5.4\times10^{-4}l,$ which is just the opposite situation from that of Alexander and Lukyanov. That being said, the Eq. (\[eq:ODE\]) was plotted for the curve and the result is shown in Fig. \[fig:lens\_cross\_section\_test\]. As expected, the resulting cross-sectional profile for the lens does not resemble simple parabolic lens. However, the curve profile for each of the twelve zones in Fig. \[fig:lens\_cross\_section\_test\] represents the portion of a parabolic curve corresponding to $\beta_{i}$ illustrated in Fig. \[fig:lens-data-0\_test\]. By superimposing the two graphs, Figs. \[fig:lens-data-0\_test\] and \[fig:lens\_cross\_section\_test\], the boundaries for each zones can be identified by kinks in Fig. \[fig:lens\_cross\_section\_test\]. The cross-sectional profile of the lens curvature illustrated in Fig. \[fig:lens\_cross\_section\_test\] was revolved about the optical axis for the three dimensional profile of the lens. This result is shown in Fig. \[fig:3D\_lens\_profile\_test\].
![\[fig:lens-data-0\_test\] Focal function $\beta$ with characteristic of a step function. ](lensdata_0_test)
$y_{i,\textup{min}},\, y_{i,\textup{max}}\,$ [\[]{}mm\] $a\,$[\[]{}1/m\] $b$ $c\,$[\[]{}m\]
--------------------------------------------------------- ------------------ ----- ----------------
0.0, 0.19182692 0.0 0.0 0.25
0.19519231, 0.27259615 0.0 0.0 0.05
0.27596154, 0.33317308 0.0 0.0 0.50
0.33653846, 0.38701923 0.0 0.0 0.10
0.39038462, 0.43413462 0.0 0.0 1.00
0.4375, 0.47451923 0.0 0.0 0.05
0.47788462, 0.51153846 0.0 0.0 2.50
0.51490385, 0.54855769 0.0 0.0 0.10
0.55192308, 0.58221154 0.0 0.0 0.50
0.58557692, 0.6125 0.0 0.0 0.10
0.61586538, 0.64278846 0.0 0.0 0.01
0.64615385, 0.67307692 0.0 0.0 1.00
: Domain $y_{i}$ and coefficients ($a,b,c$) of $\beta_{i}=ay_{i}^{2}+by_{i}+c$ for Fig. \[fig:lens-data-0\_test\]\[tab:table-test\]
![\[fig:lens\_cross\_section\_test\] The cross-sectional profile of lens curvature corresponding to the focal function defined in Fig. \[fig:lens-data-0\_test\]. ](lens_profile_nt_air_test)
![\[fig:3D\_lens\_profile\_test\] Three dimensional profile of lens satisfying the focal function defined in Fig. \[fig:lens-data-0\_test\]. The $y$ and $z$ axes have been multiplied by $\lambda=3.365385\times10^{-6}.$ ](result3D_test)
Basing on these results, it can be concluded that for the case where the lens diameter $d$ is much larger than the **EDoF** lens parameter $l,$ the profile of the lens resembles closely the profile of simple parabolic lens. Here, the word “simple” has been used to denote the parabolic curve with single focus. In the opposite situation, where the lens diameter $d$ is much smaller than the **EDoF** lens parameter $l,$ the profile of the lens no longer resembles the simple parabolic lens. Instead, in this latter case, the shape for the lens resembles superimposed, multiple number of parabolic lenses of different degrees of curvature, as illustrated in Fig. \[fig:3D\_lens\_profile\_test\].
Concluding Remarks
==================
The image processing speed and the quality of processed images are both of critical importance in software assisted imaging technology. Such a requirement calls for the optimization of image reconstruction code based on the principles of wave optics. The coding side of the **SAIT** system can be optimized if **EDoF** lens is used for the input stage.
In this presentation, the formula for the **EDoF** lens has been derived based on the knowledge of instantaneous focal function, $\beta.$ The $\beta$ information is an important tool in the design of **EDoF** lens, as this allows optical engineer to try out various mathematical curves using computers for optimization. With the knowledge of $\beta,$ this can be achieved without having to actually make **EDoF** lens prototypes, thereby saving time and the cost. Once the optimal solution for the instantaneous focal function is obtained, the physical **EDoF** lens can be manufactured based on the lens formula presented in this work.
Acknowledgments
===============
I would like to thank G. Alexander for providing the raw data for the focal function. I would also like to thank Dr. S. Lee for generating the **PSF** information for the lens designed in this work using CODE V. The author acknowledges the support for this work provided by Samsung Electronics, Ltd.
[10]{} Wikipedia, “Depth of Field,” .
V. Portney, “Multifocal Ophthalmic Lens,” U.S. Patent 4898461 (1990).
D. Mendlovic, Z. Zalevsky, G. Shabtay, U. Levy, E. Marom, and N. Konforti, “Synthesis of light beams,” U.S. Patent (2002).
E. Ben-Eliezer, Z. Zalevsky, E. Maron, N. Konforti, and D. Mendlovic, “All optical extended ’depth-of-field’ imaging system,” U.S. Patent 7158317 (2007).
E. Dowski, “Wavefront coding optics,” U.S. Patent 6842297 (2005).
S. Bradburn, W. Cathey, E. Dowski, “Realization of focus invariance in optical-digital systems with wave-front coding,” Appl. Opt. **36**(35), pp. 9157-9166 (1997).
E. Dowski, Jr., and W. Cathey, “Extended depth of field through wave-front coding,” Appl. Opt. **34** (11), pp. 1859-1866 (1995).
M. Golub, D. Leonid, N. Kazanskiy, S. Kharitonov, I. Sisakian, and V. Soifer, “Focusators at letters diffraction design,” in *Proceedings of SPIE*, Vol. **1500**, 211 (1991).
B. Forster, D. Van De Ville, J. Berent, D. Sage, and M. Unser, “Extended Depth-of-Focus for Multi-Channel Microscopy Images: A Complex Wavelet Approach,” in *Proceedings of the Second IEEE International Symposium on Biomedical Imaging: From Nano to Macro (ISBI’04)*, (Arlington VA, USA, April 15-18, 2004), pp. 660-663.
Z. Liu, A. Flores, M. Wang, and J. Yang, “Diffractive infrared lens with extended depth of focus,” Optical Engineering **46**(1), 018002 (2007).
G. Alexander and A. Lukyanov, “Lens with extended depth of focus and optical system having the same,” Korean Patent 10-2008-0043428 (2008), .
S. Cho, “Method for designing physical lens from depth of focus characteristics and lens with extended depth of focus designed by the same method,” Korean Patent 10-2008-0111002 (2008), .
W. Derrick and S. Grossman, *A First Course in Differential Equations with Applications* (West Publishing Company, St. Paul, 1987).
G. Thomas and R. Finney, *Calculus and analytic geometry,* 7th Ed, ** (Addison-Wesley, USA, 1988).
[^1]: According to the industry standard of specifications set by the flat panel display consortium, the **1080P** specification of the full-**HD** quality of liquid crystal displays (**LCD**s) process sixty image frames per second.
[^2]: Unfortunately, the work place I am affiliated with, Samsung Advanced Institute of Technology, also uses the abbreviation “SAIT” for the name. To distinguish this SAIT from software assisted imaging technology, I denote the latter with bold faced version, **SAIT**.
[^3]: CODE V is an optical design program with graphical user interface for image forming and fiber optical systems by Optical Research Associates (ORA), an organization that has been supporting customer success for over 40 years, .
[^4]: Using CODE V to obtain **PSF** information turned out to be a nontrivial task, as the built-in curve fitting functions for CODE V were not too “happy” with the obtained result for the surface profile of **EDoF** lens.
[^5]: Although the actual value varies depending on the surrounding humidity and so on, the refractive index of an air is about $n\approx1.0008.$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show that the equivariant cohomology of any hyperpolar action of a compact and connected Lie group on a symmetric space of compact type is a Cohen-Macaulay ring. This generalizes some results previously obtained by the authors.'
address:
- |
Department of Mathematics and Computer Science\
Philipps-Universität Marburg\
Germany
- |
Institut für Geometrie und Topologie\
Universität Stuttgart\
Germany
- |
Department of Mathematics and Statistics\
University of Regina\
Canada
author:
- Oliver Goertsches
- Sam Hagh Shenas Noshari
- 'Augustin-Liviu Mare'
title: On the equivariant cohomology of hyperpolar actions on symmetric spaces
---
Introduction {#intro}
============
When investigating smooth group actions on manifolds, an insightful topological invariant one associates to it is its equivariant cohomology. (The coefficient ring throughout this paper is $\bQ$, unless otherwise specified.) This remains an effective tool even in cases when concrete descriptions are hard to obtain. Geometric features of the action are reflected by algebraic properties of its equivariant cohomology. Equivariant formality is an example of such a property, which is relevant due to its simple definition (see Sect. \[sec:general\]) and to the numerous situations when it is satisfied, such as Hamiltonian actions of compact Lie groups on compact symplectic manifolds (cf. [@Ki]) or isotropy actions of compact symmetric spaces (see [@Go]). A larger class consists of the so-called Cohen-Macaulay actions, which are defined in Sect. \[sec:general\] below. Their relevance in the context of equivariant topology was noticed by several authors already in the 70s and early 80s: for example, M. F. Atiyah used the Cohen-Macaulay condition to study equivariant $K$-theory for torus actions in [@Atiyah] and his ideas were adapted to equivariant cohomology by G. E. Bredon in [@Bredon]; see also the paper [@Duflot] by J. Duflot for actions of discrete groups. More recently, Lie group actions whose equivariant cohomology satisfies the requirement above have been considered in [@FranzPuppe2003], before being thoroughly investigated in [@Go-Ro] and [@Go-To]. They are natural generalizations of equivariantly formal group actions to the more general situation when all isotropy subgroups may have their ranks strictly smaller than the rank of the acting group. Cohen-Macaulay actions are worth studying due to their interesting features: for example, just like in the equivariantly formal case, a compact Lie group action and the induced action of a maximal torus satisfy simultaneously this requirement; furthermore, Cohen-Macaulay actions are characterized by the exactness of a certain Atiyah-Bredon long sequence; for more details we refer to Sect. \[cmsec\] and the appendix below. Also, this condition is satisfied in numerous concrete situations, for example when the cohomogeneity of the action is zero or one, as it was pointed out in [@Go-Ma].
Cohomogeneity-one actions are special cases of hyperpolar actions. Recall that an isometric action of a compact connected Lie group on a Riemannian manifold is [*polar*]{} if there exists a submanifold which is intersected by each orbit of the action, the orbit being orthogonal to the submanifold at each intersection point. If the submanifold can be chosen to be flat relative to the induced Riemannian metric, we say that the action is [*hyperpolar*]{}. In this paper we will be interested in the case when the manifold on which the group acts is a Riemannian symmetric space of compact type, that is, a quotient $G/H$, where $G$ is a compact connected semisimple Lie group and $H$ a closed subgroup such that $G^\sigma_0 \subset H \subset G^\sigma$, where $\sigma$ is an involutive automorphism of $G$, $G^\sigma$ its fixed point set and $G^\sigma_0$ the identity component of the latter group. Polar actions on compact symmetric spaces have been extensively investigated by many authors, see for instance [@HPTT1], [@HPTT2], [@Ko1], [@Ko2], [@Ko3], and [@Ko]. We will prove the following:
\[thm:realmain\] Any hyperpolar action of a compact connected Lie group on a symmetric space of compact type is Cohen-Macaulay.
The proof relies essentially on the classification of the actions mentioned in the theorem, which was obtained by A. Kollross in [@Ko]. Before stating this result, we need to describe an important class of examples of hyperpolar actions on the symmetric space $G/H$ mentioned above. Let $\tau$ be another involutive automorphism of $G$ and consider its fixed point set, $G^\tau$, along with the identity component $K:=G^\tau_0$. The action of $K$ on $G/H$ by left translations turns out to be hyperpolar, see [@HPTT1 Ex. 3.1]. It is known under the generic name of a [*Hermann action*]{}, after R. Hermann, who first investigated this situation in [@Her] (note that originally, in [@Her], the group $G^\tau$ was assumed to be connected). According to Kollross’ theorem [@Ko], any hyperpolar action is orbit equivalent to a direct product of actions of one of the following types: transitive, of cohomogeneity one, or Hermann (for the notion of orbit equivalence, see Sect. \[oea\]). We already know that any action of one of the first two types is Cohen-Macaulay, see [@Go-Ma Cor. 1.2]. Thus, to prove Thm. \[thm:realmain\], there are two steps to be performed: show first that any Hermann action is Cohen-Macaulay and second that the Cohen-Macaulay property is preserved under orbit equivalence. The first goal is achieved by the following result:
\[main\] If $G$, $K$, and $H$ are as above, then the (Hermann) action of $K$ on $G/H$ by left translations is Cohen-Macaulay.
A big part of the paper is devoted to the proof of this theorem, see Sect. \[semisimple\] (also note that the result remains true even if $G$ is not necessarily semisimple, see Rem. \[nonsemi\]). It is worth recalling at this point a related result, obtained by the first-named author of this paper in [@Go], which says that the $K$-action on $G/K$ by left translations is equivariantly formal; recently, a conceptually different proof has been obtained by the second-named author in [@No]. Thm. \[main\] above is a generalization of this (taking into account Prop. \[characef\] below). Pairs $(G,K)$ with the property that the action of $K$ on $G/K$ is equivariantly formal have also been investigated in [@Go-No], [@Ca2], [@Ca1], [@Ca-Fo], and again [@No].
The second of the aforementioned goals is to show that the Cohen-Macaulay property is preserved under orbit equivalence. This turns out to be true: see Thm. \[orbit\], which represents the main result of Sect. \[oea\]. The proof of Thm. \[thm:realmain\] is also given in full detail in that section.
[**Acknowledgement.**]{} We would like to thank the referee for numerous valuable comments on a previous version of this paper.
General considerations {#sec:general}
======================
The action of $G$ on $M$ {#cmsec}
------------------------
Throughout this subsection $G$ will always be a compact and connected Lie group which acts smoothly on a closed manifold $M$, although many of the results presented below hold in a larger generality. To any such action one attaches the equivariant cohomology $H^*_G(M)$. This can be defined as the usual cohomology of the Borel construction $EG\times_G M$, where $EG\to BG$ is the classifying principal bundle of $G$. This construction originally belongs to A. Borel [@Bo] and it became gradually an important tool in the theory of transformation groups by work of M. F. Atiyah and R. Bott [@At-Bo], D. Quillen [@Quillen], and many others. The book [@AlldayPuppe] by C. Allday and V. Puppe is a useful reference for this topic. We confine ourselves to mentioning the basic fact that $H^*_G(M)$ has a canonical structure of an $H^*(BG)$-algebra.
In case $H^*_G(M)$ is free as an $H^*(BG)$-module, we say that the $G$-action is [*equivariantly formal*]{}. Here is a well-known class of examples of such actions (cf. [@GS Thm. 6.5.3]):
\[exodd\] [If $H^{\rm odd}(M)=0$ then any $G$-action on $M$ is equivariantly formal.]{}
The following criterion will be useful later, see [@Hs p. 46].
\[criteqf\] Let $T$ be a torus acting on a closed manifold $M$.
\(a) The total Betti number of the fixed point set $M^T$ is at most equal to the total Betti number of $M$.
\(b) The two numbers mentioned above are equal if and only if the $T$-action is equivariantly formal.
In what follows we will be looking in more detail at the algebraic structure of $H^*_G(M)$. Relative to its structure of an $H^*(BG)$-module, it is a positively graded module over a positively graded ring. The latter, $H^*(BG)$, is a polynomial algebra on generators of even degrees, and thus a commutative, graded, Noetherian ring. It is also a \*local ring, in the sense that it has a unique graded ideal which is maximal among all graded ideals (namely $H^{>0}(BG)$). The theory of graded modules over \*local Noetherian rings is nicely treated in [@BrunsHerzog Sect. 1.5]. Let us just briefly recall that to each such module one can associate its depth, that is, the maximal length of a regular sequence of elements in the maximal graded ideal. One can also associate its Krull dimension, which is the Krull dimension of the ring modulo the annihilator of the module. In general, the latter number is at least equal to the depth. We say that the module is Cohen-Macaulay if it is equal to 0 or its Krull dimension is equal to its depth. For a more detailed account of this topic, suitable for applications to equivariant cohomology, we refer to [@Go-To Sect. 5].
\[def:cm\] A smooth action $G\times M \to M$ is called [Cohen-Macaulay]{} if $H^*_G(M)$, considered as a module over $H^*(BG)$, is Cohen-Macaulay.
A systematic study of this type of actions was undertaken in [@Go-Ro] and [@Go-To] and we will often rely on results obtained there. However, some caution is necessary, since in the two references above the coefficient field for cohomology is $\bR$, whereas in this paper it is $\bQ$. We will next show that the two seemingly different Cohen-Macaulayness conditions are in fact equivalent. This will be done by using the following general result:
\[commalg\] Let $R$ and $S$ be two Noetherian \*local rings and $\varphi : R \to S$ a ring homomorphism which is homogeneous of degree 0 and maps the maximal graded ideal ${\mathfrak m}$ of $R$ into the maximal graded ideal ${\mathfrak n}$ of $S$. Assume that both $R/{\mathfrak m}$ and $S/{\mathfrak n}$ are fields. Let $A$ be a finitely generated non-zero $R$-module and $B$ a finitely generated non-zero $S$-module, which is flat over $R$. Then:
\(a) ${\rm depth}_S \ \! A\otimes_R B = {\rm depth}_R \ \! A + {\rm depth}_S \ \! B/{\mathfrak m}B$; (b) $\dim_S A\otimes_R B = \dim_R A + \dim_S B/{\mathfrak m}B$;
\(c) The module $A\otimes_R B$ is Cohen-Macaulay over $S$ if and only if $A$ and $B/{\mathfrak m}B$ are Cohen-Macaulay over $R$ and $S$, respectively.
In the special case when $R$ and $S$ are just local rings, both (a) and (b) are standard facts in commutative algebra, cf. e.g. Prop. 1.2.16 and Thm. A.11 in [@BrunsHerzog]. The general case can be reduced to this special situation by using localization at the maximal graded ideals. One takes into account that if $I$ is a graded ideal of $R$ then $ {\rm grade} \ \! (I, A)={\rm grade} \ \! (I_{\mathfrak m}, A_{\mathfrak m})$, see [@BrunsHerzog Prop. 1.5.15 (e)]: this takes care of the three terms involved in equation (a). For (b), one uses three times the general formula $\dim A=\dim A_{\mathfrak m}$, for which we refer to the proof of Prop. 5.1 in [@Go-To] (note that the argument found there uses the assumption that $R/{\mathfrak m}$ is a field).
A group action $G\times M \to M$ is Cohen-Macaulay in the sense of Def. \[def:cm\] if and only if $H^*_G(M; \bR)$ is Cohen-Macaulay as a module over $H^*(BG; \bR)$.
One applies Proposition \[commalg\] by taking $R=H^*(BG)$, $S=H^*(BG; \bR)$, $A=H^*_G(M)$, and $B=H^*(BG;\bR)$, in each case with the obvious module structure; also, take $\varphi : H^*(BG)\to H^*(BG; \bR)$ the inclusion map.
\[pair\] [ The same argument shows that if $N$ is a $G$-invariant subspace of $M$, then the $G$-equivariant cohomologies $H^*_G(M, N)$ and $H^*_G(M, N; \bR)$ as modules over $H^*(BG)$, respectively $H^*(BG; \bR)$, have the same depth, the same Krull dimension, and are thus simultaneously Cohen-Macaulay.]{}
An immediate observation is that any equivariantly formal action is Cohen-Macaulay. The converse implication is in general not true, as one can easily see in concrete situations (for example, by [@Go-Ma Prop. 2.6], any transitive action is Cohen-Macaulay, without being in general equivariantly formal). However, the following result is helpful in this context, see [@Go-Ro Prop. 2.5]:
\[characef\] A smooth action $G\times M \to M$ is equivariantly formal if and only if it is Cohen-Macaulay and there exists at least one point in $M$ whose isotropy subgroup has rank equal to ${\rm rank } \ \! G$.
For later use, we also mention:
\[maxtor\] Let $G\times M \to M$ be a smooth action and $T\subset G$ an arbitrary maximal torus.
\(a) The $G$-action on $M$ is equivariantly formal if and only if so is the induced $T$-action.
\(b) The $G$-action on $M$ is Cohen-Macaulay if and only if so is the induced $T$-action.
Item (a) is a standard result, see for instance [@GGK Prop. C.26]. Item (b) is the content of [@Go-Ro Prop. 2.9]. It is worth pointing out in this context that if $H\subset G$ is an arbitrary closed subgroup, then the property of being equivariantly formal is preserved when passing from $G$ to $H$: indeed, on the one hand, the $G$-action is equivariantly formal if and only if the canonical map $H^*_G(M)=H^*(E\times_G M) \to H^*(M)$ is surjective (cf., e.g., [@Go-Ma Prop. 2.1]), and on the other hand the previous map factors as $H^*_G(M) \to H^*_H(M) \to H^*(M)$. However, it is possible for the $G$-action to be Cohen-Macaualy and the $H$-action not to be like that, see [@Go-Ro Ex. 2.13].
From the previous proposition one can deduce:
\[dirprod\] If two smooth actions on closed manifolds are Cohen-Macaulay then their direct product is Cohen-Macaulay as well.
Let the two group actions be $G_i \times M_i \to M_i$ and let $T_i \subset G_i$ be maximal tori, where $i=1,2$. By Prop. \[maxtor\] above, the restricted $T_i$-action on $M_i$ is Cohen-Macaulay. By [@Go-To Rem. 2], the restricted $T_i$-action on $M_i$ is Cohen-Macaulay if and only if there exists a subtorus $S_i \subset T_i$ of rank equal to the minimal dimension of a $T_i$-orbit on $M_i$, which acts locally freely on $M_i$ such that the induced action of $T/S_i$ on the orbit space $M/S_i$ is equivariantly formal. But then $S_1 \times S_2$ acts locally freely on $M_1 \times M_2$, has rank equal to the minimal dimension of a $T_1\times T_2$-orbit on $M_1\times M_2$ and the induced action of $T_1/S_1 \times T_2/S_2$ on $M_1/S_1 \times M_2/S_2$ is equivariantly formal (note that equivariant formality is preserved under taking the direct product). Thus the $T_1\times T_2$-action on $M_1 \times M_2$ is Cohen-Macaulay, and so is the $G_1\times G_2$-action.
We will need an equivalent characterization of the Cohen-Macaulay condition above, this time exclusively in terms of the ring structure of $H^*_G(M)$. The latter is in general not a commutative ring, but it is nevertheless graded commutative, in the sense that $x\cdot y =(-1)^{({\rm deg} x)({\rm deg} y)}y\cdot x$, for all homogeneous $x, y \in H^*_G(M)$. Although a self-contained treatment of graded commutative rings is not immediately available in the literature, there is no essential difference relative to the (usual) commutative case. A systematic and thorough study of graded commutative rings has been undertaken by M. Poulsen in Appendix A of his Master Thesis [@Po]. For instance, by using the remark following Prop. A.5 in [@Po] and also taking into account that $H^*_G(M)$ is finitely generated as an algebra over its degree zero component $H^0_G(M)\simeq \bQ$ (see for instance [@Quillen Cor. 2.2]), we deduce that $H^*_G(M)$ is a Noetherian ring. For graded commutative Noetherian rings of the type $R=\oplus_{i\ge 0}R^i$ whose degree zero component $R^0$ is a field, one can see in [@Po] that the concepts of Krull dimension, depth (relative to the ideal $\oplus_{i>0}R^i$), and Cohen-Macaulayness can be defined in the same way as for commutative rings. This enables us to prove the following result. It is obtained by adapting [@Se Prop. 12, Sect. IV.B] to our current set-up.
[(J.-P. Serre)]{} \[serre\] Let $R=\oplus_{i\ge 0}R^i$ and $S=\oplus_{i\ge 0}S^i$ be two graded commutative Noetherian rings with $R^0=S^0=\bQ$ and $\varphi: R\to S$ a homomorphism of graded rings which makes $S$ into an $R$-module which is finitely generated. Let also $A$ be a finitely generated $S$-module. Then $A$ is Cohen-Macaulay as an $S$-module if and only if so is $A$ as an $R$-module.
\[cmcrit\] The $G$-action on $M$ is Cohen-Macaulay if and only if the ring $H^*_G(M)$ is Cohen-Macaulay.
The action of $K$ on $G/H$
--------------------------
Let $G$ be again a compact and connected Lie group and let $K, H\subset G$ be closed subgroups. In this section we list some results concerning the three group actions mentioned in the following proposition.
\[crit\] If $K$ and $H$ are connected, the following assertions are equivalent:
- the action of $K$ on $G/H$ by left translations is Cohen-Macaulay;
- the action of $H$ on $G/K$ by left translations is Cohen-Macaulay;
- the action of $H\times K$ on $G$ given by $$\label{hgk}(h,k)\cdot g=hgk^{-1},$$ is Cohen-Macaulay.
For the equivalence between (b) and (c) observe that we have the ring isomorphisms $$H^*_{H\times K}(G)\simeq H^*_H(G/K),$$ due to the fact that the factor $K$ of the product $H\times K$ acts freely on $G$. The equivalence between (a) and (c) can be proved in a similar way, although this time one also needs to take into account that the $H\times K$-action in the lemma is equivalent to the group action $(H\times K)\times G \to G$, $(h, k)\cdot g = kgh^{-1}.$ The result now follows from Cor. \[cmcrit\].
In the light of item (c) above, the following proposition is useful. It is a direct consequence of a certain Sullivan model obtained by J. D. Carlson and C.-K. Fok in [@Ca-Fo Sect. 3.1] (see also [@Ca1 Sect. 8.8.3] or [@Ca3 end of Sect. 2]), which describes the equivariant cohomology of actions on $G$ of type (\[hgk\]), the acting group being a general subgroup of $G\times G$. Denote by ${\mathfrak g}, {\mathfrak k}, {\mathfrak h}$ the Lie algebras of $G$, $K$, and $H$.
[(J. D. Carlson and C.-K. Fok)]{} \[carlson\] If $K$ and $H$ are connected, then the equivariant cohomology of the action of $H\times K$ on $G$ given by $(h,k).g=hgk^{-1}$ depends only on ${\mathfrak g}$, ${\mathfrak h}$, ${\mathfrak k}$, and the inclusions ${\mathfrak h}\hookrightarrow {\mathfrak g}$ and $ {\mathfrak k}\hookrightarrow {\mathfrak g}$.
Finally, we mention a result that shows how to deal with the situation when $H$ is not connected.
\[redcon\] Assume $K$ is connected and let $H_0$ be the identity component of $H$. If the $K$-action on $G/H_0$ is Cohen-Macaulay, then so is the $K$-action on $G/H$.
Use the characterization given by Prop. \[crit\], (c). Note that $H_0\times K$ is the identity component of $H\times K$, hence $H^*_{H\times K}(G)\simeq H^*_{H_0\times K}(G)^{H/H_0}.$ By hypothesis, $H^*_{H_0\times K}(G)$ is Cohen-Macaulay as a module over the ring $H^*(BH_0 \times BK)$, thus also over its subring $H^*(BH\times BK)=H^*(BH_0 \times BK)^{H/H_0}$, since the ring is a finitely generated module over the subring and one can use Lemma \[serre\]. By a standard averaging argument, the invariant module $H^*_{H_0\times K}(G)^{H/H_0}$ can be realized as a direct summand of $H^*_{H_0\times K}(G)$ in the category of $H^*(BH \times BK)$-modules, hence it is Cohen-Macaulay.
Hermann actions {#semisimple}
===============
In this section we will prove Thm. \[main\]. The meaning of $G$, $H$, and $K$ is as in Sect. \[intro\]. In particular, the Lie group $G$ is compact, connected, and semisimple (the case when $G$ is not necessarily semisimple is discussed in Rem. \[nonsemi\]). Due to Propositions \[crit\] and \[redcon\], we only need to show that the equivariant cohomology ring of the $H_0 \times K$-action on $G$ is Cohen-Macaulay. By Prop. \[carlson\], we lose no generality by assuming that $G$ is simply connected, since otherwise we can replace the triple $G$, $H_0$, $K$ with the universal cover of $G$ along with its two connected and closed subgroups that are locally isomorphic to $H_0$ and $K$ respectively. Under the aforementioned assumption, the fixed point sets of $\sigma$ and $\tau$ are connected, see [@He Thm. 1.8, Ch. VII], and thus equal to $H_0$ and $K$ respectively.
The case when $G$ is simple {#sec:simple}
---------------------------
We will prove Thm. \[main\] under the assumption that $G$ is simple and simply connected, which is valid throughout the whole subsection. The induced automorphisms of $\g$ will also be denoted by $\s$ and $\t$. In general, we will not make any notational distinction between an automorphism of $G$ and the induced Lie algebra automorphism.
\[torus\] Assume $G$ is not equal to ${\rm Spin}(8)$. If ${\rm rank} \ \! K \le {\rm rank} \ \! H$, then a maximal torus in $K$ is group-conjugate with a subgroup of $H$.
If one of $\sigma$ or $\tau$ is an inner automorphism, the claim in the lemma is clear by [@He Thm. 5.6, p. 424]. From now on we will assume that both $\sigma$ and $\tau$ are outer automorphisms of $G$. Let $T_K\subset K$ be an arbitrary maximal torus. There is a unique maximal torus $T$ in $G$ such that $T_K\subset T$; furthermore, $T$ is $\tau$-invariant and there is a Weyl chamber $C$ in ${\mathfrak t}:={\rm Lie}(T)$ which is $\tau$-invariant as well (see [@Lo Prop. 3.2, p. 125]). Let $c: G \to G$ be an inner automorphism such that the torus $S:=c(T)$ contains $T_H$, the latter being a maximal torus in $H$. As before, there exists a Weyl chamber $C'\subset c({\mathfrak t})$ which is invariant under $\sigma$. On the other hand, $c\tau c^{-1}$ is an involutive automorphism of $G$ which leaves $S$ invariant; it even leaves the chamber $c(C)$ (inside the Lie algebra of $S$) invariant. But the chambers $c(C)$ and $C'$ are conjugate under the Weyl group of $(G,S)$; that is, there exists an inner automorphism, say $c'$, such that $$c'(S) =S \ {\rm and} \ c'(C') = c(C).$$ We now compare the involutions $c\tau c^{-1}$ and $c'\sigma c'^{-1}$ of $G$: they both leave the torus $S$ invariant, and along with it, its Lie algebra and the chamber $c(C)$ inside it: both are realized in terms of permuting the simple roots that determine the chamber, the permutation being necessarily a Dynkin diagram automorphism.
Since $G$ is different from ${\rm Spin}(8)$, there is a unique such (involutive) permutation which is not the identity map. But none of the automorphisms $c\tau c^{-1}$ and $c'\sigma c'^{-1}$ is inner, thus by composing them and then restricting the result to the chamber $c(C)$, one gets the identity map; as an automorphism of $G$, this composition must consequently be an inner automorphism $c_g$ defined by some $g\in S$. That is, $$c\tau c^{-1}=c'\sigma c'^{-1} c_g.$$ Notice that the automorphism in the left hand side of the previous equation leaves $c(T_K)\subset S$ pointwise fixed. Consequently, $c'\sigma c'^{-1}$ does the same. In other words, $c(T_K)$ is contained in the fixed point set of $c'\sigma c'^{-1}$, which is just $c'(H)$. This finishes the proof.
[*Proof of Thm. \[main\] in the case $G$ simple.*]{} First assume that $G$ is different from ${\rm Spin}(8)$. By Prop. \[crit\], we may assume that ${\rm rank} \ \! K \le {\rm rank} \ \! H$. Thus, by Lemma \[torus\], there exists $T\subset K$ a maximal torus and $g_0\in G$ such that $g_0Tg_0^{-1}\subset H$. The actions $K \times G/H\to G/H$ and $g_0Kg_0^{-1} \times G/H\to G/H$ given by left translation are equivalent, relative to the map $$(K, G/H) \to (g_0Kg_0^{-1}, G/H), \quad (k, gH) \mapsto (g_0kg_0^{-1}, g_0gH).$$ Thus it is sufficient to show that the $g_0Kg_0^{-1}$-action on $G/H$ is Cohen-Macaulay. Equivalently, by means of Prop. \[maxtor\] above, that the action of $g_0Tg_0^{-1}$ on $G/H$ is so. But this is clear, because, by the main result in [@Go], the action of $H$ on $G/H$ is equivariantly formal.
If $G={\rm Spin}(8)$, then, as in the proof of Lemma \[torus\], we can assume that both $\sigma$ and $\tau$ are outer automorphisms. It is known that the group of outer automorphisms of ${\rm Spin}(8)$ is isomorphic to the symmetric group $\Sigma_3$ on three letters and the fixed point set of any order-two outer automorphism is isomorphic to ${\rm Spin}(7)$, giving rise to a copy of the symmetric space ${\rm Spin}(8)/{\rm Spin}(7) = S^7$ (cf. [@Ad1 Sect. 2]). Thus Thm. \[main\] is a consequence of Lemma \[odd\] below. $\square$
\[odd\] Any smooth action of a compact connected Lie group on a closed manifold which is a rational cohomology sphere is Cohen-Macaulay.
If the dimension of the sphere is even, the lemma is a consequence of Ex. \[exodd\]. To deal with the remaining situation, we use Prop. \[maxtor\] (b), which enables us to only consider the action of a torus $T$ on $X$, where $X$ is a closed manifold with $H^*(X)\simeq H^*(S^{2n+1})$ as vector spaces, for some $n \ge 0$. In particular, the total Betti number $\dim H^*(X)$ is equal to 2. In case the fixed point set $X^T$ is non-empty, the latter is a union of closed submanifolds of $X$. By Prop. \[criteqf\], $\dim H^*(X^T)$ is thus a strictly positive number at most equal to 2. On the other hand, the Euler-Poincaré characteristics of $X$ and $X^T$ are equal, see e.g. [@Kobayashi Thm., (4)] or [@Go-Zo Thm. 9.3]; thus $\dim H^*(X^T)$ is an even number. We conclude that the latter number is equal to 2, hence the $T$-action on $X$ is equivariantly formal and consequently Cohen-Macaulay.
Let us now consider the case when the set $X^T$ is empty. There exists a one-dimensional subtorus $S\subset T$ whose action on $X$ is locally free. One can compute the cohomology of the orbit space $X/S$ by means of the following version of the Gysin sequence (cf. [@Co Lemma 2.2]): $$\ldots \to H^j(X/S) \to H^{j+2}(X/S) \to H^{j+2}(X) \to H^{j+1}(X/S) \to \ldots .$$ It follows that $H^*(X/S)\simeq H^*(\bC P^n)$ by an isomorphism of vector spaces. Consequently, by Ex. \[exodd\], the canonical action of $T/S$ on $X/S$ is equivariantly formal, hence, by Cor. \[cmcrit\], the ring $H^*_{T/S}(X/S)$ is Cohen-Macaulay. To conclude the proof, it only remains to notice that we have the ring isomorphism $$H^*_T(X)\simeq H^*_{T/S}(X/S).$$
The case when $G$ is not simple
-------------------------------
From now on, $G$ is just simply connected, not necessarily being simple. Our main tool in dealing with this situation is the following result, see [@Ko Prop. 18]:
[(A. Kollross)]{}\[andreas\] If $G$ is simply connected then the $H$-action on $G/K$ is a direct product of actions of one of the following types:
- the action of $H'\times L^{n-1} \times K'$ on $L^n$ defined by $$\begin{aligned}
{}&(h, g_1, \ldots, g_{n-1}, k)\cdot(x_1, \ldots, x_n)\\{}&=
(hx_1g_1^{-1}, g_1x_2g_2^{-1}, \ldots, g_{n-2}x_{n-1}g_{n-1}^{-1}, g_{n-1}x_n k^{-1}),\end{aligned}$$
- the action of $H'\times L^{n-1}$ on $L^{n-1}\times L/K'$ defined by $$\begin{aligned}
{}&(h, g_1, \ldots, g_{n-1})\cdot(x_1, \ldots, x_{n-1}, x_nK')\\{}&=
(hx_1g_1^{-1}, g_1x_2g_2^{-1}, \ldots, g_{n-2}x_{n-1}g_{n-1}^{-1}, g_{n-1}x_n K'),\end{aligned}$$
- the action of $L^{n-1}$ on $H'\backslash L\times L^{n-2}\times L/K'$ defined by $$\begin{aligned}
{}&(g_1,\ldots, g_{n-1})\cdot (H'x_1, x_2, \ldots, x_{n-1}, x_n K')
\\{}&=(H'x_1g_1^{-1}, g_1x_2g_2^{-1}, \ldots, g_{n-2}x_{n-1}g_{n-1}^{-1}, g_{n-1}x_n K'),\end{aligned}$$
- the action of $L^n$ on $L^n$ defined by $$\begin{aligned}
{}&(g_1, \ldots, g_n)\cdot(x_1,\ldots, x_n)\\{}&
=(g_1x_1g_2^{-1}, g_2x_2g_3^{-1}, \ldots, g_{n-1}x_{n-1}g_n^{-1}, g_n x_n\alpha(g_1)^{-1}),\end{aligned}$$
where $L$ is a simply connected, simple and compact Lie group, $H',K'\subset L$ are fixed points of involutions of $L$, $\alpha$ is an outer or trivial automorphism of $L$, and $n$ is an arbitrary integer, at least equal to 1 in cases (i), (ii), and (iv), respectively to 2 in case (iii).
For the reader’s convenience, here are a few details concerning this result.
\(a) Proposition \[andreas\] above is just a consequence of the result actually proved in [@Ko]. Namely, we are here in the special case when $G$ is simply connected. Thus our $H$-action on $G/K$ is of the Hermann type in the sense of [@Ko Def. 15] (see the definition of locally symmetric subgroups at the beginning of [@Ko Sect. 2] and also recall the well-known fact that the identity component of the isometry group of the symmetric space $G/K$ is just $G$). Under the hypothesis mentioned above, $G/K$ is simply connected as well, hence the $H$-action on $G/K$ is not only locally conjugate to one of the types (i)-(iv), as the result in [@Ko] says, but genuinely conjugate.
\(b) We also wish to explain the meaning of the number $n$ in the proposition above. It comes from the structure of $G/K$. For example, in case (i) the pair $(G, K)$ is equal to $(L \times \cdots \times L, \Delta(L)\times \cdots \times
\Delta(L))$, where the first direct product has $2n$ factors, the second has $n$, and $\Delta(L) :=\{(g, g) | g\in L\}$. The cases (ii)-(iv) are left as an exercise to the reader.
In view of Cor. \[dirprod\], to prove Thm. \[main\] for $G$ semisimple it is sufficient to show that the actions (i)-(iv) are Cohen-Macaulay.
Let us start with (i). We use an inductive argument. Note that the first factor $L$ in $H'\times L^{n-1}\times K'$ acts freely on $L^{n}$, the orbit space being diffeomorphic to $L^{n-1}$ via $$L^{n}/L \to L^{n-1}, \quad (x_1, x_2, \ldots, x_{n}) \mapsto (x_1 x_2, x_3 \ldots, x_{n}).$$ Thus it is sufficient to prove that the action of $H'\times L^{n-2}\times K'$ on $L^{n-1}$ given by $$\begin{aligned}
{}&(h, g_2, \ldots, g_{n-1}, k)\cdot(x_2, \ldots, x_n)\\{}&=
(hx_2g_2^{-1}, g_2x_3g_3^{-1}, \ldots, g_{n-2}x_{n-1}g_{n-1}^{-1}, g_{n-1}x_n k^{-1}),\end{aligned}$$is Cohen-Macaulay. We continue the procedure and gradually drop out the $L$-factors in $H'\times L^{n-1}\times K'$ until we finally obtain the action of $H'\times K'$ on $L$ given by $(h, k)\cdot x=hxk^{-1}$. But this action is Cohen-Macaulay by the result already proved in Subsect. \[sec:simple\] (see also Prop. \[crit\]).
To deal with (ii), we start by modding out the action of the first factor $L$ in $H'\times L^{n-1}$, which is clearly a free action. In this way, we reduce the problem to showing that the action of $H'\times L^{n-2} $ on $L^{n-2}\times L/K'$ described by $$\begin{aligned}
{}&(h, g_2, \ldots, g_{n-1})\cdot(x_2, \ldots, x_{n-1}, x_nK')\\{}&=
(hx_2g_2^{-1}, g_2x_3g_3^{-1}, \ldots, g_{n-2}x_{n-1}g_{n-1}^{-1}, g_{n-1}x_n K'),\end{aligned}$$is Cohen-Macaulay. We continue the procedure until we are led to the action of $H'$ on $L/K'$ given by left translations. Again, this action is Cohen-Macaulay by the result we proved in Subsect. \[sec:simple\].
Similarly, in case (iii) we reduce the problem to the action of $L$ on $H'\backslash L \times L/K'$ described by $g\cdot (H'x_1, x_2K')=(H'x_1g^{-1}, gx_2K')$. The map $$H'\backslash L \times L/K'\to L/H'\times L/K', \quad (H'x_1, x_2K')\mapsto (x_1^{-1}H', K'x_2)$$ is an $L$-equivariant diffeomorphism. This allows us to change our focus to the action of $\Delta(L):=\{(g,g)\mid g\in L\}$ on $(L\times L)/(H'\times K')$ by left translations. By Prop. \[crit\], this is Cohen-Macaulay if and only if so is the action of $H'\times K'$ on $(L\times L)/\Delta(L)$. But the latter is just the action of $H'\times K'$ on $L$ given by $(h, k)\cdot x=hxk^{-1}$, which was discussed in Subsect. \[sec:simple\].
As about (iv), the recursive procedure already used in each of the previous cases leads us to the action of $L$ on itself given by $g\cdot x=gx\alpha(g)^{-1}$. It was proved by Baird in [@Ba p. 212] (cf. also [@Sieb p. 58]) that all isotropy groups of this action have the same rank. By [@Go-Ro Cor. 4.3], the action is thus Cohen-Macaulay.
\[nonsemi\] [Thm. \[main\] holds even when $G$ is not necessarily semisimple. To prove this, consider a finite cover of $G$ of the type $T\times G_s$, where $T$ is a torus and $G_s$ is compact, connected, and simply connected. In view of Propositions \[crit\], \[carlson\], and \[redcon\] it is sufficient to consider the case when $G$ is equal to such a direct product and $H$ is the identity component of $G^\tau$. But then both $\sigma$ and $\tau$ leave the factors $T$ and $G_s$ invariant. Their fixed point sets split as direct products of subgroups of $T$ and $G_s$ respectively; that is, $H_0=T_1\times H_s$ and $K=T_2\times K_s$, where $T_1, T_2\subset T$ are subtori and $H_s, K_s \subset G_s$. The $H_0\times K$-action on $G$ can be described as follows: $$\left((t_1, h), (t_2, k)\right). \left(t, g\right) = \left(t_1tt_2^{-1}, hgk^{-1}\right).$$ This is the direct product of the following two actions: $$\begin{aligned}
{}&(T_1 \times T_2) \times T \to T, \ (t_1, t_2)\cdot t =t_1tt_2^{-1}\\
{}& (H_s \times K_s)\times G_s \to G_s, \ (h, k)\cdot g = hgk^{-1}.\end{aligned}$$ In view of Cor. \[dirprod\], it is sufficient to observe that both of them are Cohen-Macaulay: for the first factor, one observes that the kernel of the $T_1 \times T_2$-action on $T$ is isomorphic to $T_1\cap T_2$, and an arbitrary direct complement of its identity component acts locally freely; for the second factor, one uses Thm. \[main\]. ]{}
We conclude the section with a remark concerning equivariant formality of Hermann actions:
[Let $G, K$, and $H$ be as in Thm. \[main\]. The action of $K$ on $G/H$ is equivariantly formal if and only if a maximal torus in $K$ is conjugate with a subgroup of $H$ (note that the latter condition is equivalent to the existence of a fixed point for the action of the maximal torus in $K$, which is necessary for the equivariant formality of the $K$-action; the other implication is proved by invoking the main result of [@Go] as in the proof of Thm. \[main\] in the case $G$ simple). In particular, ${\rm rank} \ \! K \le {\rm rank } \ \! H$. We note that the latter condition alone is not sufficient for equivariant formality. Take for instance two Dynkin diagram involutions of ${\rm Spin}(8)$ which are not conjugate with each other. Their fixed point sets, $H$ and $K$ respectively, are both isomorphic to ${\rm Spin}(7)$. It turns out that the action of $H$ on ${\rm Spin}(8)/K\simeq S^7$ by left translations is transitive, with isotropy subgroups isomorphic to the exceptional compact Lie group of type $G_2$, see e.g. [@Va Thm. 3]. Thus this action is not equivariantly formal. It is interesting to notice, however, that if $G$ is simple and simply connected, $G\neq {\rm Spin}(8)$, then the condition ${\rm rank} \ \! K \le {\rm rank } \ \! H$ is sufficient for the $K$-action on $G/H$ to be equivariantly formal: this follows from Lemma \[torus\] above. ]{}
Orbit equivalent actions {#oea}
========================
We start by proving the following two lemmata, whose relevance will become clear immediately:
\[prop:orbit\] Let $M$ be a closed manifold and $K, K'$ two compact and connected Lie groups that act on $M$ such that $K\subseteq K'$ and $Kp=K'p$ for all $p\in M$. Then the $K$-action on $M$ is Cohen-Macaulay if and only if so is the $K'$-action.
Let $b$ denote the maximal rank of an isotropy subgroup of the $K$-action and $M_{b,K}$ the subspace of $M$ consisting of all points whose isotropy group has rank equal to $b$. In the same way, to the action of $K'$ one assigns the number $b'$ and the subspace $M_{b', K'}$. For any $p\in M$ we have $Kp=K'p$. In general, if $G$ is a compact Lie group and $U \subset G$ a closed subgroup, then the rank $\chi\pi(G/U)$ of the homogeneous space $G/U$ is a homotopy invariant, which turns out to be equal to the difference ${\rm rank } \ \! U - {\rm rank } \ \! G$ (see [@All2 Def. Sect. 1.1 and Sect. 4.2]). In our situation, for any $p\in M$ we have $Kp = K'p$, and hence the homogeneous spaces $K/K_p$ and $K'/K'_p$ are diffeomorphic, where $K_p$ and $K'_p$ are the isotropy subgroups at $p$. We deduce $$\label{rankk}{\rm rank } \ \! K - {\rm rank } \ \! K_p={\rm rank } \ \! K' -
{\rm rank } \ \! K'_p.$$ (We note that the same argument was used in [@Go-Po Prop. 7]; we could also just apply the latter proposition to the $K$- and $K'$-actions on $Kp=K'p$ to obtain the same conclusion.) Eq. (\[rankk\]) implies: $$\label{mbh} M_{b, K} = M_{b', K'}.$$ Pick maximal tori $T$ in $K$ and $T'$ in $K'$ such that $T\subset T'$. The minimal dimension of a $T$-orbit in $M$ is ${\rm rank} \ \! K - b$, cf. e.g. [@Go-Ro Lemma 4.1]. Similarly, the minimal dimension of a $T'$-orbit in $M$ is ${\rm rank} \ \! K' - b'$. But the two aforementioned numbers are the $K$-, respectively $K'$-coranks of any point in $M_{b,K}$ respectively $M_{b',K'}$, hence, by eq. (\[mbh\]), they are equal. Thus, there exists a subtorus $S\subset T$ of rank equal to the two numbers which acts locally freely on $M$.
The sets $M_{b, T}$ and $M_{b', T'}$ defined in the same way as before are non-empty, clearly contained in $M_{b,K}$ and $M_{b', K'}$ respectively. The $K$-action on $M_{b, K}$ is Cohen-Macaulay, see [@Go-Ro Cor. 4.3]. Consequently, this time by Prop. \[maxtor\] (b) and Cor. \[cmcrit\], $H^*_T(M_{b,K})$ is a Cohen-Macaulay ring. Since $S\subset T$ acts locally freely, the latter ring is isomorphic to $H^*_{T/S}(M_{b,K}/S)$.
On the other hand, the $T/S$-action on $M_{b,K}/S$ admits points that are fixed. Namely, they are orbits of the form $Sp$ such that $Sp=Tp$, which implies that $$\label{fix}{\rm corank}_T \ \! T_p = {\rm rank} \ \! S.$$ But if $p\in M$ satisfies the latter condition, then $Sp$ is a connected and closed submanifold of $Tp$, of the same dimension as the latter, hence $Sp=Tp$. Thus condition (\[fix\]) characterizes the fixed points. Since ${\rm rank} \ \! S = {\rm rank} \ \! T -b$, that condition is actually equivalent to $p\in M_{b, T}$. We have actually proved: $$\label{movers}(M/S)^{T/S}=(M_{b,K}/S)^{T/S}=M_{b,T}/S.$$
From the previous considerations, the $T/S$-action on $M_{b,K}/S$ is equivariantly formal. Consequently, by eq. (\[movers\]), $$\label{dimh}
\dim H^*(M_{b,K}/S) =\dim H^*(M_{b,T}/S).$$ In the same way, one analyzes the $K'$-action on $M$ and obtains $$\label{movers2}(M/S)^{T'/S}=(M_{b',K'}/S)^{T'/S}=M_{b',T'}/S.$$ as well as $$\label{dimh2} \dim H^*(M_{b',K'}/S) =\dim H^*(M_{b',T'}/S).$$
We are now in a position to prove the equivalence stated in the lemma. First, the $K$-action on $M$ is Cohen-Macaulay if and only if so is the induced $T$-action, see Prop. \[maxtor\] (b). But $H^*_T(M)=H^*_{T/S}(M/S)$ and the $T/S$-action on $M/S$ admits fixed points, thus the latter Cohen-Macaulay condition is equivalent to: the $T/S$-action on $M/S$ is equivariantly formal. Equivalently, by eqs. (\[movers\]) and (\[dimh\]), $$\dim H^*(M/S)= \dim H^*(M_{b,K}/S).$$ In exactly the same way, this time by using eqs. (\[movers2\]) and (\[dimh2\]), the $K'$-action on $M$ is Cohen-Macaulay if and only if $$\dim H^*(M/S)= \dim H^*(M_{b',K'}/S).$$ The proof is completed by taking into account eq. (\[mbh\]).
\[kernel\] Let $M$ be a closed manifold and $K$ a compact and connected Lie group that acts smoothly on $M$. Let also $H\subset K$ be the kernel of the action. Then the $K$-action on $M$ is Cohen-Macaulay if and only if so is the $K/H$-action.
Let $b$ and $M_{b,K}$ be as in the proof of Lemma \[prop:orbit\]. Pick maximal tori $T$ and $T'$ in $K$ and $H$ respectively such that $T'\subset T$. Again, let $S\subset T$ be a subtorus of rank equal to ${\rm rank} \ \! T -b$ whose action on $M$ is locally free. We already noticed that the $K$-action on $M$ is Cohen-Macaulay if and only if $$\dim H^*(M/S)= \dim H^*(M_{b,T}/S).$$ Observe that there exists a maximal torus $T_1 \subset K/H$ along with a covering map $T/T' \to T_1$. Furthermore, for the $K/H$-action on $M$, the maximal rank of an isotropy subgroup is $b_1:=b - {\rm rank} \ \! H.$ Since the intersection $T'\cap S$ is finite, the image $S_1$ of $S$ under the projection $T\to T/T'\to T_1$ is a subtorus of rank equal to ${\rm rank} \ \! S={\rm rank} \ \! T_1 -b_1$ which acts locally freely on $M$. As before, the $K/H$-action on $M$ is Cohen-Macaulay if and only if $$\dim H^*(M/S_1)= \dim H^*(M_{b_1,T_1}/S_1).$$ But $M/S_1=M/S$ and $M_{b_1,T_1}=M_{b, T}$, thus the equivalence stated in the lemma is clear.
Let us now recall that two isometric actions of two connected compact Lie groups on a Riemannian manifold are [*orbit equivalent*]{} if there is an isometry of the manifold to itself which maps each orbit of the first group action to an orbit of the second one.
\[orbit\] Let $M$ be a compact Riemannian manifold and $K$ a connected and closed subgroup of the isometry group of $M$. Assume that the $K$-action on $M$ is orbit equivalent to a Cohen-Macaulay action on $M$. Then the $K$-action is Cohen-Macaulay as well.
Let $G$ be the isometry group of $M$. By hypothesis, there exists a connected and closed subgroup $K'\subset G$ whose canonical action on $M$ is Cohen-Macaulay such that the actions of $K$ and $K'$ on $M$ are orbit equivalent (notice that a priori $K'$ might not be contained in $G$; in this case, we mod out the kernel of its action on $M$ and use Lemma \[kernel\] above). Thus there exists an isometry $f : M \to M$ which maps any $K$-orbit to a $K'$-orbit. Consider the closed subgroup $K''$ of $G$ generated by $f^{-1} K' f$ and $K$. Note that $K''$ is connected. The key-observation is that for any $p\in M$, we have $$\label{hp} Kp=(f^{-1}K'f)p=K''p.$$ The first equality is clear and immediately implies the second. Due to eq. (\[hp\]), the result stated in the theorem is a direct consequence of Lemma \[prop:orbit\], used twice.
We are now in a position to prove the main result of the paper:
[*Proof of Thm. \[thm:realmain\].*]{} As already mentioned in the introduction, Kollross has shown in [@Ko] that any hyperpolar action is orbit equivalent to a direct product of actions of one of the following types: transitive, of cohomogeneity one, or Hermann. We know that an action of each of these three types is Cohen-Macaulay: for the first two, see [@Go-Ma Cor. 1.2], for the last, use Thm. \[main\]. We apply Thm. \[orbit\]: even though the acting group is not necessarily a subgroup of ${\rm Iso}(M)$, we can mod out the kernel of the action and use Lemma \[kernel\].$\square$
[It would be interesting to find a classification-free proof of Thm. \[thm:realmain\], using the very definition of a hyperpolar action.]{}
The non-abelian Atiyah-Bredon exact sequence for Cohen-Macaulay actions
=======================================================================
Although not directly related to the main topic of this paper, the following result is devoted to illustrate the importance of showing that a group action is Cohen-Macaulay. In the special case when the acting group is abelian, the result was proved by the first-named author of this paper and D. Töben in [@Go-To]. For an arbitrary (compact and connected) acting group, versions of it were obtained by M. Franz in [@Franz]. We refer to the two papers above for a discussion concerning the history of the topic, which goes back to M. Atiyah [@Atiyah] and G. E. Bredon [@Bredon].
Let $M$ be a closed manifold equipped with a smooth action of a compact and connected Lie group $G$ and let $b$ be the maximal rank of a $G$-isotropy subgroup. (Note that $b$ is in general smaller than the rank of $G$.) The corresponding orbit filtration is defined by $$M_i:=\{ p \in M \mid {\rm rank} \ \! G_p \ge i\},$$ where $i$ is arbitrary between $0$ and $b+1$. For any $1\le i \le b$, consider the map $$H_G^*(M_{i}, M_{i+1}) \to H_G^{*+1}(M_{i-1}, M_{i}),$$ which is the connecting homomorphism in the long exact sequence in $G$-equivariant cohomology of the triple $(M_{i-1}, M_i, M_{i+1})$. By concatenating these maps one obtains the following long sequence: $$\label{ab}0\to H^*_G(M)\to H^*_G(M_b)\to H^{*+1}_G(M_{b-1}, M_b)\to \ldots \to H_G^{*+b-1}(M_{1}, M_{2}) \to H^{*+b}_G(M,M_1)\to 0,$$ which we call the Atiyah-Bredon sequence.
The $G$-action on $M$ is Cohen-Macaulay if and only if the Atiyah-Bredon sequence is exact.
The result follows using the same argument as in the proof of [@Go-To Thm. 6.1], cf. also [@FranzPuppe2003 Sect. 4]. The main ingredients in the proof are:
- The exactness of the sequence (\[ab\]) is equivalent to the exactness of $$0\to H^*_G(M, M_i) \to H^*_G(M_{i-1}, M_i) \to H^{*+1}_G(M, M_{i-1}) \to 0,$$ for any $1 \le i \le b$; in other words, the long exact sequence of the triple $(M_{i-1}, M_i, M_{i+1})$ splits into short exact sequences. This can be shown going along the same lines as in the proof of [@Fr-Pu Lemma 4.1].
- For any $1\le i \le b$, the Krull dimension of $H^*_G(M, M_i)$ is at most $i-1$. This is the content of [@FranzPuppe2003 Lemma 4.4] in the case when the acting group is abelian; the arguments of the proof carry over to the general situation, see [@Go-Ma0 Lemma 2.1] (this result was proved for cohomology with coefficients in $\bR$, thus it consequently holds for coefficients in $\bQ$, see Remark \[pair\] above).
- For any $1\le i \le b$, $H^*_G(M_{i-1}, M_i)$ is Cohen-Macaulay of dimension $i-1$. For this, we refer to [@Franz Cor. 1].
[Fo-Ge-Po]{}
J. F. Adams, [*Spin(8), triality, F4, and all that*]{}, Superspace and Supergravity, 435-445, ed. S. Hawking and M. Rocek, Cambridge Univ. Press, 1981
C. Allday, [*On the rank of a space*]{}, Trans. Amer. Math. Soc. [**166**]{} (1972), 173-185
C. Allday and V. Puppe, [*Cohomological Methods in Transformation Groups*]{}, Cambridge Studies in Advanced Mathematics, Vol. 32, Cambridge University Press, Cambridge, 1993
M. Atiyah, [*Elliptic Operators and Compact Groups*]{}, Lecture Notes in Mathematics, Vol. 401, Springer-Verlag, Berlin, 1974
M. F. Atiyah and R. Bott, [*The moment map and equivariant cohomology*]{}, Topology [**23**]{} (1984), 1-28
T. J. Baird, [*Classifying spaces of twisted loop groups*]{}, Algebr. Geom. Topol. [**16**]{} (2016), no. 1, 211-229
A. Borel, [*Seminar in transformation groups*]{}, Annals of Math. Studies, Princeton University Press, 1960
G. E. Bredon, [*The free part of a torus action and related numerical equalities*]{}, Duke Math. J. [**41**]{} (1974), 843-854
W. Bruns and J. Herzog, [*Cohen-Macaulay Rings*]{}, Cambridge Studies in Advanced Mathematics, Vol. 39, Cambridge University Press, Cambridge 1993
J. D. Carlson, [*The Borel equivariant cohomology of real Grassmannians*]{}, preprint arXiv:1611.01175
J. D. Carlson, [*Equivariant formality of isotropic torus actions, I*]{}, preprint arXiv:1410.5740
J. D. Carlson, [*On the Equivariant Cohomology of Homogeneous Spaces*]{}, http://www.math.toronto.edu/jcarlson/homogeneous\_space\_book.pdf (version of Jan. 10, 2019; accessed on March 7, 2019)
J. D. Carlson and C.-K. Fok, [*Equivariant formality of isotropy actions*]{}, J. Lond. Math. Soc. [**97**]{} (2018), no. 3, 470-494
P. E. Conner, [*On the action of the circle group*]{}, Michigan Math. J. [**4**]{} (1957), no. 3, 241-247
J. Duflot, [*Depth and equivariant cohomology*]{}, Comment. Math. Helv. [**56**]{} (1981), 627-637
M. Franz, [*Syzygies in equivariant cohomology for non-abelian Lie groups*]{}, pp. 325-360 in: Filippo Callegaro et al. (eds.), Configuration spaces (Cortona, 2014), Springer INdAM Ser. 14, Springer, Cham 2016
M. Franz and V. Puppe, [*Exact cohomology sequences with integral coefficients for torus actions*]{}, Transform. Groups [**12**]{} (2007), 65-76
M. Franz and V. Puppe, [*Exact sequences for equivariantly formal spaces*]{}, C. R. Math. Acad. Sci. Soc. R. Can. [**33**]{} (2011), 1-10
O. Goertsches, [*The equivariant cohomology of isotropy actions on symmetric spaces*]{}, Documenta Math. [**17**]{} (2012), 79-94
O. Goertsches and S. Hagh Shenas Noshari, [*Equivariant formality of isotropy actions on homogeneous spaces defined by Lie group automorphisms*]{}, J. Pure Appl. Algebra [**220**]{} (2016), 2017-2028
O. Goertsches and A.-L. Mare, [*Non-abelian GKM theory*]{}, Math. Z. [**277**]{} (2014), 1-27
O. Goertsches and A.-L. Mare, [*Equivariant cohomology of cohomogeneity one actions*]{}, Top. Appl. [**167**]{} (2014), 36-52
O. Goertsches and S. Rollenske, [*Torsion in equivariant cohomology and Cohen-Macaulay actions*]{}, Transform. Groups [**16**]{} (2011), 1063-1080
O. Goertsches and D. Töben, [*Torus actions whose equivariant cohomology is Cohen-Macaulay*]{}, J. of Topology [**3**]{} (2010), 1-28
O. Goertsches and L. Zoller, [*Equivariant de Rham cohomology: theory and applications*]{}, preprint arXiv:1812.09511
C. Gorodski and F. Podesta, [*Homogeneity rank of real representations of compact Lie groups*]{}, J. Lie Theory [**15**]{} (2005), 63-77
V. W. Guillemin, V. L. Ginzburg, and Y. Karshon, [*Moment maps, cobordisms, and Hamiltonian group actions*]{}, Mathematical Surveys and Monographs, Vol. 96, Amer. Math. Soc., Providence, RI, 2002
V. Guillemin and S. Sternberg, [*Supersymmetry and Equivariant de Rham Theory*]{}, Springer-Verlag, Berlin 1999
S. Hagh Shenas Noshari, [*On the equivariant cohomology of isotropy actions*]{}, Doctoral Dissertation, University of Marburg, 2018
S. Helgason, [*Differential Geometry, Lie Groups, and Symmetric Spaces*]{}, Graduate Studies in Mathematics, Vol. 34, American Mathematical Society, Providence, Rhode Island, 2001
E. Heintze, R. Palais, G. Thorbergsson, and C.-L. Terng, [*Hyperpolar actions and k-flat homogeneous spaces*]{}, J. Reine Angew. Math. [**454**]{} (1994), 163-179
E. Heintze, R. Palais, G. Thorbergsson, and C.-L. Terng, [*Hyperpolar actions on symmetric spaces*]{}, Geometry, Topology and Physics, International Press, Cambridge, MA, 1995, 214-245
R. Hermann, [*Variational completeness for compact symmetric spaces*]{}, Proc. Amer. Math. Soc. [**11**]{} (1960), 544-546
W.-Y. Hsiang, [*Cohomology Theory of Topological Transformation Groups*]{}, Ergebnisse der Mathematik und ihrer Grenzgebiete, Vol. 85, Springer-Verlag, New-York 1975
F. C. Kirwan, [*Cohomology of Quotients in Symplectic and Algebraic Geometry*]{}, Mathematical Notes [**31**]{}, Princeton University Press, Princeton, 1984
S. Kobayashi, [*Fixed points of isometries*]{}, Nagoya Math. J. [**13**]{} (1958), 63-68
A. Kollross, [*A classification of hyperpolar and cohomogeneity one actions*]{}, Trans. Amer. Math. Soc. [**354**]{} (2002), 571-612
A. Kollross, [*Polar actions on symmetric spaces*]{}, J. Differential Geom. [**77**]{} (2007), 425-482
A. Kollross, [*Low cohomogeneity and polar actions on exceptional compact Lie groups*]{}, Transform. Groups [**14**]{} (2009), 387–415
A. Kollross, [*Hyperpolar actions on reducible symmetric spaces*]{}, Transform. Groups [**22**]{} (2017), 207-228
O. Loos, [*Symmetric Spaces II: Compact Spaces and Classification*]{}, Mathematics Lecture Notes Series, W. A. Benjamin, Inc., New York, 1969
M. Poulsen, [*Depth, detection and associated primes in the cohomology of finite groups (An introduction to Carlson’s depth conjecture)*]{}, Master Thesis, University of Copenhagen, 2007, URL: http://web.math.ku.dk/$\sim$moller/students/mortenP.pdf
D. Quillen, [*The spectrum of an equivariant cohomology ring: I*]{}, Ann. Math. [**94**]{} (1971), 549-572
J.-P. Serre, [*Algèbre locale. Multiplicités, 3 éd.*]{}, Lecture Notes in Mathematics, Vol. 11, Springer-Verlag, Berlin 1975
J. de Siebenthal, Sur les groupes de Lie compacts non connexes, Comm. Math. Helvetici [**31**]{} (1956), 41-89
V. S. Varadarajan, [*$Spin(7)$-subgroups of $SO(8)$ and $Spin(8)$*]{}, Expo. Math. [**19**]{} (2001), 163-177
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Régis [Monneau]{}$^{\hbox{\small{a}}}$, Jean-Michel [ Roquejoffre]{}$^{\hbox{\small{b}}}$,\
Violaine [Roussier-Michon]{}$^{\hbox{\small{c}}}$\
\
\
\
\
\
\
title: |
**[Travelling graphs for the forced mean curvature motion in an arbitrary space dimension\
Ondes progressives pour le mouvement par courbure moyenne forcé en toute dimension d’espace]{}**
---
Introduction {#intro}
============
Setting of the problem
----------------------
The question investigated here is the description of the travelling wave graph solutions to the forced mean curvature motion in any dimension $N \geq 2$, that is written under the general form $$\label{vitesse normale}
V_n= - c_0 + {\kappa}$$ where $V_n$ is the normal velocity of the graph, $\kappa$ its local mean curvature and $c_0$ a given strictly positive constant to be defined later. A graph satisfying can be given by the equation $z=u(t,x)$ where $u:(t,x) \in \RR^+ \times \RR^{N-1} \mapsto u(t,x) \in \RR$ is a solution to the parabolic equation $$\label{MCM0}
\frac{u_t}{\sqrt{1+\vert Du\vert^2}}= -c_0 + \mbox{ div }\biggl(\frac{Du}{\sqrt{1+\vert
Du\vert^2}}\biggl)
\, , \quad t>0 \, , \, x \in \RR^{N-1}$$ Indeed, at any time $t>0$ fixed, the outer normal to the subgraph $\{(x,z)\in \RR^{N-1}\times \RR \, | \, z\leq u(t,x)\}$ is given by $$\vec{n}=\frac{1}{\sqrt{1+|Du|^2}} \left(\begin{array}{c}
-D_x u \\
1
\end{array}\right)$$ its normal velocity $V_n$ by $(0,\partial_t u)^T \cdot \vec{n}$ while its mean curvature by $\kappa=-\div_{(x,z)} \vec{n}$, see [@GT].
A travelling wave to is a solution of the form $u(t,x)=-ct +
\phi(x)$ where $\phi:x \in \RR^{N-1} \mapsto \phi(x) \in \RR$ is the profile of the wave and $c\geq c_0$ is some given constant standing for its speed. Thus $\phi$ satisfies the following elliptic equation $$\label{eq de phi}
- \mbox{ div }\left( \frac{D\phi}{\sqrt{1+|D\phi|^2}}\right) +c_0 -
\frac{c}{\sqrt{1+|D\phi|^2}}=0 \, , \quad x \in \RR^{N-1}$$
Connection with reaction diffusion equations
--------------------------------------------
This work should provide us a better understanding of the multidimensional solutions to the non linear scalar reaction diffusion equation $$\label{reaction diffusion}
\partial_t v =\Delta v + f(v) \, , \quad t>0 \, , \, (x,z)\in \RR^{N-1} \times
\RR$$ where $v:(t,x,z) \in [0,+\infty) \times \RR^{N-1} \times \RR \mapsto v(t,x,z)
\in \RR$ and, especially the case of travelling waves in dimension $N$. In the case of a “bistable” nonlinearity $f$, that is to say when $f$ is a continuously differentiable function on $\RR$ satisfying
i.
: $f(0)=f(1)=0$
ii.
: $f'(0)<0$ and $f'(1)<0$
iii.
: there exists $\theta \in (0,1)$ such that $f(v)<0$ for $v \in
(0,\theta)$, $f(v)>0$ for $v \in (\theta, 1)$
iv.
: $\di\int_0^1 f(v) {\,\mathrm{d}}v >0$,
it is well-known [@kanel] that there exists a one-dimensional travelling front $v(t,z)=\phi_0(z+c_0 t)$ solution to with $N=1$. The speed $c_0$ is unique and strictly positive by $[\tt iv]$ while the profile $\phi_0$ is unique up to translations. This result defines the constant $c_0>0$ that appears in equation .
In the case $N=2$, multidimensional solutions to are well understood. Paper [@HMR1] proves the existence of conical travelling waves solutions to , and paper [@HMR3] classifies all possible bounded non constant travelling waves solutions under rather weak conditions at infinity. In particular, it is proved in [@HMR3] that $c \geq c_0$ and, up to a shift in $x\in \RR$, either $u$ is a planar front $\phi_0(\pm x \cos \alpha+z \sin \alpha)$ with $\alpha=\arcsin(c_0/c) \in (0,\frac{\pi}{2}]$ or $u$ is the unique conical front found in [@HMR1].
In higher dimensions, less is known. In [@HMR3], Hamel, Monneau and Roquejoffre proved the existence of conical travelling waves with cylindrical symmetry whose level sets are Lipschitz graphs moving away logarithmically from straight cones. Some special, non cylindrically symmetric pyramidal-shaped solutions (Taniguchi, [@taniguchi2]) are also known in the particular case $N=3$.
Thus, in order to get a better understanding of the mechanisms at work, we further the idea of bridging reaction-diffusion equations with geometric motions. In particular, travelling wave graph solutions to the forced mean curvature motion go back to Fife [@fife]. He proved (in a formal fashion) that reaction-diffusion travelling fronts propagate with normal velocity $$V_n=-c_0 + \frac{\kappa}{t} +O\left(\frac{1}{t^2}\right) \, , \quad t>>1.$$ For a mathematically rigorous treatment of these ideas, we refer for instance to de Mottoni, Schatzman [@dMS] - small times, smooth solutions context - and Barles, Soner, Souganidis [@BSS] - arbitrary large times, viscosity solutions context.
Related results must me mentioned in the case of a balanced bistable non-linearity $f$: assumption `iv.` is replaced by $\int_0^1 f(v) {\,\mathrm{d}}v=0$ and is called the balanced Allen-Cahn equation. In this case, $c_0=0$ and the forced mean curvature equation is replaced by the mean curvature equation. Chen, Guo, Hamel, Ninomiya and Roquejoffre [@CGHNR] proved that there exist cylindrically symmetric traveling waves with paraboloid like interfaces solutions to in dimension $N \geq 3$. Precisely, they proved that those solutions’ level sets are asymptotically given by the equation $z=\frac{c}{2(N-1)} |x|^2$. On the other hand, Clutterbuck, Schnürer and Schulze [@clut] proved that there exists a unique rotationally symmetric, strictly convex, translating graph $u(t,x)=-ct+\phi(r)$ to the mean curvature motion with $c_0=0$ and whose asymptotics is given by $$\phi(r)= \frac{c}{2(N-1)} r^2 - \ln r + C + O\left( \frac{1}{r} \right)$$ Further works have also been done in the non radial case for the mean curvature equation. For instance, Xuan Hien Nguyen [@nguyen] built non radial and non convex translating graphs solution to with $c_0=0$.
Main results
------------
Our theorem \[le resultat\] below states that, given a $1$-homogeneous solution $\phi_\infty$ to the eikonal equation derived from (i.e. the equation obtained by removing the curvature term) there exists a smooth solution $\phi$ to the forced mean curvature motion equation whose asymptotic behaviour is prescribed by $\phi_{\infty}$. Here is the precise result.
[**(Existence of solutions with prescribed asymptotics in dim. $N$)**]{} \[le resultat\]\
Let $N \in \NN \setminus \{0,1\}$, $\alpha \in (0,\frac{\pi}{2}]$, $c_0>0$ and $c=c_0/\sin \alpha$. Choose $\phi_{\infty}$ a $1$-homogeneous viscosity solution to the eikonal equation $$\label{eq de phi infty}
|D\phi_{\infty}(x)|= \cot \alpha\ \, , \quad x\in \RR^{N-1} \, .$$ Then there exists a smooth concave solution $\phi \in C^{\infty}(\RR^{N-1})$ to such that $$\label{asymptotic of phi}
\phi(x)=\phi_{\infty}(x) + o(|x|) \ \ \hbox{as $\vert x\vert\to+\infty$.}$$
This is the most possible general result. However, due to the possible complexity of a solution to the eikonal equation , it is useful to specialise our result to the particular case of a solution with a finite number of facets.
[**(Solutions with finite number of facets in dimension $N$)**]{} \[le mini resultat\]\
Let $N \in \NN \setminus \{0,1\}$, $\alpha \in (0,\frac{\pi}{2}]$, $c_0>0$ and $c=c_0/\sin \alpha$. Choose $\phi^*$ a viscosity solution to the eikonal equation given for any $x \in \RR^{N-1}$ by $$\label{eq::profil}
\phi^*(x)=\inf_{\nu \in A} \left( -(\cot\alpha)\ x\cdot \nu + \gamma_\nu \right)$$ where $A$ is a finite subset of cardinal $k\in \NN^*$ of the sphere $\S^{N-2}$ and $\gamma_\nu$ are given real numbers. Then there exists a unique smooth concave solution $\phi \in C^{\infty}(\RR^{N-1})$ to such that $$\label{eq::rr2}
\left\{\begin{array}{l}
\displaystyle -\frac{2 \ln k}{c_0 \sin \alpha}\le \phi-\phi^* \le 0 \, , \quad x \in \RR^{N-1}\\
\\
\displaystyle \lim_{l\to +\infty} \sup_{\mbox{dist}(x,E_\infty)\ge l}|\phi(x)-\phi^*(x)| =0
\end{array}\right.$$ where $E_\infty$ is the set of edges defined as $$E_\infty=\{ x \in \RR^{N-1} \, | \, \phi_\infty \mbox{ is not $C^1$ at $x$}\}$$ with the $1$-homogeneous function $$\phi_\infty(x)=\inf_{\nu \in A} \left( -(\cot\alpha)\ x\cdot \nu \right)$$
In space dimension $N=3$, we obtain a more precise result by considering solutions having a finite number of gradient jumps. Those solutions are still more complex than the infimum of a finite number of affine forms. Here is the precise result.
[**(Solutions with finite number of gradient jumps and $N=3$)**]{} \[le micro resultat\]\
Let $\alpha \in (0,\frac{\pi}{2}]$, $c_0>0$ and $c=c_0/\sin \alpha$. Choose $\phi_{\infty}$ a $1$-homogeneous viscosity solution to the eikonal equation in dimension $N=3$ with a finite number of singularities on $\S^1$. Then, there exist
i.
: a $2\pi$-periodic continuous function $\psi_{\infty}:\theta \in [0,2\pi] \mapsto \psi_{\infty}(\theta) \in [-\cot \alpha, \cot \alpha]$ and a finite number $k \in \NN \backslash \left\{0\right\}$ of angles $\theta_1<\dots < \theta_k$ in $[0,2\pi)$ such that $$\phi_{\infty}(r\cos \theta, r \sin \theta)=r \psi_{\infty}(\theta) \, , \quad (r, \theta) \in \RR^+ \times [0,2\pi)$$ Moreover, for any $i \in \{1, \dots, k\}$,
a.
: Either $\forall \theta \in [\theta_i, \theta_{i+1}]$, $\psi_{\infty}(\theta)= -(\cot \alpha)$ and we set $\sigma_i=1$
b.
: Or $$\left\{\begin{array}{ll}
\forall \theta \in \left[\theta_i, \frac{\theta_i+ \theta_{i+1}}{2}\right] \, ,
& \psi_{\infty}(\theta)= -(\cot \alpha) \cos(\theta-\theta_i) \\
\forall \theta \in \left[\frac{\theta_i+ \theta_{i+1}}{2}, \theta_{i+1}\right] \, ,
& \psi_{\infty}(\theta)= -(\cot \alpha) \cos(\theta-\theta_{i+1})
\end{array}\right.
\mbox{ and we set } \sigma_i=0$$
By convention, $\theta_{k+1}=2\pi + \theta_1$ and $\sigma_{k+1}=\sigma_1$. If $k\ge 2$, then $\sigma_i \sigma_{i+1}=0$ for any $i\in \left\{1,...,k\right\}$.
ii.
: a smooth concave function $\phi \in C^{\infty}(\RR^2)$ solution to equation such that when $|x|$ goes to infinity $$\phi(x)=\phi_*(x)+O(1)$$ where $$\label{eq::rv22}
\phi_*(x)=-\frac{2}{c_0 \sin \alpha} \ln \left( \int_{\S^1}
e^{\frac{c_0 \cos \alpha}{2} x \cdot \nu} {\,\mathrm{d}}\mu (\nu) \right)$$ and $\mu$ is the non negative measure on $\S^1$ with finite mass determined by $\psi_{\infty}$ as follows: We set $\mu=\sum_{i=1}^k \mu_i$ where for any fixed $\lambda_0>0$, we set
a.
: If $\sigma_i=1$, then $\mu_i= \indicatrice_{(\theta_i,\theta_{i+1})} {\,\mathrm{d}}\theta
+ \lambda_0 (\delta_{\theta_i} + \delta_{\theta_{i+1}})$\
(with the exception for $k=1$: $\mu_1=\indicatrice_{(\theta_1,\theta_{1}+2\pi)} {\,\mathrm{d}}\theta$).
b.
: If $\sigma_i=0$, then $\mu_i= \lambda_0 (\delta_{\theta_i} + \delta_{\theta_{i+1}})$
We plan to use our travelling graphs for the forced mean curvature motion exhibited in theorems \[le resultat\] to \[le micro resultat\] in order to construct multi-dimensional travelling fronts to the reaction diffusion equation ; we plan to do it in a forthcoming paper.
That equation prescribes the asymptotic behaviour of has nothing surprising: let $\varepsilon >0$ and denote by $\phi_{\varepsilon}$ the scaled function $$\phi_{\varepsilon}(x)=\varepsilon \phi\left(\frac{x}{\varepsilon}\right) \, ,
\quad x \in \RR^{N-1}$$ Since $\phi$ is a solution to , $\phi_{\varepsilon}$ satisfies $$- \varepsilon \mbox{ div }\left( \frac{D\phi_{\varepsilon}}{\sqrt{1+|D\phi_{\varepsilon}|^2}}\right)
+c_0 - \frac{c}{\sqrt{1+|D\phi_{\varepsilon}|^2}}=0 \, , \quad x \in \RR^{N-1}$$ Let $\varepsilon$ go to zero. If adequate estimates for $\phi_{\varepsilon}$ are known, (a subsequence of) $(\phi_\varepsilon)_{\varepsilon>0}$ converges to a function $\phi_{\infty}$ satisfying .
The proof of Theorem \[le resultat\] is done by a sub and super solutions argument. We first construct a family of smooth sub-solutions to , which will give us some better insight in the equation. This step is quite general, and works in any space dimension. Then, we will construct a Lipschitz super-solution whose rescaled asymptotics is prescribed by $\phi_{\infty}$ and this will give us a smooth solution whose asymptotic behaviour is not well precise. To get a better asymptotics of the super-solution prescribed by the sub-solution, this will require a more delicate matching procedure which will limit us, for the moment, to any space dimension $N$ with a finite number of facets (theorem \[le mini resultat\]) or to the space dimension $N=3$ and a finite number of gradient jumps (theorem \[le micro resultat\]).
The rest of this paper is organised as follows. In section \[section eikonal equation\], we build and characterise all $1$-homogeneous solutions to the eikonal equation . In section \[section perron\], we detail Perron’s method in our context, and explain why it will yield a smooth concave solution. Sub-solutions are built in section \[section subsolution\], and super-solutions in section \[section supersolution\]. Finally, section \[proof le resultat\] sums up previous constructions to prove theorems \[le resultat\] and \[le mini resultat\]. Section \[section sept\] presents a more precise approach in dimension $N=3$ and details the proof of Theorem \[le micro resultat\]. An appendix is devoted to the Laplace’s method that we use in our estimates.\
[**Acknowledgments.**]{} The first author was partly supported by the ANR project MICA, the second and third ones by the ANR project PREFERED. They acknowledge a fruitful discussion with G. Barles and thank C. Imbert for enlightening discussions on his paper [@imbert]. They are also indebted to H. Berestycki and CAMS Center of EHESS in Paris for their hospitality while preparing this work.
Eikonal equation {#section eikonal equation}
================
In this section, we classify the continuous viscosity solutions to the eikonal equation in any dimension $N\geq 2$: $$\label{eikonal}
|D \phi_{\infty}(x)|=\cot \alpha \, , \quad x \in \RR^{N-1}$$ where $\alpha \in (0,\frac{\pi}{2}]$ is some given angle. In a first subsection, we are interested in the general case. In a second one, we reduce our study to $1$-homogeneous functions and give a better description of those solutions in order to use them in both sections \[section subsolution\] and \[section supersolution\].
Characterisation of solutions to in any dimension $N$ {#solution eikonal dimension N}
-----------------------------------------------------
For any unit vector $\nu\in \S^{N-2}$ and $\gamma \in (-\infty,+\infty]$, let us define the affine map $$\phi_{\nu,\gamma}(x)=- (\cot \alpha) \ \nu\cdot x + \gamma \in (-\infty,+\infty] \, , \quad x \in \RR^{N-1}$$
[**(A Liouville theorem for the eikonal equation)**]{} \[th::1\]\
Let $\phi_{\infty} \in C(\RR^{N-1})$. Then $\phi_{\infty}$ is a viscosity solution to the eikonal equation (\[eikonal\]) if and only if there exists a lower semi-continuous map $\gamma: \S^{N-2} \to (-\infty,+\infty]$ such that $$\label{eq::5}
\phi_{\infty}(x)=\inf_{\nu\in \S^{N-2}} \phi_{\nu,\gamma(\nu)}(x)$$ Moreover $\phi_{\infty}$ is $1$-homogeneous if and only if for all $\nu\in\S^{N-2}$, $\gamma(\nu)\in \left\{0,+\infty\right\}$.
This result is most certainly known. Because we not only need the result but also an insight of the construction, we give a complete proof.
[**Proof of Proposition \[th::1\].**]{}\
We first show the direct implication. Let $\phi_{\infty} \in C(\RR^{N-1})$ be a viscosity solution to . We shall prove that $\phi_{\infty}$ is $(\cot \alpha)$-Lipschitz and concave before giving its characterisation as an infimum of affine maps.
[**Step 1: $\phi_{\infty}$ is locally Lipschitz**]{}\
Let us consider the ball $B(a,R)$ centered in $a \in \RR^{N-1}$ with radius $R>2$ and define $$C:=\sup_{|x-y|\le 1,\ (x,y)\in B(a,R)^2} |\phi_{\infty}(x)-\phi_{\infty}(y)|$$ Because $\phi_{\infty}$ is continuous, $0 \leq C <+\infty$. Denote $\bar{C}=\max(C,\cot\alpha) \in (0,+\infty)$. Then we claim that for any $(x,y) \in B(a,R-1)^2$ such that $|x-y| \leq 1$, we have $$\label{eq::1}
|\phi_{\infty}(x)-\phi_{\infty}(y)|\leq \bar{C} |x-y|$$ which asserts that $\phi_{\infty}$ is locally Lipschitz. Indeed, for any point $x_0 \in \overline{B(a,R-1)}$, any constant $\bar{\bar{C}}> \bar{C}$ and any $\lambda\geq 0$, we consider the function $\psi_{\lambda}$ defined as $$\psi_\lambda(x):=\lambda + \phi_{\infty}(x_0)+ \bar{\bar{C}}|x-x_0| \, , \quad x \in \RR^{N-1}$$ and we set $$\lambda_*=\inf\left\{\lambda \in \RR^+ \, | \, \forall \mu \geq \lambda \, , \,
\forall x \in \overline{B(x_0,1)} \, , \quad \psi_\mu(x) \geq \phi_{\infty}(x) \right\}$$ We shall prove by contradiction that $\lambda_*=0$. If not, because $\psi_0 \geq \phi_{\infty}$ on $\left\{x_0\right\}\cup \partial B(x_0,1)$, there exists a contact point $z_0$ between $\psi_{\lambda_*}$ and $\phi_{\infty}$ which satisfies $z_0\in B(x_0,1)\backslash \left\{x_0\right\}$. Then $\psi_{\lambda_*}$ is a test function for the viscosity subsolution $\phi_{\infty}$ at that point. Because $|\nabla \psi_{\lambda_*}(z_0)|= \bar{\bar{C}} > \cot\alpha$, we get a contradiction with the viscosity subsolution inequality. Therefore $\lambda_*=0$ and $\psi_0\geq \phi_{\infty}$ on $B(x_0,1)$. Because this is true for any $\bar{\bar{C}}> \bar{C}$, we deduce that this is still true for $\bar{\bar{C}}= \bar{C}$ which implies (\[eq::1\]).\
[**Step 2: $\phi_{\infty}$ is $(\cot\alpha)$-Lipschitz**]{}\
We now define $$L=\limsup_{n \to +\infty} L_n \quad \mbox{with}\quad
L_n:= \sup\left\{\frac{\phi_{\infty}(y)-\phi_{\infty}(x)}{|y-x|}, \quad x\in \overline{B(a,R-2)},\quad
|y-x|\le \frac{1}{n} \right\}$$ Notice that for any $n \in \NN^*$, $0<L_n \le \bar C$. Moreover, there exists a sequence $(x_n,y_n)_{n \in \NN^*}$ such that $$\lim\limits_{n \to +\infty}\frac{\phi_{\infty}(y_n)-\phi_{\infty}(x_n)}{|y_n-x_n|} = L
\, \mbox{ and } |y_n-x_n| \leq \frac{1}{n} \, \mbox{ with } x_n \in \overline{B(a,R-2)}$$ Define for any $x \in \RR^{N-1}$ $$\varepsilon_n=|y_n-x_n| \, , \quad \phi_n(x)=
\frac{\phi_{\infty}(x_n+ \varepsilon_n x)-\phi_{\infty}(x_n)}{\varepsilon_n}
\quad \mbox{and}\quad \nu_n = \frac{y_n-x_n}{\varepsilon_n} \in \S^{N-2}$$ Thus $\left(\phi_n(\nu_n)\right)_{n \in \NN^*}$ converges to $L$ as $n$ goes to infinity and for any $x \in \overline{B(0,1)}$, $|\phi_n(x)| \leq L_n |x|$. Because $\mbox{Lip}(\phi_n; B(0,n)) \le \bar C$, we see that up to a subsequence, $(\phi_n)_{n \in \NN^*}$ converges locally uniformly on $\RR^{N-1}$ to $\phi_0$ a viscosity solution to . Moreover, $\left(\nu_n\right)_{n \in \NN^*}$ converges to $\nu_0\in\S^{N-2}$ with $$\phi_0(\nu_0)=L \mbox{ and for any } x \in \RR^{N-1} \, , \quad \phi_0(x) \le L |x| =:\psi(x)$$ Because $\psi$ touches $\phi_0$ from above at $\nu_0$, we conclude from the viscosity inequality for subsolutions that $$L\le \cot\alpha$$ Now for any $\epsilon>0$, there exists $n_\varepsilon \in \NN^*$ such that $L_n \le L +\varepsilon$ for any $n\geq n_\varepsilon$. In particular, for any $(x,y) \in \overline{B(a,R-2)}$ we can split the segment $$\displaystyle [x,y] = \bigcup_{i=0,...,K-1} [x_i,x_{i+1}] \mbox{ with } x_0=x
\, , \, x_{K}=y \mbox{ and } |x_{i+1}-x_i|=\frac{|y-x|}{ K} \le \frac{1}{n}$$ This implies that $$|\phi_{\infty}(x)-\phi_{\infty}(y)|\le (L+\varepsilon) |x-y|$$ which is true for any $\varepsilon>0$. This implies that $\phi$ is $L$-Lipschitz on $\overline{B(a,R-2)}$ with $L\le \cot\alpha$.
[**Step 3: $\phi_{\infty}$ is concave**]{}\
Because $\phi_{\infty}$ is a Lipschitz stationary viscosity solution to the evolution equation $$u_t + H(Du)=0\, , \quad x \in \RR^{N-1} \mbox{ where }
H(p)=\frac12 (p^2- \cot^2\alpha) \, , \quad p \in \RR^{N-1}$$ we can apply Lemma 4 page 131 in [@evans], and get that $\phi_{\infty}$ satisfies for any $t>0$ $$\phi_{\infty}(x+x')-2\phi_{\infty}(x)+\phi_{\infty}(x-x')
\leq C_0\frac{|x'|^2}{t},\quad \mbox{for all } (x,x')\in\RR^{2(N-1)}$$ and we can check that we have $C_0= 1$. Letting $t$ go to infinity shows that $\phi_{\infty}$ is concave in $\RR^{N-1}$.
[**Step 4: Tangent cone**]{}\
Since $\phi_{\infty}$ is Lipschitz continuous, it is differentiable almost everywhere by Rademacher’s theorem. Let $D \subset \RR^{N-1}$ be the set of differentiability of $\phi_{\infty}$ and fix $x_0\in D$. Since $\phi_{\infty}$ is concave, for any $x \in \RR^{N-1}$, we have $$\phi_{\infty}(x) \leq \phi_{\infty}(x_0) + D\phi_{\infty}(x_0) \cdot (x-x_0)$$ Passing to the infimum on $D$, we get for any $x \in \RR^{N-1}$, $$\phi_{\infty}(x) \leq \psi(x):= \inf_{x_0 \in D} \phi_{\infty}(x_0) +
D\phi_{\infty}(x_0) \cdot (x-x_0)$$ Thus, $\psi$ and $\phi_{\infty}$ are $(\cot \alpha)$-Lipschitz functions that coincide on $D$ which is a dense set on $\RR^{N-1}$. Therefore, they are in fact equal on $\RR^{N-1}$. Using equation , we finally have $$\phi_{\infty}(x)=\inf_{x_0 \in D} -(\cot \alpha) \, \nu(x_0)\cdot x + g(x_0)$$ where for any $x_0 \in D$, $\nu(x_0)=-D\phi_{\infty}(x_0)/\cot\alpha \in \S^{N-2}$ and $g(x_0)=\phi_{\infty}(x_0) - x_0 \cdot D\phi_{\infty}(x_0) \in \RR$. Defining $\gamma$ as $$\begin{array}{cccl}
\gamma :& \S^{N-2}& \to & (-\infty,+\infty] \\
& \nu & \mapsto & \left\{ \begin{array}{l}
\inf_{x_0 \in A} g(x_0) \mbox{ if } A:= \{x_0 \in D \, | \, \nu(x_0)=\nu\} \neq \emptyset \\
+\infty \mbox{ otherwise } \end{array}\right.
\end{array}$$ we get the desired characterisation (\[eq::5\]). Since $\phi_{\infty}$ is continuous, we also deduce from (\[eq::5\]) that $\gamma$ is lower semi-continuous.
[**Step 5: The $1$-homogeneous case**]{}\
We assume that $\phi_{\infty}$ is a $1$-homogeneous continuous viscosity solution to . Then for any $x_0\in \RR^{N-1}$, there exists $p\in\RR^{N-1}$ with $|p|=\cot\alpha$ such that by $$\forall x \in \RR^{N-1} \, , \quad \phi_{\infty}(x) \leq \phi_{\infty}(x_0)+p\cdot (x-x_0)$$ On the one hand, considering $x=0$, we get $$p\cdot x_0 \leq \phi_{\infty}(x_0)$$ because $\phi_{\infty}$ is $1$-homogeneous. On the other hand considering $\lambda x$ instead of $x$ and taking the limit $\lambda \to +\infty$, we get $$\psi(x):=p\cdot x \ge \phi_{\infty}(x) \quad \mbox{with equality at}\quad x=x_0 \, .$$ Therefore if we call ${\mathcal L}_{\phi_{\infty}}$ the set of linear functions $\psi$ satisfying $\psi\ge \phi_{\infty}$ such that $|\nabla \psi|=\cot\alpha$, we have $$\phi_{\infty} = \inf_{\psi\in {\mathcal L}_{\phi_{\infty}}} \psi$$ because this is true at any point $x_0\in\RR^{N-1}$.\
[**Step 6: Conclusion**]{}\
Conversely, if a function $\phi_{\infty}$ is given by (\[eq::5\]), then it is straightforward to check that $\phi_{\infty}$ is a viscosity solution to (\[eikonal\]).
------------------------------------------------------------------------
In dimension $N=2$, the previous proposition simply reads:\
If $N=2$ and $\alpha \in (0,\frac{\pi}{2}]$, $\phi_{\infty}$ is a viscosity solution to if and only if $\phi_{\infty}$ is affine or if there exists $(x_0,y_0) \in \RR^2$ such that $$\label{valeur absolue}
\phi_{\infty}(x)=- (\cot \alpha) ~ |x-x_0|+y_0 \, , \quad x \in \RR$$ Moreover, $\phi_{\infty}$ is $1$-homogeneous if and only if $y_0=0$.
The proof of this proposition can also be done directly from definitions of viscosity solutions and we omit the details. Notice however the link with [@HMR1]: two-dimensional reaction diffusion waves are either planar fronts or the unique (up to translations) conical front whose level sets are asymptotics to the graph of $\phi_{\infty}$ just described.
The $1$-homogeneous case {#homogeneous}
------------------------
As stressed is theorem \[le resultat\], we only build solutions to the forced mean curvature motion equation whose asymptotics is prescribed by a $1$-homogeneous solution to the eikonal equation . Therefore, it is worth emphasising this particular case.
Notice however that there exist viscosity solutions to the eikonal equation defined in $\RR^{N-1}$ that are not homogeneous of order $1$. For instance, consider solutions given by with $x \in\RR^{N-1}$ and $y_0 \neq 0$. We can also consider any translation of a $1$-homogeneous solution. Another example is for instance given in dimension $N=3$ by a function $\phi_{\infty}= \inf_{i=1\dots 4} \phi_i$ where $(\phi_i)_{i \in \{1 \dots 4\}}$ are four planar solutions defined for $x=(x_1,x_2) \in \RR^2$ by $$\begin{aligned}
\phi_1(x) = - (\cot \alpha)\ x_1+2 \quad & \quad \phi_2(x)= (\cot \alpha)\ x_1+2 \\
\phi_3(x) = -(\cot\alpha) \ x_2 \quad & \quad \phi_4(x)=(\cot \alpha)\ x_2 \end{aligned}$$ It is straightforward to check that $\phi_{\infty}$ satisfies $|D\phi_{\infty}|=\cot \alpha$ in the viscosity sense and that it is not homogeneous of order $1$ since there exists $\lambda >0$ such that $\phi_{\infty}(\lambda,0)\neq \lambda \phi_{\infty}(1,0)$.
In any case, a solution $\phi_{\infty}$ to the eikonal equation is concave (see the proof of proposition \[th::1\], step 3). Therefore the function $g: \lambda \in \RR^{+*} \mapsto g(\lambda)=\phi_{\infty}(\lambda x)/(\lambda |x|) \in \RR$ is decreasing in $\lambda >0$. Since $\phi_{\infty}$ is $(\cot \alpha)$-Lipschitz, $g$ is bounded from below and for any $x \in \S^{N-2}$, the limit $$\lim_{\lambda \to+\infty}\frac{\phi_{\infty}(\lambda x)}{|\lambda|}$$ exists and $\phi_{\infty}$ is asymptotically homogeneous. Thus we have a fairly general understanding of what is going on by restricting ourselves to homogeneous solutions to equation .
\[countable\] **(A countable characterisation of homogeneous solutions)**\
Let $\phi_{\infty} \in C(\RR^{N-1})$. Then $\phi_{\infty}$ is a $1$-homogeneous viscosity solution to the eikonal equation if and only if there exists a sequence $(\nu_i)_{i \in \NN}$ of $\S^{N-2}$ such that $$\label{phi countable}
\phi_{\infty}(x)=\inf_{i \in \NN} -(\cot \alpha) \ \nu_i \cdot x$$
**Proof of Proposition \[countable\].**\
Let $\phi_{\infty} \in C(\RR^{N-1})$ be a $1$-homogeneous viscosity solution to . According to proposition \[th::1\], there exists a lower semi continuous function $\gamma$ defined from $\S^{N-2}$ to $\{0,+\infty\}$ such that $$\phi_{\infty}(x)=\inf_{\nu \in \S^{N-2}} \phi_{\nu,\gamma(\nu)}(x) \, , \quad x\in \RR^{N-1}$$ Then $K=\{\nu \in \S^{N-2} \, | \, \gamma(\nu) =0\}$ is a compact set of $\S^{N-2}$. We claim (see lemma \[cube\] and corollary \[cor cube\] below) that there exists a sequence $(\nu_i)_{i \in \NN}$ of $\S^{N-2}$ such that $$K=\overline{\bigcup_{i \in \NN} \{\nu_i\}}$$ Thus, $\phi_{\infty}(x)$ can be described as the infimum over $\nu \in K$ of the linear functions $-(\cot \alpha) \ x \cdot \nu$. Since $\cup_{i \in \NN} \{\nu_i\}$ is dense in $K$, $\phi_{\infty}(x)$ can also be written as the infimum over $i \in \NN$ of the linear functions $-(\cot \alpha) \ x \cdot \nu_i$. This ends the proof of proposition \[countable\] since the converse implication is straightforward.
------------------------------------------------------------------------
**(Decomposition of a compact set of $\S^{N-2}$ in cubes)**\
\[cube\] For any compact set $K$ of $\S^{N-2}$, there exists a countable family $(Q_i)_{i \in \NN}$ of closed cubes of $\RR^{N-1}$ such that $$\label{assum}
\left\{\begin{array}{l}
\forall n \in \NN \, , \quad \displaystyle K \subset \bigcup_{i \geq n} Q_i \\
\forall i \in \NN \, , \quad Q_i \cap K \neq \emptyset \\
\displaystyle \limsup_{i \to +\infty} \mbox{diam}(Q_i)=0
\end{array}\right.$$
**Proof of Lemma \[cube\].**\
We built this decomposition in cubes by induction. Let $C_0=[-1,1]^{N-1}$ be the first cube of width $2$. Thus $K \subset C_0$. Since $C_0 \cap K$ is not empty, we divide $C_0$ in $2^{N-1}$ smaller cubes of width $2^{0}=1$. We call $C_{1,i}$ for $i=1 \dots n_1$ those whose intersection with $K$ is not empty. Then, $1 \leq n_1 \leq 2^{N-1}$ and $$K\quad \subset \quad \bigcup_{i=1}^{n_1} C_{1,i}$$ In the same way, for $i=1,...,n_1$, we divide each cube $C_{1,i}$ in $2^{N-1}$ smaller cubes of width $2^{-1}$ and keep only those whose intersection with $K$ is not empty. We call them $C_{2,k}$ for $k=1 \dots n_2$ and $1 \leq n_2 \leq 2^{N-1} n_1$. Then, one can easily verify that $K \subset \cup_{k=1 \dots n_2} C_{2,k}$.
Assume the cubes $C_{j,i}$ are built for $j\in \NN$, $i=1 \dots n_j$ and $1 \leq n_j \leq 2^{j(N-1)}$ such that $$\left\{\begin{array}{l}
K \subset \bigcup_{i =1}^{n_j} C_{j,i} \\
\forall i =1 \dots n_j \, , \quad C_{j,i} \cap K \neq \emptyset \\
\mbox{diam}(C_{j,i}) = 2^{-j+1}
\end{array}\right.$$ Then we construct the cubes $C_{j+1,i}$ as follows. We divide each cube $C_{j,i}$ into $2^{N-1}$ smaller cubes of width $2^{-j}$ and keep only those whose intersection with $K$ is not empty. We call them $C_{j+1,i}$ for $i=1 \dots n_{j+1}$ and $1 \leq n_{j+1} \leq 2^{N-1} n_j\leq 2^{(j+1)(N-1)}$. By construction, it is easy to verify that $$\left\{\begin{array}{l}
K \subset \bigcup_{i =1}^{n_{j+1}} C_{j+1,i} \\
\forall i =1 \dots n_{j+1} \, , \quad C_{j+1,i} \cap K \neq \emptyset \\
\mbox{diam}(C_{j+1,i}) = 2^{-j}
\end{array}\right.$$ The induction is then proved. We thus construct a countable family of cubes that we recall $(Q_j)_{j \in \NN}$ for convenience with the desired assumptions . This ends the proof of lemma \[cube\].
------------------------------------------------------------------------
\[cor cube\] **(Representation of a compact set of $\S^{N-2}$)**\
For any compact set $K$ of $\S^{N-2}$, there exists a sequence $(\nu_j)_{j \in \NN}$ of $\S^{N-2}$ such that $$K=\overline{\bigcup_{j \in \NN} \{\nu_j\}}$$
**Proof of Corollary \[cor cube\].**\
For $K$ a compact set of $\S^{N-2}$, we define $(Q_j)_{j \in \NN}$ a family of cubes as proposed in lemma \[cube\]. For any $j \in \NN$, we choose $\nu_j \in K \cap Q_j$. Then, it is straightforward to check that $\overline{\cup_{j \in\NN} \{\nu_j\}} \subset K$. Regarding the converse inclusion, we fix $x_0 \in K$ and $\varepsilon >0$. By , there exists $n_{\varepsilon} \in \NN$ such that the width of cube $Q_i$ is smaller than $\varepsilon$ provided $i \geq n_{\varepsilon}$. Since $K \subset \cup_{i \geq n_{\varepsilon}} Q_i$, there exists $i_{\varepsilon} \geq n_{\varepsilon}$ such that $$x_0 \in Q_{i_{\varepsilon}} \, \mbox{ and } \, |x_0-\nu_{i_{\varepsilon}}|\leq \varepsilon \sqrt{N-1}$$ This shows the density of $\cup_{j \in \NN} \{\nu_j\}$ in $K$ and ends the proof of corollary \[cor cube\].
------------------------------------------------------------------------
Perron’s method and comparison principle {#section perron}
========================================
In this section, we are concerned with the forced mean curvature motion equation $$\label{MCM}
- \mbox{ div }\left( \frac{D\phi}{\sqrt{1+|D\phi|^2}}\right) +c_0 -
\frac{c}{\sqrt{1+|D\phi|^2}}=0 \, , \quad x \in \RR^{N-1}$$ with the condition at infinity $$\label{CL}
\phi(x) =\phi_{\infty}(x) + o(|x|) \, , \quad x \in \RR^{N-1}$$ where $\phi_{\infty}$ is a homogeneous viscosity solution to $|D\phi_{\infty}|=\cot \alpha$ found in section \[section eikonal equation\] with $\alpha =\arcsin(c_0/c) \in (0,\frac{\pi}{2}]$. We choose to solve using Perron’s method with sub and super-solutions (see [@user] or [@GT]). Let us first recall the existence process and clarify the regularity of the solution in the following
[**(Existence of a solution to in dimension $N$)**]{}\
\[perron\] Let $N \in \NN \setminus \{0,1\}$, $(c_0,c) \in \RR^2$ such that $c \geq c_0>0$. Assume that $\phi_*$ is a viscosity sub-solution and $\phi^*$ a viscosity super-solution to such that $\phi_* \leq \phi^*$ on $\RR^{N-1}$.\
Then,\
[**i)**]{} there exists a function $\phi \in [\phi_*,\phi^*]$ viscosity solution to .\
[**ii)**]{} Moreover, if $\phi^*$ is concave, and satisfies the following technical condition: $$\label{eq::tech}
\mbox{there exists } p\in\RR^{N-1} \mbox{ such that } \limsup_{|x|\to +\infty} \frac{\phi^*(x)-p\cdot x}{|x|} <0,$$ then $\phi$ can be chosen concave and smooth.
**Proof of Proposition \[perron\].**\
We build the solution $\phi$ using Perron’s method directly in the framework of viscosity solutions to , that is to say $\phi$ is chosen as the maximal sub-solution to (see the user’s guide to viscosity solutions [@user]).\
[**Step 1: Concavity**]{}\
We apply a result due to Imbert (see [@imbert]) that we first recall. Denote $F$ the following Hamiltonian $$F(p,M)=-\frac{{\mathrm {tr}}M}{\sqrt{1+\vert p\vert^2}}+\frac{{\mathrm {tr}}(M\cdot (p\otimes p))}{(1+\vert p\vert^2)^{3/2}}+c_0-\frac{c}{\sqrt{1+\vert p\vert^2}} \, , \quad (p,M) \in \RR^{N-1} \times \RR^{(N-1)\times (N-1)}_{\textup{sym}}$$ where $\RR^{(N-1)\times (N-1)}_{\textup{sym}}$ is the set of $(N-1)$-square symmetric matrices.
\[imbert\] [**(Imbert’s proposition 5 in [@imbert])**]{}\
Let $u$ be a lower semi-continuous and epi-pointed function. If $u$ is a supersolution to $$F( D u(x), D^2u(x))=0 \, , \quad x \in \RR^{N-1}$$ then, so is its convex envelope.
In our context, $-\phi^*$ is epi-pointed because of the technical condition (\[eq::tech\]). Thus, the maximal sub-solution $\phi$ to is concave (otherwise Imbert’s result stated in proposition \[imbert\] contradicts the maximal property of $\phi$). Then, $\phi$ is a concave viscosity solution to .
[**Step 2: Regularity**]{}\
Once concavity is at hand, a Lipschitz bound is automatically available from the equation itself: $$|D\phi(x)| \leq \cot \alpha \, , \quad x \in \RR^{N-1}$$ where $\alpha \in (0,\frac{\pi}{2}]$ is such that $c_0=c \sin \alpha$. Then $F$ becomes uniformly elliptic, thus allowing for $C^{1,1}$ estimates (see Theorem 4 in [@imbert]). A bootstrap argument then shows that the solution is $C^\infty$. This concludes the proof of proposition \[perron\].
------------------------------------------------------------------------
Notice that the condition (\[eq::tech\]) is hidden in the statement of Proposition 9 in [@imbert]. Thus, the proof of Proposition 5 in [@imbert] uses Proposition 9.
It now remains to find sub and super-solutions to .
Sub-solution {#section subsolution}
============
In this section we build smooth sub-solutions to the forced mean curvature equation as global solutions to a viscous eikonal equation and we do believe that they are really close to the desired solutions.
Sub-solutions as solutions to a viscous eikonal equation
--------------------------------------------------------
We have the following
\[CS\] [**(Sufficient condition for a sub-solution to )**]{}\
Fix $\alpha \in (0,\frac{\pi}{2}]$, $c_0>0$ and $c=c_0/\sin \alpha$. Let $\phi_*$ be a concave smooth solution to $$\label{MCM reg concave}
-\Delta \phi_* = \frac{c_0 \sin \alpha}{2} \left(\cot^2 \alpha-|D\phi_*|^2
\right) \, , \quad x \in \RR^{N-1}$$ such that $$\label{majoration grad}
|D\phi_*(x)| \leq \cot \alpha \, , \quad x\in \RR^{N-1}$$ Then $\phi_*$ is a smooth sub-solution to equation .
**Proof of Lemma \[CS\].**\
Let $\phi_*$ be any concave function verifying and . Since $\phi_*$ is smooth and concave, we have $$\begin{aligned}
N[\phi_*]:=- & \mbox{ div }\left( \frac{D\phi_*}{\sqrt{1+|D\phi_*|^2}}\right) +c_0 -
\frac{c}{\sqrt{1+|D\phi_*|^2}} \\
= & -\frac{\Delta \phi_*}{\sqrt{1+|D\phi_*|^2}} +
\frac{D^2\phi_*(D\phi_*,D\phi_*)}{(1+|D\phi_*|^2)^{\frac{3}{2}}} +c_0 -
\frac{c}{\sqrt{1+|D\phi_*|^2}} \\
\leq & \frac{1}{\sqrt{1+|D\phi|^2}} \left( -\Delta \phi_*
+c_0\sqrt{1+|D\phi_*|^2}-c \right)\end{aligned}$$ From and $c_0=c \sin\alpha$, we deduce that $$\begin{aligned}
\cot^2\alpha -|D\phi_*|^2& =\left(\frac{c}{c_0}\right)^2 - \left(\sqrt{1+|D\phi_*|^2}\right)^2 \\
& \leq \frac{2c}{c_0} \left(\frac{c}{c_0}-\sqrt{1+|D\phi_*|^2}\right)\end{aligned}$$ Using equation satisfied by $\phi_*$, we get $$N[\phi_*] \leq \frac{1}{\sqrt{1+|D\phi_*|^2}} \left(
\frac{c}{c_0}-\sqrt{1+|D\phi_*|^2} \right) \left( \frac{2c}{c_0}
\frac{c_0^2}{2c}-c_0 \right)=0.$$ Thus, $\phi_*$ is a sub-solution to .
------------------------------------------------------------------------
\
As it is well-known, equation is readily transformed into a linear one by the Hopf-Cole transform $$\tilde\phi_*(x)={\mathrm{exp}}\biggl(-\frac{c_0\sin\alpha}{2}
\phi_*\left(\frac{2x}{c_0\cos\alpha}\right)\biggl) \, , \quad x \in \RR^{N-1}
\, , \quad \alpha \neq \frac{\pi}{2}$$ where $\tilde \phi_*$ is a positive solution to $$\label{e4.1}
-\Delta \tilde{\phi}_*(x) + \tilde{\phi}_*(x) =0 \, , \quad x \in \RR^{N-1}$$ From [@caf], a positive solution $\tilde\phi_*$ to has the form $$\tilde{\phi}_*(x)= \int_{\S^{N-2}} e^{\nu \cdot x} {\,\mathrm{d}}\mu(\nu) \, , \quad x \in
\RR^{N-1},$$ where $\mu$ is a non negative measure on $\S^{N-2}$ with finite mass.
Now, for any non negative measure $\mu$ on the sphere $\S^{N-2}$, let us define $$\label{def de psi}
\phi_*(x)= - \frac{2}{c_0 \sin \alpha} \ln \left( \int_{\S^{N-2}} e^{\frac{c_0
\cos \alpha}{2} x \cdot \nu} {\,\mathrm{d}}\mu(\nu) \right) \, , \quad x \in \RR^{N-1}
\, , \quad \alpha \in \left(0,\frac{\pi}{2}\right]$$ By construction, $\phi_*$ is a smooth solution to . Let us now prove that $\phi_*$ is a sub-solution to equation , with all the requirements.
\[lm psi\] [**(Inequalities for the derivatives of $\phi_*$)**]{}\
Let $\mu$ be a non negative measure on $\S^{N-2}$ with finite mass, $\alpha \in (0,\frac{\pi}{2}]$, $c_0>0$ and $c=c_0/\sin\alpha$. Define $\phi_*$ as in . Then $\phi_*$ is a smooth concave solution to (\[MCM reg concave\]) and its gradient is uniformly bounded, that is to say, for any $(x,\xi) \in \RR^{N-1} \times \RR^{N-1}$, $$\label{prop de psi}
|D\phi_*(x)|\leq \cot \alpha \, , \quad D^2\phi_*(x)(\xi,\xi) \leq 0$$
**Proof of Lemma \[lm psi\]**.\
Let $\mu$ and $\phi_*$ be so defined. We have $$D\phi_*(x)=-(\cot \alpha) \, \frac{F_{\nu}(x)}{F_1(x)} \, , \quad x \in \RR^{N-1}$$ where for any continuous (scalar or vector) function $f$ defined on $\S^{N-2}$ $$F_f(x)=\int_{ \S^{N-2}} e^{\frac{c_0 \cos \alpha}{2} x \cdot \nu}f(\nu) {\,\mathrm{d}}\mu(\nu)$$ Remark that if we define for some fixed $x \in \RR^{N-1}$, $\int_{\S^{N-2}} f(\nu) {\,\mathrm{d}}\overline{\mu}_x(\nu)=\frac{F_f(x)}{F_1(x)}$, then $\overline{\mu}_x$ is a probability measure on $\S^{N-2}$. We can then apply Jensen’s inequality to the convex function $y \mapsto |y|^2$. This gives $$\label{Jensen}
\left|\frac{F_{f}(x)}{F_1(x)}\right|^2 =\left| \int_{\S^{N-2}} f(\nu) {\,\mathrm{d}}\overline{\mu}_x \right|^2
\leq \int_{\S^{N-2}} |f(\nu)|^2 {\,\mathrm{d}}\overline{\mu}_x = \frac{F_{|f|^2}(x)}{F_1(x)}$$ for any continuous function $f$ defined on $\S^{N-2}$. Applying this inequality to $f(\nu)=\nu$, we get the desired bound on the gradient of $\phi_*$: $|D\phi_*(x)|\leq \cot \alpha$.
Regarding the concavity property of $\phi_*$, we use the same type of arguments. Indeed, for any $\xi \in \RR^{N-1}$ and $x \in \RR^{N-1}$, we have $$D^2\phi_*(x)(\xi,\xi)=-\frac{c_0 \cos^2 \alpha}{2 \sin \alpha}
\left(\frac{F_{f^2}(x)}{F_1(x)} - \left(\frac{F_f(x)}{F_1(x)}\right)^2\right)$$ where $f$ is the continuous function defined on $\S^{N-2}$ by $f(\nu)=\nu \cdot \xi$. Applying again Jensen’s inequality , we conclude that $D^2\phi_*(x)(\xi,\xi) \leq 0$ for any $\xi \in \RR^{N-1}$ and $x \in \RR^{N-1}$ which shows that $\phi_*$ is concave.
------------------------------------------------------------------------
Finally, we proved the following proposition:
**(Existence of a sub-solution to )**\
\[existence subsolution\] Fix $\alpha \in (0,\frac{\pi}{2}]$, $c_0>0$ and $c=c_0/\sin \alpha$. Let $\mu$ be a non negative measure on $\S^{N-2}$ with finite mass. Define $\phi_*$ as in . Then, $\phi_*$ is a smooth concave sub-solution to .
The way we choose the measure $\mu$ is decisive in the asymptotic behaviour of the sub-solution $\phi_*$ built as in proposition \[existence subsolution\]. Indeed, if we want the subsolution (and hence the solution) to the mean curvature equation to follow asymptotically some given solution $\phi_{\infty}$ to the eikonal equation , we will have to choose the measure $\mu$ carefully. In that procedure, information collected in section \[section eikonal equation\] will help.
Of course, it will be also very interesting to assess whether each sub-solution built with a general probability measure gives rise to a solution to the mean curvature equation .
Super-solution {#section supersolution}
==============
A natural super-solution to the forced mean curvature equation is a viscosity solution $\phi_{\infty}$ to the eikonal equation . Indeed, $\phi_{\infty}$ satisfies (in the distributional and viscosity sense) $$-\mbox{ div }\left( \frac{D\phi_{\infty}}{\sqrt{1+|D\phi_{\infty}|^2}}\right)
+c_0 - \frac{c}{\sqrt{1+|D\phi_{\infty}|^2}} \geq 0 \, , \quad x \in \RR^{N-1}$$ However, this super-solution does not satisfy the right comparison with the previous sub-solution $\phi_*$. For instance, if $N=3$, $\phi_{\infty}$ is the radially symmetric viscosity solution to the eikonal equation and $\phi_*$ the sub-solution associated with the Lebesgue measure $\mu={\,\mathrm{d}}\theta$ on $[0,2\pi]$ as in , then we can compute the asymptotic behaviour of both functions $\phi_{\infty}$ and $\phi_*$ for $x \in \RR^2$ with $|x|$ large enough using Laplace’s method (see appendix \[laplace\]). We then observe that the sub-solution $\phi_*$ is above the super-solution $\phi_{\infty}$ in this area. This contradicts the crucial assumption $\phi_* \leq \phi^*$ on $\RR^2$ in the Perron’s method (see proposition \[perron\]).
Super-solutions as infimum of hyperplanes
-----------------------------------------
Since we do believe that the sub-solution is close to the viscosity solution to the forced mean curvature equation at infinity, we prefer to change the super-solution. In the general case of dimension $N$, we use the countable characterisation of the solution $\phi_{\infty}$ to the eikonal equation that we want to approach (see proposition \[countable\]).
\[supersolution dim N\] **(Existence of a super-solution to )**\
Fix $\alpha \in (0,\frac{\pi}{2}]$, $c_0>0$ and $c=c_0/\sin \alpha$. Choose $\phi_{\infty}$ a $1$-homogeneous solution to the eikonal equation . Define $(\nu_i)_{i \in \NN}$ the sequence of $\S^{N-2}$ given by its countable characterisation in proposition \[countable\].
For any sequence $(\lambda_i)_{i \in \NN}$ such that $\lambda_i>0$ and $\sum_{i \in \NN} \lambda_i <+\infty$, we set $$\phi_i(x)=-(\cot \alpha) \ x \cdot \nu_i - \frac{2}{c_0 \sin \alpha}
\ln \lambda_i \, , \quad i \in \NN \, , \quad x \in \RR^{N-1}$$ and $$\phi^*(x)= \inf_{i \in \NN} \phi_i(x) \, , \quad x \in \RR^{N-1}$$ Then, $\phi^*$ is a concave continuous super-solution to .
**Proof of Proposition \[supersolution dim N\].**\
Since $\phi_i$ are exact solutions to the forced mean curvature equation , it is clear that $\phi^*$ is a super-solution to that equation. As the infimum of affine functions, it is concave and continuous.
------------------------------------------------------------------------
This construction is very easy. However, it is not clear whether the technical condition is satisfied or not. It is even clear that when the set of $\{\nu_i\}_{i \in \NN}$ is finite of cardinal less or equal to $N-1$, this condition is NOT verified. We will see later (see Step 4 of the proof of theorem \[le resultat\]) how to modify the sub- and super-solutions in order to satisfy condition and then pass to the limit to recover the general case.
In the case when the set $\{\nu_i\}_{i \in \NN}$ is infinite, the convergence of $\sum \lambda_i$ forces $(\lambda_i)_{i \in \NN}$ to go to zero and the sequence $(-\ln \lambda_i)_{i \in \NN}$ grows as $i$ goes to infinity.
General existence results {#proof le resultat}
==========================
Now equipped with sub and super-solutions as well as a Perron’s method, we are able to prove existence results. The general case in dimension $N \geq 2$ is the easiest one since the asymptotics is less precise. Let us explain our ideas in details depending on the degree of precision we want to obtain in our construction.
Let $N \in \NN\setminus \{0,1\}$, $\alpha \in (0,\frac{\pi}{2}]$, $c_0>0$ and $c=c_0/\sin \alpha$. It is worth noticing that some of our constructions do not work for $\alpha=\pi/2$. However, this case is obvious and leads to planar fronts. Therefore, we restrict ourselves to $\alpha \in (0, \frac{\pi}{2})$.
Proof of Theorem \[le resultat\] {#demo main result}
--------------------------------
**Step 1: Sub and super-solutions**\
Choose $\phi_{\infty}$ a $1$-homogeneous continuous viscosity solution to the eikonal equation in $\RR^{N-1}$. By proposition \[countable\], there exists a sequence $(\nu_i)_{i \in \NN}$ of $\S^{N-2}$ such that $$\phi_{\infty}(x)= \inf_{i \in \NN} -(\cot \alpha) \nu_i \cdot x$$ Let $\mu$ be the probability measure on $\S^{N-2}$ be defined as $$\mu=\sum_ {i \in \NN} \lambda_i \delta_{\nu_i}$$ where $(\lambda_i)_{i \in \NN}$ are chosen so that $\lambda_i >0$ and $\sum_{i =0}^{+\infty} \lambda_i =1$.
Build the sub-solution $\phi_*$ as in with the above measure $\mu$. Then by proposition \[existence subsolution\], $\phi_*$ is a smooth concave sub-solution to the mean curvature motion equation . Build a concave continuous super-solution $\phi^*$ by proposition \[supersolution dim N\] as the infimum of hyperplanes where the $(\lambda_i)_{i \in \NN}$ and $(\nu_i)_{i \in \NN}$ are defined by the choice of $\mu$. For any $x \in \RR^{N-1}$, $$\phi_*(x) =-\frac{2}{c_0 \sin \alpha} \ln \left( \sum_{i=0}^{+\infty}
\lambda_i \ e^{\frac{c_0 \cos \alpha}{2} \nu_i \cdot x}\right)
\leq -\frac{2}{c_0 \sin \alpha} \ln
\left( \lambda_i \ e^{\frac{c_0 \cos \alpha}{2} \nu_i \cdot x}\right)$$ Since the last inequality holds for any $i \in \NN$, we have $$\phi_*(x) \leq \inf_{i \in \NN } \left(-(\cot \alpha) \nu_i \cdot x -
\frac{2 }{c_0 \sin \alpha} \ln \lambda_i \right)
= \inf_{i \in \NN} \phi_i(x)=\phi^*(x)$$ and the super-solution $\phi^*$ is above the sub-solution $\phi_*$.
**Step 2: Asymptotics of sub and super-solutions**\
Let us now precise their asymptotics: we claim that as $|x|$ goes to infinity $$\label{claim}
\phi_*(x)=\phi_{\infty}(x)+o(|x|) \mbox{ and } \phi^*(x)
=\phi_{\infty}(x) +o(|x|)$$ To prove such a claim, the idea is to compare $\phi_{\infty}(x)$ with the limits as $\varepsilon$ goes to zero of $\varepsilon \phi_*(x/\varepsilon)$ and $\varepsilon \phi^*(x/\varepsilon)$. In particular, we will prove the sequence of three inequalities: for any $x \in \RR^{N-1}$ $$\label{trois ineq}
\phi_{\infty}(x) \leq \lim\limits_{\varepsilon \to 0} \varepsilon \phi_*(x/\varepsilon)
\leq \lim\limits_{\varepsilon \to 0} \varepsilon \phi^*(x/\varepsilon) \leq \phi_{\infty}(x)$$ which proves the desired claim .
The first step of the present proof leads easily to the second inequality in since $\phi_* \leq \phi^*$ on $\RR^{N-1}$. As far as the first inequality is concerned, we have for any $i \in \NN$ and $x \in \RR^{N-1}$, $$|x \cdot \nu_i| =|x| \cos (\theta_x-\theta_i) \leq |x| \cos \delta_x$$ where $\delta_x$ is the angular distance between $x/|x|$ and $K:=\overline{\cup_{i \in \NN} \{\nu_i\}}$ . Thus, $$\begin{aligned}
\phi_*(x) &\geq -\frac{2}{c_0 \sin \alpha} \ln
\left( e^{\frac{c_0 \sin \alpha }{2} |x| \ \cos \delta_x} \mu(\S^{N-2}) \right)\\
& = - (\cot \alpha) |x| \ \cos \delta_x= \inf_{\nu \in K}
-(\cot \alpha) \ x \cdot \nu = \phi_{\infty}(x)\end{aligned}$$ where $\mu(\S^{N-2})=\sum_{i} \lambda_i=1$. Thus $\phi_{\infty} \leq \phi_*$ on $\RR^{N-1}$ and the homogeneity of $\phi_{\infty}$ gives the first inequality of .
Regarding the last inequality in , we know that for any $x \in \RR^{N-1}$, $$\phi^*(x) \leq \phi_i(x)=-(\cot \alpha) \ x \cdot \nu_i - \frac{2}{c_0 \sin \alpha} \ln \lambda_i$$ Since $\lim_{\varepsilon \to 0} \varepsilon \phi_i(x/\varepsilon)=-(\cot \alpha) \ x \cdot \nu_i$, it is clear that $$\lim_{\varepsilon \to 0}\varepsilon \phi^*\left(\frac{x}{\varepsilon}\right)
\leq \inf_{i \in\NN} -(\cot \alpha)\ x \cdot \nu_i =\phi_{\infty}(x)$$ This ends the proof of the three inequalities and hence of .
**Step 3: Existence of a solution**\
By proposition \[perron\], there exists a function $\phi \in[\phi_*,\phi^*]$ viscosity solution to and by the previous step, $\phi$ verifies the right asymptotics $$\phi(x)=\phi_{\infty}(x) + o(|x|)$$ However, in the statement of theorem \[le resultat\], we claim that there exists a smooth concave solution to and the above construction does not provide such information. By proposition \[perron\], the regularity and concavity of the solution are at hand if the super-solution $\phi^*$ satisfies the technical assumption . If it does not, we will first modify the sub and the super-solutions in order to satisfy , then get a concave solution, and in a last step pass to the limit to find a solution (still concave) between $\phi^*$ and $\phi_*$.
**Step 4: Regularity and concavity**\
Let us consider for any $\varepsilon >0$ $$\label{def de psie}
\phi_*^\varepsilon(x)= - \frac{2}{c_0 \sin \alpha} \ln \left( \int_{\S^{N-2}} e^{\frac{c_0
\cos \alpha}{2} x \cdot \nu} {\,\mathrm{d}}\mu_\varepsilon(\nu) \right) \, , \quad x \in \RR^{N-1}$$ with $\mu_\varepsilon={\mu} + {\varepsilon} {\mu}_1 $ where $${\mu}_1 = \sum_{\pm}\sum_{j=1}^{N-1} \delta_{\pm e_j}$$ denoting $(e_i)_{i\in \{1,\dots,N-1\}}$ as the canonical orthonormal basis of $\RR^{N-1}$. In the same way, we define $$\phi^{\varepsilon*}(x)=\inf_{i \in \NN ,\ j =1 \dots N-1,\ \pm}
\left( -(\cot \alpha)\ x \cdot \nu_i - \frac{2}{c_0 \sin \alpha}
\ln {\lambda}_i, -(\cot \alpha)\ x \cdot (\pm e_j) -\frac{2}{c_0 \sin \alpha}\ln {\varepsilon} \right)$$ Then, $\phi_*^{\varepsilon}$ is a sub-solution, $\phi^{\varepsilon*}$ is a super-solution and $\phi_*^{\varepsilon}\le \phi^{\varepsilon*}$. It satisfies for any $\varepsilon>0$ and for $p=0$. By proposition \[perron\], there exists a concave smooth solution $\phi^\varepsilon$ satisfying equation , with $\phi^\varepsilon$ being $(\cot \alpha)$ - Lipschitz such that $$\phi_*^{\varepsilon}(x) \leq \phi^{\varepsilon}(x)
\leq \phi^{\varepsilon *}(x) \, , \quad x \in \RR^{N-1}$$ Finally, we take the limit as $\varepsilon$ goes to zero. The sub-solutions $\phi_*^\varepsilon$ go to $\phi_*$. The super-solutions $\phi^{\varepsilon*}$ converge to $\phi^*$. This follows from the expression of super-solutions as an infimum of hyperplanes, those associated to the $\varepsilon$ weights going to $+\infty$. Moreover by Ascoli’s theorem, $(\phi^{\varepsilon})_{\varepsilon>0}$ converges (up to a subsequence) to some concave and $(\cot \alpha)$ - Lipschitz function $\phi^0$ solution to and satisfying $$\phi_* \leq \phi^0 \leq \phi^*$$ Again a bootstrap argument shows that $\phi^0$ is smooth. Therefore, $\phi^0$ is the intended solution to the mean curvature equation .
------------------------------------------------------------------------
Proof of Theorem \[le mini resultat\]
-------------------------------------
**Step 1: Existence of a solution**\
Choose $\phi^*$ the viscosity solution to the eikonal equation given by $$\label{def::31}
\phi^*(x)=\inf_{\nu \in A}( -(\cot\alpha) \ x \cdot \nu + \gamma_{\nu}) \, , \quad x \in \RR^{N-1}$$ where $A=\{\nu_1,\dots,\nu_k\}$ is a finite subset of the sphere $\S^{N-2}$, $k \in \NN^*$ and $\gamma_{\nu}$ are any given real numbers. We build a sub-solution $\phi_*$ as in proposition \[existence subsolution\] $$\label{def::32}
\phi_*(x)= -\frac{2}{c_0 \sin\alpha} \ln
\left( \sum_{i=1}^k \lambda_i \ e^{\frac{c_0 \cos \alpha}{2} x \cdot \nu_i} \right)
\, , \quad x \in \RR^{N-1}$$ where $\lambda_i$ is determined by the relation $\gamma_{\nu_i}=-\frac{2}{c_0 \sin \alpha} \ln \lambda_i $ for $i=1 \dots k$. Let us notice that in the particular case when $A$ is finite, the super-solution built in proposition \[supersolution dim N\] coincides with the solution $\phi^*$ to the eikonal equation. As in section \[demo main result\], $\phi_* \leq \phi^*$ and the assumptions of proposition \[perron\] i) are satisfied. Thus, there exists a function $ \phi \in [\phi_*,\phi^*]$ viscosity solution to . Dealing as in section \[demo main result\] step $4$, we can even find a smooth concave solution still denoted $\phi \in [\phi_*,\phi^*]$. It now remains to study $\phi_* - \phi^*$ to get a precise asymptotics of the solution $\phi$.\
**Step 2: Asymptotics (first line of (\[eq::rr2\]))**\
Setting $$\phi_i(x) = -(\cot \alpha)\ x \cdot \nu_i - \frac{2}{c_0 \sin \alpha} \ln {\lambda}_i$$ we have $$\begin{aligned}
\phi_*(x) & = -\frac{2}{c_0\sin \alpha}\ln \left(\sum_{i=1}^k e^{-\frac{c_0 \sin \alpha}{2}\phi_i(x)}\right)
\ge -\frac{2}{c_0\sin \alpha}\ln\left( k e^{-\frac{c_0 \sin \alpha}{2}
\left(\displaystyle \min_{i=1,...,k}\phi_i(x)\right)}\right)\\
& = -\frac{2}{c_0\sin \alpha}\ln\left( k e^{-\frac{c_0 \sin \alpha}{2} \phi^*(x)}\right)
= \phi^*(x) - \frac{2\ln k}{c_0 \sin \alpha}\end{aligned}$$ This implies in particular that $$\label{eq::rr1}
\displaystyle - \frac{2\ln k}{c_0 \sin \alpha} \le \phi_* - \phi^* \le 0$$ which shows the first line of (\[eq::rr2\]).\
**Step 3: Asymptotics (second line of (\[eq::rr2\]))**\
We now notice that the set $E_\infty$ of edges (where $\phi_\infty$ is not $C^1$) is characterized by $$E_{\infty}=\left\{x\in\RR^{N-1}, \quad \max_{\nu\in A}\ x\cdot \nu = x\cdot \nu_{i_0} = x\cdot \nu_{i_1},
\mbox{ with } \nu_{i_0}\not=\nu_{i_1} \mbox{ and } (\nu_{i_0}, \nu_{i_1}) \in A^2 \right\}$$ For each index $i_0\in \left\{1,...,k\right\}$, let us denote the convex set $$K_{i_0}=\left\{x\in\RR^{N-1},\quad x\cdot \nu_{i_0} = \max_{\nu\in A}\ x\cdot \nu\right\}$$ Then $$\partial K_{i_0} \subset \bigcup_{j\not=i_0} (\nu_{i_0}-\nu_j)^{\perp}$$ For $x\in \mbox{Int}(K_{i_0})$, let $x_{i_1}\in \partial K_{i_0}\subset E_{\infty}$ such that $$\mbox{dist}(x,E_\infty)= |x-x_{i_1}| \quad \mbox{with}\quad x_{i_1} \in (\nu_{i_0}-\nu_{i_1})^{\perp}.$$ For $j\not= i_0$, we define the orthogonal projection of $x$ on $(\nu_{i_0}-\nu_{j})^{\perp}$ as $$x_{j}= \mbox{Proj}_{|(\nu_{i_0}-\nu_{j})^{\perp}} (x)$$ In particular $|x-x_j|\ge |x-x_{i_1}|$. Moreover $$\begin{aligned}
x\cdot \nu_j & = (x-x_j)\cdot \nu_j + x_j \cdot \nu_j = (x-x_j)\cdot \nu_j + x_j \cdot \nu_{i_0}\\
&= (x-x_j)\cdot (\nu_j-\nu_{i_0}) + x \cdot \nu_{i_0}= x \cdot \nu_{i_0} - |\nu_j-\nu_{i_0}| |x-x_j|\\
& \leq x \cdot \nu_{i_0} - \delta\ \mbox{dist}(x, E_{\infty})\end{aligned}$$ with $$\delta = \min_{\nu\not=\nu',\ \nu,\nu'\in A} |\nu-\nu'| >0$$ Therefore $$\begin{aligned}
\phi_*(x)&= \displaystyle -\frac{2}{c_0\sin\alpha}\ln
\left(\sum_{i=1}^k\lambda_i e^{\frac{c_0\cos\alpha}{2}\ x\cdot \nu_i}\right)\\
&\ge \displaystyle -\frac{2}{c_0\sin\alpha}\ln
\left(\sum_{i=1}^k\lambda_i e^{\frac{c_0\cos\alpha}{2}\
\left(x\cdot \nu_{i_0} -\delta \ \mbox{dist}(x, E_{\infty})\right)}\right)\end{aligned}$$ and then for $x\in K_{i_0}$, we have $$\phi^*(x)\ge \phi_*(x) \ge \phi^*(x) -
\frac{2}{c_0\sin\alpha}\ln \left(1+ \sum_{i\not= i_0}\frac{\lambda_i}{\lambda_{i_0}}
e^{- \frac{c_0\cos\alpha}{2}\
\delta\ \mbox{dist}(x,{\color{red} E_\infty})}\right)$$ This shows that $$\displaystyle \lim_{l\to +\infty} \sup_{\mbox{dist}(x,E_\infty)\ge l}|\phi_*(x)-\phi^*(x)| =0$$ which implies the second line of (\[eq::rr2\]).
**Step 4: Uniqueness**\
To end the proof of theorem \[le mini resultat\], it only remains to prove uniqueness of the above smooth solution $\phi$ to the mean curvature equation with the prescribed asymptotics given by $\phi^*$. Let $\overline{\phi}$ and $\underline{\phi}$ be two solutions to with the asymptotics (\[eq::rr2\]). Let $$\varepsilon := \inf \left\{\varepsilon'>0 \, | \, \forall x \in \RR^{N-1}
\, , \, \overline{\phi}(x)+\varepsilon' \ge \underline{\phi}(x)\right\}$$ then for any $x \in \RR^{N-1}$, $$\overline{\phi}(x)+\varepsilon \ge \underline{\phi}(x)$$ and there exists a sequence of points $(x_n)_n$ such that $$\overline{\phi}(x_n)+\varepsilon - \underline{\phi}(x_n) \to 0 \, \mbox{ as $n$ goes to infinity}$$ Let us define for any $x \in \RR^{N-1}$ $$\left\{\begin{array}{l}
\overline{\phi}_n(x)=\overline{\phi}(x+x_n)-\overline{\phi}(x_n),\\
\\
\underline{\phi}_n(x)=\underline{\phi}(x+x_n)-\overline{\phi}(x_n)
\end{array}\right.$$ Then, up to the extraction of a subsequence, we have as $n$ goes to infinity $$\overline{\phi}_n\to \overline{\phi}_\infty \quad \mbox{ and } \quad
\underline{\phi}_n \to \underline{\phi}_\infty$$ with a uniform convergence on any compact sets of $\RR^{N-1}$. Moreover $\overline{\phi}_\infty$ and $\underline{\phi}_\infty$ solve equation (\[eq de phi\]) and satisfy $$\overline{\phi}_{\infty}+\varepsilon \ge \underline{\phi}_{\infty}
\quad \mbox{with equality at}\quad x=0$$ From the strong maximum principle, we deduce that for any $x \in \RR^{N-1}$, $$\label{eq::rr3}
\overline{\phi}_{\infty}(x)+\varepsilon = \underline{\phi}_{\infty}(x)$$ Let us now assume that $\varepsilon>0$. Because we have $$E_\infty\subset \bigcup_{\nu\not= \nu',\ (\nu,\nu')\in A^2} (\nu-\nu')^{\perp} =:\hat{E}_\infty$$ we deduce that there exists $C>0$ such that for any $R\ge 1$ and any $x\in\RR^{N-1}$, we have $$\label{eq::dist}
\sup_{y\in \overline{B_R(x)}} \mbox{dist}(y,E_\infty)
\ge \sup_{y\in \overline{B_R(x)}} \mbox{dist}(y,\hat{E}_\infty) =
R \sup_{y\in \overline{B_1(x/R)}} \mbox{dist}(y,\hat{E}_\infty)\ge CR$$ with $$C=\inf_{z\in \RR^{N-1}}\left(\sup_{y\in \overline{B_1(z)}} \mbox{dist}(y,\hat{E}_\infty)\right)$$ We easily check by contradiction that $C>0$. Therefore by , let us choose $R$ large enough such that $$\displaystyle \sup_{\mbox{dist}(y,E_\infty)\ge CR} |\phi(y)-\phi^*(y)|
\le \frac{\varepsilon}{4} \quad \mbox{for}\quad \phi=\overline{\phi},\underline{\phi}$$ Then using , we get for some $y_n\in\overline{B_R(x_n)}$ with $\mbox{dist}(y_n,E_\infty)\ge CR$, $$\displaystyle \inf_{y\in\overline{B_R(x_n)}} |\overline{\phi}(y)-\underline{\phi}(y)|
\le |\overline{\phi}(y_n)-\underline{\phi}(y_n)| \le |\overline{\phi}(y_n)-{\phi}^*(y_n)|
+ |{\phi}^*(y_n)-\underline{\phi}(y_n)|
\le \frac{\varepsilon}{2}$$ This implies that $$\inf_{y\in \overline{B_R(0)}} |\overline{\phi}_\infty(y)-
\underline{\phi}_\infty(y)|\le \frac{\varepsilon}{2}$$ which is in contradiction with (\[eq::rr3\]). Therefore $\varepsilon=0$ and we get $\overline{\phi}\ge \underline{\phi}$. By symmetry, we also get $\underline{\phi}\ge \overline{\phi}$, which implies $\overline{\phi} = \underline{\phi}$ and shows the uniqueness of the solution. This ends the proof of the theorem \[le mini resultat\].
------------------------------------------------------------------------
Proof of further results in dimension $N=3$ {#section sept}
===========================================
In this section, the space dimension is $N=3$ and we denote any $x \in \RR^2$ with its polar coordinates $(r,\theta_x) \in \RR^+ \times [0,2\pi)$ such that $x=r(\cos \theta_x,\sin \theta_x)$.
Classification in dimension $N=3$ of solutions to the eikonal equation with a finite number of singularities {#solution eikonal dimension 3}
------------------------------------------------------------------------------------------------------------
This subsection gives alternative statement and proof of proposition \[countable\] in dimension $N=3$, in the special case of a finite number of singularities (i.e. gradient jumps).
[**(Classification with a finite number of singularities, $N=3$)**]{} \[resolution eikonal 3D\]\
Let $\alpha \in (0,\frac{\pi}{2}]$, $c_0>0$ and $c=c_0/\sin \alpha$. Choose $\phi_{\infty}$ a $1$-homogeneous viscosity solution to the eikonal equation in dimension $N=3$ with a finite number of singularities on $\S^1$. Then the [i.]{} of Theorem \[le micro resultat\] holds.
**Proof of Proposition \[resolution eikonal 3D\].**\
From Proposition \[th::1\], we know that there exists a (non empty) compact set $K=\gamma^{-1}(\{0\}) \subset \S^1$, such that $$\label{eq::rv10}
\phi_{\infty}(x)=\inf_{\nu\in K} \left(-(\cot \alpha) \ \nu\cdot x\right)$$ Thus, for any $\theta \in [0,2\pi)$, $\psi_{\infty}(\theta)=\phi_\infty(\cos \theta, \sin \theta)$ defines a continuous function with values in $[-\cot\alpha, \cot \alpha]$. Firstly $\psi_{\infty | K}=-\cot \alpha$. Moreover for any maximal interval $(a,b)$ contained in $\S^1\backslash K$, we necessarily have $$\psi_\infty(\theta)=\left\{\begin{array}{ll}
-(\cot \alpha)\ \cos (\theta -a) & \quad \mbox{if}\quad \theta\in \left[a,\frac{a+b}{2}\right],\\
\\
-(\cot \alpha)\ \cos (\theta -b) & \quad \mbox{if}\quad \theta\in \left[\frac{a+b}{2}, b\right].
\end{array}\right.$$ Therefore $\phi_\infty$ has a singularity (gradient jump) at $\theta= \frac{a+b}{2}$. If $\psi_\infty$ only has a finite number of singularities, then we get the characterization of $\psi_\infty$ given in the [i.]{} of Theorem \[le micro resultat\]. This ends the proof of proposition \[resolution eikonal 3D\].
------------------------------------------------------------------------
Notice that without assuming that $\phi_{\infty}$ has a finite number of singularities on $\S^1$, the set $K$ could be a Cantor set in (\[eq::rv10\]).
Notice that the particular function $\phi_\infty(x) = -(\cot \alpha) |x|$ is the analogue (at the level of the eikonal equation) of the level sets of cylindrically symmetric solutions to reaction diffusion equation, constructed in [@HMR3] by Hamel, Monneau and Roquejoffre. Similarly, the particular case where the graph of $\phi_\infty$ is a pyramid is also the analogue of solutions contructed by Taniguchi in [@taniguchi2].
Explicit construction of super-solutions in dimension $N=3$ {#supersolution dim 3}
-----------------------------------------------------------
In the particular case $N=3$, we construct super-solutions by hand and try to be more precise than in section \[section supersolution\], above all when $\psi_{\infty}$ is constant and equal to $-(\cot \alpha)$ on some interval $I$. In that case, we construct our super-solution by hand. We explain our ideas on different elementary pieces that we bring together in the proof of theorem \[le micro resultat\] to build a global super-solution $\phi^*$. Those different elementary pieces are: **a cone**, **an edge** or **an arc** .
### The cone case
\[supersolution cone\] **(Radially symmetric solutions)**\
Let $\phi_{\infty}$ be the viscosity solution to eikonal equation whose graph is the straight cone i.e. $\phi_{\infty}(x)=-(\cot \alpha) |x|$ for $x \in \RR^2$. Then, there exists a unique radially symmetric solution $\phi_c$ (unique up to an additive constant) to the forced mean curvature equation , satisfying $$\phi_c'(0)=0 \mbox{ and } \, \phi_c(x)=\phi_{\infty}(x)+o(|x|)$$ Moreover $\phi_c$ is concave and $|D\phi_c|\le \cot \alpha$. In the case $\alpha=\pi/2$, $\phi_c$ is zero (up to an additive constant). Otherwise, as $|x|$ goes to infinity, its asymptotics is more precisely given (up to a constant $C \in \RR$) by $$\label{asymptotic cone}
\phi_c(x)=-(\cot \alpha) |x| + \frac{1}{c_0 \sin \alpha} \ln |x|
+ C + \frac{2-3\sin^2 \alpha}{c_0^2 \sin(2\alpha) |x|}+ O \left(\frac{1}{|x|^2}\right)
\, , \quad \alpha \neq \frac{\pi}{2}$$ Moreover, let $\phi_*$ be the sub-solution defined by with $\mu=\frac{d\theta}{2\pi}$ and $N=3$. Fix $\phi_c$ such that $C=C_0:=\frac{\ln (\pi c_0 \cos \alpha)}{c_0 \sin \alpha}$, then for any $x \in \RR^2$, $\phi_c(x) \geq \phi_*(x)$ and as $|x|$ goes to infinity $$\label{eq::rv13}
\phi_c(x)=\phi_*(x)+ O \left(\frac{1}{ \sqrt{|x|}}\right).$$
**Proof of Lemma \[supersolution cone\].**\
This result is proved using quite classical methods. The proof is sketched for the reader’s convenience. With a slight misuse of notation, we denote in the case of radially symmetric solutions $\phi_c(x)$ by $\phi_c(|x|)=\phi_c(r)$ with $r=|x|\geq 0$. Then, equation reads $$-\frac{\phi_c'}{r} - \frac{\phi_c"}{1+\phi_c^{'2}}
+ c_0 \sqrt{1+\phi_c^{'2}} -c=0 \, , \quad r>0$$ Thus, $\phi_c$ satisfies an ODE involving only its first two derivatives and it can only be defined up to constants. Setting $v=\phi_c'$, we get $$\label{eq::v}
v'=(1+v^2)\left(c_0 \sqrt{1+v^2}-c-\frac{v}{r}\right):=(1+v^2)g(v,r) \, , \quad r>0$$ The proof of lemma \[supersolution cone\] now reduces to the study of this ODE (existence, uniqueness and asymptotics).
**Step 1: Existence**\
Since for any $r>0$, $g(0,r) \leq 0$ and $g(v_0(r),r)=0$ where $$v_0(r)=-\frac{c^2-c_0^2}{\frac{c}{r} + c_0 \sqrt{ \frac{1}{r^2}+c^2-c_0^2}} \leq 0 \, ,$$ $v=0$ is a super-solution and $v=v_0$ is a negative decreasing sub-solution to the ODE . Thus, for every $r_1>0$, there exist $r_2>r_1$ and a solution $v \in C^{\infty}((r_1,r_2),\RR)$ to the ODE $v'=(1+v^2)g(v,r)$ satisfying $v_0 \leq v \leq 0$ for any $r \in (r_1,r_2)$. Moreover we have $g'_v(v,r) \le 0$ for $v\in [v_0,0]$, and then we conclude that $$\label{eq::rv11}
v'\le 0\quad \mbox{for}\quad r\in(r_1,r_2)$$
**Step 2: Qualitative properties**\
Since for any $r>0$, $v_0(r) \in (-\cot \alpha, 0)$, the bounds of $v(r)$ by $v_0(r)$ and zero force $v$ to exist globally for $r>0$. Moreover, as $\lim\limits_{r \to 0} v_0(r)=0$, $v$ satisfies the same limit and we can extend $v$ to $0$ by continuity as $v(0)=0$. This proves that $v$ is a global smooth solution to with initial condition $v(0)=0$. Thus it is easy to check that any primitive function $\phi_c$ to $v$ is a smooth radially symmetric solution to satisfying $\phi_c'(0)=0$. From (\[eq::rv11\]), we conclude that $\phi_c$ is concave.
**Step 3: Asymptotics**\
Since $v$ is strictly decreasing on $\RR^+$ and bounded from below, it converges to a finite limit $-\cot \alpha \leq l <0$ as $r$ goes to infinity. Since $v$ is uniformly bounded in $[-\cot \alpha,0]$, $l$ must satisfy $g(l,+\infty)=0$ which leads to $l=-\cot \alpha$.
Linearising equation around $-\cot \alpha$, we set $w=v+\cot\alpha$. As $w$ is uniformly bounded on $\RR^+$ and goes to zero at infinity, equation reads $$w'(r)=-c \ (\cot\alpha) w + \tilde{g}(w,r) \, , \quad r>0$$ where $\tilde{g}(w,r)=O(w^2) + O(1/r)$ as $r$ goes to infinity. By the Duhamel’s formula, $w$ follows exponentially fast the behaviour of the slowest term of $\tilde{g}$. Thus $w\sim C/r$ as $r$ goes to infinity and a straight calculation gives $C=1/(c_0 \sin \alpha)$. Repeating this method up to order $2$, one gets $$v(r)=-\cot \alpha + \frac{1}{c_0 (\sin \alpha) r}
+ \frac{3\sin^2 \alpha -2}{c_0^2 \sin(2\alpha) r^2}
+ O \left(\frac{1}{r^3}\right)$$ This gives the desired asymptotics for $\phi_c$ up to constants.
**Step 4: Uniqueness**\
Let $\phi_c^1$ and $\phi_c^2$ be two smooth radially symmetric solutions to . From step $3$, we know that they satisfy the same asymptotic expansion as $r$ goes to infinity and we assume the constants $C$ are the same. Since $\phi_c^1-\phi_c^2$ solves an elliptic equation with smooth coefficients and no zero order term, the classical maximum principle applies. Hence $\phi_c^1-\phi_c^2=0$ because $\lim_{r\to \infty} (\phi_c^1-\phi_c^2)(r)=0$. This proves the uniqueness of $\phi_c$ up to constants.\
**Step 5: Comparison with $\phi_*$**\
Using Lemma \[l6.1\], we can check (\[eq::rv13\]) with the suitable value of the constant $C=C_0$ (see for instance the computation (\[eq::rv14\]) with $N_0(x)\simeq 1/\sqrt{\pi}$). Finally, using the comparison principle (as in Step 4), we deduce that $\phi_*\le \phi_c$. This ends the proof of lemma \[supersolution cone\].
------------------------------------------------------------------------
### The edge case
\[supersolution edge\] **(Edge super-solution)**\
Assume $\phi_*$ is given by where the measure $\mu$ is the sum of two Dirac masses $$\mu= \mu_{\{\theta_1,\theta_2\}} = \lambda_1 \delta_{\theta_1} + \lambda_2 \delta_{\theta_2}$$ with $\lambda_i >0$, $\theta_i \in [0,2\pi)$ for $i=1,2$ such that $\theta_1< \theta_2$ and $\delta_{\theta_i}$ the Dirac mass in $\theta_i$. In the case $\alpha \neq \pi/2$, define $\phi_e$ for any $x \in \RR^2$ by $$\label{def phie}
\phi_e(x)=\min(p_1(x), p_2(x)) \mbox{ with }
p_i(x)= - (\cot \alpha)\ x\cdot \nu_i - \frac{2}{c_0 \sin \alpha} \ln \lambda_i$$ where $\nu_i=(\cos \theta_i,\sin \theta_i)$. Then, $\phi_e$ is a Lipschitz and piecewise smooth global super-solution to verifying $\phi_* \leq \phi_e$ on $\RR^2$. Moreover, as $|x|$ goes to infinity, $$\label{eq::rv12}
\phi_e(x)=\phi_*(x) + O(1) \, , \quad x \in \RR^2$$
**Proof of Lemma \[supersolution edge\].**\
Notice first that $\phi_e=\phi^*$ with $\phi^*$ defined as a special case of Proposition \[supersolution dim N\]. This shows that $\phi_e$ is a concave (Lipschitz) supersolution. Finally (\[eq::rv12\]) follows from (\[eq::rr1\]). This ends the proof of the lemma.
------------------------------------------------------------------------
### The arc case
Here we wish to describe a super-solution to which, from above, looks like an arc, i.e. is made up of two non parallel straight lines connected by a circle.
\[supersolution arc\] **(Arc super-solution)**\
Assume $\phi_*$ is given by where the measure $\mu$ is the sum of two Dirac masses and a Lebesgue measure $$\mu= \mu_{[\theta_1,\theta_2]}
= \lambda \delta_{\theta_1} + \lambda \delta_{\theta_2}
+ \indicatrice_{(\theta_1,\theta_2)} {{\,\mathrm{d}}\theta},$$ where $\lambda >0$, $\theta_i \in [0,2\pi)$ for $i=1,2$ and $\theta_1<\theta_2$.
Define $\phi_e$ as in the edge case with $\lambda_1=\lambda_2=\lambda$. Define $\phi_c$ as in the cone case (lemma \[supersolution cone\]) where the constant $C \in \RR$ in is chosen such that $\phi_c(0)=\phi_e(0)= -\frac{2}{c_0 \sin \alpha} \ln \lambda$.
Finally, define $\phi_a$ on $\RR^2$ by $$\label{e5.3}
\forall x \in \RR^2 \, , \quad \phi_a(x)=\left\{\begin{array}{ll}
\displaystyle -\frac{2}{c_0 \sin \alpha} \ln \lambda & \mbox{ if } x=0\\
\min (\phi_c(x) , \phi_e(x))& \mbox{ if }\theta_x \in (\theta_1,\theta_2)\\
\phi_e(x) & \mbox{ otherwise, }
\end{array}\right.$$ Then $\phi_a $ is a Lipschitz continuous global super-solution to . Moreover, as $|x|$ goes to infinity $$\phi_a(x)=\phi_*(x)+ O(1)$$
The shape of $\phi_a$ is sketched on Figures \[F1\],\[F2\], \[F3\].\
**Proof of Lemma \[supersolution arc\]**\
**Step 1: $\phi_a$ is a global continuous super-solution**\
By definition and lemmas \[supersolution cone\] and \[supersolution edge\], $\phi_a$ is a super-solution to where it is locally the minimum of supersolutions, i.e. everywhere except on the two half lines $\theta_x =\theta_i$ for $i=1,2$. However, we have $\phi_c(0)=\phi_e(0)$ and $D\phi_e(x)=-(\cot \alpha) \nu_i$ while $\phi_c'(r)\in (- \cot \alpha,0]$ for any $x \in \RR^2$ with $\theta_x=\theta_i$, $i=1$ or $2$. Thus, $\phi_e(x) \leq \phi_c(x)$ on a neighborhood ${\mathcal N}$ (not containing the origin) of the two half lines $\theta_x = \theta_i$ for $i=1,2$. This implies $\phi_a = \phi_e$ on ${\mathcal N}$ and then $\phi_a$ is at least a supersolution on $\RR^2\backslash \left\{0\right\}$. Moreover, $\phi_a$ is a supersolution on the whole $\RR^2$. Indeed, $\phi_a=\phi_e$ for $\theta_x\not\in [\theta_1,\theta_2]$, then $\phi_a$ has a gradient jump along the edge $\theta=(\theta_1+\theta_2)/2 +\pi$ up to the origin. And this gradient jump implies that there is no $C^2$ test function touching $\phi_a$ from below at $x=0$.
**Step 2: Relative positions of $\phi_e$ and $\phi_c$**\
Let us now study the relative positions of both graphs of $\phi_e$ and $\phi_c$. Since $\phi_c(x) \in (-|x| \cot \alpha + \phi_c(0) ,\phi_c(0)]$, we have $$\begin{aligned}
p_1(x)=\phi_c(x) & \Longleftrightarrow -r (\cot\alpha) \cos( \theta_x-\theta_1)+\phi_c(0)=\phi_c(r) \\
& \Longleftrightarrow \theta_x= \theta_1 \pm
\arccos\left( \frac{\phi_c(r)-\phi_c(0)}{-r\cot \alpha} \right) := \theta_1 \pm \bar{\theta}(r)\end{aligned}$$ where $ \bar{\theta}(r) \in (0,\pi/2)$ for $r>0$ (see Figure \[F4\]). Notice that from the concavity of $\phi_c$, we deduce that the set $$\left\{x,\quad p_1(x) \le \phi_c(x)\right\}=\left\{x,\quad \theta_x\in [\theta_1-\overline{\theta}(r),\theta_1+\overline{\theta}(r)]\right\}$$ is a convex set. Therefore we deduce that $$\phi_e(x) \leq \phi_c(x) \Leftrightarrow \theta_x \in
[\theta_1-\bar{\theta}(r), \theta_1+\bar{\theta}(r)] \cup
[\theta_2-\bar{\theta}(r), \theta_2+\bar{\theta}(r)]$$ and $\phi_a(x)=\phi_c(x)$ if and only if $\theta_x \in
I_r= [ \theta_1+\bar{\theta}(r), \theta_2-\bar{\theta}(r)]$. Since we choose $\phi_c$ such that $\phi_c'(0)=0$, we get $$\lim_{r \to 0} \bar{\theta}(r)= + \frac{\pi}{2}$$ This forces both curves $\theta_x=\theta_1+\bar{\theta}(r)$ and $\theta_x=\theta_2 -\bar{\theta}(r)$ to intersect at some point $(x,z) \neq ( 0, \phi_c(0) )$ as soon as $\theta_2-\theta_1 < \pi$. In that case, it is worth noticing that the above interval $I_r$ is empty for sufficiently small $r$ (see Figures \[F5\],\[F6\]).
On the other hand, using the asymptotics of $\phi_c$ found in lemma \[supersolution cone\], we get as $r$ goes to infinity $$\label{asymptotic cosine}
\cos \bar{\theta}(r) =1- \frac{1}{c_0 \cos \alpha} \frac{\ln r}{r} +
\frac{\phi_c(0)-C}{r\cot \alpha} + O\left(\frac{1}{r^2}\right) \, , \quad \alpha \neq \frac{\pi}{2}$$ where $C$ is the constant given by and fixed by the choice $\phi_c(0)=\phi_e(0)$.
**Step 3: $\phi_*\le \phi_a + constant$**\
Let $\phi_*$ be the sub-solution given by where the measure $\mu$ is $\mu_{[\theta_1,\theta_2]}$, i.e. with $b =c_0\cos \alpha$: $$\phi_*(x)=-\frac{2}{c_0\sin \alpha}
\ln\left(
\lambda e^{\frac{br}{2}\cos(\theta_x -\theta_1)}
+\lambda e^{\frac{br}{2}\cos(\theta_x -\theta_2)}
+\int_{\theta_1}^{\theta_2}e^{\frac{br}{2}\cos(\theta_x -\theta)} \ d\theta \right)$$ Each term in the Logarithm being non negative, we have $$\label{eq::rv16}
\phi_*\leq \phi_e \quad \mbox{on}\quad \RR^2.$$ To prove that $\phi_a$ is above the sub-solution $\phi_*$ up to an additive constant, it remains to compare $\phi_* $ and $\phi_c$ when $\theta_x \in
I_r= [ \theta_1+\bar{\theta}(r), \theta_2-\bar{\theta}(r)]$ and $r$ sufficiently large.
According to lemma \[l6.1\], one gets that for any $x \in \RR^2$ for $r=|x|$ sufficiently large and uniformly in $\theta_x \in [\theta_1,\theta_2]$ $$\begin{aligned}
\label{eq::rv14}
& \phi_*(x)=-(\cot \alpha) r + \frac{\ln r}{c_0 \sin \alpha}
- \frac{2}{c_0 \sin \alpha} \ln (\Phi(x))\\
\notag \mbox{ with } \quad & \Phi(x):=\frac{2\pi N_0(x)}{\sqrt{b}}
+ \lambda \sqrt{r} e^{\frac{br}{2} (\cos (\theta_x-\theta_1)-1) }
+ \lambda \sqrt{r} e^{\frac{br}{2} (\cos (\theta_x-\theta_2)-1) }
+ O\left(\frac{1}{\sqrt{r}}\right)\end{aligned}$$ where $$N_0(x)= \int_{\sqrt{r}g(\theta_1-\theta_x)}^{\sqrt{r}g(\theta_2-\theta_x)}
e^{-\frac{u^2}{4}} \frac{{\,\mathrm{d}}u}{2\pi}$$ as defined in lemma \[l6.1\]. Since $g$ is odd and $\theta_x\in [\theta_1,\theta_2]$, we see that $$N_0(x)\ge \int_{0}^{\sqrt{r}g\left(\frac{\theta_2-\theta_1}{2}\right)}
e^{-\frac{u^2}{4}} \frac{{\,\mathrm{d}}u}{2\pi} = \frac{1}{2\sqrt{\pi}} + o_r(1)$$ We deduce in particular that for $r$ large enough and uniformly in $\theta_x\in[\theta_1,\theta_2]$: $$\Phi(x)\ge \frac{\sqrt{\pi}}{2\sqrt{b}}$$ Therefore, from the asymptotics (\[asymptotic cone\]) of $\phi_c$, we deduce that there exist $r_1>0$, $C_1>0$ such that $$\label{eq::rv15}
\forall r\geq 0 \, , \quad \forall \theta_x \in [\theta_1,\theta_2]\, , \quad r \geq r_1 \Rightarrow \phi_*(x)\le \phi_c(x) + C_1$$ Now from (\[eq::rv16\]) and (\[eq::rv15\]), we deduce that (up to increasing the constant $C_1$), $$\forall r\geq 0 \, , \quad \forall \theta_x \in [\theta_1,\theta_2]\, , \quad r \geq r_1 \Rightarrow \phi_*(x)\le \phi_a(x) + C_1$$
**Step 4: $\phi_*\ge \phi_a - constant$**\
**Case 1: $\theta_x\in I_r$**\
We start with the asymptotics (\[eq::rv14\]). Using (\[asymptotic cosine\]), we see that there exist $r_2>0$, $C_2>0$ such that $$\forall i=1,2 \, , \quad \forall r\geq 0 \, , \quad \forall \theta_x \in I_r \, , \quad
r \geq r_2 \Rightarrow\sqrt{r} e^{\frac{br}{2} (\cos (\theta_x-\theta_i)-1)}
\le \sqrt{r} e^{\frac{br}{2} (\cos (\overline{\theta}(r))-1)}\le C_2$$ Using also the fact that $N_0(x)\le 1/(2\sqrt{\pi})$, we deduce that $\Phi$ is bounded for $r$ large enough and then (up to increasing $r_2$ and $C_2$) $$\label{eq::rv17}
\forall r\geq 0 \, , \quad \forall \theta_x \in I_r \, , \quad
r \geq r_2 \Rightarrow \phi_*(x) \ge \phi_c(x) -C_2$$
**Case 2: $\theta_x\in [\theta_1,\theta_2]\backslash I_r$**\
Let us assume that $\theta_x \in [\theta_1,\theta_1+\overline{\theta}(r))$ (the symmetric case is similar). Then there exist $r_3>0$, $C_3>0$ such that $$\forall r \geq 0 \, , \quad \forall \theta_x \in [\theta_1,\theta_1+\overline{\theta}(r)) \, , \quad r \geq r_3 \Rightarrow
\sqrt{r} e^{\frac{br}{2} (\cos (\theta_x-\theta_1)-1)}\ge \sqrt{r} e^{\frac{br}{2} (\cos (\overline{\theta}(r))-1)} \ge C_3>0$$ Therefore,( up to increasing the constants $r_3$ and $C_3$) for any $r\geq 0$ and any $\theta_x \in [\theta_1,\theta_1+\overline{\theta}(r))$, $$r \geq r_3 \Rightarrow \Phi(x)\le C_3 \left(\lambda \sqrt{r} e^{\frac{br}{2} (\cos (\theta_x-\theta_1)-1) }
+ \lambda \sqrt{r} e^{\frac{br}{2} (\cos (\theta_x-\theta_2)-1) }\right)$$ and then $$\begin{aligned}
\notag
\phi_*(x) & \ge \displaystyle - \frac{2}{c_0 \sin \alpha} \ln \left\{ \lambda e^{\frac{br}{2}\cos(\theta_x -\theta_1)}
+\lambda e^{\frac{br}{2}\cos(\theta_x -\theta_2)} \right\} - \frac{2\ln C_3}{c_0 \sin \alpha}\\
\notag & \geq -\frac{2}{c_0 \sin \alpha} \ln \left( 2 e^{\frac{c_0 \sin \alpha}{2} \phi_e(x) }\right) - \frac{2\ln C_3}{c_0 \sin \alpha}\\
\label{eq::rv20} & = \displaystyle \phi_e(x) - \frac{2\ln (2C_3)}{c_0 \sin \alpha}\end{aligned}$$
**Case 3: $\theta_x\in \S^1\backslash [\theta_1,\theta_2]$**\
Notice that the set $\S^1\backslash [\theta_1,\theta_2]$ is not empty because $\theta_2-\theta_1< 2\pi$ (as a consequence of $\theta_1,\theta_2\in [0,2\pi)$). In that case, we define $$\theta_m'=\theta_m -\pi\quad \mbox{with}\quad \theta_m=\frac{\theta_1+\theta_2}{2}.$$ Then, it satisfies $\theta_2-2\pi < \theta_m' < \theta_1$. Let us assume that (the other case is similar): $$\label{eq::rv19}
\theta_x\in [\theta_m',\theta_1).$$ We also define $$\theta_x'= \theta_x+\pi$$ Then we have $$\int_{\theta_1}^{\theta_2}e^{\frac{br}{2}\cos(\theta_x -\theta)} \ d\theta
= \int_{[\theta_1,\theta_2]\cap [\theta_x,\theta_x']} (...)\ d\theta + \int_{[\theta_1,\theta_2]\backslash [\theta_x,\theta_x']} (...)\ d\theta$$ We have $$\int_{[\theta_1,\theta_2]\cap [\theta_x,\theta_x']}e^{\frac{br}{2}\cos(\theta_x -\theta)} \ d\theta \le \pi \ e^{\frac{br}{2}\cos(\theta_x -\theta_1)}$$ Using (\[eq::rv19\]), we also see that $$\int_{[\theta_1,\theta_2]\backslash [\theta_x,\theta_x']}e^{\frac{br}{2}\cos(\theta_x -\theta)} \ d\theta \le \pi \ e^{\frac{br}{2}\cos(\theta_x -\theta_1)}$$ Therefore, we conclude that in this third case, $$\begin{aligned}
\notag \phi_*(x)
& \ge \displaystyle - \frac{2}{c_0 \sin \alpha} \ln \left\{ (\lambda+2\pi) e^{\frac{br}{2}\cos(\theta_x -\theta_1)}
+(\lambda+2\pi) e^{\frac{br}{2}\cos(\theta_x -\theta_2)} \right\}\\
\label{eq::rv21} & \ge \displaystyle \phi_e(x) - \frac{2\ln \left(\frac{2(\lambda+2\pi)}{\lambda}\right)}{c_0 \sin \alpha} \end{aligned}$$
**Conclusion**\
Putting (\[eq::rv17\]), (\[eq::rv20\]) and (\[eq::rv21\]) together, we get that there exists a constant $C>0$ such that for $r$ large enough and uniformly in $\theta_x\in \S^1$. $$\phi_*(x) \ge \phi_a(x) -C$$ The functions $\phi_*$ and $\phi_a$ being continuous, the result still holds for any $r \geq 0$ (up to increasing the constant $C$). This concludes the proof of lemma \[supersolution arc\].
------------------------------------------------------------------------
Proof of Theorem \[le micro resultat\]
--------------------------------------
**Step 1: Existence of a solution**\
Let $\alpha \in (0,\frac{\pi}{2}]$, $c_0>0$ and $c=c_0/\sin \alpha$. The case $\alpha=\pi/2$ is obvious and we omit it. Choose $\phi_{\infty}$ a $1$-homogeneous viscosity solution to the eikonal equation (\[eikonal\]) in dimension $N=3$ with a finite number $m$ of singularities. By proposition \[resolution eikonal 3D\], the [i.]{} of Theorem \[le micro resultat\] is already established, and we can consider the measure $\mu$ given in the [ii.]{} of Theorem \[le micro resultat\]. Then Proposition \[existence subsolution\] implies that the function $\phi_*$ given by (\[eq::rv22\]) is a smooth concave subsolution of (\[MCM\]).
If $k=1$, $\phi_{\infty}$ has no gradient jump and the corresponding measure is $\mu={\,\mathrm{d}}\theta$ or $\mu=\lambda_0 (\delta_{\theta_1}+\delta_{\theta_1+2\pi})$. In the first case, we saw in lemma \[supersolution cone\] that $\phi_c$, to which a suitable constant is added, is a smooth solution to with the right asymptotics at infinity. In the second one, $\phi=\phi_{\infty}$ is a suitable solution to .
We now turn to the case $k \ge 2$. For any $i \in \{1, \dots ,k\}$, choose $\lambda_0>0$ a given positive constant. We have $$\mu=\sum_j \mu_j \ge \tilde{\mu}_i$$ with $$\label{def de mu i tilde}
\tilde{\mu}_i=
2\lambda_0 (\delta_{\theta_i} + \delta_{\theta_{i+1}}) +\sigma_i\ \indicatrice_{(\theta_i,\theta_{i+1})} {\,\mathrm{d}}\theta$$ Let $\tilde{\phi}_{i*}$ be the subsolution defined in with the measure $\tilde{\mu}_i$. If $\tilde{\mu}_i$ corresponds to an arc ($\sigma_i=1$), denote $\tilde{\phi}_i^*$ the global supersolution defined in lemma \[supersolution arc\] with $\lambda=2\lambda_0$. Notice that there is a constant $C>0$ (that can be chosen independently of the index $i$) such that $$\label{eq::rv25}
\tilde{\phi}^*_{i} -C\le \tilde{\phi}_{i*} \le \tilde{\phi}^*_{i} +C$$ If $\tilde{\mu}_i$ corresponds to an edge ($\sigma_i=0$), denote $\tilde{\phi}_i^*$ the global supersolution defined in lemma \[supersolution edge\] with $\lambda_1=\lambda_2=2\lambda_0$, which satisfies in particular (\[eq::rv25\]). Finally, define on $\RR^2$ the function $\tilde{\phi}^*$ as the infimum over $i \in \{1, \dots ,k\} $ of $\tilde{\phi}_i^*$. Notice that, by construction, we have $$\label{eq::rv40}
\tilde{\phi}^*(x)=\tilde{\phi}_i^*(x) \quad \mbox{if}\quad \theta_x\in [\theta_i,\theta_{i+1}]$$ We also have in particular $$\phi_*\le \tilde{\phi}_{i*} \le \tilde{\phi}^*_{i} +C \le \tilde{\phi}^* +C =: \phi^*$$ We claim that at infinity $$\label{eq::rv26}
\phi^*(x) = \phi_*(x) +O(1).$$ We shall first finish the proof of theorem \[le micro resultat\] and come back to the proof of that claim in a second step. Thus $\phi^*$ is a global supersolution above the subsolution $\phi_*$. Moreover, either there exists $\sigma_i=1$ and then we have (see in particular Figures \[F1\], \[F2\], \[F3\]) $$\label{eq::rv45}
\mbox{there exists } p\in\RR^{N-1} \mbox{ such that } \limsup_{|x|\to +\infty} \frac{\phi^*(x)-p\cdot x}{|x|} <0.$$ Or $\sigma_i=0$ for any $i$, and condition (\[eq::rv45\]) is satisfied if $k\ge 3$. The special case $k=2$ and $\sigma_1=\sigma_2=0$ corresponds to an edge for which we already know the existence of a smooth concave solution, by Theorem \[le mini resultat\]. In the other cases, condition (\[eq::rv45\]) and Proposition \[perron\] imply the existence of a smooth concave solution $\phi \in [\phi_*,\phi^*]$.
**Step 2: Proof of (\[eq::rv26\]) in the case $k \ge 2$**\
Let $x \in \RR^2$, then there exists $i \in \{1, \dots, k \}$ such that $\theta_x \in [\theta_i,\theta_{i+1}]$.\
We can write: $$\mu = \tilde{\mu}_i +\bar\mu_i$$ where $\tilde{\mu}_i$ is defined by . So $$\label{eq::rv31}
\mbox{supp}(\bar\mu_i)\quad \subset \quad \S^1\backslash (\theta_i,\theta_{i+1}) \quad \not=\emptyset$$ Then $k\ge 2$ implies that $$l_x:=\min\left\{\theta_x -\theta_i, \theta_{i+1}-\theta_x\right\}< \pi$$ and $$\int_{\S^1}e^{\frac{br}{2}\cos(\theta_x-\theta)}\ d\bar \mu_i(\theta)
\le \bar \mu_i(\S^1) e^{\frac{br}{2}\cos l_x}\le \frac{\bar \mu_i(\S^1)}{2\lambda_0}\
\int_{\S^1}e^{\frac{br}{2}\cos(\theta_x-\theta)}\ d\tilde{\mu}_i(\theta)$$ with $$\int_{\S^1}e^{\frac{br}{2}\cos(\theta_x-\theta)}\ d\tilde{\mu}_i(\theta) =
2\lambda_0 e^{ \frac{br}{2} \cos (\theta_x-\theta_i)}
+ 2\lambda_0 e^{ \frac{br}{2} \cos (\theta_x-\theta_{i+1})}
+ \sigma_i \int_{\theta_i}^{\theta_{i+1}} e^{ \frac{br}{2} \cos (\theta_x-\theta)} {\,\mathrm{d}}\theta$$ Therefore we have $$\begin{aligned}
\phi_*(x)& \ge -\frac{2}{c_0 \sin \alpha} \ln \left\{\left(1+\frac{\bar \mu_i(\S^1)}{2\lambda_0}\right)
\int_{\S^1}e^{\frac{br}{2}\cos(\theta_x-\theta)}\ d\tilde{\mu}_i(\theta)\right\}\\
& = \tilde{\phi}_{i*}(x) -\frac{2}{c_0 \sin \alpha} \ln \left(1+\frac{\bar \mu_i(\S^1)}{2\lambda_0}\right)\\
& \ge \tilde{\phi}^*_{i}(x) -C'\end{aligned}$$ where we have used (\[eq::rv25\]) in the last line. Using (\[eq::rv40\]), we see that this implies $$\label{eq::rv41}
\phi_*(x) \ge {\phi}^*(x) -C'' \quad \mbox{for}\quad \theta_x\in [\theta_i,\theta_{i+1}]$$ Finally, this implies (\[eq::rv26\]) and ends the proof of theorem \[le micro resultat\].
------------------------------------------------------------------------
Appendix: Laplace’s method {#laplace}
==========================
For the reader’s convenience, we reproduce here Laplace’s method. It investigates asymptotics as $r$ goes to infinity of integrals involving expressions of the form $e^{-rJ}$, $J$ denoting some given function. Our interest is to find uniform estimates as $x=r(\cos \theta_x,\sin \theta_x)$ lies in a given angle sector $[\theta_1,\theta_2]$. The proof develops ideas that can be found for a simpler case in [@evans], chapter 4.5.2 page 204. Lemma \[l6.1\] below is only used in Step 3 and 4 of the proof of Lemma \[supersolution arc\].
\[l6.1\] **(Uniform asymptotics in a sector $[\theta_1,\theta_2]$)**\
Define for any $x=r(\cos\theta_x,\sin\theta_x) \in \RR^2$ with $\theta_x \in[0,2\pi)$ $$F(x)= \lambda_1 \, e^{\frac{br}{2} \cos(\theta_1-\theta_x)}
+ \lambda_2 \, e^{\frac{br}{2} \cos(\theta_2-\theta_x)}
+ \int_{\theta_1}^{\theta_2} e^{\frac{br}{2} \cos(\theta-\theta_x)}
f(\theta) \frac{{\,\mathrm{d}}\theta}{2\pi}$$ where
i.
: $b=c_0\cos\alpha >0$, $\lambda_i \in \RR$, $\theta_i \in [0,2\pi]$ for $i=1,2$ and $\theta_1< \theta_2$
ii.
: $f \in C^1([0,2\pi],\CC)$ is $2\pi$-periodic.
As $r$ goes to infinity, we have the following asymptotics uniform in the angular sector $\theta_x \in [\theta_1,\theta_2]$ $$F(x)= \lambda_1 \, e^{\frac{br}{2} \cos(\theta_1-\theta_x)}
+ \lambda_2 \, e^{\frac{br}{2} \cos(\theta_2-\theta_x)}
+e^{\frac{br}{2}} \left( \frac{f(\theta_x)}{\sqrt{br}} N_0 (x)
+ \frac{R(x)}{r} \right)$$ where $$\displaystyle N_0(x)=\int_{\sqrt{r}g(\theta_1-\theta_x)}^{\sqrt{r}g(\theta_2-\theta_x)}
e^{-\frac{u^2}{4}} \frac{{\,\mathrm{d}}u}{2\pi} \quad \in [0,1/ \sqrt{\pi}]$$ and $$\label{def de g}
g(\theta)=\left\{\begin{array}{ll}
\mbox{ sign}(\theta) \sqrt{2b(1-\cos \theta)} & \quad \mbox{for}\quad \theta\in [-\pi,\pi]\\
\\
\mbox{ sign}(\theta) 2 \sqrt{b} & \quad \mbox{for}\quad \theta\in \RR\backslash [-\pi,\pi]
\end{array}\right.$$ Moreover, there exists a constant $C>0$ such that for any $x \in \RR^2$, if $r>1$ and $\theta_x \in [\theta_1,\theta_2]$ then $|R(x)| \leq C$.
**Proof of Lemma \[l6.1\].**\
It is straightforward to check that $g$ defined by is an odd $C^3$-diffeomorphism from $[-\pi,\pi]$ to $[-2\sqrt{b},2\sqrt{b}]$ satisfying $g(0)=0$, $g'(0)=\sqrt{b}$ and $g"(0)=0$. We have also chosen to extend $g$ to the real line by continuity. However, when we speak about $g^{-1}$, it has to be understood as the inverse of $g$ on $[-\pi,\pi]$.
Afterwards, for any $x \in \RR^2$, we define $$I(x):= \int_{\theta_1}^{\theta_2} e^{\frac{br}{2} \cos(\theta-\theta_x)} f(\theta)\frac{ {\,\mathrm{d}}\theta}{2\pi}=\int_{\theta_1-\theta_x}^{\theta_2-\theta_x} e^{\frac{br}{2} \cos\theta}
f(\theta+\theta_x) \frac{{\,\mathrm{d}}\theta}{2\pi}$$ Assume $\theta_x \in [\theta_1,\theta_2]$. In order to get a bound on $I$ uniform in the angle $\theta_x$, we fix some $\delta >0$ and set $$\theta_*=\left\{ \begin{array}{ll}
\theta_1-\theta_x & \mbox{ if } \theta_1-\theta_x \geq -\pi+\delta \\
-\pi + \delta & \mbox{ otherwise}
\end{array}\right.
\quad
\theta^*=\left\{ \begin{array}{ll}
\theta_2-\theta_x & \mbox{ if } \theta_2-\theta_x \leq \pi-\delta \\
\pi - \delta & \mbox{ otherwise}
\end{array}\right.$$ We then cut the integral $I$ into three parts, integrating between $\theta_1-\theta_x$ and $\theta_*$, between $\theta_*$ and $\theta^*$ and finally between $\theta^*$ and $\theta_2-\theta_x$. We call those three integrals $I_1$, $I_2$ and $I_3$ respectively.
Regarding $I_1$ and $I_3$, $\cos \theta$ can be bounded in both cases by $\cos (\pi-\delta)$ and $f$ by its $L^{\infty}$ norm on the compact set $[0,2\pi]$. Thus, there exists a constant $C>0$ such that for any $x \in \RR^2$ with $\theta_x \in [\theta_1,\theta_2]$, $$I_1+I_3 \leq C e^{\frac{br}{2} \cos(\pi-\delta)}$$ For sufficiently small $\delta>0$, the right hand term decreases exponentially fast and the contribution of $I_1$ and $I_3$ in $I$ is exponentially small as $r$ goes to infinity uniformly in $\theta_x$.
Using the change of variables $u=\sqrt{r}g(\theta)$, we rewrite $I_2$ as $$I_2(x)=\frac{e^{\frac{br}{2}}}{\sqrt{br}}
\int_{\sqrt{r}g(\theta_*)}^{\sqrt{r}g(\theta^*)} e^{-\frac{u^2}{4}}
\, h\left(\frac{u}{\sqrt{r}}\right) \frac{{\,\mathrm{d}}u}{2\pi}$$ where $h(t)=f(\theta_x+g^{-1}(t))/\sqrt{1-(t^2/(4b))}$ . Since $h(t)=h(0)+\int_0^t h'(s) {\,\mathrm{d}}s$, we have $$I_2(x)=e^{\frac{br}{2}} \left( \frac{f(\theta_x)}{\sqrt{br}} N_0^*(x)+ \frac{R(x)}{r} \right)$$ where $N_0^*$ is defined as in lemma \[l6.1\] with $\theta_*$ or $\theta^*$ when needed, but it only changes the desired asymptotics with an exponentially small correction as above. The remainder term $R$ is defined as $$R(x)=r^{\frac{1}{2}}
\int_{\sqrt{r}g(\theta_*)}^{\sqrt{r}g(\theta^*)} e^{-\frac{u^2}{4}}
\int_0^{\frac{u}{\sqrt{r}}} \frac{h'(s)}{\sqrt{b}} {\,\mathrm{d}}s \, \frac{{\,\mathrm{d}}u}{2\pi}$$ Since $h$ is smooth, $h'$ is uniformly bounded on $[g(\theta_*),g(\theta^*)]$ and the bound only depends on $\delta$. A straight calculation then shows that there exists $C>0$ such that $$R(x) \leq C \int_{\sqrt{r}g(\theta_*)}^{\sqrt{r}g(\theta^*)} |u| e^{-\frac{u^2}{4}} {\,\mathrm{d}}u
\leq C \int_{-\infty}^{+\infty} |u| e^{-\frac{u^2}{4}} {\,\mathrm{d}}u$$ Putting finally $I_1$, $I_2$ and $I_3$ together, we get the desired asymptotics.
------------------------------------------------------------------------
[9]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We review some developments on clustering stochastic processes and come with the conclusion that asymptotically consistent clustering algorithms can be obtained when the processes are ergodic and the dissimilarity measure satisfies the triangle inequality. Examples are provided when the processes are distribution ergodic, covariance ergodic and locally asymptotically self-similar, respectively.
**Keywords:** stochastic process, unsupervised clustering, stationary ergodic processes, local asymptotic self-similarity
author:
- 'Qidi Peng[^1]'
- 'Nan Rao[^2]'
- 'Ran Zhao[^3]'
bibliography:
- 'references.bib'
title: Some Developments in Clustering Analysis on Stochastic Processes
---
Introduction
============
A stochastic process is an infinite sequence of random variables indexed by “time”. The time indexes can be either discrete or continuous. Stochastic process type data have been broadly explored in biological and medical research [@Damian2007; @Zhao2014; @Jaaskinen2014; @McDowell2018]. Unsupervised learning on stochastic processes (or time series) has increasingly attracted people from various areas of research and practice. Among the above unsupervised learning problems, one subject, called cluster analysis, aims to discover patterns of stochastic process type data. There is a rich literature in bioinformatics, biostatistics and genetics on clustering stochastic process type data. We refer the readers to the review of Aghabozorgi et al. [@Aghabozorgi2015] for updates of cluster analysis on stochastic processes til 2015. Recently Khaleghi et al. [@khaleghi2016; @khaleghi2014; @khaleghi2012a] obtained asymptotically consistent algorithms for clustering distribution stationary ergodic processes, in the case where the correct number of clusters is known or unknown. In their framework, the key idea is to define a proper dissimilarity measure $d$ between any 2 observed processes, which characterizes the features of stationarity and ergodicity. Further Peng et al. [@Peng2018-1; @Peng2018-2] derived consistent algorithms for clustering covariance stationary ergodic processes and locally asymptotically self-similar processes.
In this framework we review the recent developments in cluster analysis on the following 3 types of stochastic processes:
Type (1)
: distribution stationary ergodic processes;
Type (2)
: covariance stationary ergodic processes;
Type (3)
: locally asymptotically self-similar processes.
According to the nature of each type of processes, the ground-truths in the 3 clustering problems are defined differently. In the ground-truth of Type $(1)$, two processes of identical process distributions are in one cluster; in the ground-truth of Type $(2)$, two processes having the same means and covariance structures are in one cluster; for the third type, the pattern is the means and covariance structures of the tangent processes.
From the summary we conclude that a sufficient condition for the clustering algorithms (provided below) being consistent, is that the corresponding dissimilarity measure and its estimates satisfy the triangle inequality and its estimator are consistent (they converge to the theoretical dissimilarity as the path length goes up to the infinity).
Asymptotically Consistent Algorithms
====================================
In [@khaleghi2016], assuming the correct number of clusters $\kappa$ is known, two types of datasets are considered in the cluster analysis: offline dataset and online dataset. In the offline dataset, the number of sample paths and the length of each sample path do not vary with respect to time. However in the online dataset, both can vary. In [@khaleghi2016] for each type of datasets, by using a particular dissimilarity measure, asymptotically consistent algorithms (Algorithm 1 for offline dataset and Algorithm 2 for online dataset) have been derived, aiming to cluster distribution stationary ergodic processes. Here asymptotic consistency means the output clusters from the algorithm converge to the ground-truths either in probability (weak sense) or almost surely (strong sense). Based on Khaleghi et al.’s approaches, Peng et al. [@Peng2018-1; @Peng2018-2] suggested asymptotically consistent algorithms that are valid for a more general class of processes and dissimilarity measures.
Let $X_1,X_2$ be one of the 3 types of processes in the above section. We denote by $d(X_1,X_2)$ a dissimilarity measure between 2 stochastic processes $X_1,X_2$, which satisfies the triangle inequality. And we denote by $\widehat d(\mathbf x_1,\mathbf x_2)$ the estimate of $d(X_1,X_2)$, where for $i=1,2$, $\mathbf x_i=(x_1^{(i)},\ldots,x_{n_i}^{(i)})$ is an observed sample path of the process $X_i$, with length $n_i$. Moreover, assume that $\widehat d$ also verifies the triangle inequality and it is consistent: for all $\mathbf x_1,\mathbf x_2$, sampled from $X_1,X_2$ respectively, $$\widehat d(\mathbf x_1,\mathbf x_2)\xrightarrow[\min\{n_1,n_2\}\to\infty]{\mathbb P~\mbox{or}~a.s.}d(X_1,X_2),$$ where $\xrightarrow[]{\mathbb P}$ and $\xrightarrow[]{a.s.}$ denote the convergence in probability and almost sure convergence, respectively.
The clustering algorithms suggested by Peng et al. [@Peng2018-1; @Peng2018-2] are given below.
$(c_1,c_2) \longleftarrow \operatorname*{argmax}\limits_{(i,j)\in\{1,\ldots,N\}^2, i<j}\widehat{d}(\mathbf z_i,\mathbf z_j)$ $C_1 \longleftarrow \left \{ c_1 \right \}$; $C_2\longleftarrow\{c_2\}$ ***Assign the remaining points to the nearest centers*:**
\[thm:offline1\] Algorithms \[algo::offline\_known\_k\] and \[algo::online\_known\_k\] are asymptotically consistent for the processes of Types $(1)$ and $(2)$ respectively, provided that the correct number $\kappa$ of clusters is known, and the sample dissimilarity measure $\widehat d$ is consistent and both $\widehat d$ and $d$ satisfy the triangle inequality.
**Proof.** The consistency of Algorithms \[algo::offline\_known\_k\] and \[algo::online\_known\_k\] applied for clustering processes of Type $(1)$ is proved in [@khaleghi2016]; the consistency of the two algorithms applied for clustering processes of Type $(2)$ in proved in [@Peng2018-1]. $\square$
It is worth noting that in the above proof, the key features used are the fact that both $d$ and $\widehat d$ verify the triangle inequality and $\widehat d$ is a consistent estimator of $d$.
For clustering the processes of Type $(3)$, an additional assumption is needed, which will be introduced in Section \[local\_self\_similar\].
For clustering the processes of Type $(1)$, the specific form of $d$ and $\widehat d$ are given in [@khaleghi2016]. Then we mainly introduce the other 2 pairs of $(d,
\widehat d)$ for clustering analysis on the processes of Types $(2)$ and $(3)$.
Dissimilarity Measure for Covariance Stationary Ergodic Processes {#d:covariance}
=================================================================
The definition of covariance stationary ergodic process is given below.
A stochastic process $\{X_t\}_{t\in\mathbb N}$ is covariance stationary ergodic if:
- its mean and covariance are invariant subject to any time shift;
- any of its sample autocovariance converges in probability to the theoretical autocovariance function as the sample length goes to $+\infty$.
The dissimilarity measure $d$ and its sample estimate $\widehat d$ suggested in Peng et al. [@Peng2018-1] to measure the distance between 2 covariance stationary ergodic processes are given below:
\[definition1\] The dissimilarity measure $d$ between a pair of covariance stationary ergodic processes $X^{(1)}$, $X^{(2)}$ is defined as follows: $$\begin{aligned}
\label{def:d}
&&d(X^{(1)},X^{(2)})\nonumber\\
&&:= \displaystyle\sum_{m,l = 1}^{\infty} w_m w_l \rho\left(Cov(X_l^{(1)},\ldots,X_{l+m-1}^{(1)}),Cov(X_l^{(2)},\ldots,X_{l+m-1}^{(2)})\right),\end{aligned}$$ where:
- The sequence of positive weights $\{w_j\}$ is chosen such that $d(X^{(1)},X^{(2)})$ is finite.
- The distance $\rho$ between 2 equal-sized covariance matrices $M_1,M_2$ is defined to be $
\rho(M_1,M_2):=\|M_1-M_2\|_F$, with $\|\cdot\|_F$ being the *Frobenius norm*.
For two processes’ paths $\mathbf x_j=(X_1^{(j)},\ldots,X_{n_j}^{(j)})$ for $j=1,2$, let $n=\min\{n_1,n_2\}$, then the empirical covariance-based dissimilarity measure between $\mathbf x_1$ and $\mathbf x_2$ is given by $$\label{dxx}
\widehat{d}(\mathbf x_{1},\mathbf x_{2}):=\sum_{m= 1}^{m_n} \sum_{l= 1}^{n-m+1} w_m w_l \rho\left (\nu(X^{(1)}_{l\ldots n},m), \nu(X^{(2)}_{l\ldots n},m)\right),$$ where:
- $m_n$, chosen to be $o(n)$, denotes the size of covariance matrices considered in the estimator.
- $ \nu(\mathbf x,l,m):=\frac{\sum_{i=l}^{n-m+1}(X_i\ldots X_{i+m-1})^T(X_i\ldots X_{i+m-1})}{n-m-l+2}$ are the estimators of stationary covariance matrices.
Dissimilarity Measure for Locally Asymptotically Self-similar Processes {#local_self_similar}
=======================================================================
In this section we review the work on clustering processes of Type $(3)$. Locally asymptotically self-similar processes play a key role in the study of fractal geometry and wavelet analysis. They are generally not covariance stationary, however, one can still apply the dissimilarity measure $d$ introduced in Section \[d:covariance\] under some assumption (see [@Peng2018-2]).
The following definition of locally asymptotically self-similar process is given in [@Boufoussi2008].
\[locally\_asymptotically\_self\_similar\] A continuous-time stochastic process $\left\{Z_t^{(H(t))}\right\}_{t\ge0}$ with its index $H(\cdot)$ being a continuous function valued in $(0,1)$, is called locally asymptotically self-similar, if for each $t\ge0$, there exists a non-degenerate self-similar process $\left\{Y_u^{(H(t))}\right\}_{u\ge0}$ with self-similarity index $H(t)$, such that $$\label{local_self}
\left\{\frac{Z_{t+\tau u}^{(H(t+\tau u))}-Z_t^{(H(t))}}{\tau^{H(t)}}\right\}_{u\ge0}\xrightarrow[\tau\to0^+]{\mbox{f.d.d.}}\left\{Y_u^{(H(t))}\right\}_{u\ge0},$$ where the convergence $\xrightarrow[]{\mbox{\textit{f.d.d.}}}$ is in the sense of all the finite dimensional distributions.
Here $\{Y_u^{(H(t))}\}_u$ is called the *tangent process* of $\{Z_t^{(H(t))}\}_t$ at $t$ (see [@Falconer2002]). Throughout [@Peng2018-2] it is assumed that the datasets are sampled from a known number of processes satisfying the following condition: **Assumption $(\mathcal A)$:** The processes are locally asymptotically self-similar with distinct functional indexes $H(\cdot)$; their tangent processes’ increment processes are *covariance stationary ergodic*.
Then from (\[local\_self\]), Peng et al. [@Peng2018-2] showed the following: under Assumption ($\mathcal A$), for $\tau$ being sufficiently small, $$\label{local_self_2}
\left\{Z_{t+ \tau(u+h)}^{(H(t+\tau(u+h)))}-Z_{t+ \tau u}^{(H(t+\tau u))}\right\}_{u\in[0,Kh]}\stackrel{\mbox{f.d.d.}}{\approx}\left\{\tau^{H(t)}X_u^{(H(t))}\right\}_{u\in[0,Kh]},$$ where $K$ is an arbitrary positive integer. Statistically, (\[local\_self\_2\]) can be interpreted as: given a discrete-time path $Z_{t_1}^{(H(t_1))},\ldots,Z_{t_n}^{(H(t_n))}$ with $t_i=ih\Delta t$ for each $i\in\{1,\ldots,n\}$, sampled from a locally asymptotically self-similar process $\{Z_t^{(H(t))}\}$, its localized increment path with time index around $t_i$, i.e., $$\label{increment:z}
\mathbf z^{(i)}:=\left(Z_{t_{i+1}}^{(H(t_{i+1}))}-Z_{t_i}^{(H(t_i))},\ldots,Z_{t_{i+1+K}}^{(H(t_{i+1+K}))}-Z_{t_{i+K}}^{(H(t_{i+K}))}\right),$$ is *approximately* distributed as a covariance stationary ergodic increment process of the self-similar process $\left\{{\Delta t}^{H(t_i)}X_u^{(H(t_i))}\right\}_{u\in[0,Kh]}$. This fact inspires one to define the sample dissimilarity measure between two paths of locally asymptotically self-similar processes $\mathbf z_1$ and $\mathbf z_2$ as below: $$\label{dissimilairty_local}
\widehat{d^{*}}(\mathbf z_1,\mathbf z_2):=\frac{1}{n-K-1}\sum_{i=1}^{n-K-1}\widehat{d}(\mathbf z_1^{(i)},\mathbf z_2^{(i)}),$$ where $\mathbf z_1^{(i)}$, $\mathbf z_2^{(i)}$ are the localized increment paths defined as in (\[increment:z\]).
Accordingly, the consistency of Algorithms \[algo::offline\_known\_k\] and \[algo::online\_known\_k\] can be expressed in the following way:
\[thm:offline2\] Under Assumption $(\mathcal A)$, Algorithms \[algo::offline\_known\_k\] and \[algo::online\_known\_k\] are approximately asymptotically consistent, if $\widehat{d}$ is replaced with $\widehat{d^{*}}$.
In Theorem \[thm:offline2\], “approximately” is in the sense of Eq. (\[local\_self\_2\]).
Simulation Study
================
In the frameworks of khaleghi et al. [@khaleghi2016], Peng et al. [@Peng2018-1] and Peng et al. [@Peng2018-2], simulation study are provided. In [@khaleghi2016], a distribution stationary ergodic process is simulated based on random walk; in [@Peng2018-1] the increment process of fractional Brownian motion [@Ayache2004] is picked as an example of covariance stationary ergodic process; in [@Peng2018-2], simulation study is performed on the so-called multifractional Brownian motion [@Peltier1995], which is an excellent example of locally asymptotically self-similar process. The simulation study results for clustering distribution stationary ergodic processes are given in [@khaleghi2016]. Here we summarize the results for clustering the processes of Types $(2)$ and $(3)$, from Peng et al. [@Peng2018-1] and [@Peng2018-2] respectively.
Clustering Processes of Type $(2)$: Fractional Brownian Motion
--------------------------------------------------------------
Fractional Brownian motion (fBm) $\{B^H(s)\}_{s\ge0}$, as a natural extension of the Brownian motion, was first introduced by Kolmogorov in 1940 and then popularized by Mandelbrot and Taqqu [@Mandelbrot1968; @Taqqu2013]. The influences made by the fractional Brownian motion model have been on a great many fields such as biological science, physical sciences and economics (see [@Hofling2013]). As a stationary increment process, it is shown that the increment process of the fBm is covariance stationary ergodic (see [@Maruyama1970; @Slezak2017]).
In [@Peng2018-1], the clustering algorithms are performed on a dataset of $100$ paths of fBms with $\kappa=5$ clusters. In the sample dissimilarity measure the so-called $\log^*$-transformation is applied to increase the efficiency of the algorithms. One considers the mis-clustering rates to be the measure of fitting errors. The top figure in Figure \[fig::empirical\_results\] presents the comparison results of the offline and online algorithms, based on the behavior of mis-clustering rates as time increases. Both algorithms show to be asymptotically consistent as the mis-clustering rates decrease.
Clustering Processes of Type $(3)$: Multifractional Brownian Motion
-------------------------------------------------------------------
Multifractional Brownian motion (mBm) $\{W_{H(t)}(t)\}_{t\ge0}$, as a natural generalization of the fBm, was introduced in [@Peltier1995; @ACLV00]. Then it was quickly applied to describe phenomena in for instance molecular biology [@Marquez2012], biomedical engineering [@Buard2010] and biophysics [@Humeau2007].
It can be obtained from [@Boufoussi2008] that the mBm is locally asymptotically self-similar satisfying Assumption $(\mathcal A)$.
The datasets of mBms for testing the 2 clustering algorithms are similar to those of fBms. The performance of the algorithms are shown in the bottom figure in Figure \[fig::empirical\_results\]. Similar conclusion can be drawn that both offline and online algorithms are approximately asymptotically consistent.
[^1]: Institute of Mathematical Sciences, Claremont Graduate University, Claremont, CA 91711. Email: [email protected].
[^2]: School of Mathematical Sciences, Shanghai Jiao Tong University, Shanghai, China. Email: [email protected].
[^3]: Institute of Mathematical Sceinces and Drucker School of Management, Claremont Graduate University, Claremont, CA 91711. Email: [email protected].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Driven two-level system is a useful model to describe many quantum objects, particularly in quantum information processing. However, the exact master equation for such a system is barely explored. Making use of the Feynman-Vernon influence functional theory, we derive an exact non-Markovian master equation for the driven two-level system and show the lost feature in the perturbative treatment for this system. The perturbative treatment leads to the time-convolutionless (TCL) and the Nakajima-Zwanzig (NZ) master equations. So to this end, we derive the time-convolutionless (TCL) and the Nakajima-Zwanzig (NZ) master equations for the system and compare the dynamics given by the three master equations. We find the validity condition for the TCL and NZ master equations. Based on the exact non-Markovian master equation, we analyze the regime of validity for the secular approximation in the time-convolutionless master equation and discuss the leading corrections of the nonsecular terms to the quantum dynamics, significant effects are found in the dynamics of the driven system.'
author:
- 'H. Z. Shen$^{1}$, M. Qin$^{1}$, Xiao-Ming Xiu$^{1,2}$, and X. X. Yi$^3$ [^1]'
title: 'Exact non-Markovian master equation for a driven damped two-level system'
---
Introduction
============
The dynamics of open quantum systems [@Alicki2007717; @Breuer2002; @Weiss2008] has attracted much attention and becomes active again in recent years due to its possible applications in quantum information science [@DiVincenzo1998393; @Knill2001409; @Cirac199959; @DiVincenzo200048; @Cirac199778; @Duan200367]. Indeed the study of coupled system-environment system is an long standing endeavor in many fields of physics including quantum optics [@Walls1994; @Scully1997; @Gardiner2000; @Carmichael1993], atomic optics [@Mandel1995; @Weissbluth1988; @Vogel1994; @Compagno1995; @Carmichael1993] and condensed matter physics [@Caldeiral1983149; @Leggett198759; @Weiss2008]. The coupling of the system to its environment leads to dissipation and dephasing with flows of energy or information from the system to the environment [@Breuer2002; @Weiss2008]. The back flowing of information from the environment to the system determines the Markovianity of the dynamics.
Driven two-level model is available to effectively describe many actual physical systems, for example, a quantum bit in quantum information processing. Thus the theoretical analysis as well as the practical implementation of the driven two-level systems brings us a renewed topic. There are several ways to create a driven two-level system (or qubit) today by current quantum technologies, each exploits different approaches or in different quantum systems. For instance, by means of quantum optics and in microscopic quantum objects (electrons, ions, atoms) in traps, quantum dots, and quantum circuits [@Majer200594; @Berkley2003300; @Pashkin2003421; @Bellomo200799]. Different implementations of qubit [@Nielsen2000; @Barenco199574] are subjected to different types of environmental noise [@Biercuk2009458], most environments are assumed Markovian [@Das200942; @Sinaysky200878] and the dynamics of system was studied perturbatively in the literatures.
In recent years, an increasing interest has been paid to developing a non-Markovian generalization for open quantum system theory, some of them are formulated in terms of non-local time evolutions. There exist diverse formalisms for describing memory effects, including the generalization of the Lindblad master equation from time-independent dissipative rates to time-convoluted kernel functions. A wide class of both phenomenological and theoretical approaches were formulated for building and characterizing this type of master equations, which in turn lead to a completely positive map.
By means of the Feynman-Vernon influence functional theory [@Feynamn196324118; @Caldeira1983121587; @Hu1992452843; @zhang199062867; @Anastopoulos200062], exact master equations describing the general non-Markovian dynamics of a wide range of open quantum system have been recently developed, e.g. quantum Brownian motion [@Hu1992452843; @Karrlein199755153; @Haake198532], single-mode cavity [@An20093241737] and two entangled cavities [@An200776042127; @An200990317] with vacuum fluctuations, **spin-boson model** [@Lucke199911110843], coupled harmonic oscillators [@Chow200877011112; @Paz2008100220401; @Paz200979032102], quantum dot in nanostructures [@Tu200878235311; @Tu20098631], various nanodevices with time-dependent external control field [@Jin201012083013], nanocavity systems including initial system-reservoir correlations [@Tan201183], and photonic networks imbedded in photonic crystals [@Lei20123271408; @zhang2012109170402]. However, an exact master equation for driven systems are very rare.
Projection operator technique is other mean to study the open quantum system, both the time-convolutionless (TCL) [@Shibata197717; @Prataviera201387; @Chaturvedi197935] master equation and the Nakajima-Zwanzig (NZ) [@Nakajima195820; @Zwanzig196033; @zhang201387032117] master equation can be derived by this approach. The NZ approach provides us with a generalized master equation in which the time derivative of the density operator is connected to the past of the reduced density matrix through the convolution of the density operator and an appropriate integral kernel. While the TCL approach leads to a generalized master equation which is local in time. It seems that the NZ should run better than the TCL approach in describing the non-Markovian effect, since it takes into account the history of the reduced density matrix. However this is not the case as we will show later, examples in [@Breuer2002; @Breuer1999591633; @Yan1998582721; @Ferraro200980042112; @Xu20011143868; @Schroder2006124084903; @Liu200776022312; @Haikka201081052103; @Haikka2010014047] confirm this point, namely, the exact dynamics of the open system can be described via a master equation with time-dependent decay rate, as in the well-known case of the Hu-Paz-Zhang generalized master equation[@Breuer2002; @Hu1992452843].
In the weak coupling limit, the non-Markovian master equation for a driven two-level system coupled to a bosonic reservoir at zero temperature has been derived and discussed in Ref.[@Haikka201081052103]. This derivation treat the system-environment coupling perturbatively, and hence it is available for weak system-environment couplings. In this paper, exploiting the Feynman-Vernon influence theory in the coherent state path integral formalism, we derive an exact non-Markovian master equation for the driven two-level system. The Feynman-Vernon influence theory enables us to treat the environment-system coupling non-perturbatively. The dynamics of the driven open two-level system, going beyond the TCL, NZ, and Markovian approximations, is governed by an effective action associated with the influence functional containing all the influences of the environment on the system. The exact master equation is available to examine the validity of those perturbative approaches applied to the TCL and NZ techniques. We show that the TCL approach works better than the NZ one, since the latter does not guarantee the positivity of the density matrix when the correlations in the reservoir become strong, while the former is available for a wider range of values of reservoir memory time.
The remainder of the paper is organized as follows. In Sec. [II]{}, we introduce a model to describe a driven two-level system subject to reservoir and give a detailed derivation of the influence functional for the model in the coherent state representation. In Sec. [III]{}, an exact non-Markovian master equation describing the evolution of the driven open two-level system is derived. In Sec. [IV]{}, a derivation of the second-order NZ master equation is presented and the characteristics of the second-order TCL derived in Ref. [@Breuer2002] are discussed, and then we give a comparison among the exact, TCL, and NZ master equations. In Sec. [V]{}, we investigate the validity of the secular approximation in Markovian and non-Markovian regimes, respectively. Discussions and conclusions are given in Sec. [VI]{}.
atomic coherent state path-integrate approach to the driven open two-level system
=================================================================================
Model Hamiltonian
-----------------
**We start by considering a two-level system with Rabi frequency $\omega _0$ driven by an external laser of frequency $\omega_L$. The two-level atom is embedded in a bosonic reservoir at zero-temperature modeled by a set of infinite harmonic oscillators. In a rotating frame, the Hamiltonian of such a system (system plus environment) takes** $$\begin{aligned}
H = {H_S} + {H_E} + {H_I},\label{hHh12}\end{aligned}$$ with $$\begin{aligned}
{H_S} =& \Delta \cdot {\sigma _ + }\sigma_- + \Omega {\sigma _ x },\\
{H_E} =& \sum\limits_k {{\Omega _k}a_k^\dag {a_k}} ,\\
{H_I} =& \sum\limits_k {{g_k}{\sigma _ + }{a_k} + H.c.} ,
\label{H}
\end{aligned}$$ **where $\Delta = {\omega _0} - {\omega _L}$, ${\Omega _k} =
{\omega _k} - {\omega _L}$, and ${\sigma _x} = {\sigma _ +}
+\sigma_-$. $\Omega$ is the driven strength, and H.c. stands for the Hermitian conjugation. $\sigma _+= \left| e \right\rangle
\left\langle g \right|$ is the Pauli matrix. ${a_k}$ and ${g_k}$ are the annihilation operator and coupling constants, respectively. In the following we shall start with this Hamiltonian (\[hHh12\]) and derive all master equations in this paper.**
Coherent state representation
-----------------------------
The starting point of analysis is to observe that the lowing and raising operators of the atomic transition operators ${\sigma _ + }
= \left| e \right\rangle \left\langle g \right|$ and $\sigma_- =
\left| g \right\rangle \left\langle e \right|$ satisfy anticommutation rules similar to those of fermions, i.e., $$\begin{aligned}
\left\{ {\sigma_-,{\sigma _ + }} \right\} =&
\left| e \right\rangle \left\langle e \right| + \left| g \right\rangle \left\langle g \right| \equiv 1,\\
\left\{ {\sigma_- ,\sigma_-} \right\} = & \left\{ {{\sigma _ +
},{\sigma _ + }} \right\} = 0, \label{antio}
\end{aligned}$$ where $\{ A,B\} = AB + BA$. Identifying the ground state $\left| g
\right\rangle$ with the fermionic vacuum, we can therefore treat ${\sigma _ + }$ and $\sigma_-$ as fermionic creation and annihilation operators, respectively. Following Ref. [@Cahill1999591538], we introduce a couple of conjugate Grassmann variables $\zeta $ and ${\bar{\zeta}}$ imposing standard anticorrelation with the annihilation and creation operators of the system.
Therefore, coherent states are defined as a tensor product of states generated by exponentiated operation of a creation operator and a suitable label on a chosen fiducial state [@Glauber1963131; @Shresta200571022109; @Anastopoulos200062; @zhang199062867; @Ghosh201286011138] $$\begin{aligned}
\left| \mathbf{z} \right\rangle = \prod\limits_k {\left| {{z_k}}
\right\rangle } ,\left| {{z_k}} \right\rangle = \exp (a_k^\dag
{z_k})\left| {{0_k}} \right\rangle , \label{coherentstate1}\end{aligned}$$ and $$\begin{aligned}
\left| \zeta \right\rangle = \exp ({\sigma _ + }\zeta )\left| g \right\rangle.
\label{coherentstate2}\end{aligned}$$ For bosonic coherent states defined in Eq. (\[coherentstate1\]), the label ${z_k}$ is a complex number, and for atomic coherent states defined in Eq. (\[coherentstate2\]), the label $\zeta $ is a Grassmannian or anticommuting number. A state of the combined atom-field system can be expanded in a direct product of coherent state $$\begin{aligned}
\left| \mathbf{z}\zeta \right\rangle = \left| \mathbf{z}
\right\rangle \otimes \left| \zeta \right\rangle.\end{aligned}$$ Atomic and bosonic coherent states possess the well-known properties such as being nonorthogonal $$\begin{aligned}
\begin{aligned}
\left\langle {\mathbf{z}}
\mathrel{\left | {\vphantom {z {z'}}}
\right. \kern-\nulldelimiterspace}
{{\mathbf{z'}}} \right\rangle = \exp (\sum\limits_k {\bar{z}_k {z'_k}} ),\left\langle {\zeta }
\mathrel{\left | {\vphantom {\zeta {\zeta '}}}
\right. \kern-\nulldelimiterspace}
{{\zeta '}} \right\rangle = \exp ({\bar{\zeta} }\zeta '),\\
\end{aligned}\end{aligned}$$ $$\begin{aligned}
\begin{aligned}
{a_k}\left| {{z_k}} \right\rangle = {z_k}\left| {{z_k}} \right\rangle ,\sigma_-\left| \zeta \right\rangle = \zeta \left| \zeta \right\rangle ,
\label{ortho2}
\end{aligned}\end{aligned}$$ where $\bar{z}_k$ and $\bar{\zeta}$ denote the conjugation of $z_k$ and $\zeta$, respectively. Despite their nonorthogonality, both types of coherent states form an over-complete basis set $$\begin{aligned}
\int {d\varphi ({\mathbf{z}})\left| \mathbf{z} \right\rangle \langle \mathbf{z}|} = \int {d\varphi (\zeta )\left| \zeta \right\rangle \langle {\zeta}|} = 1,
\label{unity}\end{aligned}$$ where the integral measures are defined by $d\varphi ({\mathbf{z}})
= \prod\limits_k {\frac{{\exp ( - {{\bar z}_k}{z_k}){d^2}{z_k}}}{\pi
},}$ and $d\varphi (\zeta ) = {\exp ( - \bar \zeta \zeta ){d^2}\zeta
}$. As shown, the bosonic coherent states we use here are not normalized, and the normalization factors are moved into the integration measures, which is similar to the Bargmann representation of the complex space. The application of the coherent state representation makes the evaluation of path integrals extremely simple. In the coherent state representation, the Hamiltonians of the system, the environment, and the interaction between them are expressed as, respectively $$\begin{aligned}
\begin{aligned}
{H_S}({\bar{\zeta} },\zeta ) =& \Delta {\bar{\zeta} }\zeta + \Omega ({\bar{\zeta} } + \zeta ),\\
{H_E}({{\rm{\bar{\mathbf{z}}}} },\mathbf{z}) =& \sum\limits_k {{\Omega _k}\bar{z}_k {z_k}},\\
{H_I}({\bar{\mathbf{z}} },\mathbf{z},{\bar{\zeta} },\zeta ) =& ({g_k}{\bar{\zeta} }{z_k} + g_k^ * \bar{z}_k \zeta ).
\label{representation}
\end{aligned}\end{aligned}$$ With these notations, we will present a detailed derivation of the exact master equation for the reduced density matrix of the system in the following sections.
The influence functional in coherent state representation
---------------------------------------------------------
Explicitly, the density matrix of the whole system (the system plus the environment) obeys the quantum Liouville equation, $i\partial {\rho _T}(t)/\partial t = [H,{\rho _T}(t)]$, which gives the formal solution $$\begin{aligned}
{\rho _T}(t) = \exp ( - iHt){\rho _T}(0)\exp (iHt).
\label{rhoT }\end{aligned}$$ In the coherent state representation, by use of Eq. (\[unity\]), ${\rho _T}(t)$ can be expressed as $$\begin{aligned}
&\langle {\zeta _f},{{\bf{z}}_f}|{\rho _T}(t) \left| {{{\zeta '}_f},{{\bf{z}}_f}} \right\rangle\\
=& \int {d\varphi ({\mathbf{z}_i})d\varphi ({\zeta _i})d\varphi ({\mathbf{z}'_i})d\varphi ({\zeta '_i})\left\langle {{{\zeta _f},{\mathbf{z}_f};t}}
\mathrel{\left | {\vphantom {{{\zeta _f},{\mathbf{z}_f};t} {{\zeta _i},{\mathbf{z}_i};0}}}
\right. \kern-\nulldelimiterspace}
{{{\zeta _i},{\mathbf{z}_i};0}} \right\rangle } \\
&\times \left\langle {{\zeta _i},{\mathbf{z}_i}} \right|{\rho _T}(0)\left| {{\zeta '_i},{\mathbf{z}'_i}} \right\rangle \left\langle {{{\zeta '_i},{\mathbf{z}'_i};0}}
\mathrel{\left | {\vphantom {{{\zeta '_i},{\mathbf{z}'_i};0} {{\zeta '_i},{\mathbf{z}_i};t}}}
\right. \kern-\nulldelimiterspace}
{{{\zeta '_f},\mathbf{{z}}_f;t}} \right\rangle.
\label{repre}
\end{aligned}$$ Assume the initial density matrix be factorized into a direct product of the system and the environment state, i.e., ${\rho
_T}(0) = \rho (0) \otimes {\rho _E}(0)$ [@Leggett198759], the reduced density matrix of the system is then given by $$\begin{aligned}
&\rho (\bar{\zeta} _f ,{\zeta '_f};t) = \int {d\varphi ({\mathbf{z}_f})\left\langle {{\zeta _f},{\mathbf{z}_f}} \right|{\rho _T}(t)\left| {{\zeta '_f},{\mathbf{z}_f}} \right\rangle } \\
& = \int {d\varphi ({\zeta _i})d\varphi ({\zeta '_i})\rho (\bar{\zeta} _i ,{\zeta '_i};0) \cdot J(\bar{\zeta} _f ,{\zeta '_f};t|\bar{\zeta} _i ,{\zeta '_i};0)}.
\label{reducedmatrix}
\end{aligned}$$ The next task is to determine the effective propagating function for the reduced density matrix [@Feynamn196324118; @Anastopoulos200062; @Ishizaki2008347185], $$\begin{aligned}
J(\bar{\zeta} _f ,{\zeta '_f};t|\bar{\zeta} _i ,{\zeta '_i};0) =& \int {{D^2}\zeta {D^2}\zeta '\exp \{ i({S_S}[{\bar{\zeta}},\zeta ]} \\
&- S_S^ * [\bar{\zeta}^\prime ,\zeta '])\} F[{\bar{\mathbf{\zeta}}},\mathbf{\zeta},\bar{\mathbf{\zeta}} ^\prime ,\mathbf{\zeta}'],
\label{JJ}
\end{aligned}$$ with ${{S_S}[{\bar{\zeta}},\zeta ]}$ being the action of the system in the atomic coherent state representation, see Eq. (\[dfg\]). $F[{\bar{\mathbf{\zeta}}},\mathbf{\zeta},\bar{\mathbf{\zeta}}^\prime
,\mathbf{\zeta}']$ is the influence functional which takes into account the back-action (in Eq. (\[FJ\])) of the environment on the system.
Assume the environment be initially at zero temperature, i.e., the initial state of the environment takes, $$\begin{aligned}
{\rho _E} = {\left| 0 \right\rangle _{BB}}\langle 0|,
\label{rhoE}\end{aligned}$$ then the influence functional can be solved exactly and we have $$\begin{aligned}
F[{\bar{\mathbf{\zeta}}},\mathbf{\zeta},\bar{\mathbf{\zeta}}^\prime ,\mathbf{\zeta}'] =& \exp \{ \int_{t_0}^t {d\tau \int_{t_0}^\tau {d\tau '[f(\tau - \tau ')} } \\
&(\bar{\zeta}^\prime (\tau ) - {\bar{\zeta}}(\tau ))\zeta (\tau ') + {f^ * }(\tau - \tau ')\\
&\bar{\zeta}^\prime (\tau ')(\zeta (\tau ) - {\zeta ^\prime }(\tau ))]\},
\label{finally F}
\end{aligned}$$ where $$\begin{aligned}
f(\tau - \tau ') =& \sum\limits_k {{{\left| {{g_k}} \right|}^2}{e^{ - i{\Omega _k}(\tau - \tau ')}}} \\
{\rm{ = }}&\int {d\omega J(\omega ){e^{ - i(\omega - {\omega _L})(\tau - \tau ')}}}
\label{ft}
\end{aligned}$$ is called the dissipation-fluctuation kernel.
The details of derivation of Eq. (\[finally F\]) can be found in Appendix.
The exact non-Markovian master equation
=======================================
We now derive the master equation for the reduced density matrix of the system. Since the effective action after tracing/integrating out the environmental degrees of freedom, (i.e., combining Eqs. (\[JJ\]) and (\[finally F\]) together) is in a quadratic form of the dynamical variables, the path integral (\[JJ\]) can be calculated exactly by making use of the stationary path method and Gaussian integrals [@Faddeev1980; @Feynman1965]. Substituting Eq. (\[dfg\]) into Eq. (\[JJ\]), we have $$\begin{aligned}
J({{\bar \zeta}_f},{\zeta '_f};t|{\zeta _i},{{\bar \zeta '}_i};0) =& \int {{D^2}\zeta {D^2}\zeta '\exp \{ \frac{1}{2}[{{\bar \zeta }_f}\zeta (t) + \bar \zeta ({t_0}){\zeta _i}} \\
& + \bar \zeta '(t){\zeta'_f} + {{\bar \zeta '}_i}\zeta '({t_0})] - \int_{t_0}^t {d\tau \frac{1}{2}[\bar \zeta \dot \zeta } \\
&- \dot{\bar{\zeta}} \zeta + \dot{\bar{\zeta}}'\zeta ' - \bar \zeta '\dot \zeta '] + i{H_S}(\bar \zeta ,\zeta )\\
&-i{H_S}(\bar \zeta ',\zeta ')\} F[\bar \zeta ,\zeta ,\bar \zeta ',\zeta '].
\label{FINALLYJ}
\end{aligned}$$ To calculate the path integral in Eq. (\[FINALLYJ\]), we use the stationary phase method [@Klauder1979192349; @zhang199062867], which yields the equations of motion $$\begin{aligned}
&\dot \zeta (\tau ) + i[\Omega + \Delta \cdot \zeta (\tau )] + \int_{t_0}^\tau {d\tau 'f(\tau - \tau ')\zeta (\tau ')} = 0,\\
&\dot \zeta '(\tau ) + i[\Omega + \Delta \cdot \zeta '(\tau )] - \int_\tau ^t {d\tau 'f(\tau - \tau ')\zeta '(\tau ')}\\
& + \int_{t_0}^t {d\tau 'f(\tau - \tau ')\zeta (\tau ')} = 0,
\label{fstationary path}
\end{aligned}$$ subject to the boundary conditions $\zeta ({t_0}) = {\zeta _i}$ and $\zeta '(t) = {\zeta '_f}$, respectively. **$\bar \zeta
'(\tau )$ and $\bar \zeta (\tau )$ denote the conjugates of $ \zeta
'(\tau )$ and $\zeta (\tau )$, respectively. The equations for these conjugations can be obtained by first exchanging $\zeta (\tau )$ and $\zeta '(\tau )$ in Eq. (\[fstationary path\]) and taking then a complex conjugate to these equations. The corresponding boundary conditions are $\bar \zeta '({t_0}) \equiv \bar \zeta '_i$ and $\bar
\zeta (t) \equiv {\bar \zeta _f}$. With these boundary conditions, we can get the solution of $\zeta (\tau )$ and $\zeta
'(\tau )$. For clarity, we illustrate these notations in Fig. \[bianjietiaojian\]. Noticing ${t_0} \le \tau \le t$, we keep in mind that $\zeta (t)$ in Fig. \[bianjietiaojian\] (a) can be obtained by setting $\tau =
t$ and $\zeta '({t_0})$ in Fig. \[bianjietiaojian\] (b) can be obtained by $\tau = {t_0}$. Fig. \[bianjietiaojian\] (c) and (d) is similar, namely, $\bar \zeta '(t)$ and $\bar \zeta '({t_0})$ can be obtained with $\tau = t$ and $\tau = {t_0}$, respectively.**
![(Color online) Schematic illustration of the four independent paths denoted by $\zeta (\tau )$, $\zeta '(\tau )$, $\bar \zeta '(\tau )$ and $\bar \zeta (\tau )$, respectively.[]{data-label="bianjietiaojian"}](boundaryendpoints.eps){width="48.00000%"}
The solution of the integro-differential Eq. (\[fstationary path\]) can be expressed in terms of two complex functions $u(\tau)$ and $u_1(\tau)$ as $$\begin{aligned}
\zeta '(\tau )=& {u_1}(\tau )[{\zeta '_f} - \zeta (t)]+\zeta (\tau ) ,\\
\zeta (\tau ) =& u(\tau ){\zeta _i} + h(\tau ),
\label{transformation}
\end{aligned}$$ a similar transformation can be written down for their conjugate variables with the exchange of $\zeta $ with $\zeta'$ for the boundary values $\bar{\zeta}(t) = \bar{\zeta}_f$ and $\bar{\zeta}'(t_0)=\bar{\zeta}'_i$. Substituting Eq. (\[transformation\]) into Eq. (\[fstationary path\]), we can obtain the equations of motion for $u(\tau ), {u_1}(\tau )$ and $h(\tau )$ $$\begin{aligned}
\dot u(\tau ) + i\Delta \cdot u(\tau ) + \int_{t_0}^\tau d\tau 'f(\tau - \tau ')u(\tau ') &= 0, \\
{{\dot u}_1}(\tau ) + i\Delta \cdot{u_1}(\tau ) - \int_\tau ^t d \tau 'f(\tau - \tau '){u_1}(\tau '){\rm{ }} &= 0, \\
\dot h(\tau ) + i\Delta \cdot h(\tau ) + \int_{t_0}^\tau d\tau 'f(\tau - \tau ')h(\tau ') &= - i\Omega ,
\label{uu}
\end{aligned}$$ subject to the boundary conditions ${u_1}(t)=1$, $u(t_0)=1$ and $h(t_0)=0$ with $t_0 \le \tau ,\tau ' \le t.$ By means of Laplace transform to Eq. (\[uu\]), we can easily find that $$\begin{aligned}
{u_1}(\tau) = {u^ * }(t - \tau ),h(\tau ) = - i\Omega \int_{t_0}^\tau {d\tau 'u(\tau - \tau ')}.
\label{solution for u}
\end{aligned}$$ Now, we set $\tau = t_0$ in the first equation and $\tau = t$ in the second equation of Eq. (\[transformation\]), $\zeta (t)$ and $\zeta '({t_0})$ can be expressed in terms of the boundary conditions ${\zeta _i}$ and $\zeta'_{f}$ $$\begin{aligned}
\zeta (t) =& u(t){\zeta _i} + h(t),\\
\zeta '(t_0) =& {u^*}(t)[{{\zeta '}_f} - h(t)] + n(t){\zeta _i},
\label{inversezeta}
\end{aligned}$$ where $n(t) = 1 - {\left| {u(t)} \right|^2}$. Similarly, ${\bar{\zeta}}(t_0)$ and $\bar{\zeta}'(t)$ can be obtained by exchanging $\zeta $ and $\zeta'$ in Eq. (\[inversezeta\]) and by taking a complex conjugate to these equations. Finally, substituting these results with Eq. (\[transformation\]) into Eq. (\[FINALLYJ\]), we obtain the form of the propagating function for the reduced density matrix $$\begin{aligned}
J({{\bar \zeta }_f},{\zeta '_f};t|{\zeta _i},{{\bar \zeta '}_i};0)=& \exp \{ u(t)[\bar{\zeta}_{f} - {h^ * }(t)]{\zeta _i} +{u^ * }(t)\bar{\zeta}'_{i}\\
& \times [\zeta'_{f} - h(t)] +n(t)\bar{\zeta}'_{i}{\zeta _i} + h(t)\bar{\zeta }_f\\
& + {h^ *(t) }\zeta'_{f} - {\left| {h(t)} \right|^2}\}.
\label{finallyJJ}
\end{aligned}$$ **Notice that the pre-exponential factor in Eq. (\[finallyJJ\]) is one, this is due to the fact that Eq. (\[finallyJJ\]) is the result of integrating out fluctuations around the stationary path.** Now we can derive the master equation by computing the time derivative of Eq. (\[reducedmatrix\]). First, from Eq. (\[finallyJJ\]), we can write down the following identities $$\begin{aligned}
{\zeta _i}J = \frac{1}{u}\left( {\frac{{\delta J}}{{\overline{\zeta} _f}} - hJ} \right),\bar{\zeta}'_{i}J = \frac{1}{{{u^ * }}}\left( {\frac{{\delta J}}{{{\zeta '_f}}} - {h^ * }J} \right),
\label{zetai}\end{aligned}$$ which will be used to remove ${\zeta _i}$ and $\bar{\zeta}'_{i}$ from the time derivative of $J$. After taking time derivative of Eq. (\[reducedmatrix\]) and substituting Eqs. (\[finallyJJ\]) and (\[zetai\]) into it, we obtain the evolution equation $$\begin{aligned}
\frac{{\partial \rho (\overline{\zeta} _f ,{\zeta '_f})}}{{\partial t}} =& m\bar{\zeta} _f{{\rm P}_1} + {m^ * }{\zeta '_f}{{\rm P}_2} - (m + {m^ * }){{\rm P}_3}\\
& + {m^ * }{h^ * }{{\rm P}_1} + mh{{\rm P}_2} - {{\dot h}^ * }{{\rm P}_1} - \dot h{{\rm P}_2}\\
& - m h\bar{\zeta} _f\rho - {m^ * }{h^ * }\rho {\zeta '_f} + \dot h\bar{\zeta} _f\rho\\ &+ {{\dot h}^ * }\rho {\zeta '_f},
\label{representationrho}
\end{aligned}$$ where, $m(t) \equiv \frac{{\dot u(t)}}{{u(t)}}$, ${{\rm P}_1}
\equiv \frac{\partial{\rho}}{\partial{\overline\zeta_f}}$, ${{\rm
P}_2} \equiv \frac{{\partial \rho }}{{\partial {\zeta '_f}}}$, ${{\rm P}_3} \equiv \frac{{{\delta ^2}\rho }}{\partial ^2
\overline\zeta _f\zeta_f'}$. By introducing the following functional differential relations in the coherent state representation [@Anastopoulos200062; @Tu200878235311] $$\begin{aligned}
&\bar{\zeta} _f{{\rm P}_1} \leftrightarrow {\sigma _ + }
\sigma_-\rho (t),{{\rm P}_2}{\zeta '_f} \leftrightarrow
\rho (t){\sigma _ + }\sigma_-,{{\rm P}_3}
\leftrightarrow \sigma_-\rho (t){\sigma _ + },
\label{234}
\end{aligned}$$ we arrive at an exact non-Markovian master equation $$\begin{aligned}
\frac{{d\rho (t)}}{{dt}} = - i[H(t),\rho (t)]
+ \gamma (t)[2\sigma_-\rho (t){\sigma _ + }
- \{ {\sigma _ + }\sigma_-,\rho (t)\} ],
\label{finallyrhoo}
\end{aligned}$$ with the effective Hamiltonian containing the classical driven field $$\begin{aligned}
H(t) = s(t){\sigma _ + }\sigma_- + r(t){\sigma _ + } + {r^ * }(t)\sigma_-.
\label{H(t)}\end{aligned}$$ The renormalized frequency $s(t)$ and the renormalized driving field $r(t)$ are results of the back-action of the environment. The time dependent dissipative coefficient $\gamma (t)$ describes the dissipative non-Markovian dynamics due to the interaction between the system and environment. All these time-dependent coefficients can be given explicitly, $$\begin{aligned}
s(t) =& \frac{i}{2}[m(t) - c.c.],\\
\gamma (t){\rm{ = }}& - \frac{1}{2}[m(t) + c.c.],\\
r(t) =& i[\dot h(t) -h(t)m(t)],
\label{dphiin}
\end{aligned}$$ where $u(t)$ and $h(t)$ are determined by the integro-differential equations of Eq. (\[uu\]). The non-Markovian effect is fully manifested in the integral kernels in Eq. (\[uu\]), which include the non-local time-correlation function $f(t)$ of the environment. The non-Markovian memory effect is coded into the homogenous non-local time integrals with the integral kernel. In addition, our derivation of the master equation is fully non-perturbative, which goes beyond the TCL, NZ and Markovian approximations and includes all effects resulting from the environment-system couplings.
comparison between the exact and approximate master equations
=============================================================
The Nakajima-Zwanzig and time-convolutionless master equations
--------------------------------------------------------------
To derive the second-order perturbative master equation, we first go to the interaction picture, in which the effective Hamiltonian ${H_I}(t)$ in Eq. (\[H\]) can be rewritten as $$\begin{aligned}
{H_I}(t) = \sigma_-(t){a^\dag }(t) + H.c.,
\label{HI}\end{aligned}$$ where $\sigma_-(t) = {U^\dag }(t)\sigma_-U(t)$, $U(t) = {e^{ -
i{H_S}t}}, $ ${a^\dag }(t) = \sum\limits_k {{g_k}a_k^\dag
{e^{i{\Omega _k}t}}} $. The density operator ${{\bar \rho }_T}(t)$ of the whole system including the system and environment satisfies the following Liouville equation $$\begin{aligned}
\dot{\bar{\rho}}_T(t) = - i[{H_I}(t),{{\bar \rho }_T}(t)].
\label{rhoI}\end{aligned}$$ Integrating the left and right sides of Eq. (\[rhoI\]), we have $$\begin{aligned}
{{\bar \rho }_T}(t) = {{\bar \rho }_T}(t_0) - i\int_{t_0}^t
{dt'[{H_I}(t'),{{\bar \rho }_T}(t')]} . \label{rhI}\end{aligned}$$ Substituting Eq. (\[rhI\]) into Eq. (\[rhoI\]), we obtain $$\begin{aligned}
\dot{\bar{\rho}}_T(t)=& - i[{H_I}(t),{{\bar \rho }_T}(0)]\\
&- \int_{t_0}^t {dt'[{H_I}(t),[{H_I}(t'),{{\bar \rho }_T}(t')]]} .
\label{rhoI2}
\end{aligned}$$ Tracing over the degrees of freedom of the environment, we can obtain the dynamical equation for the system density matrix $\bar
\rho (t) = T{r_B}{{\bar \rho }_T}(t)$ $$\begin{aligned}
\dot{\bar{\rho}}(t)=& - iT{r_R}[{H_I}(t),{{\bar\rho}_T}(t_0)]\\
&- T{r_R}\int_{t_0}^t {dt'[{H_I}(t),[{H_I}(t'),{{\bar \rho }_T}(t')]]} .
\label{rhosi}
\end{aligned}$$ Let us apply the Born approximation and assume that the reservoir stays in the vacuum state (\[rhoE\]) in the dynamics, then we have $$\begin{aligned}
\dot{\bar{\rho}} (t) = - T{r_R}\int_{t_0}^t
{dt'[{H_I}(t),[{H_I}(t'),\bar \rho (t') \otimes {\rho _E}]]}.
\label{ohbmnb}\end{aligned}$$ Notice that $$\begin{aligned}
\left\langle {a(t){a^\dag }({t_1})} \right\rangle =& f(t - {t_1}),\\
\left\langle {{a^\dag }(t){a^\dag }({t_1})} \right\rangle
=& \left\langle {a(t)a({t_1})} \right\rangle
= \left\langle {{a^\dag }(t)a({t_1})} \right\rangle = 0,
\label{dgbbvx}
\end{aligned}$$ where $\left\langle A \right\rangle {\rm{ =
T}}{{\rm{r}}_B}\left\langle {A{\rho _E}} \right\rangle =
\left\langle 0 \right|A\left| 0 \right\rangle_B$, and substituting Eq. (\[HI\]) into Eq. (\[ohbmnb\]), we have $$\begin{aligned}
\dot{\bar{\rho}}(t) = \int_{t_0}^t {dt'f(t - t')[} {\rm{ }}
\sigma_- (t')\bar \rho (t'),{\sigma _+}(t)] + H.c..
\label{NZ2}
\end{aligned}$$ By transforming Eq. (\[NZ2\]) back into the Schrödinger picture, we obtain $$\begin{aligned}
{{\dot \rho }_{NZ}} =& - i[{H_S},{\rho _{NZ}}(t)]
+ \int_{{t_0}}^t {dt'\{ f(t - t')} [U(t - t')\\
& \times {\sigma _ - }{\rho _{NZ}}(t'){U^\dag }(t - t'),{\sigma _ +
}] + H.c.\}. \label{RHOnz2}
\end{aligned}$$ The non-Markovian master equation (\[RHOnz2\]) is in the standard form of the Nakajima-Zwanzig (NZ) equation $\dot \rho (t) = \int_0^t
{dt'f(t,t')\rho (t')}$ [@Nakajima195820; @Zwanzig196033], where the NZ kernel $f(t,t')$ is of the time-translationally-invariant form $f(t-t')$.
Note that Eq. (\[ohbmnb\]) is in a form of delayed integro-differential equation and thus it is a time-nonlocal master equation. It is worth reminding that the other systematically perturbative non-Markovian master equation that is local in time can be derived from the time-convolutionless projection operator formalism [@Breuer2002; @Breuer1999591633; @Yan1998582721]. Now, we go to the details. Under a similar assumption, i.e., the factorized initial system-reservoir density matrix, the second-order time-convolutionless master equation in the interaction picture can be obtained [@Breuer2002; @Breuer1999591633; @Yan1998582721; @Ferraro200980042112; @Xu20011143868; @Schroder2006124084903; @Liu200776022312; @Haikka201081052103; @Haikka2010014047] $$\begin{aligned}
\dot{\bar{\rho}}(t)= - T{r_R}\int_{t_0}^t {dt'[{H_I}(t),[{H_I}(t'),
\bar \rho (t) \otimes {\rho _E}]]}.
\label{tcl2}
\end{aligned}$$ Substituting Eq. (\[HI\]) into Eq. (\[tcl2\]) and using Eq. (\[dgbbvx\]), we transform Eq. (\[NZ2\]) back into the Schrödinger picture and obtain, $$\begin{aligned}
{{\dot \rho }_{TCL}} =& - i[{H_S},{\rho _{TCL}}(t)] + \int_{t_0}^t {dt'\{ f(t - t')} \\
&\times[\sigma_- (t' - t){\rho _{TCL}}(t){\sigma_+} - {\sigma_+}\\
&\times\sigma_- (t' - t){\rho _{TCL}}(t)]+ H.c.\}.
\label{tcll22}
\end{aligned}$$ We note here that obtaining the time-convolutionless non-Markovian master equation perturbatively up to second order in the coupling by the use of the time-convolutionless projection operator technique is equivalent to obtaining it by replacing $\bar \rho
(t')$ with $\bar \rho (t)$ in Eq. (\[ohbmnb\]) [@Breuer2002; @Breuer1999591633; @Yan1998582721; @Ferraro200980042112; @Xu20011143868; @Schroder2006124084903; @Liu200776022312; @Haikka201081052103; @Haikka2010014047]. One may wonder if the second order time-nonlocal master equation (\[RHOnz2\]) is more accurate than the second-order time-convolutionless master equation (\[tcll22\]). In the following, using the exact master equation, we show that the TCL approach (\[tcll22\]) works better than the NZ one (\[RHOnz2\]) for a wide range of parameters.
Comparison to the Nakajima-Zwanzig and time-convolutionless master equations
----------------------------------------------------------------------------
We now analyze the characteristics of the damped driven two-level systems, by comparing the exact dynamics with that from the NZ and TCL master equations. Our purpose is to shed light on the performances of two master equations and to point out their ranges of validity. As stressed in the introduction, without the exact master equation, it is difficult to examine the range of validity for these master equations.
We assume that the system couples to a reservoir with detuning and the reservoir has a Lorentzian spectral density [@Li201081062124; @Breuer2002; @Shen201388033835; @Haikka201081052103] $$\begin{aligned}
J(\omega ) = \frac{\Gamma }{{2\pi }}\frac{{{\lambda ^2}}}{{{{(\omega - {\omega _0} + \delta )}^2} + {\lambda ^2}}},
\label{spectral density}\end{aligned}$$ where $\delta = {\omega _0} - {\omega _c}$ is the detuning of $\omega _c$ to $\omega _0$, and $\omega _c$ is the center frequency of the cavity. It is worth noting that the parameter $\lambda $ defines the spectral width of the reservoir and is connected to the reservoir correlation time ${\tau _R} = {\lambda ^{ - 1}}$. The parameter $\Gamma $ can be shown to be related to the decay of the system in the Markovian limit with a flat spectrum. The relaxation time scale is ${\tau _L} = {\Gamma ^{ - 1}}$.
**The Markovian dynamics usually describes a situation where the coupling strength between the system and the environment is very weak**, and the characteristic correlation time ${\tau _R}$ of the environment is sufficiently shorter than that of the system ${\tau _L}$, i.e., $$\begin{aligned}
{\tau _R} \ll {\tau _L},
\label{markovlimit}\end{aligned}$$ equivalently, the spectrum of the reservoir takes $J(\omega ) =
\frac{\Gamma }{{2\pi }}$, which leads to a Markovian dynamics. The reservoir has no memory effect on the evolution of the system. Then according to Eq. (\[ft\]), we have $$\begin{aligned}
f(t) = \Gamma \delta (t).
\label{markovainft}\end{aligned}$$ Substituting Eq. (\[markovainft\]) into the first equation of Eq. (\[uu\]), we reduce the solution of $u(t)$ to $$\begin{aligned}
u(t) = {e^{ - i\Delta t - \frac{\Gamma }{2}t}},
\label{uut}\end{aligned}$$ i.e., all the coefficients in Eq. ( \[dphiin\]) are constants, $$\begin{aligned}
s(t) = \Delta ,r(t) = \Omega ,\gamma (t) = \Gamma.
\label{reducessrt}\end{aligned}$$ The exact master equation (\[finallyrhoo\]) is then reduced to Markovian master equation [@Breuer2002; @Gardiner2000; @Shatokhin2000174157] $$\begin{aligned}
\frac{{d\rho (t)}}{{dt}} =& - i[\Delta {\sigma_+}\sigma_- + \Omega {\sigma _x},\rho (t)] + \frac{\Gamma }{2}[2\sigma_- \rho (t){\sigma _ + }\\
&- {\rm{\{ }}{\sigma _ + }\sigma_- ,\rho (t)\} ],
\label{Markovianmaster}
\end{aligned}$$ where the decoherence rates are time independent. This gives the standard Lindblad form for the Markovian dynamics. When $$\begin{aligned}
{\tau _R} \ge {\tau _L}
\label{nonmarkovlimit}\end{aligned}$$ is satisfied, the strong non-Markovian effect plays an important role and the dynamics must be described by the exact master equation (\[finallyrhoo\]).
Now we calculate the two-time correlation functions $f(t - t')$ by substituting Eq. (\[spectral density\]) into Eq. (\[ft\]) $$\begin{aligned}
f(t - t') = \frac{1}{2}\lambda \Gamma \exp [ - (\lambda + i\Delta - i\delta )(t - t')].
\label{lorenzft}\end{aligned}$$ It is clear that the bandwidth $\lambda$ is inversely proportional to the memory time of reservoir. For this correlation function $f(t
- t')$, Eq. (\[uu\]) can be easily solved by use of Eq. (\[lorenzft\]), the solution reads, $$\begin{aligned}
u(t) = k(t) \times \left[ {\cosh \left( {\frac{{dt}}{2}} \right) +
\frac{{\lambda - i\delta }}{d}\sinh \left( {\frac{{dt}}{2}}
\right)} \right],\end{aligned}$$ where $k(t) = {e^{ - (\lambda + 2i\Delta - i\delta )t/2}}$ and $d
= \sqrt {{{(\lambda - i\delta )}^2} - 2\Gamma \lambda } $.
In order to calculate $U(t)$ and $\sigma_-(t)$ in Eqs. (\[RHOnz2\]) and (\[tcll22\]), we calculate the eigenstates of the free system Hamiltonian ${H_S}$, $$\begin{aligned}
\left| {{\phi _{\lambda 1}}} \right\rangle = \frac{1}{{\sqrt 2 }}(\sqrt {1 + \sin \theta } \left| e \right\rangle + \sqrt {1 - \sin \theta } \left| g \right\rangle ),\\
\left| {{\phi _{\lambda 2}}} \right\rangle = \frac{1}{{\sqrt 2 }}(\sqrt {1 - \sin \theta } \left| e \right\rangle - \sqrt {1 + \sin \theta } \left| g \right\rangle ),
\label{twobasic}
\end{aligned}$$ the corresponding eigenvalues are ${\lambda _1} = (\Delta +
{W_0})/2$ and ${\lambda _2} = (\Delta - {W_0})/2$. Here ${W_0} =
\sqrt {{\Delta ^2} + 4{\Omega ^2}} $, $\theta = acr\tan (\Delta
/2\Omega )$. Straightforward algebra yields, $$\begin{aligned}
\sigma_-(t) =& {e^{i{H_S}t}}\sigma_- {e^{ - i{H_S}t}} = \sum\limits_{j,k = 1}^2 {{\sigma _{jk}}{e^{it({\lambda _j} - {\lambda _k})}}\left| {{\phi _{\lambda j}}} \right\rangle \left\langle {{\phi _{\lambda k}}} \right|} ,\\
U(t) =& \sum\limits_{j = 1}^2 {{e^{i{\lambda _j}t}}\left| {{\phi _{\lambda j}}} \right\rangle \left\langle {{\phi _{\lambda j}}} \right|} ,
\label{Usigma}
\end{aligned}$$ where ${\sigma _{jk}} = \left\langle {{\phi _{\lambda j}}}
\right|\sigma_- \left| {{\phi _{\lambda k}}} \right\rangle $. Now let us concentrate on the average $\left\langle {{\sigma _z}}
\right\rangle $, i.e., on the probability difference of finding the system in the atomic excited and ground levels. To examine the validity of the two approximate approaches we explore three different regimes by changing the width $\lambda$ of the Lorentzian spectral density. This investigation will allow us to estimate in which cases the non-Markovian master equations are efficient in the description of the system dynamics.
![(Color online) The time evolution of the population difference $\left\langle {{\sigma _z}} \right\rangle $ for the system initially in the excited state $\left| e \right\rangle$ versus the dimensionless parameter $\Gamma t$. The red line, black-dashed line, and blue-dashed-dotted line denote the exact Eq. (\[finallyrhoo\]), TCL Eq. (\[tcll22\]), and NZ Eq. (\[RHOnz2\]) master quations, respectively. The width of the Lorentzian spectral density is $\lambda = 25\Gamma$. The other parameters chosen are $\Delta = 0.3\Gamma,\Omega = 0.02\Gamma,\delta = 0.01\Gamma$ for (a), $\Delta = 0.3\Gamma,\Omega = \Gamma,\delta = 0.01\Gamma$ for (b), $\Delta
= 5\Gamma,\Omega = \Gamma,\delta = 0.01\Gamma$ for (c), $\Delta = \Gamma,\Omega =
\Gamma,\delta = 10\Gamma$ for (d).[]{data-label="nonmar:"}](nonmar.eps){width="48.00000%"}
Fig. \[nonmar:\] shows a comparison among the exact, TCL, and NZ master equations with large bandwidth $\lambda = 25\Gamma .$ We find that the results given by the TCL (\[tcll22\]) and NZ (\[RHOnz2\]) are in good agreement with those obtained by the exact master equation (\[finallyrhoo\]) for any time scales. In this case, both TCL and NZ give a very good description for the dynamics. They indeed provide us with the same results, which are very close to the Markovian dynamics; see the discussion in Eq. (\[Markovianmaster\]). In addition, in such cases the TCL master equation which is easier to solve might be preferred to use because it is a time-local first order differential equations.
![(Color online) $\left\langle {{\sigma _z}} \right\rangle $ versus the dimensionless parameter $\Gamma t$. The width of the Lorentzian spectrum is $\lambda =\Gamma$. The results are obtained by the exact (red line), TCL (black-dashed line), and NZ (blue-dashed-dotted line) solutions. The other parameters chosen are $\Delta = 0.3\Gamma,\Omega = 0.02\Gamma,\delta = 0.01\Gamma$ for (a), $\Delta =
10\Gamma,\Omega = \Gamma,\delta = 0.01\Gamma$ for (b), $\Delta = 10\Gamma,\Omega =
0.02\Gamma,\delta = 0.2\Gamma$ for (c), $\Delta = 10\Gamma,\Omega = \Gamma,\delta =
0.2\Gamma$ for (d).[]{data-label="litnonmar:"}](litnonmar.eps){width="48.00000%"}
We set the same quantity $\lambda = \Gamma$ in Fig. \[litnonmar:\]. Clearly, the results given by the TCL (\[tcll22\]) and NZ (\[RHOnz2\]) are in good agreement with those obtained by the exact expression (\[finallyrhoo\]) in a short-time scale, but they deviate from each other in a long-time scale. Especially considering the long-time behavior, the NZ equation leads to a non-physical result. For times longer than some critical values, the solution for the population difference $\left\langle {{\sigma _z}} \right\rangle $ cannot represent a physical result, because the absolute value of $\left\langle
{{\sigma _z}} \right\rangle $ is larger than $1$. We therefore can conclude that for this range of parameters the TCL equation gives a better description of the dynamics because it reflects all the qualitative characteristics of the exact expression.
![(Color online) Comparison of the density matrices obtained by solving the TCL and NZ master equation with the one by exact master equation. We quantify the difference by the fidelity defined by $F(\rho_1,\rho_2)=Tr \sqrt{\rho_1^{\frac 1 2}\rho_2\rho_1^{\frac 1 2
}}$. The results show that the density matrix given by TCL is always better than that given by NZ master equation. The parameters in (a),(b) and (c) are chosen as the same as in Fig.\[nonmar:\]-(a), Fig.\[litnonmar:\]-(a) and Fig.\[strnonmar:\]-(a), respectively. []{data-label="fig3:"}](fidelity.eps)
One may wonder if this observation depends on the quantity plotted. To clarify this point, we plot the fidelity of the density matrix from the exact master equation to these from TCL and NZ master equations in Fig. \[fig3:\]. The results suggest that the TCL master equation is indeed better than the NZ for a wide range of parameters.
In Fig. \[strnonmar:\], we choose the parameter $\lambda =
0.05\Gamma $, which, according to Eq. (\[lorenzft\]), corresponds to very strong reservoir correlations and very long memory effect. We find again that a good agreement among all the three approaches in the short-time scale, but in this case the TCL approximation works not so good. The dynamics of the TCL master equation (black-dashed line) does not succeed to follow the oscillations given by the exact expression (red line). The NZ approach has the same problem that it can not conserve the positivity of the density matrix (i.e., the absolute value of $\left\langle {{\sigma
_z}} \right\rangle $ exceeds $1$). Thus in this case two approximate methods are not suitable to describe the dynamics of the driven two-level system.
![(Color online) $\left\langle {{\sigma _z}} \right\rangle $ versus time $\Gamma t$. The results are obtained by the exact (red line), TCL (black-dashed line), and NZ (blue-dashed-dotted line) solutions. The parameters chosen are $\lambda =0.05\Gamma$, $\Delta
= 0.3\Gamma,\Omega = 0.02\Gamma,\delta = 0.01\Gamma$ for (a), $\Delta = 3.5\Gamma,\Omega
= 0.4\Gamma,\delta = 0.01\Gamma$ for (b), $\Delta = 10\Gamma,\Omega = 0.02\Gamma,\delta
= 0.08\Gamma$ for (c), $\Delta = 0.3\Gamma,\Omega = 0.02\Gamma,\delta = 0.14\Gamma$ for (d).[]{data-label="strnonmar:"}](strnonmar.eps){width="48.30000%"}
**Before closing this section, we present a discussion on the function $f\left( {\tau - \tau '} \right)$ in Eq. (\[ft\]). Concretely, we examine mathematically the validity to extend the lower limit of the integration from $0$ to $ - \infty $. We will explore three different regimes characterized by the width $\lambda
$ in the spectral density in the following.**
**In Fig. \[zero-infinity:\], we show a comparison between results with two different lower limits in the integration (\[ft\]) with the spectral density given in Eq. (\[spectral density\]), the simulation is performed for the exact dynamics described by Eq. (\[finallyrhoo\]). Fig. \[zero-infinity:\] (a) is for the integration with lower limit $- \infty$, which is slightly different from that with lower limit 0. In Fig. \[zero-infinity:\] (b) and (c), the results with lower limit $- \infty$ are in good agreement with that obtained with lower limit 0.**
**This numerical result can be explained as follows. When we change $\omega \to \omega - {\omega _L}$, Eq. (\[ft\]) becomes** $$\begin{aligned}
\begin{aligned}
f\left( {\tau - \tau '} \right) =
\int_{ - {\omega _L}}^\infty {d\omega J\left( \omega \right)}
{e^{ - i\omega (\tau - \tau ')}}
\label{fttt}
\end{aligned}\end{aligned}$$ **with** $$\begin{aligned}
J\left( \omega \right) = \frac{\Gamma }{{2\pi }}
\frac{{{\lambda ^2}}}{{{{\left( {\omega - \Delta + \delta } \right)}^2} + {\lambda ^2}}},
\label{jwww}\end{aligned}$$ **this tells us that the frequency ${{\omega _L}}$ affects only the lower limit of the integral (\[fttt\]) when $\Delta$ is fixed. Define ${x_ \pm }= \Delta - \delta \pm \lambda $ representing the position of half-height of the Lorentzian spectral density (\[jwww\]), we thought that the integral of $J(\omega)$ over $\omega$ from $-\infty$ to $\infty$ can be approximately replaced by the same integral but from $x_-$ to $x_+$. With this approximation, we find that ${x_ - } = -24.71\Gamma$ and $\lambda =
25\Gamma$ in Fig. \[zero-infinity:\] (a). Clearly, $x_ -$ is much smaller than $-{{\omega _L}}$, thus the integral of $J(\omega)$ over $\omega$ from $x_-$ to $-{{\omega _L}}$ can not be ignored \[see Fig. \[zero-infinity:\] (d)\]. This explains the difference of the two curves in Fig. \[zero-infinity:\] (a). On the contrary, $\lambda = \Gamma$, ${x_ - } = -0.71\Gamma$ in Fig. \[zero-infinity:\] (b), and $\lambda = 0.05\Gamma$, ${x_ -
} = 0.24\Gamma$ in Fig. \[zero-infinity:\] (c). $x_-$ is larger than $-{{\omega _L}}$ in both cases of (b) and (c). Thus, the integral from $-{{\omega _L}}$ to $x_-$ can be ignored \[see Fig. \[zero-infinity:\] (e) and (f)\]. As a result, the two lines in both (b) and (c) are in good agreement.**
**The above discuss suggests that it is reasonable to extend the lower limit of the integral of Lorentzian spectral $J\left(
\omega \right)$ from $0$ to $ - \infty $.**
![(Color online) **$\left\langle {{\sigma _z}}
\right\rangle $ given by the exact master equation (\[finallyrhoo\]) as a function of time. The purpose of this figure is to show the difference in $\left\langle {{\sigma _z}}
\right\rangle$ caused by different lower limits of the integral of kernel (\[fttt\]). The red and blue-dashed lines correspond to lower limits $-\omega_L$ and $-\infty$, respectively. The Lorentzian spectral density $J(\omega )$ (in units of $\Gamma/{2\pi}$) in (d), (e) and (f) correspond respectively to results shown in (a), (b) and (c). $x_-$ denotes the left location of the half-height of the spectral density. The parameters in (a), (b) and (c) are chosen as the same as in Fig.\[nonmar:\]-(a), Fig.\[litnonmar:\]-(a) and Fig.\[strnonmar:\]-(a), respectively. Notice that $\Delta=\omega_0-\omega_L=0.3\Gamma$ in Eq. (\[H\]), we set $\omega_0=1.3\Gamma$ and $\omega_L=\Gamma$.**[]{data-label="zero-infinity:"}](zero-infinity.eps){width="48.90000%"}
validity of secular approximation in time-convolutionless master equations
==========================================================================
Taking advantage of the exact expression for the dissipative dynamics of the open driven two-level system, we have shown that the TCL approach can reveal all the characteristics of the non-Markovian dynamics for a range of parameters much wider than the results that the NZ equation gives, this is physically reasonable, since the latter may violate the positivity condition on the density matrix for the reservoir correlations which are not very strong. Therefore through comparing with the exact non-Markovian master equation (\[finallyrhoo\]), we can investigate the validity of the secular approximation based on time-convolutionless master equation (\[tcll22\]).
We now use the orthonormalized basis (\[twobasic\]) and these relations (\[Usigma\]) to derive explicitly the time-convolutionless master equations (\[tcll22\]) as follows $$\begin{aligned}
\dot \rho = - i[{H_S} - H_1,\rho ] + D(\rho ) + {D_1}(\rho ),
\label{tcl2expend}\end{aligned}$$ with $$\begin{aligned}
H_1 =g_0^2{Q_0}(t)S_z^2 +g_2^2{Q_{ + 1}}(t)S_-{S_+}+g_1^2{Q_{ - 1}}(t){S_+}S_-,
\label{Hprimee}
\end{aligned}$$ which describes a small shift in the energy of the two-level system. The above new operators are defined as $S_-{\rm{ = }}\left| {{\phi _{\lambda 2}}} \right\rangle
\left\langle {{\phi _{\lambda 1}}} \right|$, ${S_ + }{\rm{ =
}}\left| {{\phi _{\lambda 1}}} \right\rangle \langle {\phi _{\lambda
2}}|,$ and ${S_z}{\rm{ = }}\left| {{\phi _{\lambda 1}}}
\right\rangle \left\langle {{\phi _{\lambda 1}}} \right| - \left|
{{\phi _{\lambda 2}}} \right\rangle \left\langle {{\phi _{\lambda
2}}} \right|$, then the dissipative superoperator $D(\rho)$ in Eq. (\[tcl2expend\]) can be written in a Lindblad form $$\begin{aligned}
D(\rho ) =& g_1^2{P_{ - 1}}(t)[2S_- \rho {S_+} - \{ {S_+}S_-,\rho \} ]\\
&+ g_2^2{P_{ + 1}}(t)[2{S_+}\rho S_- - \{ S_-{S_+},\rho \} ]\\
&+ g_0^2{P_{ 0}}(t)[2{S_z}\rho {S_z} - \{ S_z^2,\rho \} ],
\label{domig}
\end{aligned}$$ where the coefficients ${g_0} = \Omega /{W_0}$, ${g_1} = ({W_0} +
\Delta )/(2{W_0})$, ${g_2} = ({W_0} - \Delta )/(2{W_0})$, ${W_0} =
\sqrt {{\Delta ^2} + 4{\Omega ^2}} $. The second dissipator $D_1(\rho)$ in Eq. (\[tcl2expend\]) has a more complicated form and contains the contribution of the so-called nonsecular terms, $$\begin{aligned}
{D_1}(\rho ) =&{g_0}{R_0}(t)[{g_2}({S_z}\rho S_- - S_-{S_z}\rho ) + {g_1}({S_ + }\\
&\times{S_z}\rho- {S_z}\rho{S_ +})]+ {g_2}{R_1}(t)[{g_0}({S_ + }\rho {S_z}\\
&- {S_z}{S_ + }\rho )- {g_1}{S_ + }\rho {S_ + }]+ {g_1}{R_{ - 1}}(t)[{g_0}\\
&\times({S_z}S_-\rho - S_-\rho {S_z})- {g_2} S_- \rho S_-]+ H.c..
\label{nonsecular}
\end{aligned}$$ For TCL master equations, the non-Markovian effects are contained in the time-dependent coefficients ${P_m}(t),{Q_m}(t),$ and ${R_m}(t)$, with $m \in \{ {\rm{ + ,0,}} - \}$. The time-dependent coefficient reads $$\begin{aligned}
{R_m}(t) =& \int_0^t {dt'\int {d\omega J(\omega )\exp [i({M_m} - \omega )(t-t')]} },
\label{coefficients}
\end{aligned}$$ where ${M_m} = {\omega _L} - m{W_0}$. The other coefficients take ${P_m}(t) = {\rm{Re[}}{R_m}(t)]$ and ${Q_m}(t) = -
{\rm{Im[}}{R_m}(t)].$ Conventionally, the nonsecular terms included in the dissipator ${D_1}(\rho )$ are neglected in the secular approximation. In order to investigate the effects of the nonsecular terms on the non-Markovian dynamics, we focus on two regimes identified by the mutual relationship between the system characteristic time and the reservoir correlation time.
The time-dependent coefficient (\[coefficients\]) for the driven two-level system in a Lorentzian reservoir can be calculated explicitly using Eq. (\[lorenzft\]) $$\begin{aligned}
{R_m}(t) = \frac{{\Gamma \lambda }}{{\lambda + i{N_m}}}\left\{ {1
- \exp [ - (\lambda + i{N_m})t]} \right\}, \label{rmt}\end{aligned}$$ with $$\begin{aligned}
{N_m}{\rm{ = }}\Delta {\rm{ - }}\delta {\rm{ + }}m{W_0}.
\label{nm}\end{aligned}$$ We can see from Eq. (\[rmt\]) that when $\min \left[ {\left| {{N_
+ }} \right|,\left| {{N_0}} \right|,\left| {{N_ - }} \right|}
\right] \gg \lambda,$ namely, the relaxation time ${\tau _R} =
{\lambda ^{ - 1}}$ of the reservoir correlation is very large compared to the typical timescale defined as ${\tau _S} = {[\min
(\left| {{N_ + }} \right|,\left| {{N_0}} \right|,\left| {{N_ - }}
\right|)]^{ - 1}}$, i.e. $$\begin{aligned}
{\tau _R} \gg {\tau _S}
\label{tairtaus}\end{aligned}$$ is satisfied, oscillating terms (\[nonsecular\]) (that containing ${R_m}(t)$) may be neglected as $t$ increases, since rapid oscillations have no contribution to the dynamics on the timescale of the relaxation, this constitutes the secular approximation.
When $$\begin{aligned}
{\tau _R} \le {\tau _S},
\label{nonsecularcondition}\end{aligned}$$ we cannot neglect the nonsecular terms (\[nonsecular\]) in the master equation (\[tcl2expend\]) in the dynamics of the driven two-level system. Therefore in this case, we can no longer obtain a simple expression for the system. The master equation of the system is no longer in the time-dependent Lindblad form.
Examining Eqs. (\[markovlimit\]) and (\[tairtaus\]), we can summarize the comparison of the nonsecular with the secular approximation in the following Table \[table:\], which shows the validity regimes for secular and nonsecular approximation in TCL, Markovian and non-Markovian regimes, respectively.
 \[table:\]
From Table \[table:\], we can divide the time dependent dynamics into two regimes, labeled by $\alpha$ and $\beta$, i.e., Markovian and non-Markovian regimes, respectively.
![(Color online) This plot shows the comparison of the secular approximation (regime $\rm I$) \[(a), (c), and (e)\] and nonsecular terms (regime $\rm II$) \[(b), (d), and (f)\] in Markovian regime $\alpha$ in Table \[table:\]. The red line , blue-dashed line, and black dashed-dotted line denote the exact expression Eq. (\[finallyrhoo\]), the secular approximation Eq. (\[tcl2expend\]) neglecting the nonsecular terms (\[nonsecular\]), and the nonsecular Eq. (\[tcl2expend\]) containing (\[nonsecular\]), respectively. Parameters chosen are $\lambda=10\Gamma, \Delta = 0,\Omega = 0.5\Gamma,\delta = 40\Gamma$ for (a), $\Delta = 0.5\Gamma,\Omega = 0.2\Gamma,\delta = 10\Gamma$ for (b), $\Delta =
10\Gamma,\Omega = 2\Gamma,\delta = 60\Gamma$ for (c), $\Delta = 0.1\Gamma,\Omega =
0.2\Gamma,\delta = 5\Gamma$ for (d), $\Delta = 10\Gamma,\Omega = 0.2\Gamma,\delta = 60\Gamma$ for (e), $\Delta = \Gamma,\Omega = 0.5\Gamma,\delta = 10\Gamma$ for (f).[]{data-label="Marksecu:"}](Marksecu.eps)
In regime $\alpha$, i.e., Markovian regime, we can see that the results given by the regime $\rm I$ under the secular approximation in the TCL Eq. (\[tcl2expend\]) are in good agreement with those obtained by the exact master equation Eq. (\[finallyrhoo\]) when the weak coupling condition (\[markovlimit\]) and the secular approximation (\[tairtaus\]) are simultaneously satisfied \[see Figs. \[Marksecu:\] (a), \[Marksecu:\] (c), and \[Marksecu:\] (e)\]. When the parameters simultaneously satisfy Eqs. (\[markovlimit\]) and (\[nonsecularcondition\]) \[see Figs. \[Marksecu:\] (b), \[Marksecu:\] (d), and \[Marksecu:\] (f)\], i.e., the regime (II), the dynamics of the TCL master equation (\[tcl2expend\]) involving the nonsecular terms Eq. (\[nonsecular\]) are in good agreement with those obtained by the exact expression (\[finallyrhoo\]), but the results obtained by the secular approximation have serious deviations from those obtained by the exact solution Eq. (\[finallyrhoo\]). This difference comes from the nonsecular terms (\[nonsecular\]), which are ignored in the regime (II).
![(Color online) This plot shows the comparison of the secular approximation (regime $\rm III$) \[(a), (c), and (e)\] with nonsecular terms (regime $\rm IV$) \[(b), (d), and (f)\] in non-Markovian regime $\beta$ in Table \[table:\]. The red line, blue-dashed line, and black dashed-dotted line denote the exact master equation Eq. (\[finallyrhoo\]), the secular approximation Eq. (\[tcl2expend\]) neglecting the nonsecular terms (\[nonsecular\]), and the nonsecular Eq. (\[tcl2expend\]) containing (\[nonsecular\]), respectively. Parameters chosen are $\lambda=0.8\Gamma, \Delta = 2\Gamma,\Omega = 0.2\Gamma,\delta = 15\Gamma$ for (a), $\Delta = 0.04\Gamma,\Omega = 0.06\Gamma,\delta = 0.4\Gamma$ for (b), $\Delta =
0,\Omega = 0.2\Gamma,\delta = 10\Gamma$ for (c), $\Delta = 0.05\Gamma,\Omega =
0.1\Gamma,\delta = 1.8\Gamma$ for (d), $\Delta = 20\Gamma,\Omega = \Gamma,\delta = 5\Gamma$ for (e), $\Delta = 0.5\Gamma,\Omega = 0.2\Gamma,\delta = 2.5\Gamma$ for (f).[]{data-label="nonmarsecu:"}](nonmarsecu.eps)
Examining the non-Markovian regime labeled by $\beta$ in Table \[table:\], we find that the results given by the secular approximation Eq. (\[domig\]) in the regime $\rm III$ are in good agreement with those obtained by the exact expression Eq. (\[finallyrhoo\]) when the strong coupling condition (\[nonmarkovlimit\]) and the secular approximation (\[tairtaus\]) are simultaneously satisfied \[see Figs. \[nonmarsecu:\] (a), \[nonmarsecu:\] (c), and \[nonmarsecu:\] (e)\]. When the parameters satisfy simultaneously Eqs. (\[nonmarkovlimit\]) and (\[nonsecularcondition\]) \[see Figs. \[nonmarsecu:\] (b), \[nonmarsecu:\] (d), and \[nonmarsecu:\] (f)\], i.e., in the regime IV, the dynamics of the TCL master equation (\[tcl2expend\]) involving the nonsecular terms Eq. (\[nonsecular\]) are in good agreement with those obtained by the exact one (\[finallyrhoo\]). However, the results obtained by the secular approximation have serious deviations from the exact solution Eq. (\[finallyrhoo\]). The same observation can be found in the regime $\rm II$.
From Figs. \[Marksecu:\] and \[nonmarsecu:\], we can learn that the non-Markovian effect occurs when $\lambda $ is small. The non-Markovian regime $\beta$ transits to the Markovian regime $\alpha$ when $\lambda$ is large. Therefore by manipulating $\lambda$ we can control the crossover from non-Markovian to Markovian processes and vice versa. This provides us with a method to manipulate the non-Markovian dynamics in the driven two-level system.
Now we turn to discuss the positivity and complete positivity of the reduced dynamics given by the TCL master equation. The non-Markovian TCL master equation derived in this paper is not of the Lindblad form, even in the secular regime discussed in Sec.V, therefore, both the positivity and the complete positivity of the reduced dynamics can not be guaranteed. In other words, the Lindblad-Gorini-Kossakowski-Sudarshan theorem[@Lindblad481191976; @Gorini178211976] that ensures the positivity can not be satisfied in general, indicating that the dynamics given by the TCL master equation might not be physical for all range of parameters.
Nevertheless, the parameters chosen (in fact, it is wide range of parameters) in this paper assure the positivity of the reduced dynamics given by the TCL master equations. This can be understood as follows. For the driven qubit in the TCL approximation, the necessary and sufficient condition for complete positivity and positivity is given by (for details, see Ref.[@Haikka201081052103]) $$\begin{aligned}
2\alpha (t) + \beta (t) \ge 0, \label{cpp}\end{aligned}$$ where $$\begin{aligned}
\alpha (t) =& 2\int_0^t {d\tau [g_1^2{P_{ - 1}}(\tau ) + g_2^2{P_{ + 1}}(\tau ) + 4g_0^2{P_0}(\tau )]} ,\\
\beta (t) =& 4\int_0^t {d\tau [g_1^2{P_{ - 1}}(\tau ) + g_2^2{P_{ + 1}}(\tau )]} .
\label{Hprimgee}
\end{aligned}$$ Now back to the Sec. IV, we stress that the necessary and sufficient condition (\[cpp\]) for complete positivity is satisfied for the parameters chosen in Fig. (\[litnonmar:\]) and (\[strnonmar:\]) (not for a very long time). Therefore, for a wide range of parameters, the complete positivity of the reduced dynamics is guaranteed. Hence our conclusion, i.e., the TCL equation gives a better description of the dynamics, holds true for a wide range of parameters. It is important to remind that theoretical descriptions of non-Markovian open quantum systems are often based on a series of assumptions and approximations without which it would not be possible to tackle the problem of the description of the dynamics in simple analytic terms. But those approximations plague almost all approximated reduced dynamics and lead them to break the complete positivity required for reduced dynamics. Therefore the observation here is available for short times and certain ranges of parameters.
The case with non-Lorentzian spectrum
=====================================
**Note that the spectral density ${J_{SB}}(\omega ) $ is proportional to the imaginary part of the dynamical susceptibility $\tilde \chi (\omega )$ of a damped harmonic oscillator, in this section, we present a numerical simulation for $\left\langle\sigma_z(t)\right\rangle$ adapting a different spectral density, e.g., spin-boson spectral density [@Weiss2008; @Thoss20011152991],** $$\begin{aligned}
{J_{SB}}(\omega ) = \frac{1}{M}\frac{{\omega \lambda }}{{{{({\omega
^2} - \omega _0^2)}^2} + {\omega ^2}{\lambda ^2}}}. \label{JSB}\end{aligned}$$ **In Fig. (\[spin-boson:\]), we plot the time evolution of the population difference $\left\langle {{\sigma _z}} \right\rangle$ for three typical spectral width $\lambda$. Interestingly, in Fig. \[spin-boson:\] (a), i.e., for large $\lambda=25\Gamma$ the population difference $\left\langle {{\sigma _z}} \right\rangle $ decays monotonically for both spin-boson and Lorentzian spectral density, the difference is that the former decay more slowly than the latter. This corresponds to the Markovian case, see the discussion in Eq. (\[Markovianmaster\]). For $\lambda=\Gamma$, small oscillations can be observed in the case with spin-boson spectral density, while it is not obvious in the case with Lorentzian spectral density \[see Fig. \[spin-boson:\] (b)\]. For small $\lambda=0.05\Gamma$, oscillations in the population difference can be found in both cases with spin-boson spectral density and Lorentzian spectral density\[see Fig. \[spin-boson:\] (c)\]. These oscillations correspond to a rapid exchange of energy and information between the two-level atom and reservoir.**
![(Color online) This plot shows the comparison of the exact dynamics (\[finallyrhoo\]) for Lorentzian (red-line) and spin-boson(blue-dashed line) spectral density. The parameters in (a), (b) and (c) are chosen as the same as in Fig.\[nonmar:\]-(a), Fig.\[litnonmar:\]-(a) and Fig.\[strnonmar:\]-(a), respectively. The other parameter chosen is $M=5\Gamma$.[]{data-label="spin-boson:"}](spin-boson.eps)
**Spectral density is a key feature for environments. It characterizes the correlation among the particles in the environment and determines the dynamics of open system, as we show in this section.**
Conclusion
==========
For a driven two-level quantum system, secular and weak coupling approximations break down when the system-environment coupling varies significantly on the scale of the Rabi frequency. In this paper, we avoid these approximations and have studied the non-Markovian dynamics of the driven two-level system coupled to a bosonic reservoir at zero temperature. Making use of the Feynman-Vernon influence functional theory in the coherent state representation, we derive an exact non-Markovian master equation for the driven two-level system. We compare this exact master equation with the other equations describing non-Markovian dynamics, i.e., the Nakajima-Zwanzig and the time-convolutionless non-Markovian master equation, it is found that the TCL approach is valid for a range of parameters much wider than the NZ master equation. This is reasonable since the latter may violate the positivity of dynamical map when the correlation in the reservoir is strong. By using the exact master equation, we also have given the analytical condition of validity of the secular approximation and show how it depends on the environmental spectral density, we found that the nonsecular terms have significant corrections to results obtained by the secular approximation when the relaxation time of the environment is less than or equal to that of the system, i.e. ${\tau _R} \le
{\tau _S}$.
The limitation of this representation is the state of the bath, here we only consider the bath initially at vacuum. Although the zero temperature case is problematic for getting reduced dynamics as the bath correlation functions may decay slowly, the zero-temperature reservoir is a good approximation for many problems in physics. For the reservoir initially at thermal states, the question becomes complicated, since the influence functional in the Feynman-Vernon influence functional theory is very involved.
ACKNOWLEDGMENTS {#acknowledgments .unnumbered}
===============
This work is supported by the NSF of China under Grants No. 11175032.
DERIVATION OF THE INFLUENCE FUNCTIONAL
======================================
The propagating function controlling the time evolution of the reduced density matrix is given by Eq. (\[JJ\]), where the generalized Feynman-Vernon influence functional is defined by $$\begin{aligned}
F[{\bar{\zeta}},\zeta,{\bar{\zeta}}^\prime ,\zeta']
{\rm{ = }}&\int {d\varphi ({\mathbf{z}_f})d\varphi
({\mathbf{z}_i})d\varphi ({\mathbf{z}'_i})} {D^2}\mathbf{z}{D^2}\mathbf{z}'\\
&\times {\rho _E}(\bar{\mathbf{z}}_i ,{\mathbf{z}'_i};0)\exp
\{ i({S_E}[{\bar{\mathbf{z}}},\mathbf{z}]\\
&- S_E^ * [\bar{\mathbf{z}}',\mathbf{z}'] + {S_I}
[{\bar{\mathbf{z}}},\mathbf{z},{\bar{\zeta} },\zeta ]\\
& -{S_I^*}[\bar{\mathbf{z}}',\mathbf{z}',\bar{\zeta}',\zeta'])\},
\label{FJ}
\end{aligned}$$ where ${S_S}, {S_I}$ and ${S_E}$ are the actions corresponding to ${H_S}, {H_I}$ and ${H_E}$, respectively, $$\begin{aligned}
{S_S}[{\bar{\zeta}},\zeta ] =& - i[\bar{\zeta} _f\zeta (t)
+ {\bar{\zeta}}(t_0){\zeta _i}]/2 + \int_{t_0}^t {d\tau \{ }
i[{\bar{\zeta}}(\tau )\\
& \times \dot \zeta (\tau ) - {\dot{\bar{\zeta}} }(\tau )
\zeta (\tau )]/2 - {H_S}({\bar{\zeta} },\zeta )\},\\
{S_E}[{\bar{\mathbf{z}}},\mathbf{z}] =& \sum\limits_k { - i}
\bar{z}_k{z_k}(t) + \int_{t_0}^t {d\tau [i\bar{z}_k{{\dot z}_k}(\tau)} \\
& - {H_E}({\bar{\mathbf{z}}},\mathbf{z})],\\
{S_I}[{\bar{\mathbf{z}}},\mathbf{z},{\bar{\zeta}},\zeta ]
=& - \int_{t_0}^t {d\tau {H_I}[{\bar{\mathbf{z}}},
\mathbf{z},{\bar{\zeta}},\zeta ]} .\label{dfg}
\end{aligned}$$ All the functional integrations are worked out over paths ${\bar{\mathbf{z}}}(\tau ),\mathbf{z}(\tau ),{\bar{\zeta}}(\tau )$, and $\zeta (\tau )$, the endpoints are ${\bar{\mathbf{z}}}(t)
\equiv \bar{\mathbf{z}}_f,\mathbf{z}(t_0) \equiv
{\mathbf{z}_i},{\bar{\zeta}}(t) \equiv {\zeta _f}$, and $\zeta (t_0)
\equiv {\zeta _i}.$
Now we can calculate explicitly the influence functional of our model using the coherent state path-integral formalism. Substituting Eq. (\[representation\]) into the actions of Eq. (\[dfg\]), we obtain the explicit form of the propagator. The path integral of the environmental part in the propagator can be exactly done by the stationary phase method [@Klauder1979192349; @zhang199062867] with the boundary conditions ${z_k}(t_0) = {z_{ki}}$ and $\bar{z}_k(t) =
\bar{z}_{kf}$. This method needs the equations of motion of the path, $$\begin{aligned}
{{\dot z}_k}+i{\Omega _k}{z_k}=-ig_k^ * \zeta,
{{\dot{\bar{z}}}}_k-i{\Omega_k}\bar{z}_k=i{g_k}{\bar{\zeta}},
\label{station phase}
\end{aligned}$$ where $\zeta $ and ${\bar{\zeta}}$ are treated as external sources. By formally integrating Eq. (\[station phase\]), we obtain (\[station phase\]) $$\begin{aligned}
{z_k}(\tau ) =& {z_{ki}}{e^{ - i{\Omega _k}\tau }} - ig_k^ * \int_0^\tau {d\tau '{e^{ - i{\Omega _k}(\tau - \tau ')}}\zeta (\tau ')} ,\\
\bar{z}_k(\tau ) =& \bar{z}_{kf}{e^{i{\Omega _k}(\tau - t)}} + i{g_k}\int_\tau ^t {d\tau '{e^{i{\Omega _k}(\tau - \tau ')}}\bar{\zeta} (\tau ')}.
\label{integratezk}
\end{aligned}$$ By taking the reservoir to be initially at zero temperature (\[rhoE\]), i.e., ${\rho _E}(\bar{\mathbf{z}}_i
,{\mathbf{z}'_i};0)=1$, we finally can obtain Eq. (\[finally F\]) after substituting the result and Eq. (\[integratezk\]) into Eq. (\[FJ\]).
R. Alicki and K. Lendi, *Quantum Dynamical Semigroups and Applications*, Lecture Notes in physics, Vol. **717**, 2nd ed. (Springer, Berlin, 2007)
H.-P. Breuer and F. Petruccione, *The Theory of Open Quantum Systems* (Oxford University Press, Oxford, UK, 2002).
U. Weiss, *Quantum Dissipative Systems*, 3rd ed. (World Scientific Press, Singapore, 2008)
D. P. DiVincenzo, Nature **393**, 113 (1998).
E. Knill, R. Laflamme, and G. J. Milburn, Nature **409**, 46 (2001).
J. I. Cirac, A. K. Ekert, S. F. Huelga, and C. Macchiavello, Phys. Rev. A **59**, 4249 (1999).
D. P. DiVincenzo, Fortschr. Phys. **48**, 771 (2000).
J. I. Cirac, P. Zoller, H. J. Kimble, and H. Mabuchi, Phys. Rev. Lett. **78**, 3221 (1997).
L.-M. Duan, A. Kuzmich, and H. J. Kimble, Phys, Rev, A **67**, 032305 (2003).
C. W. Gardiner and P. Zoller, *Quantum Noise* (Springer-Verlag, Berlin, Germany, 2000).
M. O. Scully and M. S. Zubairy, *Quantum Optics* (Cambridge University Press, Cambridge, UK, 1997).
D. F. Walls and G. J. Milburn, *Quantum Optics* (Springer-Verlag, Berlin, 1994).
H. J. Carmichael, *An Open Systems Approach to Quantum Optics*, Lecture Notes in Physics m18 (Springer-Verlag, Berlin, 1993).
L. Mandel and E. Wolf, *Optical Coherence and Quantum Optics* (Cambridge University Press, England, 1995).
M. Weissbluth, *Photon-Atom Interactions* (Academic Press, Boston, 1989).
W. Vogel and D. G. Welsch, *Lectures on Quantum Optics* (Akademie Verlag, Berlin, 1994).
G. Compagno, R. Passante, and F. Persico, *Atom-Field Interactions and Dressed Atom* (Cambridge University Press, Cambridge, 1995).
A. O. Caldeira and A. J. Leggett, Ann. Phys. (N.Y.) **149**, 374 (1983).
A. J. Leggett, S. Chakravarty, A. T. Dorsey, M. P. A. Fisher, A. Garg, W. Zwerger, Rev. Mod. Phys. **59**, 1 (1987).
J. B. Majer, F. G. Paauw, A. C. J. ter Haar, C. J. P. M. Harmans, and J. E. Mooij, Phys. Rev. Lett. **94**, 090501 (2005).
A. J. Berkley, H. Xu, R. C. Ramos, M. A. Gubrud, F. W. Strauch, P. R. Johnson, J. R. Anderson, A. J. Dragt, C. J. Lobb, and F. C. Wellstood, Science **300**, 1548 (2003).
Y. A. Pashkin, T. Yamamoto, O. Astafiev, Y. Nakamura, D. V. Averin, and J. S. Tsai, Nature (London) 421, 823 (2003).
B. Bellomo, R. Lo Franco, and G. Compagno, Phys. Rev. Lett. **99**, 160502 (2007).
M. A. Nielsen and I. L. Chuang, *Quantum Computation and Quantum Information* (Cambridge University Press, Cambridge, UK, 2000).
A. Barenco, D. Deutsch, A. Ekert, and R. Jozsa, Phys. Rev. Lett. **74**, 4083 (1995).
M. J. Biercuk, H. Uys, A. P. VanDevender, N. Shiga, W. M. Itano, and J. J. Bollinger, Nature (London) **458**, 996 (2009).
S. Das and G S Agarwal, J. Phys. B **42** 205502 (2009).
I. Sinaysky, F. Petruccione, and D. Burgarth, Phys. Rev. A **78**, 062301 (2008).
C. Anastopoulos and B. L. Hu, Phys. Rev. A **62**, 033821 (2000). R. P. Feynman and F. L. Vernon, Ann. Phys. (N.Y.) **24**, 118 (1963).
A. O. Caldeira and A. J. Leggett, Physica A **121**, 587 (1983).
B. L. Hu, J. P. Paz, and Y. Zhang, Phys. Rev. D **45**, 2843 (1992).
W. M. Zhang, D. H. Feng, and R. Gilmore, Rev. Mod. Phys. **62**, 867 (1990).
J. R. Klauder, Phys. Rev. D **19**, 2349 (1979).
R. Karrlein and H. Grabert, Phys. Rev. E **55**, 153 (1997).
F. Haake and R. Reibold, Phys. Rev. A **32**, 2462 (1985).
J.-H. An, Y. Yeo, and C. H. Oh, Ann. Phys. (NY) **324**, 1737 (2009).
J. H. An and W. M. Zhang, Phys. Rev. A **76**, 042127 (2007).
J. H. An, M. Feng, and W. M. Zhang, Quantum. Inf. Comput. **9**, 0317 (2009).
A. Lucke, C. H. Mak, and J. T. Stockburger£¬ J. Chem. Phys. **111**, 10843 (1999).
C.-H. Chou, T. Yu, and B. L. Hu, Phys. Rev. E **77**, 011112 (2008).
J. P. Paz and A. J. Roncaglia, Phys. Rev. Lett. **100**, 220401 (2008).
J. P. Paz and A. J. Roncaglia, Phys. Rev. A 79, 032102 (2009).
M. W. Y. Tu and W. M. Zhang, Phys. Rev. B **78**, 235311 (2008).
M. W. Y. Tu, M. T. Lee, and W. M. Zhang, Quant. Info. Proc. **8**, 631 (2009).
J. S. Jin, M. T.W. Tu,W. M. Zhang, and Y. J. Yan, New J. Phys. **12**, 083013 (2010).
H. T. Tan and W. M. Zhang, Phys. Rev. A **83**, 032102 (2011).
C. U. Lei and W. M. Zhang, Ann. Phys. **327**, 1408 (2012).
W.-M. Zhang, P.-Y. Lo, H.-N. Xiong, M.W.-Y. Tu, and F. Nori, Phys. Rev. Lett. **109**, 170402 (2012).
S. Chaturvedi and F. Shibata, Z. Phys. B **35**, 297 (1979).
R. Shibata, Y. Takahashi and N. Hashitsume, J. Stat. Phys. **17** 171 (1977).
G. A. Prataviera, A. C. Yoshida, and S. S. Mizrahi, Phys. Rev. A **87**, 043831 (2013).
S. Nakajima, Prog. Theor. Phys. **20**, 948 (1958).
R. Zwanzig, J. Chem. Phys. **33**, 1338 (1960).
J. Zhang, Y.-X. Liu, R.-B. Wu, K. Jacobs, and F. Nori, Phys. Rev. A **87**, 032117 (2013).
H.-P. Breuer, B. Kappler, and F. Petruccione, Phys. Rev. A **59**, 1633 (1999).
Y. J. Yan, Phys. Rev. A **58**, 2721 (1998).
E. Ferraro, M. Scala, R. Migliore, and A. Napoli, Phys. Rev. A **80**, 042112 (2009).
R. X. Xu and Y. J. Yan, J. Chem. Phys. **114**, 3868 (2001).
M. Schröer, U. Kleinekathöfer, and M. Schreiber, J. Chem. Phys. **124**, 084903 (2006).
K.-L. Liu and H.-S. Goan, Phys. Rev. A **76**, 022312 (2007).
P. Haikka and S. Maniscalco, Phys. Rev. A **81**, 052103 (2010).
P. Haikka, Phys. Scr. **2010**, 014047 (2010).
K. E. Cahill and R. J. Glauber, Phys. Rev. A **59**, 1538 (1999).
R. J. Glauber, Phys. Rev. **131**, 2766 (1963).
S. Shresta, C. Anastopoulos, A. Dragulescu, and B. L. Hu, Phys. Rev. A **71**, 022109 (2005).
A. Ghosh, S. S. Sinha, and D. S. Ray, Phys. Rev. E **86**, 011138 (2012).
A. Ishizaki and Y. Tanimura, Chem. Phys. **347**, 185 (2008).
L. D. Faddeev and A. A. Slavnov, *Gauge Fields: Introduction to Quantum Theory* (Benjamin-Cummings, Reading, MA, 1980).
R. P. Feynman and A. R. Hibbs, *Quantum Mechanics and Path Integrals* (McGraw-Hill, New York, 1965).
J.-G. Li, J. Zhou, and B. Shao, Phys. Rev. A **81**, 062124 (2010).
H. Z. Shen, M. Qin, and X. X. Yi, Phys. Rev. A **88**, 033835 (2013).
V. N. Shatokhin, S. Ya. Kilin, Opt. Commun. **174**, 157 (2000).
G. Lindblad, Commun. Math. Phys. **48**, 119 (1976).
V. Gorini, A. Kossakowski, and E. Sudarshan, J. Math. Phys. **17**, 821 (1976).
M. Thoss, H. Wang, and W. H. Miller, J. Chem. Phys. **115**, 2991 (2001).
[^1]: Corresponding address: [email protected]
| {
"pile_set_name": "ArXiv"
} |
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.